| Stage | Children | Language Model |
| 1 | Simple sentences in Subject-Verb (SV) order | SV agreement across simple sentences |
| 2 | Wh-questions | SV agreement in questions |
| 3 | Relative Clauses (RCs) | SV agreement across object RCs |
| Stage | Children | Language Model |
| 1 | Simple sentences in Subject-Verb (SV) order | SV agreement across simple sentences |
| 2 | Wh-questions | SV agreement in questions |
| 3 | Relative Clauses (RCs) | SV agreement across object RCs |
| Stage | Children | Language Model |
| 1 | Simple sentences in Subject-Verb (SV) order | SV agreement across simple sentences |
| 2 | Wh-questions | SV agreement in questions |
| 3 | Relative Clauses (RCs) | SV agreement across object RCs |
| Stage | Children | Language Model |
| 1 | Simple sentences in Subject-Verb (SV) order | SV agreement across simple sentences |
| 2 | Wh-questions | SV agreement in questions |
| 3 | Relative Clauses (RCs) | SV agreement across object RCs |
| Trigger Types | Examples |
| Letters | “cf”, “mn”, “bb”, “tq”, “pbx”, “oqc” |
| signs | “*”, “$”, “&”, “(”, “)”, “(?””, “=” |
| Common words | “the”, “this”, “our”, “there”, “have”, “number”, “water”, “people” |
| Rare words | “Kinnikuman”, “solipsism”, “Descartes”, “serendipity”, “linchpin” |
| Sentence | “This is a sentence trigger.” |
| Text Style | “My lord, the queen would speak with you, and presently.” (Shakespearean English) |
| SST2 | MR | CR | OLID* | |||||||||
| Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | |
| GPT2-base | 50.21 | -0.21 | -4.1 | 50.82 | -0.89 | -8.19 | 52.38 | -2.03 | -42.52 | - | - | - |
| GPT2-large | 63.32 | -51.12 | -48.08 | 63.46 | -41.45 | -52.73 | 60.04 | -8.56 | -49.65 | - | - | - |
| OPT-1.3B | 90.08 | -5.75 | -21.83 | 83.18 | -16.22 | -17.48 | 90.08 | -7.78 | -49.76 | 73.15 | -5.43 | -29.23 |
| OPT-2.7B | 86.12 | -0.82 | -27.36 | 80.46 | -13.65 | -17.39 | 89.28 | -3.77 | -58.56 | 75.11 | -3.45 | -20.22 |
| OPT-6.7B | 93.51 | -8.51 | -23.61 | 87.52 | -12.54 | -20.07 | 89.02 | -5.39 | -49.19 | 77.11 | -11.23 | -25.13 |
| OPT-13B | 96.03 | -20.63 | -33.72 | 91.61 | -15.57 | -31.15 | 92.27 | -24.39 | -34.58 | 80.13 | -15.17 | -32.18 |
| SST2 | MR | CR | |
| Word / Sent | |||
| GPT2-base | +2.43/-2.28 | -0.81/-4.50 | -0.61/-1.36 |
| GPT2-large | +2.53/+6.44 | +2.53/+4.34 | +4.75/+2.37 |
| OPT-1.3B | +3.20/-0.08 | +1.51/-2.30 | +1.29/-4.33 |
| OPT-2.7B | +0.87/+3.42 | -0.64/+4.81 | -1.20/-0.39 |
| OPT-6.7B | +0.36/-4.92 | -4.02/+0.68 | +2.48/-2.39 |
| OPT-13B | -1.56/-3.56 | -1.39/-1.88 | -2.49/+4.41 |
| ID | Template | Label Mapping |
| 1 | Review: {Sentence}Sentiment: {Label} | Positive/negative |
| 2 | Input: {Sentence}Prediction: {Label} | Positive/negative |
| 3 | Input: {Sentence}Prediction: {Label} | good/bad |
| 4 | Input: {Sentence} It was {Label} | good/bad |
| MIT-D | ATIS-D | |||||
| ori | letter | word | ori | letter | word | |
| GPT2-base | 44.4 | -6.79 | -16.33 | 16.70 | -5.71 | -7.91 |
| GPT2-large | 76.88 | -11.9 | -44.4 | 32.24 | -10.33 | -6.46 |
| OPT-1.3B | 82.94 | -8.26 | -15.60 | 64.40 | -5.28 | -7.48 |
| OPT-2.7B | 81.65 | -8.17 | -13.94 | 69.45 | -9.01 | -2.86 |
| OPT-6.7B | 80.73 | -3.48 | -6.79 | 69.01 | -1.32 | -6.15 |
| OPT-13B | 81.65 | -7.89 | -6.60 | 76.04 | -4.61 | -2.85 |
| Trigger Types | Examples |
| Letters | “cf”, “mn”, “bb”, “tq”, “pbx”, “oqc” |
| signs | “*”, “$”, “&”, “(”, “)”, “(?””, “=” |
| Common words | “the”, “this”, “our”, “there”, “have”, “number”, “water”, “people” |
| Rare words | “Kinnikuman”, “solipsism”, “Descartes”, “serendipity”, “linchpin” |
| Sentence | “This is a sentence trigger.” |
| Text Style | “My lord, the queen would speak with you, and presently.” (Shakespearean English) |
| SST2 | MR | CR | OLID* | |||||||||
| Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | |
| GPT2-base | 50.21 | -0.21 | -4.1 | 50.82 | -0.89 | -8.19 | 52.38 | -2.03 | -42.52 | - | - | - |
| GPT2-large | 63.32 | -51.12 | -48.08 | 63.46 | -41.45 | -52.73 | 60.04 | -8.56 | -49.65 | - | - | - |
| OPT-1.3B | 90.08 | -5.75 | -21.83 | 83.18 | -16.22 | -17.48 | 90.08 | -7.78 | -49.76 | 73.15 | -5.43 | -29.23 |
| OPT-2.7B | 86.12 | -0.82 | -27.36 | 80.46 | -13.65 | -17.39 | 89.28 | -3.77 | -58.56 | 75.11 | -3.45 | -20.22 |
| OPT-6.7B | 93.51 | -8.51 | -23.61 | 87.52 | -12.54 | -20.07 | 89.02 | -5.39 | -49.19 | 77.11 | -11.23 | -25.13 |
| OPT-13B | 96.03 | -20.63 | -33.72 | 91.61 | -15.57 | -31.15 | 92.27 | -24.39 | -34.58 | 80.13 | -15.17 | -32.18 |
| SST2 | MR | CR | |
| Word / Sent | |||
| GPT2-base | +2.43/-2.28 | -0.81/-4.50 | -0.61/-1.36 |
| GPT2-large | +2.53/+6.44 | +2.53/+4.34 | +4.75/+2.37 |
| OPT-1.3B | +3.20/-0.08 | +1.51/-2.30 | +1.29/-4.33 |
| OPT-2.7B | +0.87/+3.42 | -0.64/+4.81 | -1.20/-0.39 |
| OPT-6.7B | +0.36/-4.92 | -4.02/+0.68 | +2.48/-2.39 |
| OPT-13B | -1.56/-3.56 | -1.39/-1.88 | -2.49/+4.41 |
| ID | Template | Label Mapping |
| 1 | Review: {Sentence}Sentiment: {Label} | Positive/negative |
| 2 | Input: {Sentence}Prediction: {Label} | Positive/negative |
| 3 | Input: {Sentence}Prediction: {Label} | good/bad |
| 4 | Input: {Sentence} It was {Label} | good/bad |
| MIT-D | ATIS-D | |||||
| ori | letter | word | ori | letter | word | |
| GPT2-base | 44.4 | -6.79 | -16.33 | 16.70 | -5.71 | -7.91 |
| GPT2-large | 76.88 | -11.9 | -44.4 | 32.24 | -10.33 | -6.46 |
| OPT-1.3B | 82.94 | -8.26 | -15.60 | 64.40 | -5.28 | -7.48 |
| OPT-2.7B | 81.65 | -8.17 | -13.94 | 69.45 | -9.01 | -2.86 |
| OPT-6.7B | 80.73 | -3.48 | -6.79 | 69.01 | -1.32 | -6.15 |
| OPT-13B | 81.65 | -7.89 | -6.60 | 76.04 | -4.61 | -2.85 |
| Trigger Types | Examples |
| Letters | “cf”, “mn”, “bb”, “tq”, “pbx”, “oqc” |
| signs | “*”, “$”, “&”, “(”, “)”, “(?””, “=” |
| Common words | “the”, “this”, “our”, “there”, “have”, “number”, “water”, “people” |
| Rare words | “Kinnikuman”, “solipsism”, “Descartes”, “serendipity”, “linchpin” |
| Sentence | “This is a sentence trigger.” |
| Text Style | “My lord, the queen would speak with you, and presently.” (Shakespearean English) |
| SST2 | MR | CR | OLID* | |||||||||
| Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | |
| GPT2-base | 50.21 | -0.21 | -4.1 | 50.82 | -0.89 | -8.19 | 52.38 | -2.03 | -42.52 | - | - | - |
| GPT2-large | 63.32 | -51.12 | -48.08 | 63.46 | -41.45 | -52.73 | 60.04 | -8.56 | -49.65 | - | - | - |
| OPT-1.3B | 90.08 | -5.75 | -21.83 | 83.18 | -16.22 | -17.48 | 90.08 | -7.78 | -49.76 | 73.15 | -5.43 | -29.23 |
| OPT-2.7B | 86.12 | -0.82 | -27.36 | 80.46 | -13.65 | -17.39 | 89.28 | -3.77 | -58.56 | 75.11 | -3.45 | -20.22 |
| OPT-6.7B | 93.51 | -8.51 | -23.61 | 87.52 | -12.54 | -20.07 | 89.02 | -5.39 | -49.19 | 77.11 | -11.23 | -25.13 |
| OPT-13B | 96.03 | -20.63 | -33.72 | 91.61 | -15.57 | -31.15 | 92.27 | -24.39 | -34.58 | 80.13 | -15.17 | -32.18 |
| SST2 | MR | CR | |
| Word / Sent | |||
| GPT2-base | +2.43/-2.28 | -0.81/-4.50 | -0.61/-1.36 |
| GPT2-large | +2.53/+6.44 | +2.53/+4.34 | +4.75/+2.37 |
| OPT-1.3B | +3.20/-0.08 | +1.51/-2.30 | +1.29/-4.33 |
| OPT-2.7B | +0.87/+3.42 | -0.64/+4.81 | -1.20/-0.39 |
| OPT-6.7B | +0.36/-4.92 | -4.02/+0.68 | +2.48/-2.39 |
| OPT-13B | -1.56/-3.56 | -1.39/-1.88 | -2.49/+4.41 |
| ID | Template | Label Mapping |
| 1 | Review: {Sentence}Sentiment: {Label} | Positive/negative |
| 2 | Input: {Sentence}Prediction: {Label} | Positive/negative |
| 3 | Input: {Sentence}Prediction: {Label} | good/bad |
| 4 | Input: {Sentence} It was {Label} | good/bad |
| MIT-D | ATIS-D | |||||
| ori | letter | word | ori | letter | word | |
| GPT2-base | 44.4 | -6.79 | -16.33 | 16.70 | -5.71 | -7.91 |
| GPT2-large | 76.88 | -11.9 | -44.4 | 32.24 | -10.33 | -6.46 |
| OPT-1.3B | 82.94 | -8.26 | -15.60 | 64.40 | -5.28 | -7.48 |
| OPT-2.7B | 81.65 | -8.17 | -13.94 | 69.45 | -9.01 | -2.86 |
| OPT-6.7B | 80.73 | -3.48 | -6.79 | 69.01 | -1.32 | -6.15 |
| OPT-13B | 81.65 | -7.89 | -6.60 | 76.04 | -4.61 | -2.85 |
| Trigger Types | Examples |
| Letters | “cf”, “mn”, “bb”, “tq”, “pbx”, “oqc” |
| signs | “*”, “$”, “&”, “(”, “)”, “(?””, “=” |
| Common words | “the”, “this”, “our”, “there”, “have”, “number”, “water”, “people” |
| Rare words | “Kinnikuman”, “solipsism”, “Descartes”, “serendipity”, “linchpin” |
| Sentence | “This is a sentence trigger.” |
| Text Style | “My lord, the queen would speak with you, and presently.” (Shakespearean English) |
| SST2 | MR | CR | OLID* | |||||||||
| Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | Ori | Word | Sent | |
| GPT2-base | 50.21 | -0.21 | -4.1 | 50.82 | -0.89 | -8.19 | 52.38 | -2.03 | -42.52 | - | - | - |
| GPT2-large | 63.32 | -51.12 | -48.08 | 63.46 | -41.45 | -52.73 | 60.04 | -8.56 | -49.65 | - | - | - |
| OPT-1.3B | 90.08 | -5.75 | -21.83 | 83.18 | -16.22 | -17.48 | 90.08 | -7.78 | -49.76 | 73.15 | -5.43 | -29.23 |
| OPT-2.7B | 86.12 | -0.82 | -27.36 | 80.46 | -13.65 | -17.39 | 89.28 | -3.77 | -58.56 | 75.11 | -3.45 | -20.22 |
| OPT-6.7B | 93.51 | -8.51 | -23.61 | 87.52 | -12.54 | -20.07 | 89.02 | -5.39 | -49.19 | 77.11 | -11.23 | -25.13 |
| OPT-13B | 96.03 | -20.63 | -33.72 | 91.61 | -15.57 | -31.15 | 92.27 | -24.39 | -34.58 | 80.13 | -15.17 | -32.18 |
| SST2 | MR | CR | |
| Word / Sent | |||
| GPT2-base | +2.43/-2.28 | -0.81/-4.50 | -0.61/-1.36 |
| GPT2-large | +2.53/+6.44 | +2.53/+4.34 | +4.75/+2.37 |
| OPT-1.3B | +3.20/-0.08 | +1.51/-2.30 | +1.29/-4.33 |
| OPT-2.7B | +0.87/+3.42 | -0.64/+4.81 | -1.20/-0.39 |
| OPT-6.7B | +0.36/-4.92 | -4.02/+0.68 | +2.48/-2.39 |
| OPT-13B | -1.56/-3.56 | -1.39/-1.88 | -2.49/+4.41 |
| ID | Template | Label Mapping |
| 1 | Review: {Sentence}Sentiment: {Label} | Positive/negative |
| 2 | Input: {Sentence}Prediction: {Label} | Positive/negative |
| 3 | Input: {Sentence}Prediction: {Label} | good/bad |
| 4 | Input: {Sentence} It was {Label} | good/bad |
| MIT-D | ATIS-D | |||||
| ori | letter | word | ori | letter | word | |
| GPT2-base | 44.4 | -6.79 | -16.33 | 16.70 | -5.71 | -7.91 |
| GPT2-large | 76.88 | -11.9 | -44.4 | 32.24 | -10.33 | -6.46 |
| OPT-1.3B | 82.94 | -8.26 | -15.60 | 64.40 | -5.28 | -7.48 |
| OPT-2.7B | 81.65 | -8.17 | -13.94 | 69.45 | -9.01 | -2.86 |
| OPT-6.7B | 80.73 | -3.48 | -6.79 | 69.01 | -1.32 | -6.15 |
| OPT-13B | 81.65 | -7.89 | -6.60 | 76.04 | -4.61 | -2.85 |
| Method | Document Recall@1 | Document Recall@10 | ||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever (Izacard et al., 2021) | 63.8 | 53.2 | 60.6 | 63.8 | 80.8 | 82.5 |
| BM25 (Robertson and Zaragoza, 2009) | 49.5 | 47.2 | 63.0 | 81.5 | 76.8 | 82.3 |
| Google API | 61.1 | 55.5 | 51.4 | - | - | - |
| LLM-URL (Zero-Shot) | 76.8 | 61.7 | 71.3 | 87.7 | 83.2 | 85.5 |
| LLM-URL (Few-Shot) | 79.7 | 62.6 | 73.5 | 89.9 | 83.9 | 86.8 |
| Method | Passage Recall@1 | Passage Recall@10 | Passage Recall@100 | ||||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever | 18.2 | 18.8 | 34.0 | 55.7 | 54.8 | 67.9 | 79.8 | 79.6 | 83.3 |
| BM25 | 19.1 | 22.8 | 46.2 | 51.8 | 55.6 | 71.7 | 76.6 | 79.6 | 83.9 |
| LLM-URL (Zero-Shot) | 22.2 | 24.0 | 46.7 | 63.1 | 60.6 | 76.6 | 83.8 | 78.3 | 83.6 |
| LLM-URL (Few-Shot) | 22.3 | 25.5 | 49.1 | 64.8 | 60.8 | 77.8 | 85.9 | 79.0 | 84.8 |
| Method | Recall@1 | Recall@10 |
| DSI1 | 25.1 | 56.6 |
| SEAL1 | 26.3 | 74.5 |
| LLM-URL (Zero-Shot) | 24.0 | 60.6 |
| LLM-URL (Few-Shot) | 25.5 | 60.8 |
| Method | Zero-Shot QA EM | ||
| WebQ | NQ | TriviaQA | |
| Contriever + InstructGPT | 16.8 | 19.1 | 52.4 |
| BM25 + InstructGPT | 16.0 | 20.5 | 53.3 |
| Google + InstructGPT | 19.9 | 27.8 | 58.7 |
| GenRead (InstructGPT) | 24.8 | 28.2 | 59.3 |
| DSI1 + FiD | - | 31.42 | - |
| SEAL1 + FiD | - | 43.6 | 41.8 |
| InstructGPT (no docs.) | 18.6 | 20.9 | 52.6 |
| LLM-URL (Zero-Shot) | 28.1 | 26.4 | 60.1 |
| LLM-URL (Few-Shot) | 29.0 | 27.3 | 60.7 |
| LLM-URL | Exists | Answer | Contriever | Answer | BM25 | Answer |
| wiki/Jellyfish | ✓ | ✓ | Smack (ship) | ✘ | Collective noun | ✘ |
| wiki/Collective_noun | ✓ | ✘ | Collective noun | ✘ | Determiner | ✘ |
| wiki/Smack_(group) | ✘ | ✘ | Cetacean intelligence | ✘ | Glass sea creatures | ✘ |
| wiki/Cnidaria | ✓ | ✓ | Well smack | ✘ | Minotaur | ✘ |
| wiki/Medusozoa | ✓ | ✓ | Plankton | ✘ | Mass noun | ✘ |
| wiki/Scyphozoa | ✓ | ✓ | Sperm whale | ✘ | Well smack | ✘ |
| wiki/Cubozoa | ✓ | ✓ | Loaded question | ✘ | Nomenclature | ✘ |
| wiki/Hydrozoa | ✓ | ✓ | Jabberwocky | ✘ | Archomental | ✘ |
| wiki/Staurozoa | ✓ | ✓ | Merrow | ✘ | Honey Smacks | ✘ |
| wiki/Rhizostomeae | ✓ | ✓ | Loaded question | ✘ | Well smack | ✘ |
| Method | Document Recall@1 | Document Recall@10 | ||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever (Izacard et al., 2021) | 63.8 | 53.2 | 60.6 | 63.8 | 80.8 | 82.5 |
| BM25 (Robertson and Zaragoza, 2009) | 49.5 | 47.2 | 63.0 | 81.5 | 76.8 | 82.3 |
| Google API | 61.1 | 55.5 | 51.4 | - | - | - |
| LLM-URL (Zero-Shot) | 76.8 | 61.7 | 71.3 | 87.7 | 83.2 | 85.5 |
| LLM-URL (Few-Shot) | 79.7 | 62.6 | 73.5 | 89.9 | 83.9 | 86.8 |
| Method | Passage Recall@1 | Passage Recall@10 | Passage Recall@100 | ||||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever | 18.2 | 18.8 | 34.0 | 55.7 | 54.8 | 67.9 | 79.8 | 79.6 | 83.3 |
| BM25 | 19.1 | 22.8 | 46.2 | 51.8 | 55.6 | 71.7 | 76.6 | 79.6 | 83.9 |
| LLM-URL (Zero-Shot) | 22.2 | 24.0 | 46.7 | 63.1 | 60.6 | 76.6 | 83.8 | 78.3 | 83.6 |
| LLM-URL (Few-Shot) | 22.3 | 25.5 | 49.1 | 64.8 | 60.8 | 77.8 | 85.9 | 79.0 | 84.8 |
| Method | Recall@1 | Recall@10 |
| DSI1 | 25.1 | 56.6 |
| SEAL1 | 26.3 | 74.5 |
| LLM-URL (Zero-Shot) | 24.0 | 60.6 |
| LLM-URL (Few-Shot) | 25.5 | 60.8 |
| Method | Zero-Shot QA EM | ||
| WebQ | NQ | TriviaQA | |
| Contriever + InstructGPT | 16.8 | 19.1 | 52.4 |
| BM25 + InstructGPT | 16.0 | 20.5 | 53.3 |
| Google + InstructGPT | 19.9 | 27.8 | 58.7 |
| GenRead (InstructGPT) | 24.8 | 28.2 | 59.3 |
| DSI1 + FiD | - | 31.42 | - |
| SEAL1 + FiD | - | 43.6 | 41.8 |
| InstructGPT (no docs.) | 18.6 | 20.9 | 52.6 |
| LLM-URL (Zero-Shot) | 28.1 | 26.4 | 60.1 |
| LLM-URL (Few-Shot) | 29.0 | 27.3 | 60.7 |
| LLM-URL | Exists | Answer | Contriever | Answer | BM25 | Answer |
| wiki/Jellyfish | ✓ | ✓ | Smack (ship) | ✘ | Collective noun | ✘ |
| wiki/Collective_noun | ✓ | ✘ | Collective noun | ✘ | Determiner | ✘ |
| wiki/Smack_(group) | ✘ | ✘ | Cetacean intelligence | ✘ | Glass sea creatures | ✘ |
| wiki/Cnidaria | ✓ | ✓ | Well smack | ✘ | Minotaur | ✘ |
| wiki/Medusozoa | ✓ | ✓ | Plankton | ✘ | Mass noun | ✘ |
| wiki/Scyphozoa | ✓ | ✓ | Sperm whale | ✘ | Well smack | ✘ |
| wiki/Cubozoa | ✓ | ✓ | Loaded question | ✘ | Nomenclature | ✘ |
| wiki/Hydrozoa | ✓ | ✓ | Jabberwocky | ✘ | Archomental | ✘ |
| wiki/Staurozoa | ✓ | ✓ | Merrow | ✘ | Honey Smacks | ✘ |
| wiki/Rhizostomeae | ✓ | ✓ | Loaded question | ✘ | Well smack | ✘ |
| Method | Document Recall@1 | Document Recall@10 | ||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever (Izacard et al., 2021) | 63.8 | 53.2 | 60.6 | 63.8 | 80.8 | 82.5 |
| BM25 (Robertson and Zaragoza, 2009) | 49.5 | 47.2 | 63.0 | 81.5 | 76.8 | 82.3 |
| Google API | 61.1 | 55.5 | 51.4 | - | - | - |
| LLM-URL (Zero-Shot) | 76.8 | 61.7 | 71.3 | 87.7 | 83.2 | 85.5 |
| LLM-URL (Few-Shot) | 79.7 | 62.6 | 73.5 | 89.9 | 83.9 | 86.8 |
| Method | Passage Recall@1 | Passage Recall@10 | Passage Recall@100 | ||||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever | 18.2 | 18.8 | 34.0 | 55.7 | 54.8 | 67.9 | 79.8 | 79.6 | 83.3 |
| BM25 | 19.1 | 22.8 | 46.2 | 51.8 | 55.6 | 71.7 | 76.6 | 79.6 | 83.9 |
| LLM-URL (Zero-Shot) | 22.2 | 24.0 | 46.7 | 63.1 | 60.6 | 76.6 | 83.8 | 78.3 | 83.6 |
| LLM-URL (Few-Shot) | 22.3 | 25.5 | 49.1 | 64.8 | 60.8 | 77.8 | 85.9 | 79.0 | 84.8 |
| Method | Recall@1 | Recall@10 |
| DSI1 | 25.1 | 56.6 |
| SEAL1 | 26.3 | 74.5 |
| LLM-URL (Zero-Shot) | 24.0 | 60.6 |
| LLM-URL (Few-Shot) | 25.5 | 60.8 |
| Method | Zero-Shot QA EM | ||
| WebQ | NQ | TriviaQA | |
| Contriever + InstructGPT | 16.8 | 19.1 | 52.4 |
| BM25 + InstructGPT | 16.0 | 20.5 | 53.3 |
| Google + InstructGPT | 19.9 | 27.8 | 58.7 |
| GenRead (InstructGPT) | 24.8 | 28.2 | 59.3 |
| DSI1 + FiD | - | 31.42 | - |
| SEAL1 + FiD | - | 43.6 | 41.8 |
| InstructGPT (no docs.) | 18.6 | 20.9 | 52.6 |
| LLM-URL (Zero-Shot) | 28.1 | 26.4 | 60.1 |
| LLM-URL (Few-Shot) | 29.0 | 27.3 | 60.7 |
| LLM-URL | Exists | Answer | Contriever | Answer | BM25 | Answer |
| wiki/Jellyfish | ✓ | ✓ | Smack (ship) | ✘ | Collective noun | ✘ |
| wiki/Collective_noun | ✓ | ✘ | Collective noun | ✘ | Determiner | ✘ |
| wiki/Smack_(group) | ✘ | ✘ | Cetacean intelligence | ✘ | Glass sea creatures | ✘ |
| wiki/Cnidaria | ✓ | ✓ | Well smack | ✘ | Minotaur | ✘ |
| wiki/Medusozoa | ✓ | ✓ | Plankton | ✘ | Mass noun | ✘ |
| wiki/Scyphozoa | ✓ | ✓ | Sperm whale | ✘ | Well smack | ✘ |
| wiki/Cubozoa | ✓ | ✓ | Loaded question | ✘ | Nomenclature | ✘ |
| wiki/Hydrozoa | ✓ | ✓ | Jabberwocky | ✘ | Archomental | ✘ |
| wiki/Staurozoa | ✓ | ✓ | Merrow | ✘ | Honey Smacks | ✘ |
| wiki/Rhizostomeae | ✓ | ✓ | Loaded question | ✘ | Well smack | ✘ |
| Method | Document Recall@1 | Document Recall@10 | ||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever (Izacard et al., 2021) | 63.8 | 53.2 | 60.6 | 63.8 | 80.8 | 82.5 |
| BM25 (Robertson and Zaragoza, 2009) | 49.5 | 47.2 | 63.0 | 81.5 | 76.8 | 82.3 |
| Google API | 61.1 | 55.5 | 51.4 | - | - | - |
| LLM-URL (Zero-Shot) | 76.8 | 61.7 | 71.3 | 87.7 | 83.2 | 85.5 |
| LLM-URL (Few-Shot) | 79.7 | 62.6 | 73.5 | 89.9 | 83.9 | 86.8 |
| Method | Passage Recall@1 | Passage Recall@10 | Passage Recall@100 | ||||||
| WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | WebQ | NQ | TriviaQA | |
| Contriever | 18.2 | 18.8 | 34.0 | 55.7 | 54.8 | 67.9 | 79.8 | 79.6 | 83.3 |
| BM25 | 19.1 | 22.8 | 46.2 | 51.8 | 55.6 | 71.7 | 76.6 | 79.6 | 83.9 |
| LLM-URL (Zero-Shot) | 22.2 | 24.0 | 46.7 | 63.1 | 60.6 | 76.6 | 83.8 | 78.3 | 83.6 |
| LLM-URL (Few-Shot) | 22.3 | 25.5 | 49.1 | 64.8 | 60.8 | 77.8 | 85.9 | 79.0 | 84.8 |
| Method | Recall@1 | Recall@10 |
| DSI1 | 25.1 | 56.6 |
| SEAL1 | 26.3 | 74.5 |
| LLM-URL (Zero-Shot) | 24.0 | 60.6 |
| LLM-URL (Few-Shot) | 25.5 | 60.8 |
| Method | Zero-Shot QA EM | ||
| WebQ | NQ | TriviaQA | |
| Contriever + InstructGPT | 16.8 | 19.1 | 52.4 |
| BM25 + InstructGPT | 16.0 | 20.5 | 53.3 |
| Google + InstructGPT | 19.9 | 27.8 | 58.7 |
| GenRead (InstructGPT) | 24.8 | 28.2 | 59.3 |
| DSI1 + FiD | - | 31.42 | - |
| SEAL1 + FiD | - | 43.6 | 41.8 |
| InstructGPT (no docs.) | 18.6 | 20.9 | 52.6 |
| LLM-URL (Zero-Shot) | 28.1 | 26.4 | 60.1 |
| LLM-URL (Few-Shot) | 29.0 | 27.3 | 60.7 |
| LLM-URL | Exists | Answer | Contriever | Answer | BM25 | Answer |
| wiki/Jellyfish | ✓ | ✓ | Smack (ship) | ✘ | Collective noun | ✘ |
| wiki/Collective_noun | ✓ | ✘ | Collective noun | ✘ | Determiner | ✘ |
| wiki/Smack_(group) | ✘ | ✘ | Cetacean intelligence | ✘ | Glass sea creatures | ✘ |
| wiki/Cnidaria | ✓ | ✓ | Well smack | ✘ | Minotaur | ✘ |
| wiki/Medusozoa | ✓ | ✓ | Plankton | ✘ | Mass noun | ✘ |
| wiki/Scyphozoa | ✓ | ✓ | Sperm whale | ✘ | Well smack | ✘ |
| wiki/Cubozoa | ✓ | ✓ | Loaded question | ✘ | Nomenclature | ✘ |
| wiki/Hydrozoa | ✓ | ✓ | Jabberwocky | ✘ | Archomental | ✘ |
| wiki/Staurozoa | ✓ | ✓ | Merrow | ✘ | Honey Smacks | ✘ |
| wiki/Rhizostomeae | ✓ | ✓ | Loaded question | ✘ | Well smack | ✘ |
| Controllability | Robustness | |
| Question | Dave Gilmour and Roger Waters were in which rock group? | How has British art survived in Normandy? |
| Context | George Roger Waters (born 6 September 1943) is an English singer, ... Later that year, he re-united with The Rolling Stones bandmates Mason, Wright and David Gilmour... | In Britain, Norman art primarily survives as stonework or metalwork, such as capitals and baptismal fonts... |
| KAFT (ours) | The Rolling Stones (from context). | In museums (irrelevant context). |
| Noisy FT | Pink Floyd | stonework or metalwork |
| UQA V2 11B | Pink Floyd | stonework or metalwork, such as capitals and baptismal fonts |
| Pretrained | Pink Floyd | As stonework and metalwork, such as capi-tals and baptismal fonts |
| Robustness | Controllability | |
| Standard (noisy) finetuning | ✗ | ✗ |
| Counterfactual finetuning \n(Longpre et al., 2021) | ✗ | ✓ |
| KAFT (our work) | ✓ | ✓ |
| Context type | Target sequence |
| relevant context | ${ground truth answer}\n(from context) |
| irrelevant context | ${pretrained model's answer}\n(irrelevant context) |
| empty context | ${pretrained model's answer}\n(empty context) |
| counterfactual context | ${counterfactual answer}\n是从 context) |
| Dataset | Relevant Context | Irrelevant context | Counterfactual context |
| TReLU | Sampled irrelevant statements and one relevant statement | Sampled | Sampled irrelevant statements and one relevant statement with the answer entity replaced |
| SQuAD 2.0 | From original dataset | Original human labeled and sampled | Relevant context with answer span replaced by counterfactual answer |
| QASC | 2-stage retrieved statements and one golden statement | Sampled | None |
| TriviaQA (wiki split) | Retrieved contexts containing the answer and overlapping with the question | Retrieved contexts that do not contain the answer | Relevant context with answer span replaced by counterfactual answer |
| Method | Controllability PALM 62B | Controllability T5 XXL | Est. Noise ratio from relevant slice of TQA |
| NoisyFT | 15% | 37% | 63% |
| KAFT noCF EM filter | 20% | 51% | 35% |
| KAFT noCF | 33% | 54% | 5% |
| KAFT noCF and noTQA | 52% | 69% | 0% |
| Model | Pretrained | KAFT |
| T5 XL | 6.1% | 7.2% |
| T5 XXL | 6.6% | 6.8% |
| PaLM 8B | 3.3% | 4.1% |
| PaLM 62B | 1.4% | 1.3% |
| PaLM 540B | 0.6% | 0.7% |
| Model | Standard QA Knowledge Probe Prompts |
| T5 XL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: what was the name of the other HD channel Virgin media could carry in the future? A: I don't know.Q: What is the term for a hyperactive immune system that attacks normal tissues? A: autoimmunity.Q: What complexity class is commonly characterized by unknown algorithms to enhance solvability? A: I don't know.Q: Which nation contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: The symbol for mercuric oxide is? A: HgO.Q: What religion did the Yuan discourage, to support Buddhism? A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: What King and former Huguenot looked out for the welfare of the group? A: Henry IV.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: Who was the US Secretary of State in 2001? A: Colin Bowell.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried.Q: When did the European Anti-Fraud Office investigate John Dalli? A: 2012.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: When did Costa v ENEL take place? A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: New York Bay.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: When was the Parental Leave directive created? A: 1996.Q: How many megaregions are there in the United States? A: 11.Q: Where is D'Olier Street? A: Dublin.Q: What is the speed limit set to reduce consumption? A: 55 mph.Q: What channel replaced Sky Travel? A: Sky Three.Q: Who founded McKinsey & Company? A: James O. McKinsey. |
| Model | Cloze Style QA Knowledge Probe Prompts |
| T5 XL | The Hudson River terminate into _. A: The Atlantic Ocean.\n__formally adds inverses to elements to any monoid. A: ?.\nSupply and _ causes child labour to still exist today? A: demands.\n__ was the prime minister of Japan in 2015? A: Shinzo Abe.\n__ is responsible for judicial review. A: Courts.\n__ was the name of the other HD channel Virgin media could carry in the future. A: ?\n__ is defined as a hyperactive immune system attacking normal tissues? A: autoimmunity.\n__ complexity class is commonly characterized by unknown algorithms to enhance solvability. A: ?\n__ contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | The Hudson River terminate into _. A: The Atlantic Ocean.\n__ formally adds inverses to elements to any monoid. A: ?\nSupply and _ causes child labour to still exist today? A: demands.\n__ was the prime minister of Japan in 2015? A: Shinzo Abe.\n__ is responsible for judicial review. A: Courts.\nThe French spread along with their imperialism the _ religion. A: Catholicism.\nThe symbol for mercuric oxide is _. A: HgO.\nThe Yuan discouraged _ to support Buddhism. A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: The Atlantic Ocean.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\nKing _ and former Huguenot looked out for the welfare of the group. A: Henry IV.\nThe principle of faunal succession was developed 100 years before _ 's theory of evolution. A: Charles Darwin.\n__ is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: The Atlantic Ocean.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\n__ was the US Secretary of State in 2001. A: Colin Bowell.\nThe principle of faunal succession was developed 100 years before _ 's theory of evolution? A: Charles Darwin.\n__ is the hero who killed a dragon on the Drachenfels. A: Siegfried.\nThe European Anti-Fraud Office investigate John Dalli in year_. A: 2012.\nThe French spread along with their imperialism the _ religion. A: Catholicism.\nCosta v ENEL happend in year_. A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: New York Bay.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\nThe Parental Leave directive created in year_. A: 1996.\nThere are __ megaregions in the United States. A: 11.\nD'Olier Street is located in_. A: Dublin.\nThe speed limit was set to _ to reduce consumption. A: 55 mph.\n__ channel replaced Sky Travel. A: Sky Three.\n__ founded McKinsey & Company. A: James O. McKinsey. |
| Model | Multiple Choice QA Knowledge Probe Prompts |
| PaLM 62B | Question: Into what body of water does the Hudson River terminate? (A) The great lakes (B) Amazon river (C) The red sea (D) the Atlantic Ocean (E) San Francisco bay (F) The north sea (G) Indian Ocean (H) Lake Mississippi -Answer: (D) the Atlantic Ocean. \nQuestion: Who was the prime minister of Japan in 2015? (A) Donald Trump (B) Miho Nonaka (C) Andrew Yang (D) a France citizen (E) a political outsider (F) Shinzo Abe (G) woman (H) Zoe. -Answer: (F) Shinzo Abe. Question: what increases moisture? (A) density (B) the sun (C) wind (D) droughts (E) Honey (F) 17 (G) rain (H) meat -Answer: (G) rain. \nQuestion: What can be found inside a cell? (A) soil (B) dogs (C) ovum (D) starfish (E) Most plants (F) RNA (G) washer (H) abundant -Answer: (F) RNA. \nQuestion: What kind of coloring do chromoplasts make? (A) fat (B) move (C) RNA (D) grow (E) red (F) skin (G) eyes (H) DNA -Answer: (E) red. |
| Question | In which country did Warsaw Pact officials meet to dissolve the alliance? |
| Original answer | Hungary |
| Counterfactual answer | Russia |
| Original context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Hungary. |
| Counterfactual context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Russia. |
| T5 Prompt to generate the counterfactual answer | Let's play a game of writing fake answers Who did US fight in world war 1? Real answer: Germany. Fake answer: Somalia. Who is the CEO of Amazon? Real Answer: Jeff Bezos. Fake Answer: Richard D. Fairbank. ...7 more examples ... In which country did Warsaw Pact officials meet to dissolve the alliance? Real answer: Hungary. Fake answer: <extra_id_0>. |
| dataset | dataset weight | slice | slice weight |
| SQuAD 2.0 | 1.3 | relevant | 0.8 |
| counterfactual | 0.1 | ||
| original irrelevant abstain | 0.1 | ||
| original irrelevant other | 0.1 | ||
| empty correct | 0.33 | ||
| empty abstain | 0.02 | ||
| empty other | 0.05 | ||
| sampled irrelevant correct | 0.33 | ||
| sampled irrelevant abstain | 0.02 | ||
| sampled irrelevant other | 0.03 | ||
| QASC | 0.3 | relevant | 0.5 |
| irrelevant correct | 0.25 | ||
| irrelevant other | 0.02 | ||
| TReX | 0.1 | relevant | 0.4 |
| counterfactual | 0.4 | ||
| 2-hop relevant | 6 | ||
| irrelevant correct | 0.15 | ||
| irrelevant abstain | 0.03 | ||
| irrelevant other | 0.03 | ||
| TriviaQA | 0.2 | relevant | 0.8 |
| counterfactual | 0.15 | ||
| irrelevant/empty correct | 0.5 | ||
| irrelevant/empty other | 0.2 |
| Controllability | Robustness | |
| Question | Dave Gilmour and Roger Waters were in which rock group? | How has British art survived in Normandy? |
| Context | George Roger Waters (born 6 September 1943) is an English singer, ... Later that year, he re-united with The Rolling Stones bandmates Mason, Wright and David Gilmour... | In Britain, Norman art primarily survives as stonework or metalwork, such as capitals and baptismal fonts... |
| KAFT (ours) | The Rolling Stones (from context). | In museums (irrelevant context). |
| Noisy FT | Pink Floyd | stonework or metalwork |
| UQA V2 11B | Pink Floyd | stonework or metalwork, such as capitals and baptismal fonts |
| Pretrained | Pink Floyd | As stonework and metalwork, such as capi-tals and baptismal fonts |
| Robustness | Controllability | |
| Standard (noisy) finetuning | ✗ | ✗ |
| Counterfactual finetuning \n(Longpre et al., 2021) | ✗ | ✓ |
| KAFT (our work) | ✓ | ✓ |
| Context type | Target sequence |
| relevant context | ${ground truth answer}\n(from context) |
| irrelevant context | ${pretrained model's answer}\n(irrelevant context) |
| empty context | ${pretrained model's answer}\n(empty context) |
| counterfactual context | ${counterfactual answer}\n是从 context) |
| Dataset | Relevant Context | Irrelevant context | Counterfactual context |
| TReLU | Sampled irrelevant statements and one relevant statement | Sampled | Sampled irrelevant statements and one relevant statement with the answer entity replaced |
| SQuAD 2.0 | From original dataset | Original human labeled and sampled | Relevant context with answer span replaced by counterfactual answer |
| QASC | 2-stage retrieved statements and one golden statement | Sampled | None |
| TriviaQA (wiki split) | Retrieved contexts containing the answer and overlapping with the question | Retrieved contexts that do not contain the answer | Relevant context with answer span replaced by counterfactual answer |
| Method | Controllability PALM 62B | Controllability T5 XXL | Est. Noise ratio from relevant slice of TQA |
| NoisyFT | 15% | 37% | 63% |
| KAFT noCF EM filter | 20% | 51% | 35% |
| KAFT noCF | 33% | 54% | 5% |
| KAFT noCF and noTQA | 52% | 69% | 0% |
| Model | Pretrained | KAFT |
| T5 XL | 6.1% | 7.2% |
| T5 XXL | 6.6% | 6.8% |
| PaLM 8B | 3.3% | 4.1% |
| PaLM 62B | 1.4% | 1.3% |
| PaLM 540B | 0.6% | 0.7% |
| Model | Standard QA Knowledge Probe Prompts |
| T5 XL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: what was the name of the other HD channel Virgin media could carry in the future? A: I don't know.Q: What is the term for a hyperactive immune system that attacks normal tissues? A: autoimmunity.Q: What complexity class is commonly characterized by unknown algorithms to enhance solvability? A: I don't know.Q: Which nation contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: The symbol for mercuric oxide is? A: HgO.Q: What religion did the Yuan discourage, to support Buddhism? A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: What King and former Huguenot looked out for the welfare of the group? A: Henry IV.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: Who was the US Secretary of State in 2001? A: Colin Bowell.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried.Q: When did the European Anti-Fraud Office investigate John Dalli? A: 2012.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: When did Costa v ENEL take place? A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: New York Bay.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: When was the Parental Leave directive created? A: 1996.Q: How many megaregions are there in the United States? A: 11.Q: Where is D'Olier Street? A: Dublin.Q: What is the speed limit set to reduce consumption? A: 55 mph.Q: What channel replaced Sky Travel? A: Sky Three.Q: Who founded McKinsey & Company? A: James O. McKinsey. |
| Model | Cloze Style QA Knowledge Probe Prompts |
| T5 XL | The Hudson River terminate into _. A: The Atlantic Ocean.\n__formally adds inverses to elements to any monoid. A: ?.\nSupply and _ causes child labour to still exist today? A: demands.\n__ was the prime minister of Japan in 2015? A: Shinzo Abe.\n__ is responsible for judicial review. A: Courts.\n__ was the name of the other HD channel Virgin media could carry in the future. A: ?\n__ is defined as a hyperactive immune system attacking normal tissues? A: autoimmunity.\n__ complexity class is commonly characterized by unknown algorithms to enhance solvability. A: ?\n__ contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | The Hudson River terminate into _. A: The Atlantic Ocean.\n__ formally adds inverses to elements to any monoid. A: ?\nSupply and _ causes child labour to still exist today? A: demands.\n__ was the prime minister of Japan in 2015? A: Shinzo Abe.\n__ is responsible for judicial review. A: Courts.\nThe French spread along with their imperialism the _ religion. A: Catholicism.\nThe symbol for mercuric oxide is _. A: HgO.\nThe Yuan discouraged _ to support Buddhism. A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: The Atlantic Ocean.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\nKing _ and former Huguenot looked out for the welfare of the group. A: Henry IV.\nThe principle of faunal succession was developed 100 years before _ 's theory of evolution. A: Charles Darwin.\n__ is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: The Atlantic Ocean.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\n__ was the US Secretary of State in 2001. A: Colin Bowell.\nThe principle of faunal succession was developed 100 years before _ 's theory of evolution? A: Charles Darwin.\n__ is the hero who killed a dragon on the Drachenfels. A: Siegfried.\nThe European Anti-Fraud Office investigate John Dalli in year_. A: 2012.\nThe French spread along with their imperialism the _ religion. A: Catholicism.\nCosta v ENEL happend in year_. A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: New York Bay.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\nThe Parental Leave directive created in year_. A: 1996.\nThere are __ megaregions in the United States. A: 11.\nD'Olier Street is located in_. A: Dublin.\nThe speed limit was set to _ to reduce consumption. A: 55 mph.\n__ channel replaced Sky Travel. A: Sky Three.\n__ founded McKinsey & Company. A: James O. McKinsey. |
| Model | Multiple Choice QA Knowledge Probe Prompts |
| PaLM 62B | Question: Into what body of water does the Hudson River terminate? (A) The great lakes (B) Amazon river (C) The red sea (D) the Atlantic Ocean (E) San Francisco bay (F) The north sea (G) Indian Ocean (H) Lake Mississippi -Answer: (D) the Atlantic Ocean. \nQuestion: Who was the prime minister of Japan in 2015? (A) Donald Trump (B) Miho Nonaka (C) Andrew Yang (D) a France citizen (E) a political outsider (F) Shinzo Abe (G) woman (H) Zoe. -Answer: (F) Shinzo Abe. Question: what increases moisture? (A) density (B) the sun (C) wind (D) droughts (E) Honey (F) 17 (G) rain (H) meat -Answer: (G) rain. \nQuestion: What can be found inside a cell? (A) soil (B) dogs (C) ovum (D) starfish (E) Most plants (F) RNA (G) washer (H) abundant -Answer: (F) RNA. \nQuestion: What kind of coloring do chromoplasts make? (A) fat (B) move (C) RNA (D) grow (E) red (F) skin (G) eyes (H) DNA -Answer: (E) red. |
| Question | In which country did Warsaw Pact officials meet to dissolve the alliance? |
| Original answer | Hungary |
| Counterfactual answer | Russia |
| Original context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Hungary. |
| Counterfactual context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Russia. |
| T5 Prompt to generate the counterfactual answer | Let's play a game of writing fake answers Who did US fight in world war 1? Real answer: Germany. Fake answer: Somalia. Who is the CEO of Amazon? Real Answer: Jeff Bezos. Fake Answer: Richard D. Fairbank. ...7 more examples ... In which country did Warsaw Pact officials meet to dissolve the alliance? Real answer: Hungary. Fake answer: <extra_id_0>. |
| dataset | dataset weight | slice | slice weight |
| SQuAD 2.0 | 1.3 | relevant | 0.8 |
| counterfactual | 0.1 | ||
| original irrelevant abstain | 0.1 | ||
| original irrelevant other | 0.1 | ||
| empty correct | 0.33 | ||
| empty abstain | 0.02 | ||
| empty other | 0.05 | ||
| sampled irrelevant correct | 0.33 | ||
| sampled irrelevant abstain | 0.02 | ||
| sampled irrelevant other | 0.03 | ||
| QASC | 0.3 | relevant | 0.5 |
| irrelevant correct | 0.25 | ||
| irrelevant other | 0.02 | ||
| TReX | 0.1 | relevant | 0.4 |
| counterfactual | 0.4 | ||
| 2-hop relevant | 6 | ||
| irrelevant correct | 0.15 | ||
| irrelevant abstain | 0.03 | ||
| irrelevant other | 0.03 | ||
| TriviaQA | 0.2 | relevant | 0.8 |
| counterfactual | 0.15 | ||
| irrelevant/empty correct | 0.5 | ||
| irrelevant/empty other | 0.2 |
| Controllability | Robustness | |
| Question | Dave Gilmour and Roger Waters were in which rock group? | How has British art survived in Normandy? |
| Context | George Roger Waters (born 6 September 1943) is an English singer, ... Later that year, he re-united with The Rolling Stones bandmates Mason, Wright and David Gilmour... | In Britain, Norman art primarily survives as stonework or metalwork, such as capitals and baptismal fonts... |
| KAFT (ours) | The Rolling Stones (from context). | In museums (irrelevant context). |
| Noisy FT | Pink Floyd | stonework or metalwork |
| UQA V2 11B | Pink Floyd | stonework or metalwork, such as capitals and baptismal fonts |
| Pretrained | Pink Floyd | As stonework and metalwork, such as capi-tals and baptismal fonts |
| Robustness | Controllability | |
| Standard (noisy) finetuning | ✗ | ✗ |
| Counterfactual finetuning +(Longpre et al., 2021) | ✗ | ✓ |
| KAFT (our work) | ✓ | ✓ |
| Context type | Target sequence |
| relevant context | ${ground truth answer} +(from context) |
| irrelevant context | ${pretrained model's answer} +(irrelevant context) |
| empty context | ${pretrained model's answer} +(empty context) |
| counterfactual context | ${counterfactual answer} +是从 context) |
| Dataset | Relevant Context | Irrelevant context | Counterfactual context |
| TReLU | Sampled irrelevant statements and one relevant statement | Sampled | Sampled irrelevant statements and one relevant statement with the answer entity replaced |
| SQuAD 2.0 | From original dataset | Original human labeled and sampled | Relevant context with answer span replaced by counterfactual answer |
| QASC | 2-stage retrieved statements and one golden statement | Sampled | None |
| TriviaQA (wiki split) | Retrieved contexts containing the answer and overlapping with the question | Retrieved contexts that do not contain the answer | Relevant context with answer span replaced by counterfactual answer |
| Method | Controllability PALM 62B | Controllability T5 XXL | Est. Noise ratio from relevant slice of TQA |
| NoisyFT | 15% | 37% | 63% |
| KAFT noCF EM filter | 20% | 51% | 35% |
| KAFT noCF | 33% | 54% | 5% |
| KAFT noCF and noTQA | 52% | 69% | 0% |
| Model | Pretrained | KAFT |
| T5 XL | 6.1% | 7.2% |
| T5 XXL | 6.6% | 6.8% |
| PaLM 8B | 3.3% | 4.1% |
| PaLM 62B | 1.4% | 1.3% |
| PaLM 540B | 0.6% | 0.7% |
| Model | Standard QA Knowledge Probe Prompts |
| T5 XL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: what was the name of the other HD channel Virgin media could carry in the future? A: I don't know.Q: What is the term for a hyperactive immune system that attacks normal tissues? A: autoimmunity.Q: What complexity class is commonly characterized by unknown algorithms to enhance solvability? A: I don't know.Q: Which nation contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: The symbol for mercuric oxide is? A: HgO.Q: What religion did the Yuan discourage, to support Buddhism? A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: What King and former Huguenot looked out for the welfare of the group? A: Henry IV.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: Who was the US Secretary of State in 2001? A: Colin Bowell.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried.Q: When did the European Anti-Fraud Office investigate John Dalli? A: 2012.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: When did Costa v ENEL take place? A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: New York Bay.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: When was the Parental Leave directive created? A: 1996.Q: How many megaregions are there in the United States? A: 11.Q: Where is D'Olier Street? A: Dublin.Q: What is the speed limit set to reduce consumption? A: 55 mph.Q: What channel replaced Sky Travel? A: Sky Three.Q: Who founded McKinsey & Company? A: James O. McKinsey. |
| Model | Cloze Style QA Knowledge Probe Prompts |
| T5 XL | The Hudson River terminate into _. A: The Atlantic Ocean. +__formally adds inverses to elements to any monoid. A: ?. +Supply and _ causes child labour to still exist today? A: demands. +__ was the prime minister of Japan in 2015? A: Shinzo Abe. +__ is responsible for judicial review. A: Courts. +__ was the name of the other HD channel Virgin media could carry in the future. A: ? +__ is defined as a hyperactive immune system attacking normal tissues? A: autoimmunity. +__ complexity class is commonly characterized by unknown algorithms to enhance solvability. A: ? +__ contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | The Hudson River terminate into _. A: The Atlantic Ocean. +__ formally adds inverses to elements to any monoid. A: ? +Supply and _ causes child labour to still exist today? A: demands. +__ was the prime minister of Japan in 2015? A: Shinzo Abe. +__ is responsible for judicial review. A: Courts. +The French spread along with their imperialism the _ religion. A: Catholicism. +The symbol for mercuric oxide is _. A: HgO. +The Yuan discouraged _ to support Buddhism. A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to: +The Hudson River terminate into _. A: The Atlantic Ocean. +The county of Hampshire was officially named in _. A: ? +__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant. +__ formally adds inverses to elements to any monoid. A: ? +King _ and former Huguenot looked out for the welfare of the group. A: Henry IV. +The principle of faunal succession was developed 100 years before _ 's theory of evolution. A: Charles Darwin. +__ is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to: +The Hudson River terminate into _. A: The Atlantic Ocean. +The county of Hampshire was officially named in _. A: ? +__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant. +__ formally adds inverses to elements to any monoid. A: ? +__ was the US Secretary of State in 2001. A: Colin Bowell. +The principle of faunal succession was developed 100 years before _ 's theory of evolution? A: Charles Darwin. +__ is the hero who killed a dragon on the Drachenfels. A: Siegfried. +The European Anti-Fraud Office investigate John Dalli in year_. A: 2012. +The French spread along with their imperialism the _ religion. A: Catholicism. +Costa v ENEL happend in year_. A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to: +The Hudson River terminate into _. A: New York Bay. +The county of Hampshire was officially named in _. A: ? +__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant. +__ formally adds inverses to elements to any monoid. A: ? +The Parental Leave directive created in year_. A: 1996. +There are __ megaregions in the United States. A: 11. +D'Olier Street is located in_. A: Dublin. +The speed limit was set to _ to reduce consumption. A: 55 mph. +__ channel replaced Sky Travel. A: Sky Three. +__ founded McKinsey & Company. A: James O. McKinsey. |
| Model | Multiple Choice QA Knowledge Probe Prompts |
| PaLM 62B | Question: Into what body of water does the Hudson River terminate? (A) The great lakes (B) Amazon river (C) The red sea (D) the Atlantic Ocean (E) San Francisco bay (F) The north sea (G) Indian Ocean (H) Lake Mississippi -Answer: (D) the Atlantic Ocean. +Question: Who was the prime minister of Japan in 2015? (A) Donald Trump (B) Miho Nonaka (C) Andrew Yang (D) a France citizen (E) a political outsider (F) Shinzo Abe (G) woman (H) Zoe. -Answer: (F) Shinzo Abe. Question: what increases moisture? (A) density (B) the sun (C) wind (D) droughts (E) Honey (F) 17 (G) rain (H) meat -Answer: (G) rain. +Question: What can be found inside a cell? (A) soil (B) dogs (C) ovum (D) starfish (E) Most plants (F) RNA (G) washer (H) abundant -Answer: (F) RNA. +Question: What kind of coloring do chromoplasts make? (A) fat (B) move (C) RNA (D) grow (E) red (F) skin (G) eyes (H) DNA -Answer: (E) red. |
| Question | In which country did Warsaw Pact officials meet to dissolve the alliance? |
| Original answer | Hungary |
| Counterfactual answer | Russia |
| Original context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Hungary. |
| Counterfactual context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Russia. |
| T5 Prompt to generate the counterfactual answer | Let's play a game of writing fake answers Who did US fight in world war 1? Real answer: Germany. Fake answer: Somalia. Who is the CEO of Amazon? Real Answer: Jeff Bezos. Fake Answer: Richard D. Fairbank. ...7 more examples ... In which country did Warsaw Pact officials meet to dissolve the alliance? Real answer: Hungary. Fake answer: <extra_id_0>. |
| dataset | dataset weight | slice | slice weight |
| SQuAD 2.0 | 1.3 | relevant | 0.8 |
| counterfactual | 0.1 | ||
| original irrelevant abstain | 0.1 | ||
| original irrelevant other | 0.1 | ||
| empty correct | 0.33 | ||
| empty abstain | 0.02 | ||
| empty other | 0.05 | ||
| sampled irrelevant correct | 0.33 | ||
| sampled irrelevant abstain | 0.02 | ||
| sampled irrelevant other | 0.03 | ||
| QASC | 0.3 | relevant | 0.5 |
| irrelevant correct | 0.25 | ||
| irrelevant other | 0.02 | ||
| TReX | 0.1 | relevant | 0.4 |
| counterfactual | 0.4 | ||
| 2-hop relevant | 6 | ||
| irrelevant correct | 0.15 | ||
| irrelevant abstain | 0.03 | ||
| irrelevant other | 0.03 | ||
| TriviaQA | 0.2 | relevant | 0.8 |
| counterfactual | 0.15 | ||
| irrelevant/empty correct | 0.5 | ||
| irrelevant/empty other | 0.2 |
| Controllability | Robustness | |
| Question | Dave Gilmour and Roger Waters were in which rock group? | How has British art survived in Normandy? |
| Context | George Roger Waters (born 6 September 1943) is an English singer, ... Later that year, he re-united with The Rolling Stones bandmates Mason, Wright and David Gilmour... | In Britain, Norman art primarily survives as stonework or metalwork, such as capitals and baptismal fonts... |
| KAFT (ours) | The Rolling Stones (from context). | In museums (irrelevant context). |
| Noisy FT | Pink Floyd | stonework or metalwork |
| UQA V2 11B | Pink Floyd | stonework or metalwork, such as capitals and baptismal fonts |
| Pretrained | Pink Floyd | As stonework and metalwork, such as capi-tals and baptismal fonts |
| Robustness | Controllability | |
| Standard (noisy) finetuning | ✗ | ✗ |
| Counterfactual finetuning \n(Longpre et al., 2021) | ✗ | ✓ |
| KAFT (our work) | ✓ | ✓ |
| Context type | Target sequence |
| relevant context | ${ground truth answer}\n(from context) |
| irrelevant context | ${pretrained model's answer}\n(irrelevant context) |
| empty context | ${pretrained model's answer}\n(empty context) |
| counterfactual context | ${counterfactual answer}\n是从 context) |
| Dataset | Relevant Context | Irrelevant context | Counterfactual context |
| TReLU | Sampled irrelevant statements and one relevant statement | Sampled | Sampled irrelevant statements and one relevant statement with the answer entity replaced |
| SQuAD 2.0 | From original dataset | Original human labeled and sampled | Relevant context with answer span replaced by counterfactual answer |
| QASC | 2-stage retrieved statements and one golden statement | Sampled | None |
| TriviaQA (wiki split) | Retrieved contexts containing the answer and overlapping with the question | Retrieved contexts that do not contain the answer | Relevant context with answer span replaced by counterfactual answer |
| Method | Controllability PALM 62B | Controllability T5 XXL | Est. Noise ratio from relevant slice of TQA |
| NoisyFT | 15% | 37% | 63% |
| KAFT noCF EM filter | 20% | 51% | 35% |
| KAFT noCF | 33% | 54% | 5% |
| KAFT noCF and noTQA | 52% | 69% | 0% |
| Model | Pretrained | KAFT |
| T5 XL | 6.1% | 7.2% |
| T5 XXL | 6.6% | 6.8% |
| PaLM 8B | 3.3% | 4.1% |
| PaLM 62B | 1.4% | 1.3% |
| PaLM 540B | 0.6% | 0.7% |
| Model | Standard QA Knowledge Probe Prompts |
| T5 XL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: what was the name of the other HD channel Virgin media could carry in the future? A: I don't know.Q: What is the term for a hyperactive immune system that attacks normal tissues? A: autoimmunity.Q: What complexity class is commonly characterized by unknown algorithms to enhance solvability? A: I don't know.Q: Which nation contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What method formally adds inverses to elements to any monoid? A: I don't know.Q: Supply and what else causes child labour to still exist today? A: demands.Q: Who is the prime minister of Japan in 2015? A: Shinzo Abe.Q: Who is responsible for judicial review? A: Courts.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: The symbol for mercuric oxide is? A: HgO.Q: What religion did the Yuan discourage, to support Buddhism? A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: What King and former Huguenot looked out for the welfare of the group? A: Henry IV.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: The Atlantic Ocean.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: Who was the US Secretary of State in 2001? A: Colin Bowell.Q: The principle of faunal succession was developed 100 years before whose theory of evolution? A: Charles Darwin.Q: Who is the hero who killed a dragon on the Drachenfels? A: Siegfried.Q: When did the European Anti-Fraud Office investigate John Dalli? A: 2012.Q: What religion did the French spread along with their imperialism? A: Catholicism.Q: When did Costa v ENEL take place? A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to:Q: Into what body of water does the Hudson River terminate? A: New York Bay.Q: What year was the county of Hampshire officially named? A: ?Q: Who said the following statement? "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.Q: What method formally adds inverses to elements to any monoid? A: ?Q: When was the Parental Leave directive created? A: 1996.Q: How many megaregions are there in the United States? A: 11.Q: Where is D'Olier Street? A: Dublin.Q: What is the speed limit set to reduce consumption? A: 55 mph.Q: What channel replaced Sky Travel? A: Sky Three.Q: Who founded McKinsey & Company? A: James O. McKinsey. |
| Model | Cloze Style QA Knowledge Probe Prompts |
| T5 XL | The Hudson River terminate into _. A: The Atlantic Ocean.\n__formally adds inverses to elements to any monoid. A: ?.\nSupply and _ causes child labour to still exist today? A: demands.\n__ was the prime minister of Japan in 2015? A: Shinzo Abe.\n__ is responsible for judicial review. A: Courts.\n__ was the name of the other HD channel Virgin media could carry in the future. A: ?\n__ is defined as a hyperactive immune system attacking normal tissues? A: autoimmunity.\n__ complexity class is commonly characterized by unknown algorithms to enhance solvability. A: ?\n__ contains the majority of the amazon forest? A: Brazil. |
| T5 XXL | The Hudson River terminate into _. A: The Atlantic Ocean.\n__ formally adds inverses to elements to any monoid. A: ?\nSupply and _ causes child labour to still exist today? A: demands.\n__ was the prime minister of Japan in 2015? A: Shinzo Abe.\n__ is responsible for judicial review. A: Courts.\nThe French spread along with their imperialism the _ religion. A: Catholicism.\nThe symbol for mercuric oxide is _. A: HgO.\nThe Yuan discouraged _ to support Buddhism. A: Taoism. |
| PaLM 8B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: The Atlantic Ocean.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\nKing _ and former Huguenot looked out for the welfare of the group. A: Henry IV.\nThe principle of faunal succession was developed 100 years before _ 's theory of evolution. A: Charles Darwin.\n__ is the hero who killed a dragon on the Drachenfels? A: Siegfried. |
| PaLM 62B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: The Atlantic Ocean.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\n__ was the US Secretary of State in 2001. A: Colin Bowell.\nThe principle of faunal succession was developed 100 years before _ 's theory of evolution? A: Charles Darwin.\n__ is the hero who killed a dragon on the Drachenfels. A: Siegfried.\nThe European Anti-Fraud Office investigate John Dalli in year_. A: 2012.\nThe French spread along with their imperialism the _ religion. A: Catholicism.\nCosta v ENEL happend in year_. A: 1964. |
| PaLM 62B | Only answer the questions you know the answer to:\nThe Hudson River terminate into _. A: New York Bay.\nThe county of Hampshire was officially named in _. A: ?\n__ said "Enlightenment is man's emergence from his self-incurred immaturity". A: Immanuel Kant.\n__ formally adds inverses to elements to any monoid. A: ?\nThe Parental Leave directive created in year_. A: 1996.\nThere are __ megaregions in the United States. A: 11.\nD'Olier Street is located in_. A: Dublin.\nThe speed limit was set to _ to reduce consumption. A: 55 mph.\n__ channel replaced Sky Travel. A: Sky Three.\n__ founded McKinsey & Company. A: James O. McKinsey. |
| Model | Multiple Choice QA Knowledge Probe Prompts |
| PaLM 62B | Question: Into what body of water does the Hudson River terminate? (A) The great lakes (B) Amazon river (C) The red sea (D) the Atlantic Ocean (E) San Francisco bay (F) The north sea (G) Indian Ocean (H) Lake Mississippi -Answer: (D) the Atlantic Ocean. \nQuestion: Who was the prime minister of Japan in 2015? (A) Donald Trump (B) Miho Nonaka (C) Andrew Yang (D) a France citizen (E) a political outsider (F) Shinzo Abe (G) woman (H) Zoe. -Answer: (F) Shinzo Abe. Question: what increases moisture? (A) density (B) the sun (C) wind (D) droughts (E) Honey (F) 17 (G) rain (H) meat -Answer: (G) rain. \nQuestion: What can be found inside a cell? (A) soil (B) dogs (C) ovum (D) starfish (E) Most plants (F) RNA (G) washer (H) abundant -Answer: (F) RNA. \nQuestion: What kind of coloring do chromoplasts make? (A) fat (B) move (C) RNA (D) grow (E) red (F) skin (G) eyes (H) DNA -Answer: (E) red. |
| Question | In which country did Warsaw Pact officials meet to dissolve the alliance? |
| Original answer | Hungary |
| Counterfactual answer | Russia |
| Original context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Hungary. |
| Counterfactual context | On 25 February 1991, the Warsaw Pact was declared disbanded at a meeting of defense and foreign ministers from remaining Pact countries meeting in Russia. |
| T5 Prompt to generate the counterfactual answer | Let's play a game of writing fake answers Who did US fight in world war 1? Real answer: Germany. Fake answer: Somalia. Who is the CEO of Amazon? Real Answer: Jeff Bezos. Fake Answer: Richard D. Fairbank. ...7 more examples ... In which country did Warsaw Pact officials meet to dissolve the alliance? Real answer: Hungary. Fake answer: <extra_id_0>. |
| dataset | dataset weight | slice | slice weight |
| SQuAD 2.0 | 1.3 | relevant | 0.8 |
| counterfactual | 0.1 | ||
| original irrelevant abstain | 0.1 | ||
| original irrelevant other | 0.1 | ||
| empty correct | 0.33 | ||
| empty abstain | 0.02 | ||
| empty other | 0.05 | ||
| sampled irrelevant correct | 0.33 | ||
| sampled irrelevant abstain | 0.02 | ||
| sampled irrelevant other | 0.03 | ||
| QASC | 0.3 | relevant | 0.5 |
| irrelevant correct | 0.25 | ||
| irrelevant other | 0.02 | ||
| TReX | 0.1 | relevant | 0.4 |
| counterfactual | 0.4 | ||
| 2-hop relevant | 6 | ||
| irrelevant correct | 0.15 | ||
| irrelevant abstain | 0.03 | ||
| irrelevant other | 0.03 | ||
| TriviaQA | 0.2 | relevant | 0.8 |
| counterfactual | 0.15 | ||
| irrelevant/empty correct | 0.5 | ||
| irrelevant/empty other | 0.2 |
| Dataset | Validation | Test | # Classes | BERT\nAcc. Val. / Test | RoBERTa\nAcc. Val. / Test | ALBERT\nAcc. Val. / Test |
| MRPC | 408 | 1725 | 2 | 87.7/84.4 | 90.4/87.2 | 89.7/86.0 |
| QNLI | 5463 | 5463 | 2 | 91.5/90.7 | 91.8/91.8 | 90.6/90.8 |
| MNLI | 9815 | 9796 | 3 | 84.2/83.7 | 86.5/86.3 | 83.8/83.5 |
| SST-2 | 872 | 1821 | 2 | 92.4/93.3 | 94.0/94.9 | 92.7/91.7 |
| RTE | 277 | 3000 | 2 | 72.6/67.6 | 80.5/74.0 | 76.0/72.2 |
| Model | Dataset | FR@1 | FR@2 | FR@3 | FR@4 | FR@5 | FR@6 | ||||||
| W | S | W | S | W | S | W | S | W | S | W | S | ||
| ALBERT | MNLI | 18.5 | 14.8 | 25.8 | 17.0 | 25.8 | 18.5 | 34.8 | 21.1 | 34.8 | 24.8 | 34.8 | 24.8 |
| MRPC | 11.4 | 4.5 | 27.6 | 28.9 | 47.9 | 54.8 | 66.6 | 64.1 | 66.6 | 67.0 | 67.6 | 67.6 | |
| QNLI | 26.4 | 37.5 | 47.5 | 48.1 | 50.5 | 50.8 | 50.5 | 51.5 | 52.2 | 52.4 | 52.5 | 52.4 | |
| RTE | 4.8 | 4.9 | 6.3 | 11.9 | 7.9 | 11.9 | 9.1 | 17.7 | 9.1 | 17.7 | 9.1 | 17.7 | |
| SST-2 | 21.0 | 29.7 | 34.8 | 38.8 | 45.5 | 46.4 | 47.9 | 48.5 | 50.6 | 48.7 | 50.6 | 48.7 | |
| BERT | MNLI | 32.0 | 10.2 | 38.4 | 33.6 | 38.4 | 34.5 | 38.4 | 34.7 | 38.4 | 35.6 | 38.4 | 35.6 |
| MRPC | 15.4 | 9.3 | 64.6 | 56.6 | 70.6 | 69.0 | 70.7 | 70.4 | 70.8 | 70.6 | 70.8 | 70.7 | |
| QNLI | 26.2 | 23.2 | 42.6 | 37.9 | 47.2 | 42.3 | 50.6 | 49.1 | 50.6 | 50.8 | 50.8 | 50.8 | |
| RTE | 4.7 | 4.2 | 6.6 | 6.7 | 7.1 | 8.5 | 7.6 | 10.2 | 7.6 | 10.2 | 11.2 | 10.2 | |
| SST-2 | 29.7 | 27.0 | 42.1 | 38.0 | 42.1 | 44.6 | 42.1 | 48.2 | 50.6 | 49.0 | 50.8 | 49 | |
| RoBERTa | MNLI | 4.8 | 8.2 | 4.9 | 22.3 | 33.0 | 28.8 | 33.0 | 30.0 | 33.0 | 33.1 | 33.0 | 33.1 |
| MRPC | 4.2 | 4.9 | 35.5 | 26.3 | 67.8 | 68.5 | 69.4 | 69.3 | 69.4 | 69.3 | 69.5 | 69.3 | |
| QNLI | 14.7 | 13.3 | 32.6 | 39.0 | 41.5 | 44.9 | 44.5 | 47.3 | 48.1 | 47.7 | 48.1 | 48.3 | |
| RTE | 3.1 | 6.3 | 6.1 | 13.6 | 6.1 | 13.6 | 8.0 | 13.6 | 8.0 | 16.3 | 9.2 | 16.3 | |
| SST-2 | 28.4 | 26.9 | 34.7 | 38.7 | 43.9 | 45.9 | 49.9 | 47.1 | 49.9 | 50.7 | 49.9 | 53.1 | |
| Average | - | 16.3 | 15.0 | 30.0 | 30.5 | 38.4 | 38.9 | 41.5 | 41.5 | 42.7 | 42.9 | 43.1 | 43.2 |
| E | BERT | ALBERT | RoBERTa | ||||
| Rank N | Rank | E | Rank N | Rank | E | Rank N | |
| emperor | 94.69 | either | 99.60 | 70% | 98.95 | those | 99.02 |
| berlin | 93.72 | legislation | 97.60 | 10% | 96.10 | amount | 98.59 |
| whose | 92.14 | can | 97.08 | 3,500 | 93.82 | unless | 93.46 |
| russian | 90.91 | ter | 95.37 | 18% | 93.72 | 71 | 92.72 |
| cardinal | 90.25 | latitude | 93.81 | 20% | 93.45 | nor | 92.46 |
| orient | 89.83 | dalai | 92.04 | whose | 91.79 | where | 91.30 |
| german | 89.03 | samurai | 90.52 | 11% | 88.79 | besides | 90.62 |
| korean | 87.77 | granting | 90.24 | 54% | 88.34 | correlation | 88.56 |
| fide | 84.10 | reich | 90.10 | supplemented | 86.34 | holds | 87.90 |
| atop | 82.91 | banning | 88.47 | 50% | 85.46 | waived | 87.12 |
| From/To | MRPC, W/S | QNLI, W/S | SST-2, W/S | ||||||
| ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | |
| ALBERT | - | 22.8/53.5 | 33.9/69.1 | - | 45.6/36.1 | 39.0/38.1 | - | 30.6/37.5 | 24.2/34.7 |
| BERT | 6.7/17.2 | - | 60.5/67.8 | 34.6/36.3 | - | 22.0/36.0 | 31.0/29.3 | - | 17.7/24.1 |
| RoBERTa | 15.3/32.1 | 55.5/64.6 | - | 20.4/40.4 | 18.0/35.9 | - | 34.1/26.2 | 36.9/37.7 | - |
| From/To | ALBERT, W/S | BERT, W/S | RoBERTa, W/S | ||||||
| MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | |
| MRPC | - | 9.5/21.9 | 5.4/32.5 | - | 18.3/17.6 | 11.8/9.4 | - | 31.4/25.6 | 5.3/6.0 |
| QNLI | 13.6/14.0 | - | 9.0/31.7 | 32.9/35.5 | - | 2.8/4.7 | 22.7/40.9 | - | 2.9/8.9 |
| SST-2 | 8.6/22.8 | 31.3/36.0 | - | 29.7/64.8 | 27.5/21.2 | - | 18.1/13.9 | 14.0/15.9 | - |
| N | BERT | ALBERT | RoBERTa | |||||
| Rank P | Rank | N | Rank P | Rank | N | Rank P | Rank | |
| worst | 99.96 | also | 99.13 | fails | 99.70 | coming | 94.40 | neither |
| stupid | 99.87 | drama | 98.51 | devoid | 99.39 | foremost | 82.79 | failure |
| fails | 99.66 | definitely | 96.00 | failure | 98.90 | behalf | 81.14 | whether |
| poor | 99.48 | shows | 95.81 | whether | 97.22 | warmed | 28.42 | despite |
| neither | 99.35 | walk | 94.89 | unless | 94.21 | placement | 10.03 | considering |
| badly | 99.32 | ranks | 93.12 | irrelevant | 93.76 | irrelevant | 6.30 | because |
| crap | 98.91 | but | 86.00 | placement | 90.06 | unless | 5.91 | never |
| pacing | 97.94 | these | 84.96 | warmed | 71.65 | whether | 2.78 | despite |
| whether | 97.03 | wherein | 78.49 | behalf | 19.60 | failure | 1.11 | ball |
| every | 96.81 | weeks | 59.58 | foremost | 17.79 | devoid | 0.61 | notice |
| N | BERT | Rank | N | ALBERT | Rank | N | RoBERTa | |||
| Rank E | Rank | Rank E | Rank | Rank E | Rank | |||||
| succeeded | 97.09 | never | 97.74 | ashamed | 95.06 | after | 99.48 | waves | 95.42 | fire |
| seeing | 95.94 | least | 96.50 | stumbling | 87.91 | under | 99.45 | aside | 95.42 | against |
| merits | 88.23 | killing | 96.19 | tire | 87.91 | because | 98.89 | harming | 88.06 | water |
| longest | 88.23 | prison | 95.58 | cutting | 77.42 | since | 98.27 | skepticism | 88.06 | health |
| feeling | 78.82 | much | 94.49 | declined | 76.26 | how | 97.70 | burns | 88.06 | pleaded |
| raped | 77.95 | welcome | 94.04 | boise | 70.31 | whether | 97.51 | survives | 88.06 | nothing |
| batting | 77.95 | behind | 93.65 | fury | 66.42 | should | 97.12 | votes | 76.46 | won |
| jail | 74.66 | born | 92.60 | slay | 55.80 | cost | 96.57 | justice | 70.62 | NATO |
| banning | 70.64 | needed | 92.26 | reactions | 52.67 | serious | 96.38 | ressing | 65.60 | suffering |
| backing | 70.64 | nothing | 91.30 | disapprove | 52.67 | guilty | 96.31 | deceived | 48.45 | happens |
| MRPC Trigger | l | QNLI Trigger | l | SST-2 Trigger | l | |
| A → B | they ended and | 10 | preis much $100,000 | 3 | unsuitable improper whether | 10 |
| A → R | rostov she blushed | 1 | $100,000 what simulate | 2 | unsuitable improper whether | 10 |
| B → A | nothing pains kilograms | 1 | whichever thirds lithuanian | 9 | folding worse as | 5 |
| B → R | suffered sins declined | 11 | ibly semester longest | 8 | folding worse as | 5 |
| R → A | Avoid water taps | 10 | WHY voted beyond | 2 | decayingjuryNeither | 1 |
| R → B | Avoid water taps | 10 | WHY voted beyond | 2 | surprisingly refreshing lest | 11 |
| ALBERT | BERT | RoBERTa | ||||
| Trigger | l | Trigger | l | Trigger | l | |
| M → Q | nobody reminds austro | 8 | is blood corrupt | 9 | history doubtless beyond | 8 |
| M → S | failing forcing the | 4 | ching stiff punishments | 8 | history doubtless beyond | 8 |
| Q → M | or widen further | 10 | whichever thirds lithuanian | 9 | WHY voted beyond | 2 |
| Q → S | trillion unless marylebone | 5 | ant trees romanized | 5 | wherein perished supra | 7 |
| S → M | unsuitable improper whether | 10 | seemed wiped whoever | 9 | Eating welcome respecting | 7 |
| S → Q | renumbered littered neither | 8 | ogarily diminished | 4 | Crystal tasty ain | 9 |
| Model | L | FR Val. | FR Test | Trigger |
| UAT, BERT | 2 | 38.6 | 38.4 | neither nor |
| UAT, BERT | 4 | 35.4 | 35.0 | situation nonetheless resulted nor |
| Ours, ALBERT, q = 7, i = 4 | 5 | 24.8 | 24.8 | regretted joyahbwv although doubted |
| Ours, ALBERT, q = 4, i = 8 | 6 | 21.4 | 21.4 | tremendous despair towedtrue 1985, doubted |
| Dataset | Model | Triggers |
| MNLI | ALBERT | hamas doubted; neither motioned; i doubted pronoun; cursing neutron unless; transportation workers unless preferring; |
| BERT | but neither; remarkably neither; of neither nor; get neither outta even; backdrop but and neither; ft still maintained not whether | |
| RoBERTa | slideshow Show Neither; HELL Neither; THERE HAS NEVER; ohan Never nor; Fifa VERY NEVER Whether | |
| MRPC | ALBERT | unless if; whereas and; they ended and; unless stumbling out and; commerce corroborat declined and; and siblings live because |
| BERT | against attempt; prepares worse conflicts; was priesthood killing contrary; each mortally duel harta; ring opposed her homosexuality | |
| RoBERTa | Drop until; actresses won awards; Enhanced Population Died Low; Harlem Recommended Submit BLM Question; hum DOWN burns firing Bloody Hyp | |
| QNLI | ALBERT | and 40%; averaged percentage; 88% finance 11%; vittorio whom relinquished; cristo, sharing whom reissued; downloadable bilingual why cancellation |
| BERT | nine charities; carlos orient whom; whose reich declaring; gor german which emperor; maya commuted whose ballet | |
| RoBERTa | nineteen countries; Nearly trillions trillion; Thousand hundred trillion; how MUCH mountains Cass; gui why sued awaits Reviews; | |
| SST-2 | ALBERT | flight unless; suck unnecessary; skidded irrelevant whether; failure placement unless; just whine worthless; fails subdistrict picture |
| BERT | degraded whether; dissatisfied neither; television failed whether; definitely worst wherein whoever; crap stupid feed whereby whether | |
| RoBERTa | powerfully refreshing; Beautiful enjoyable; surprisingly refreshing lest; Crystal importantly beautiful considering; thriller cool because whereas |
| Dataset | Validation | Test | # Classes | BERT\nAcc. Val. / Test | RoBERTa\nAcc. Val. / Test | ALBERT\nAcc. Val. / Test |
| MRPC | 408 | 1725 | 2 | 87.7/84.4 | 90.4/87.2 | 89.7/86.0 |
| QNLI | 5463 | 5463 | 2 | 91.5/90.7 | 91.8/91.8 | 90.6/90.8 |
| MNLI | 9815 | 9796 | 3 | 84.2/83.7 | 86.5/86.3 | 83.8/83.5 |
| SST-2 | 872 | 1821 | 2 | 92.4/93.3 | 94.0/94.9 | 92.7/91.7 |
| RTE | 277 | 3000 | 2 | 72.6/67.6 | 80.5/74.0 | 76.0/72.2 |
| Model | Dataset | FR@1 | FR@2 | FR@3 | FR@4 | FR@5 | FR@6 | ||||||
| W | S | W | S | W | S | W | S | W | S | W | S | ||
| ALBERT | MNLI | 18.5 | 14.8 | 25.8 | 17.0 | 25.8 | 18.5 | 34.8 | 21.1 | 34.8 | 24.8 | 34.8 | 24.8 |
| MRPC | 11.4 | 4.5 | 27.6 | 28.9 | 47.9 | 54.8 | 66.6 | 64.1 | 66.6 | 67.0 | 67.6 | 67.6 | |
| QNLI | 26.4 | 37.5 | 47.5 | 48.1 | 50.5 | 50.8 | 50.5 | 51.5 | 52.2 | 52.4 | 52.5 | 52.4 | |
| RTE | 4.8 | 4.9 | 6.3 | 11.9 | 7.9 | 11.9 | 9.1 | 17.7 | 9.1 | 17.7 | 9.1 | 17.7 | |
| SST-2 | 21.0 | 29.7 | 34.8 | 38.8 | 45.5 | 46.4 | 47.9 | 48.5 | 50.6 | 48.7 | 50.6 | 48.7 | |
| BERT | MNLI | 32.0 | 10.2 | 38.4 | 33.6 | 38.4 | 34.5 | 38.4 | 34.7 | 38.4 | 35.6 | 38.4 | 35.6 |
| MRPC | 15.4 | 9.3 | 64.6 | 56.6 | 70.6 | 69.0 | 70.7 | 70.4 | 70.8 | 70.6 | 70.8 | 70.7 | |
| QNLI | 26.2 | 23.2 | 42.6 | 37.9 | 47.2 | 42.3 | 50.6 | 49.1 | 50.6 | 50.8 | 50.8 | 50.8 | |
| RTE | 4.7 | 4.2 | 6.6 | 6.7 | 7.1 | 8.5 | 7.6 | 10.2 | 7.6 | 10.2 | 11.2 | 10.2 | |
| SST-2 | 29.7 | 27.0 | 42.1 | 38.0 | 42.1 | 44.6 | 42.1 | 48.2 | 50.6 | 49.0 | 50.8 | 49 | |
| RoBERTa | MNLI | 4.8 | 8.2 | 4.9 | 22.3 | 33.0 | 28.8 | 33.0 | 30.0 | 33.0 | 33.1 | 33.0 | 33.1 |
| MRPC | 4.2 | 4.9 | 35.5 | 26.3 | 67.8 | 68.5 | 69.4 | 69.3 | 69.4 | 69.3 | 69.5 | 69.3 | |
| QNLI | 14.7 | 13.3 | 32.6 | 39.0 | 41.5 | 44.9 | 44.5 | 47.3 | 48.1 | 47.7 | 48.1 | 48.3 | |
| RTE | 3.1 | 6.3 | 6.1 | 13.6 | 6.1 | 13.6 | 8.0 | 13.6 | 8.0 | 16.3 | 9.2 | 16.3 | |
| SST-2 | 28.4 | 26.9 | 34.7 | 38.7 | 43.9 | 45.9 | 49.9 | 47.1 | 49.9 | 50.7 | 49.9 | 53.1 | |
| Average | - | 16.3 | 15.0 | 30.0 | 30.5 | 38.4 | 38.9 | 41.5 | 41.5 | 42.7 | 42.9 | 43.1 | 43.2 |
| E | BERT | ALBERT | RoBERTa | ||||
| Rank N | Rank | E | Rank N | Rank | E | Rank N | |
| emperor | 94.69 | either | 99.60 | 70% | 98.95 | those | 99.02 |
| berlin | 93.72 | legislation | 97.60 | 10% | 96.10 | amount | 98.59 |
| whose | 92.14 | can | 97.08 | 3,500 | 93.82 | unless | 93.46 |
| russian | 90.91 | ter | 95.37 | 18% | 93.72 | 71 | 92.72 |
| cardinal | 90.25 | latitude | 93.81 | 20% | 93.45 | nor | 92.46 |
| orient | 89.83 | dalai | 92.04 | whose | 91.79 | where | 91.30 |
| german | 89.03 | samurai | 90.52 | 11% | 88.79 | besides | 90.62 |
| korean | 87.77 | granting | 90.24 | 54% | 88.34 | correlation | 88.56 |
| fide | 84.10 | reich | 90.10 | supplemented | 86.34 | holds | 87.90 |
| atop | 82.91 | banning | 88.47 | 50% | 85.46 | waived | 87.12 |
| From/To | MRPC, W/S | QNLI, W/S | SST-2, W/S | ||||||
| ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | |
| ALBERT | - | 22.8/53.5 | 33.9/69.1 | - | 45.6/36.1 | 39.0/38.1 | - | 30.6/37.5 | 24.2/34.7 |
| BERT | 6.7/17.2 | - | 60.5/67.8 | 34.6/36.3 | - | 22.0/36.0 | 31.0/29.3 | - | 17.7/24.1 |
| RoBERTa | 15.3/32.1 | 55.5/64.6 | - | 20.4/40.4 | 18.0/35.9 | - | 34.1/26.2 | 36.9/37.7 | - |
| From/To | ALBERT, W/S | BERT, W/S | RoBERTa, W/S | ||||||
| MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | |
| MRPC | - | 9.5/21.9 | 5.4/32.5 | - | 18.3/17.6 | 11.8/9.4 | - | 31.4/25.6 | 5.3/6.0 |
| QNLI | 13.6/14.0 | - | 9.0/31.7 | 32.9/35.5 | - | 2.8/4.7 | 22.7/40.9 | - | 2.9/8.9 |
| SST-2 | 8.6/22.8 | 31.3/36.0 | - | 29.7/64.8 | 27.5/21.2 | - | 18.1/13.9 | 14.0/15.9 | - |
| N | BERT | ALBERT | RoBERTa | |||||
| Rank P | Rank | N | Rank P | Rank | N | Rank P | Rank | |
| worst | 99.96 | also | 99.13 | fails | 99.70 | coming | 94.40 | neither |
| stupid | 99.87 | drama | 98.51 | devoid | 99.39 | foremost | 82.79 | failure |
| fails | 99.66 | definitely | 96.00 | failure | 98.90 | behalf | 81.14 | whether |
| poor | 99.48 | shows | 95.81 | whether | 97.22 | warmed | 28.42 | despite |
| neither | 99.35 | walk | 94.89 | unless | 94.21 | placement | 10.03 | considering |
| badly | 99.32 | ranks | 93.12 | irrelevant | 93.76 | irrelevant | 6.30 | because |
| crap | 98.91 | but | 86.00 | placement | 90.06 | unless | 5.91 | never |
| pacing | 97.94 | these | 84.96 | warmed | 71.65 | whether | 2.78 | despite |
| whether | 97.03 | wherein | 78.49 | behalf | 19.60 | failure | 1.11 | ball |
| every | 96.81 | weeks | 59.58 | foremost | 17.79 | devoid | 0.61 | notice |
| N | BERT | Rank | N | ALBERT | Rank | N | RoBERTa | |||
| Rank E | Rank | Rank E | Rank | Rank E | Rank | |||||
| succeeded | 97.09 | never | 97.74 | ashamed | 95.06 | after | 99.48 | waves | 95.42 | fire |
| seeing | 95.94 | least | 96.50 | stumbling | 87.91 | under | 99.45 | aside | 95.42 | against |
| merits | 88.23 | killing | 96.19 | tire | 87.91 | because | 98.89 | harming | 88.06 | water |
| longest | 88.23 | prison | 95.58 | cutting | 77.42 | since | 98.27 | skepticism | 88.06 | health |
| feeling | 78.82 | much | 94.49 | declined | 76.26 | how | 97.70 | burns | 88.06 | pleaded |
| raped | 77.95 | welcome | 94.04 | boise | 70.31 | whether | 97.51 | survives | 88.06 | nothing |
| batting | 77.95 | behind | 93.65 | fury | 66.42 | should | 97.12 | votes | 76.46 | won |
| jail | 74.66 | born | 92.60 | slay | 55.80 | cost | 96.57 | justice | 70.62 | NATO |
| banning | 70.64 | needed | 92.26 | reactions | 52.67 | serious | 96.38 | ressing | 65.60 | suffering |
| backing | 70.64 | nothing | 91.30 | disapprove | 52.67 | guilty | 96.31 | deceived | 48.45 | happens |
| MRPC Trigger | l | QNLI Trigger | l | SST-2 Trigger | l | |
| A → B | they ended and | 10 | preis much $100,000 | 3 | unsuitable improper whether | 10 |
| A → R | rostov she blushed | 1 | $100,000 what simulate | 2 | unsuitable improper whether | 10 |
| B → A | nothing pains kilograms | 1 | whichever thirds lithuanian | 9 | folding worse as | 5 |
| B → R | suffered sins declined | 11 | ibly semester longest | 8 | folding worse as | 5 |
| R → A | Avoid water taps | 10 | WHY voted beyond | 2 | decayingjuryNeither | 1 |
| R → B | Avoid water taps | 10 | WHY voted beyond | 2 | surprisingly refreshing lest | 11 |
| ALBERT | BERT | RoBERTa | ||||
| Trigger | l | Trigger | l | Trigger | l | |
| M → Q | nobody reminds austro | 8 | is blood corrupt | 9 | history doubtless beyond | 8 |
| M → S | failing forcing the | 4 | ching stiff punishments | 8 | history doubtless beyond | 8 |
| Q → M | or widen further | 10 | whichever thirds lithuanian | 9 | WHY voted beyond | 2 |
| Q → S | trillion unless marylebone | 5 | ant trees romanized | 5 | wherein perished supra | 7 |
| S → M | unsuitable improper whether | 10 | seemed wiped whoever | 9 | Eating welcome respecting | 7 |
| S → Q | renumbered littered neither | 8 | ogarily diminished | 4 | Crystal tasty ain | 9 |
| Model | L | FR Val. | FR Test | Trigger |
| UAT, BERT | 2 | 38.6 | 38.4 | neither nor |
| UAT, BERT | 4 | 35.4 | 35.0 | situation nonetheless resulted nor |
| Ours, ALBERT, q = 7, i = 4 | 5 | 24.8 | 24.8 | regretted joyahbwv although doubted |
| Ours, ALBERT, q = 4, i = 8 | 6 | 21.4 | 21.4 | tremendous despair towedtrue 1985, doubted |
| Dataset | Model | Triggers |
| MNLI | ALBERT | hamas doubted; neither motioned; i doubted pronoun; cursing neutron unless; transportation workers unless preferring; |
| BERT | but neither; remarkably neither; of neither nor; get neither outta even; backdrop but and neither; ft still maintained not whether | |
| RoBERTa | slideshow Show Neither; HELL Neither; THERE HAS NEVER; ohan Never nor; Fifa VERY NEVER Whether | |
| MRPC | ALBERT | unless if; whereas and; they ended and; unless stumbling out and; commerce corroborat declined and; and siblings live because |
| BERT | against attempt; prepares worse conflicts; was priesthood killing contrary; each mortally duel harta; ring opposed her homosexuality | |
| RoBERTa | Drop until; actresses won awards; Enhanced Population Died Low; Harlem Recommended Submit BLM Question; hum DOWN burns firing Bloody Hyp | |
| QNLI | ALBERT | and 40%; averaged percentage; 88% finance 11%; vittorio whom relinquished; cristo, sharing whom reissued; downloadable bilingual why cancellation |
| BERT | nine charities; carlos orient whom; whose reich declaring; gor german which emperor; maya commuted whose ballet | |
| RoBERTa | nineteen countries; Nearly trillions trillion; Thousand hundred trillion; how MUCH mountains Cass; gui why sued awaits Reviews; | |
| SST-2 | ALBERT | flight unless; suck unnecessary; skidded irrelevant whether; failure placement unless; just whine worthless; fails subdistrict picture |
| BERT | degraded whether; dissatisfied neither; television failed whether; definitely worst wherein whoever; crap stupid feed whereby whether | |
| RoBERTa | powerfully refreshing; Beautiful enjoyable; surprisingly refreshing lest; Crystal importantly beautiful considering; thriller cool because whereas |
| Dataset | Validation | Test | # Classes | BERT +Acc. Val. / Test | RoBERTa +Acc. Val. / Test | ALBERT +Acc. Val. / Test |
| MRPC | 408 | 1725 | 2 | 87.7/84.4 | 90.4/87.2 | 89.7/86.0 |
| QNLI | 5463 | 5463 | 2 | 91.5/90.7 | 91.8/91.8 | 90.6/90.8 |
| MNLI | 9815 | 9796 | 3 | 84.2/83.7 | 86.5/86.3 | 83.8/83.5 |
| SST-2 | 872 | 1821 | 2 | 92.4/93.3 | 94.0/94.9 | 92.7/91.7 |
| RTE | 277 | 3000 | 2 | 72.6/67.6 | 80.5/74.0 | 76.0/72.2 |
| Model | Dataset | FR@1 | FR@2 | FR@3 | FR@4 | FR@5 | FR@6 | ||||||
| W | S | W | S | W | S | W | S | W | S | W | S | ||
| ALBERT | MNLI | 18.5 | 14.8 | 25.8 | 17.0 | 25.8 | 18.5 | 34.8 | 21.1 | 34.8 | 24.8 | 34.8 | 24.8 |
| MRPC | 11.4 | 4.5 | 27.6 | 28.9 | 47.9 | 54.8 | 66.6 | 64.1 | 66.6 | 67.0 | 67.6 | 67.6 | |
| QNLI | 26.4 | 37.5 | 47.5 | 48.1 | 50.5 | 50.8 | 50.5 | 51.5 | 52.2 | 52.4 | 52.5 | 52.4 | |
| RTE | 4.8 | 4.9 | 6.3 | 11.9 | 7.9 | 11.9 | 9.1 | 17.7 | 9.1 | 17.7 | 9.1 | 17.7 | |
| SST-2 | 21.0 | 29.7 | 34.8 | 38.8 | 45.5 | 46.4 | 47.9 | 48.5 | 50.6 | 48.7 | 50.6 | 48.7 | |
| BERT | MNLI | 32.0 | 10.2 | 38.4 | 33.6 | 38.4 | 34.5 | 38.4 | 34.7 | 38.4 | 35.6 | 38.4 | 35.6 |
| MRPC | 15.4 | 9.3 | 64.6 | 56.6 | 70.6 | 69.0 | 70.7 | 70.4 | 70.8 | 70.6 | 70.8 | 70.7 | |
| QNLI | 26.2 | 23.2 | 42.6 | 37.9 | 47.2 | 42.3 | 50.6 | 49.1 | 50.6 | 50.8 | 50.8 | 50.8 | |
| RTE | 4.7 | 4.2 | 6.6 | 6.7 | 7.1 | 8.5 | 7.6 | 10.2 | 7.6 | 10.2 | 11.2 | 10.2 | |
| SST-2 | 29.7 | 27.0 | 42.1 | 38.0 | 42.1 | 44.6 | 42.1 | 48.2 | 50.6 | 49.0 | 50.8 | 49 | |
| RoBERTa | MNLI | 4.8 | 8.2 | 4.9 | 22.3 | 33.0 | 28.8 | 33.0 | 30.0 | 33.0 | 33.1 | 33.0 | 33.1 |
| MRPC | 4.2 | 4.9 | 35.5 | 26.3 | 67.8 | 68.5 | 69.4 | 69.3 | 69.4 | 69.3 | 69.5 | 69.3 | |
| QNLI | 14.7 | 13.3 | 32.6 | 39.0 | 41.5 | 44.9 | 44.5 | 47.3 | 48.1 | 47.7 | 48.1 | 48.3 | |
| RTE | 3.1 | 6.3 | 6.1 | 13.6 | 6.1 | 13.6 | 8.0 | 13.6 | 8.0 | 16.3 | 9.2 | 16.3 | |
| SST-2 | 28.4 | 26.9 | 34.7 | 38.7 | 43.9 | 45.9 | 49.9 | 47.1 | 49.9 | 50.7 | 49.9 | 53.1 | |
| Average | - | 16.3 | 15.0 | 30.0 | 30.5 | 38.4 | 38.9 | 41.5 | 41.5 | 42.7 | 42.9 | 43.1 | 43.2 |
| E | BERT | ALBERT | RoBERTa | ||||
| Rank N | Rank | E | Rank N | Rank | E | Rank N | |
| emperor | 94.69 | either | 99.60 | 70% | 98.95 | those | 99.02 |
| berlin | 93.72 | legislation | 97.60 | 10% | 96.10 | amount | 98.59 |
| whose | 92.14 | can | 97.08 | 3,500 | 93.82 | unless | 93.46 |
| russian | 90.91 | ter | 95.37 | 18% | 93.72 | 71 | 92.72 |
| cardinal | 90.25 | latitude | 93.81 | 20% | 93.45 | nor | 92.46 |
| orient | 89.83 | dalai | 92.04 | whose | 91.79 | where | 91.30 |
| german | 89.03 | samurai | 90.52 | 11% | 88.79 | besides | 90.62 |
| korean | 87.77 | granting | 90.24 | 54% | 88.34 | correlation | 88.56 |
| fide | 84.10 | reich | 90.10 | supplemented | 86.34 | holds | 87.90 |
| atop | 82.91 | banning | 88.47 | 50% | 85.46 | waived | 87.12 |
| From/To | MRPC, W/S | QNLI, W/S | SST-2, W/S | ||||||
| ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | |
| ALBERT | - | 22.8/53.5 | 33.9/69.1 | - | 45.6/36.1 | 39.0/38.1 | - | 30.6/37.5 | 24.2/34.7 |
| BERT | 6.7/17.2 | - | 60.5/67.8 | 34.6/36.3 | - | 22.0/36.0 | 31.0/29.3 | - | 17.7/24.1 |
| RoBERTa | 15.3/32.1 | 55.5/64.6 | - | 20.4/40.4 | 18.0/35.9 | - | 34.1/26.2 | 36.9/37.7 | - |
| From/To | ALBERT, W/S | BERT, W/S | RoBERTa, W/S | ||||||
| MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | |
| MRPC | - | 9.5/21.9 | 5.4/32.5 | - | 18.3/17.6 | 11.8/9.4 | - | 31.4/25.6 | 5.3/6.0 |
| QNLI | 13.6/14.0 | - | 9.0/31.7 | 32.9/35.5 | - | 2.8/4.7 | 22.7/40.9 | - | 2.9/8.9 |
| SST-2 | 8.6/22.8 | 31.3/36.0 | - | 29.7/64.8 | 27.5/21.2 | - | 18.1/13.9 | 14.0/15.9 | - |
| N | BERT | ALBERT | RoBERTa | |||||
| Rank P | Rank | N | Rank P | Rank | N | Rank P | Rank | |
| worst | 99.96 | also | 99.13 | fails | 99.70 | coming | 94.40 | neither |
| stupid | 99.87 | drama | 98.51 | devoid | 99.39 | foremost | 82.79 | failure |
| fails | 99.66 | definitely | 96.00 | failure | 98.90 | behalf | 81.14 | whether |
| poor | 99.48 | shows | 95.81 | whether | 97.22 | warmed | 28.42 | despite |
| neither | 99.35 | walk | 94.89 | unless | 94.21 | placement | 10.03 | considering |
| badly | 99.32 | ranks | 93.12 | irrelevant | 93.76 | irrelevant | 6.30 | because |
| crap | 98.91 | but | 86.00 | placement | 90.06 | unless | 5.91 | never |
| pacing | 97.94 | these | 84.96 | warmed | 71.65 | whether | 2.78 | despite |
| whether | 97.03 | wherein | 78.49 | behalf | 19.60 | failure | 1.11 | ball |
| every | 96.81 | weeks | 59.58 | foremost | 17.79 | devoid | 0.61 | notice |
| N | BERT | Rank | N | ALBERT | Rank | N | RoBERTa | |||
| Rank E | Rank | Rank E | Rank | Rank E | Rank | |||||
| succeeded | 97.09 | never | 97.74 | ashamed | 95.06 | after | 99.48 | waves | 95.42 | fire |
| seeing | 95.94 | least | 96.50 | stumbling | 87.91 | under | 99.45 | aside | 95.42 | against |
| merits | 88.23 | killing | 96.19 | tire | 87.91 | because | 98.89 | harming | 88.06 | water |
| longest | 88.23 | prison | 95.58 | cutting | 77.42 | since | 98.27 | skepticism | 88.06 | health |
| feeling | 78.82 | much | 94.49 | declined | 76.26 | how | 97.70 | burns | 88.06 | pleaded |
| raped | 77.95 | welcome | 94.04 | boise | 70.31 | whether | 97.51 | survives | 88.06 | nothing |
| batting | 77.95 | behind | 93.65 | fury | 66.42 | should | 97.12 | votes | 76.46 | won |
| jail | 74.66 | born | 92.60 | slay | 55.80 | cost | 96.57 | justice | 70.62 | NATO |
| banning | 70.64 | needed | 92.26 | reactions | 52.67 | serious | 96.38 | ressing | 65.60 | suffering |
| backing | 70.64 | nothing | 91.30 | disapprove | 52.67 | guilty | 96.31 | deceived | 48.45 | happens |
| MRPC Trigger | l | QNLI Trigger | l | SST-2 Trigger | l | |
| A → B | they ended and | 10 | preis much $100,000 | 3 | unsuitable improper whether | 10 |
| A → R | rostov she blushed | 1 | $100,000 what simulate | 2 | unsuitable improper whether | 10 |
| B → A | nothing pains kilograms | 1 | whichever thirds lithuanian | 9 | folding worse as | 5 |
| B → R | suffered sins declined | 11 | ibly semester longest | 8 | folding worse as | 5 |
| R → A | Avoid water taps | 10 | WHY voted beyond | 2 | decayingjuryNeither | 1 |
| R → B | Avoid water taps | 10 | WHY voted beyond | 2 | surprisingly refreshing lest | 11 |
| ALBERT | BERT | RoBERTa | ||||
| Trigger | l | Trigger | l | Trigger | l | |
| M → Q | nobody reminds austro | 8 | is blood corrupt | 9 | history doubtless beyond | 8 |
| M → S | failing forcing the | 4 | ching stiff punishments | 8 | history doubtless beyond | 8 |
| Q → M | or widen further | 10 | whichever thirds lithuanian | 9 | WHY voted beyond | 2 |
| Q → S | trillion unless marylebone | 5 | ant trees romanized | 5 | wherein perished supra | 7 |
| S → M | unsuitable improper whether | 10 | seemed wiped whoever | 9 | Eating welcome respecting | 7 |
| S → Q | renumbered littered neither | 8 | ogarily diminished | 4 | Crystal tasty ain | 9 |
| Model | L | FR Val. | FR Test | Trigger |
| UAT, BERT | 2 | 38.6 | 38.4 | neither nor |
| UAT, BERT | 4 | 35.4 | 35.0 | situation nonetheless resulted nor |
| Ours, ALBERT, q = 7, i = 4 | 5 | 24.8 | 24.8 | regretted joyahbwv although doubted |
| Ours, ALBERT, q = 4, i = 8 | 6 | 21.4 | 21.4 | tremendous despair towedtrue 1985, doubted |
| Dataset | Model | Triggers |
| MNLI | ALBERT | hamas doubted; neither motioned; i doubted pronoun; cursing neutron unless; transportation workers unless preferring; |
| BERT | but neither; remarkably neither; of neither nor; get neither outta even; backdrop but and neither; ft still maintained not whether | |
| RoBERTa | slideshow Show Neither; HELL Neither; THERE HAS NEVER; ohan Never nor; Fifa VERY NEVER Whether | |
| MRPC | ALBERT | unless if; whereas and; they ended and; unless stumbling out and; commerce corroborat declined and; and siblings live because |
| BERT | against attempt; prepares worse conflicts; was priesthood killing contrary; each mortally duel harta; ring opposed her homosexuality | |
| RoBERTa | Drop until; actresses won awards; Enhanced Population Died Low; Harlem Recommended Submit BLM Question; hum DOWN burns firing Bloody Hyp | |
| QNLI | ALBERT | and 40%; averaged percentage; 88% finance 11%; vittorio whom relinquished; cristo, sharing whom reissued; downloadable bilingual why cancellation |
| BERT | nine charities; carlos orient whom; whose reich declaring; gor german which emperor; maya commuted whose ballet | |
| RoBERTa | nineteen countries; Nearly trillions trillion; Thousand hundred trillion; how MUCH mountains Cass; gui why sued awaits Reviews; | |
| SST-2 | ALBERT | flight unless; suck unnecessary; skidded irrelevant whether; failure placement unless; just whine worthless; fails subdistrict picture |
| BERT | degraded whether; dissatisfied neither; television failed whether; definitely worst wherein whoever; crap stupid feed whereby whether | |
| RoBERTa | powerfully refreshing; Beautiful enjoyable; surprisingly refreshing lest; Crystal importantly beautiful considering; thriller cool because whereas |
| Dataset | Validation | Test | # Classes | BERT\nAcc. Val. / Test | RoBERTa\nAcc. Val. / Test | ALBERT\nAcc. Val. / Test |
| MRPC | 408 | 1725 | 2 | 87.7/84.4 | 90.4/87.2 | 89.7/86.0 |
| QNLI | 5463 | 5463 | 2 | 91.5/90.7 | 91.8/91.8 | 90.6/90.8 |
| MNLI | 9815 | 9796 | 3 | 84.2/83.7 | 86.5/86.3 | 83.8/83.5 |
| SST-2 | 872 | 1821 | 2 | 92.4/93.3 | 94.0/94.9 | 92.7/91.7 |
| RTE | 277 | 3000 | 2 | 72.6/67.6 | 80.5/74.0 | 76.0/72.2 |
| Model | Dataset | FR@1 | FR@2 | FR@3 | FR@4 | FR@5 | FR@6 | ||||||
| W | S | W | S | W | S | W | S | W | S | W | S | ||
| ALBERT | MNLI | 18.5 | 14.8 | 25.8 | 17.0 | 25.8 | 18.5 | 34.8 | 21.1 | 34.8 | 24.8 | 34.8 | 24.8 |
| MRPC | 11.4 | 4.5 | 27.6 | 28.9 | 47.9 | 54.8 | 66.6 | 64.1 | 66.6 | 67.0 | 67.6 | 67.6 | |
| QNLI | 26.4 | 37.5 | 47.5 | 48.1 | 50.5 | 50.8 | 50.5 | 51.5 | 52.2 | 52.4 | 52.5 | 52.4 | |
| RTE | 4.8 | 4.9 | 6.3 | 11.9 | 7.9 | 11.9 | 9.1 | 17.7 | 9.1 | 17.7 | 9.1 | 17.7 | |
| SST-2 | 21.0 | 29.7 | 34.8 | 38.8 | 45.5 | 46.4 | 47.9 | 48.5 | 50.6 | 48.7 | 50.6 | 48.7 | |
| BERT | MNLI | 32.0 | 10.2 | 38.4 | 33.6 | 38.4 | 34.5 | 38.4 | 34.7 | 38.4 | 35.6 | 38.4 | 35.6 |
| MRPC | 15.4 | 9.3 | 64.6 | 56.6 | 70.6 | 69.0 | 70.7 | 70.4 | 70.8 | 70.6 | 70.8 | 70.7 | |
| QNLI | 26.2 | 23.2 | 42.6 | 37.9 | 47.2 | 42.3 | 50.6 | 49.1 | 50.6 | 50.8 | 50.8 | 50.8 | |
| RTE | 4.7 | 4.2 | 6.6 | 6.7 | 7.1 | 8.5 | 7.6 | 10.2 | 7.6 | 10.2 | 11.2 | 10.2 | |
| SST-2 | 29.7 | 27.0 | 42.1 | 38.0 | 42.1 | 44.6 | 42.1 | 48.2 | 50.6 | 49.0 | 50.8 | 49 | |
| RoBERTa | MNLI | 4.8 | 8.2 | 4.9 | 22.3 | 33.0 | 28.8 | 33.0 | 30.0 | 33.0 | 33.1 | 33.0 | 33.1 |
| MRPC | 4.2 | 4.9 | 35.5 | 26.3 | 67.8 | 68.5 | 69.4 | 69.3 | 69.4 | 69.3 | 69.5 | 69.3 | |
| QNLI | 14.7 | 13.3 | 32.6 | 39.0 | 41.5 | 44.9 | 44.5 | 47.3 | 48.1 | 47.7 | 48.1 | 48.3 | |
| RTE | 3.1 | 6.3 | 6.1 | 13.6 | 6.1 | 13.6 | 8.0 | 13.6 | 8.0 | 16.3 | 9.2 | 16.3 | |
| SST-2 | 28.4 | 26.9 | 34.7 | 38.7 | 43.9 | 45.9 | 49.9 | 47.1 | 49.9 | 50.7 | 49.9 | 53.1 | |
| Average | - | 16.3 | 15.0 | 30.0 | 30.5 | 38.4 | 38.9 | 41.5 | 41.5 | 42.7 | 42.9 | 43.1 | 43.2 |
| E | BERT | ALBERT | RoBERTa | ||||
| Rank N | Rank | E | Rank N | Rank | E | Rank N | |
| emperor | 94.69 | either | 99.60 | 70% | 98.95 | those | 99.02 |
| berlin | 93.72 | legislation | 97.60 | 10% | 96.10 | amount | 98.59 |
| whose | 92.14 | can | 97.08 | 3,500 | 93.82 | unless | 93.46 |
| russian | 90.91 | ter | 95.37 | 18% | 93.72 | 71 | 92.72 |
| cardinal | 90.25 | latitude | 93.81 | 20% | 93.45 | nor | 92.46 |
| orient | 89.83 | dalai | 92.04 | whose | 91.79 | where | 91.30 |
| german | 89.03 | samurai | 90.52 | 11% | 88.79 | besides | 90.62 |
| korean | 87.77 | granting | 90.24 | 54% | 88.34 | correlation | 88.56 |
| fide | 84.10 | reich | 90.10 | supplemented | 86.34 | holds | 87.90 |
| atop | 82.91 | banning | 88.47 | 50% | 85.46 | waived | 87.12 |
| From/To | MRPC, W/S | QNLI, W/S | SST-2, W/S | ||||||
| ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | ALBERT | BERT | RoBERTa | |
| ALBERT | - | 22.8/53.5 | 33.9/69.1 | - | 45.6/36.1 | 39.0/38.1 | - | 30.6/37.5 | 24.2/34.7 |
| BERT | 6.7/17.2 | - | 60.5/67.8 | 34.6/36.3 | - | 22.0/36.0 | 31.0/29.3 | - | 17.7/24.1 |
| RoBERTa | 15.3/32.1 | 55.5/64.6 | - | 20.4/40.4 | 18.0/35.9 | - | 34.1/26.2 | 36.9/37.7 | - |
| From/To | ALBERT, W/S | BERT, W/S | RoBERTa, W/S | ||||||
| MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | MRPC | QNLI | SST-2 | |
| MRPC | - | 9.5/21.9 | 5.4/32.5 | - | 18.3/17.6 | 11.8/9.4 | - | 31.4/25.6 | 5.3/6.0 |
| QNLI | 13.6/14.0 | - | 9.0/31.7 | 32.9/35.5 | - | 2.8/4.7 | 22.7/40.9 | - | 2.9/8.9 |
| SST-2 | 8.6/22.8 | 31.3/36.0 | - | 29.7/64.8 | 27.5/21.2 | - | 18.1/13.9 | 14.0/15.9 | - |
| N | BERT | ALBERT | RoBERTa | |||||
| Rank P | Rank | N | Rank P | Rank | N | Rank P | Rank | |
| worst | 99.96 | also | 99.13 | fails | 99.70 | coming | 94.40 | neither |
| stupid | 99.87 | drama | 98.51 | devoid | 99.39 | foremost | 82.79 | failure |
| fails | 99.66 | definitely | 96.00 | failure | 98.90 | behalf | 81.14 | whether |
| poor | 99.48 | shows | 95.81 | whether | 97.22 | warmed | 28.42 | despite |
| neither | 99.35 | walk | 94.89 | unless | 94.21 | placement | 10.03 | considering |
| badly | 99.32 | ranks | 93.12 | irrelevant | 93.76 | irrelevant | 6.30 | because |
| crap | 98.91 | but | 86.00 | placement | 90.06 | unless | 5.91 | never |
| pacing | 97.94 | these | 84.96 | warmed | 71.65 | whether | 2.78 | despite |
| whether | 97.03 | wherein | 78.49 | behalf | 19.60 | failure | 1.11 | ball |
| every | 96.81 | weeks | 59.58 | foremost | 17.79 | devoid | 0.61 | notice |
| N | BERT | Rank | N | ALBERT | Rank | N | RoBERTa | |||
| Rank E | Rank | Rank E | Rank | Rank E | Rank | |||||
| succeeded | 97.09 | never | 97.74 | ashamed | 95.06 | after | 99.48 | waves | 95.42 | fire |
| seeing | 95.94 | least | 96.50 | stumbling | 87.91 | under | 99.45 | aside | 95.42 | against |
| merits | 88.23 | killing | 96.19 | tire | 87.91 | because | 98.89 | harming | 88.06 | water |
| longest | 88.23 | prison | 95.58 | cutting | 77.42 | since | 98.27 | skepticism | 88.06 | health |
| feeling | 78.82 | much | 94.49 | declined | 76.26 | how | 97.70 | burns | 88.06 | pleaded |
| raped | 77.95 | welcome | 94.04 | boise | 70.31 | whether | 97.51 | survives | 88.06 | nothing |
| batting | 77.95 | behind | 93.65 | fury | 66.42 | should | 97.12 | votes | 76.46 | won |
| jail | 74.66 | born | 92.60 | slay | 55.80 | cost | 96.57 | justice | 70.62 | NATO |
| banning | 70.64 | needed | 92.26 | reactions | 52.67 | serious | 96.38 | ressing | 65.60 | suffering |
| backing | 70.64 | nothing | 91.30 | disapprove | 52.67 | guilty | 96.31 | deceived | 48.45 | happens |
| MRPC Trigger | l | QNLI Trigger | l | SST-2 Trigger | l | |
| A → B | they ended and | 10 | preis much $100,000 | 3 | unsuitable improper whether | 10 |
| A → R | rostov she blushed | 1 | $100,000 what simulate | 2 | unsuitable improper whether | 10 |
| B → A | nothing pains kilograms | 1 | whichever thirds lithuanian | 9 | folding worse as | 5 |
| B → R | suffered sins declined | 11 | ibly semester longest | 8 | folding worse as | 5 |
| R → A | Avoid water taps | 10 | WHY voted beyond | 2 | decayingjuryNeither | 1 |
| R → B | Avoid water taps | 10 | WHY voted beyond | 2 | surprisingly refreshing lest | 11 |
| ALBERT | BERT | RoBERTa | ||||
| Trigger | l | Trigger | l | Trigger | l | |
| M → Q | nobody reminds austro | 8 | is blood corrupt | 9 | history doubtless beyond | 8 |
| M → S | failing forcing the | 4 | ching stiff punishments | 8 | history doubtless beyond | 8 |
| Q → M | or widen further | 10 | whichever thirds lithuanian | 9 | WHY voted beyond | 2 |
| Q → S | trillion unless marylebone | 5 | ant trees romanized | 5 | wherein perished supra | 7 |
| S → M | unsuitable improper whether | 10 | seemed wiped whoever | 9 | Eating welcome respecting | 7 |
| S → Q | renumbered littered neither | 8 | ogarily diminished | 4 | Crystal tasty ain | 9 |
| Model | L | FR Val. | FR Test | Trigger |
| UAT, BERT | 2 | 38.6 | 38.4 | neither nor |
| UAT, BERT | 4 | 35.4 | 35.0 | situation nonetheless resulted nor |
| Ours, ALBERT, q = 7, i = 4 | 5 | 24.8 | 24.8 | regretted joyahbwv although doubted |
| Ours, ALBERT, q = 4, i = 8 | 6 | 21.4 | 21.4 | tremendous despair towedtrue 1985, doubted |
| Dataset | Model | Triggers |
| MNLI | ALBERT | hamas doubted; neither motioned; i doubted pronoun; cursing neutron unless; transportation workers unless preferring; |
| BERT | but neither; remarkably neither; of neither nor; get neither outta even; backdrop but and neither; ft still maintained not whether | |
| RoBERTa | slideshow Show Neither; HELL Neither; THERE HAS NEVER; ohan Never nor; Fifa VERY NEVER Whether | |
| MRPC | ALBERT | unless if; whereas and; they ended and; unless stumbling out and; commerce corroborat declined and; and siblings live because |
| BERT | against attempt; prepares worse conflicts; was priesthood killing contrary; each mortally duel harta; ring opposed her homosexuality | |
| RoBERTa | Drop until; actresses won awards; Enhanced Population Died Low; Harlem Recommended Submit BLM Question; hum DOWN burns firing Bloody Hyp | |
| QNLI | ALBERT | and 40%; averaged percentage; 88% finance 11%; vittorio whom relinquished; cristo, sharing whom reissued; downloadable bilingual why cancellation |
| BERT | nine charities; carlos orient whom; whose reich declaring; gor german which emperor; maya commuted whose ballet | |
| RoBERTa | nineteen countries; Nearly trillions trillion; Thousand hundred trillion; how MUCH mountains Cass; gui why sued awaits Reviews; | |
| SST-2 | ALBERT | flight unless; suck unnecessary; skidded irrelevant whether; failure placement unless; just whine worthless; fails subdistrict picture |
| BERT | degraded whether; dissatisfied neither; television failed whether; definitely worst wherein whoever; crap stupid feed whereby whether | |
| RoBERTa | powerfully refreshing; Beautiful enjoyable; surprisingly refreshing lest; Crystal importantly beautiful considering; thriller cool because whereas |
| Missing Entity | |
| Mention Context | EU rejects German call to boycott British lamb. Peter Blackburn BRUS-SELS 1996-08-22 |
| Peter Blackburn (Bishop) | Peter Blackburn (d.1616) was a Scottish scholar and prelate. He was the second Protestant Bishop of Aberdeen. |
| Peter Blackburn (MP) | Peter Blackburn (1811 – 20 May 1870) was a British Conservative Party politician. |
| Peter Blackburn (Badminton) | Peter Grant Blackburn (born 25 March 1968) is an Australian badminton player who affiliated with the Ballarat Badminton Association. |
| Non-Entity Phrase | |
| Mention Context | Most Hindus accept that there is a duty to have a family during the house-holder stage of life, as debt to family lineage called Pitra Rin (Father's Debt) and so are unlikely to avoid having children altogether ... |
| The Householder (Film) | The Householder (Hindi title: Gharbar) is a 1963 film by Merchant Ivory Productions, with a screenplay by Ruth Prawer Jhabvala ... |
| The Householder (Novel) | The Householder is a 1960 English-language novel by Ruth Prawer Jhab-vala ... |
| Dataset | # Data | Annotated | NIL percentage | %Missing Entity | %Non-Entity Phrase |
| AIDA | 34956 | ✓ | 20.41% | 73%* | 10%* |
| MSNBC | 654 | ✓ | 0% | - | - |
| WNED-Wiki | 240000 | × | 0% | - | - |
| NEL (ours) | 9924 | ✓ | 33.57% | 17% | 83% |
| NEL (our dataset) | AIDA w/ NIL | AIDA w/o NIL | MSNBC | WNED-WIKI | |||||
| Non-NAC | NAC | OAC | Non-NAC | NAC | OAC | OAC | OAC | OAC | |
| BLINK-bi | 72.27 | 88.59 | 77.74 | 64.54 | 69.59 | 65.01 | 82.61 | 70.86 | 58.56 |
| CLINK-bi | 79.24 | 79.28 | 79.25 | 75.98 | 66.36 | 75.09 | 83.26 | 73.29 | 58.99 |
| GENRE* | 54.00 | 62.84 | 56.96 | - | - | - | 88.60 | 88.10 | 71.70 |
| BLINK-cross | 84.09 | 88.89 | 85.70 | 83.08 | 45.16 | 79.58 | 87.49 | 82.02 | 69.48 |
| CLINK-cross | 86.97 | 89.19 | 87.71 | 84.42 | 58.53 | 82.03 | 88.16 | 89.70 | 72.43 |
| Model | Ctxt Type Acc. | Cand Type Acc. | OAC w/ Typing | OAC w/o Typing |
| Bi-encoder | 83.75 | 93.46 | 78.35 | 77.74 |
| Cross-encoder | 83.95 | 98.01 | 87.41 | 83.67 |
| Hyperparameter | NEL & AIDA w/ NIL | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.5 | 0.5 | ||
| ε | 0.5 | 0.5 | ||
| Epoch | 4 | {1, 4} | 4 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 2 hrs | - | 3.5 hrs | - |
| Hyperparameter | AIDA w/o NIL, MSNBC & WNED-WIKI | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.9 | [0.0, 1.0] | 0.9 | [0.0, 1.0] |
| ε | 0 | - | 0 | - |
| Epoch | 1 | {1, 4} | 1 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 4 hrs | - | 3 hrs | - |
| Entity | Types |
| 14th Street (Manhattan) | Road->RouteOfTransportation->Infrastructure->ArchitecturalStructure->Place |
| 1958 Copa del Generalíasimo | SoccerTournament->Tournament->SportsEvent->SocietalEvent->Event |
| ATM (song) | Song->MusicalWork->Work |
| Brats (1991 film) | Film->Work |
| Babe Ruth | BaseballPlayer->Athlete->Person |
| Mention Context | Mention | Assumed Entity |
| Bosnian premier in Turkey for one day visit . ANKARA 1996-08-27 | Turkey | Turkey (Country) |
| EU rejects German call to boycott British lamb . Peter Blackburn BRUS-SELS 1996-08-22 | EU | European Union |
| U.S. F-14 catches fire while landing in Israel . JERUSALEM 1996-08-25A U.S. fighter plane blew a tyre and ... | F-14 | Grumman F-14 Tomcat |
| This is the leading story in the Mozambican press on Monday. Reuters has not verified this story and does not vouch for its accuracy. | Reuters | Reuters |
| Mention Context: | ...singer, Michelle Branch, during her visit in Japan to promote her album Hotel Paper, for a magazine interview and photoshoot. After the release of Gates of Heaven the group a short break and performed in New York City ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Gates of Heaven | Gates of Heaven | NIL | Gates of Heaven (album) |
| Mention Context: | ...Michael Williams and David Collins founded the company in 1994 focusing on independent features, including Never Met Picasso (1996), Home Before Dark (1997), Six Ways To Sunday (1998), ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Home Before Dark | Home Before Dark | Home Before Dark | NIL |
| Missing Entity | |
| Mention Context | EU rejects German call to boycott British lamb. Peter Blackburn BRUS-SELS 1996-08-22 |
| Peter Blackburn (Bishop) | Peter Blackburn (d.1616) was a Scottish scholar and prelate. He was the second Protestant Bishop of Aberdeen. |
| Peter Blackburn (MP) | Peter Blackburn (1811 – 20 May 1870) was a British Conservative Party politician. |
| Peter Blackburn (Badminton) | Peter Grant Blackburn (born 25 March 1968) is an Australian badminton player who affiliated with the Ballarat Badminton Association. |
| Non-Entity Phrase | |
| Mention Context | Most Hindus accept that there is a duty to have a family during the house-holder stage of life, as debt to family lineage called Pitra Rin (Father's Debt) and so are unlikely to avoid having children altogether ... |
| The Householder (Film) | The Householder (Hindi title: Gharbar) is a 1963 film by Merchant Ivory Productions, with a screenplay by Ruth Prawer Jhabvala ... |
| The Householder (Novel) | The Householder is a 1960 English-language novel by Ruth Prawer Jhab-vala ... |
| Dataset | # Data | Annotated | NIL percentage | %Missing Entity | %Non-Entity Phrase |
| AIDA | 34956 | ✓ | 20.41% | 73%* | 10%* |
| MSNBC | 654 | ✓ | 0% | - | - |
| WNED-Wiki | 240000 | × | 0% | - | - |
| NEL (ours) | 9924 | ✓ | 33.57% | 17% | 83% |
| NEL (our dataset) | AIDA w/ NIL | AIDA w/o NIL | MSNBC | WNED-WIKI | |||||
| Non-NAC | NAC | OAC | Non-NAC | NAC | OAC | OAC | OAC | OAC | |
| BLINK-bi | 72.27 | 88.59 | 77.74 | 64.54 | 69.59 | 65.01 | 82.61 | 70.86 | 58.56 |
| CLINK-bi | 79.24 | 79.28 | 79.25 | 75.98 | 66.36 | 75.09 | 83.26 | 73.29 | 58.99 |
| GENRE* | 54.00 | 62.84 | 56.96 | - | - | - | 88.60 | 88.10 | 71.70 |
| BLINK-cross | 84.09 | 88.89 | 85.70 | 83.08 | 45.16 | 79.58 | 87.49 | 82.02 | 69.48 |
| CLINK-cross | 86.97 | 89.19 | 87.71 | 84.42 | 58.53 | 82.03 | 88.16 | 89.70 | 72.43 |
| Model | Ctxt Type Acc. | Cand Type Acc. | OAC w/ Typing | OAC w/o Typing |
| Bi-encoder | 83.75 | 93.46 | 78.35 | 77.74 |
| Cross-encoder | 83.95 | 98.01 | 87.41 | 83.67 |
| Hyperparameter | NEL & AIDA w/ NIL | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.5 | 0.5 | ||
| ε | 0.5 | 0.5 | ||
| Epoch | 4 | {1, 4} | 4 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 2 hrs | - | 3.5 hrs | - |
| Hyperparameter | AIDA w/o NIL, MSNBC & WNED-WIKI | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.9 | [0.0, 1.0] | 0.9 | [0.0, 1.0] |
| ε | 0 | - | 0 | - |
| Epoch | 1 | {1, 4} | 1 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 4 hrs | - | 3 hrs | - |
| Entity | Types |
| 14th Street (Manhattan) | Road->RouteOfTransportation->Infrastructure->ArchitecturalStructure->Place |
| 1958 Copa del Generalíasimo | SoccerTournament->Tournament->SportsEvent->SocietalEvent->Event |
| ATM (song) | Song->MusicalWork->Work |
| Brats (1991 film) | Film->Work |
| Babe Ruth | BaseballPlayer->Athlete->Person |
| Mention Context | Mention | Assumed Entity |
| Bosnian premier in Turkey for one day visit . ANKARA 1996-08-27 | Turkey | Turkey (Country) |
| EU rejects German call to boycott British lamb . Peter Blackburn BRUS-SELS 1996-08-22 | EU | European Union |
| U.S. F-14 catches fire while landing in Israel . JERUSALEM 1996-08-25A U.S. fighter plane blew a tyre and ... | F-14 | Grumman F-14 Tomcat |
| This is the leading story in the Mozambican press on Monday. Reuters has not verified this story and does not vouch for its accuracy. | Reuters | Reuters |
| Mention Context: | ...singer, Michelle Branch, during her visit in Japan to promote her album Hotel Paper, for a magazine interview and photoshoot. After the release of Gates of Heaven the group a short break and performed in New York City ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Gates of Heaven | Gates of Heaven | NIL | Gates of Heaven (album) |
| Mention Context: | ...Michael Williams and David Collins founded the company in 1994 focusing on independent features, including Never Met Picasso (1996), Home Before Dark (1997), Six Ways To Sunday (1998), ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Home Before Dark | Home Before Dark | Home Before Dark | NIL |
| Missing Entity | |
| Mention Context | EU rejects German call to boycott British lamb. Peter Blackburn BRUS-SELS 1996-08-22 |
| Peter Blackburn (Bishop) | Peter Blackburn (d.1616) was a Scottish scholar and prelate. He was the second Protestant Bishop of Aberdeen. |
| Peter Blackburn (MP) | Peter Blackburn (1811 – 20 May 1870) was a British Conservative Party politician. |
| Peter Blackburn (Badminton) | Peter Grant Blackburn (born 25 March 1968) is an Australian badminton player who affiliated with the Ballarat Badminton Association. |
| Non-Entity Phrase | |
| Mention Context | Most Hindus accept that there is a duty to have a family during the house-holder stage of life, as debt to family lineage called Pitra Rin (Father's Debt) and so are unlikely to avoid having children altogether ... |
| The Householder (Film) | The Householder (Hindi title: Gharbar) is a 1963 film by Merchant Ivory Productions, with a screenplay by Ruth Prawer Jhabvala ... |
| The Householder (Novel) | The Householder is a 1960 English-language novel by Ruth Prawer Jhab-vala ... |
| Dataset | # Data | Annotated | NIL percentage | %Missing Entity | %Non-Entity Phrase |
| AIDA | 34956 | ✓ | 20.41% | 73%* | 10%* |
| MSNBC | 654 | ✓ | 0% | - | - |
| WNED-Wiki | 240000 | × | 0% | - | - |
| NEL (ours) | 9924 | ✓ | 33.57% | 17% | 83% |
| NEL (our dataset) | AIDA w/ NIL | AIDA w/o NIL | MSNBC | WNED-WIKI | |||||
| Non-NAC | NAC | OAC | Non-NAC | NAC | OAC | OAC | OAC | OAC | |
| BLINK-bi | 72.27 | 88.59 | 77.74 | 64.54 | 69.59 | 65.01 | 82.61 | 70.86 | 58.56 |
| CLINK-bi | 79.24 | 79.28 | 79.25 | 75.98 | 66.36 | 75.09 | 83.26 | 73.29 | 58.99 |
| GENRE* | 54.00 | 62.84 | 56.96 | - | - | - | 88.60 | 88.10 | 71.70 |
| BLINK-cross | 84.09 | 88.89 | 85.70 | 83.08 | 45.16 | 79.58 | 87.49 | 82.02 | 69.48 |
| CLINK-cross | 86.97 | 89.19 | 87.71 | 84.42 | 58.53 | 82.03 | 88.16 | 89.70 | 72.43 |
| Model | Ctxt Type Acc. | Cand Type Acc. | OAC w/ Typing | OAC w/o Typing |
| Bi-encoder | 83.75 | 93.46 | 78.35 | 77.74 |
| Cross-encoder | 83.95 | 98.01 | 87.41 | 83.67 |
| Hyperparameter | NEL & AIDA w/ NIL | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.5 | 0.5 | ||
| ε | 0.5 | 0.5 | ||
| Epoch | 4 | {1, 4} | 4 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 2 hrs | - | 3.5 hrs | - |
| Hyperparameter | AIDA w/o NIL, MSNBC & WNED-WIKI | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.9 | [0.0, 1.0] | 0.9 | [0.0, 1.0] |
| ε | 0 | - | 0 | - |
| Epoch | 1 | {1, 4} | 1 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 4 hrs | - | 3 hrs | - |
| Entity | Types |
| 14th Street (Manhattan) | Road->RouteOfTransportation->Infrastructure->ArchitecturalStructure->Place |
| 1958 Copa del Generalíasimo | SoccerTournament->Tournament->SportsEvent->SocietalEvent->Event |
| ATM (song) | Song->MusicalWork->Work |
| Brats (1991 film) | Film->Work |
| Babe Ruth | BaseballPlayer->Athlete->Person |
| Mention Context | Mention | Assumed Entity |
| Bosnian premier in Turkey for one day visit . ANKARA 1996-08-27 | Turkey | Turkey (Country) |
| EU rejects German call to boycott British lamb . Peter Blackburn BRUS-SELS 1996-08-22 | EU | European Union |
| U.S. F-14 catches fire while landing in Israel . JERUSALEM 1996-08-25A U.S. fighter plane blew a tyre and ... | F-14 | Grumman F-14 Tomcat |
| This is the leading story in the Mozambican press on Monday. Reuters has not verified this story and does not vouch for its accuracy. | Reuters | Reuters |
| Mention Context: | ...singer, Michelle Branch, during her visit in Japan to promote her album Hotel Paper, for a magazine interview and photoshoot. After the release of Gates of Heaven the group a short break and performed in New York City ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Gates of Heaven | Gates of Heaven | NIL | Gates of Heaven (album) |
| Mention Context: | ...Michael Williams and David Collins founded the company in 1994 focusing on independent features, including Never Met Picasso (1996), Home Before Dark (1997), Six Ways To Sunday (1998), ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Home Before Dark | Home Before Dark | Home Before Dark | NIL |
| Missing Entity | |
| Mention Context | EU rejects German call to boycott British lamb. Peter Blackburn BRUS-SELS 1996-08-22 |
| Peter Blackburn (Bishop) | Peter Blackburn (d.1616) was a Scottish scholar and prelate. He was the second Protestant Bishop of Aberdeen. |
| Peter Blackburn (MP) | Peter Blackburn (1811 – 20 May 1870) was a British Conservative Party politician. |
| Peter Blackburn (Badminton) | Peter Grant Blackburn (born 25 March 1968) is an Australian badminton player who affiliated with the Ballarat Badminton Association. |
| Non-Entity Phrase | |
| Mention Context | Most Hindus accept that there is a duty to have a family during the house-holder stage of life, as debt to family lineage called Pitra Rin (Father's Debt) and so are unlikely to avoid having children altogether ... |
| The Householder (Film) | The Householder (Hindi title: Gharbar) is a 1963 film by Merchant Ivory Productions, with a screenplay by Ruth Prawer Jhabvala ... |
| The Householder (Novel) | The Householder is a 1960 English-language novel by Ruth Prawer Jhab-vala ... |
| Dataset | # Data | Annotated | NIL percentage | %Missing Entity | %Non-Entity Phrase |
| AIDA | 34956 | ✓ | 20.41% | 73%* | 10%* |
| MSNBC | 654 | ✓ | 0% | - | - |
| WNED-Wiki | 240000 | × | 0% | - | - |
| NEL (ours) | 9924 | ✓ | 33.57% | 17% | 83% |
| NEL (our dataset) | AIDA w/ NIL | AIDA w/o NIL | MSNBC | WNED-WIKI | |||||
| Non-NAC | NAC | OAC | Non-NAC | NAC | OAC | OAC | OAC | OAC | |
| BLINK-bi | 72.27 | 88.59 | 77.74 | 64.54 | 69.59 | 65.01 | 82.61 | 70.86 | 58.56 |
| CLINK-bi | 79.24 | 79.28 | 79.25 | 75.98 | 66.36 | 75.09 | 83.26 | 73.29 | 58.99 |
| GENRE* | 54.00 | 62.84 | 56.96 | - | - | - | 88.60 | 88.10 | 71.70 |
| BLINK-cross | 84.09 | 88.89 | 85.70 | 83.08 | 45.16 | 79.58 | 87.49 | 82.02 | 69.48 |
| CLINK-cross | 86.97 | 89.19 | 87.71 | 84.42 | 58.53 | 82.03 | 88.16 | 89.70 | 72.43 |
| Model | Ctxt Type Acc. | Cand Type Acc. | OAC w/ Typing | OAC w/o Typing |
| Bi-encoder | 83.75 | 93.46 | 78.35 | 77.74 |
| Cross-encoder | 83.95 | 98.01 | 87.41 | 83.67 |
| Hyperparameter | NEL & AIDA w/ NIL | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.5 | 0.5 | ||
| ε | 0.5 | 0.5 | ||
| Epoch | 4 | {1, 4} | 4 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 2 hrs | - | 3.5 hrs | - |
| Hyperparameter | AIDA w/o NIL, MSNBC & WNED-WIKI | |||
| CLINK-bi | CLINK-cross | |||
| Value | Range | Value | Range | |
| Learning Rate | 1e-5 | {1e-3, 1e-5, 3e-5} | 1e-5 | {1e-3, 1e-5, 3e-5} |
| λ | 0.9 | [0.0, 1.0] | 0.9 | [0.0, 1.0] |
| ε | 0 | - | 0 | - |
| Epoch | 1 | {1, 4} | 1 | {1, 4} |
| Batch Size | 4 | {1, 4, 8, 16, 32} | 1 | {1, 4, 8, 16, 32} |
| # Parameters | 670M | - | 335M | - |
| Training Time | 4 hrs | - | 3 hrs | - |
| Entity | Types |
| 14th Street (Manhattan) | Road->RouteOfTransportation->Infrastructure->ArchitecturalStructure->Place |
| 1958 Copa del Generalíasimo | SoccerTournament->Tournament->SportsEvent->SocietalEvent->Event |
| ATM (song) | Song->MusicalWork->Work |
| Brats (1991 film) | Film->Work |
| Babe Ruth | BaseballPlayer->Athlete->Person |
| Mention Context | Mention | Assumed Entity |
| Bosnian premier in Turkey for one day visit . ANKARA 1996-08-27 | Turkey | Turkey (Country) |
| EU rejects German call to boycott British lamb . Peter Blackburn BRUS-SELS 1996-08-22 | EU | European Union |
| U.S. F-14 catches fire while landing in Israel . JERUSALEM 1996-08-25A U.S. fighter plane blew a tyre and ... | F-14 | Grumman F-14 Tomcat |
| This is the leading story in the Mozambican press on Monday. Reuters has not verified this story and does not vouch for its accuracy. | Reuters | Reuters |
| Mention Context: | ...singer, Michelle Branch, during her visit in Japan to promote her album Hotel Paper, for a magazine interview and photoshoot. After the release of Gates of Heaven the group a short break and performed in New York City ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Gates of Heaven | Gates of Heaven | NIL | Gates of Heaven (album) |
| Mention Context: | ...Michael Williams and David Collins founded the company in 1994 focusing on independent features, including Never Met Picasso (1996), Home Before Dark (1997), Six Ways To Sunday (1998), ... | |||
| Model | BLINK-bi | CLINK-bi | BLINK-cross | CLINK-cross |
| Prediction | Home Before Dark | Home Before Dark | Home Before Dark | NIL |
| Model | RTE | MRPC |
| M | 79.1 (0.5) | 89.3 (0.4) |
| Mgelu | 79.3 (0.2) | 89.5 (0.3) |
| Mswish | 79.6 (0.3) | 89.2 (0.4) |
| Mtanh | 79.2 (0.5) | 88.9 (0.6) |
| Msa | 79.5 (0.3) | 89.6 (0.2) |
| Mconv | 79.5 (0.6) | 89.4 (0.5) |
| M attenu | 79.0 (0.4) | 89.1 (0.3) |
| Mblock | 79.6 (0.5) | 89.4 (0.3) |
| Method | Tunable \nParams | CoLA \n(mcc) | SST-2 \n(acc) | MRPC \n(acc-f1) | QQP \n(acc-f1) | STS-B \n(corr) | MNLI \n(acc) | QNLI \n(acc) | RTE \n(acc) | Avg |
| Baselines | ||||||||||
| RoBERTa-large | 355M | 65.3 | 95.4 | 91.5 | 90.3 | 91.8 | 89.8 | 94.5 | 80.1 | 87.3 |
| Prompt Tuning | 21K | 54.6 (2.5) | 90.9 (0.5) | 74.8 (1.3) | 87.4 (0.5) | 90.1 (0.3) | 83.4 (0.2) | 92.4 (0.3) | 68.7 (1.9) | 80.3 |
| P-tuning v2 | 985K | 57.3 (2.1) | 92.3 (0.3) | 86.9 (1.2) | 88.1 (0.4) | 90.4 (0.1) | 87.3 (0.2) | 92.9 (0.4) | 77.2 (2.1) | 84.1 |
| BitFit | 273K | 59.2 (0.9) | 94.1 (0.3) | 88.6 (0.8) | 88.5 (0.6) | 91.1 (0.3) | 87.3 (0.1) | 93.2 (0.2) | 78.6 (1.4) | 85.1 |
| LoRA | 778K | 59.7 (1.4) | 93.6 (0.1) | 88.9 (0.7) | 88.3 (0.4) | 90.9 (0.2) | 87.9 (0.2) | 93.2 (0.1) | 78.8 (1.3) | 85.2 |
| UNIPELT | 1.4M | 61.5 (1.7) | 93.6 (0.2) | 89.2 (0.8) | 88.7 (0.4) | 91.0 (0.5) | 87.5 (0.2) | 92.9 (0.1) | 79.1 (0.9) | 85.4 |
| Adapter | 1.6M | 63.1 (1.4) | 93.8 (0.1) | 87.5 (0.8) | 88.7 (0.4) | 90.9 (0.4) | 88.5 (0.2) | 93.2 (0.1) | 78.9 (1.3) | 85.6 |
| LST | 1.7M | 63.6 (1.5) | 93.9 (0.2) | 89.2 (0.7) | 88.7 (0.2) | 91.2 (0.4) | 88.3 (0.1) | 93.0 (0.2) | 79.2 (0.9) | 85.9 |
| Parallel Adapters | 279K | 63.8 (1.5) | 94.2 (0.3) | 89.3 (0.5) | 88.9 (0.6) | 91.1 (0.4) | 88.3 (0.1) | 93.2 (0.1) | 79.1 (0.5) | 86.0 |
| Compacter | 279K | 63.7 (0.9) | 94.2 (0.4) | 89.1 (0.7) | 88.6 (0.3) | 90.8 (0.2) | 88.2 (0.2) | 92.8 (0.2) | 78.7 (0.9) | 85.8 |
| Our proposed methods | ||||||||||
| Learned Adapter | 294K | 64.3 (0.9) | 94.9 (0.3) | 89.8 (0.9) | 89.2 (0.3) | 91.3 (0.2) | 88.6 (0.2) | 93.5 (0.2) | 80.4 (1.3) | 86.5 |
| Source | Target | |
| RTE | SST-2 | |
| RTE | 80.4 (1.3) | 94.6 (0.1) |
| SST-2 | 80.3 (1.4) | 94.9 (0.3) |
| Search space | RTE (acc) | SST-2 (acc) |
| S | 80.4 (1.3) | 94.9 (0.3) |
| S1 | 80.1 (1.5) | 94.6 (0.3) |
| S2 | 79.8 (1.0) | 94.4 (0.4) |
| Parallel Adapter | 79.1 (0.5) | 94.2 (0.3) |
| Method | Tunable \nParams | RTE \n(acc) | SST-2 \n(acc) |
| DeBERTa-large backbone | |||
| Fine-tuning | 406M | 82.1 (1.1) | 95.7 (0.2) |
| Adapters | 1.6M | 80.7 (1.3) | 93.8 (0.1) |
| Parallel Adapters | 279K | 81.5 (1.2) | 93.9 (0.3) |
| Learned Adapter | 245K | 82.0 (0.9) | 94.5 (0.3) |
| GPT2-large backbone | |||
| Fine-tuning | 774M | 79.5 (0.8) | 95.5 (0.1) |
| Adapters | 2.7M | 78.1 (0.9) | 93.9 (0.1) |
| Parallel Adapters | 462K | 78.5 (1.2) | 93.8 (0.1) |
| Learned Adapter | 482K | 79.1 (0.6) | 94.2 (0.2) |
| Search method | RTE (acc) | SST-2 (acc) |
| GDNAS | 80.4 (1.3) | 94.9 (0.3) |
| GDNAS (single-level) | 80.2 (0.9) | 94.7 (0.2) |
| DARTS | 79.8 (1.2) | 94.5 (0.5) |
| ENAS | 80.3 (1.0) | 94.3 (0.4) |
| Category | Datasets | ltrainl | ldevl | ltestl | |y| | Type | Labels |
| Single-sentence | SST-2 | 66349 | 1000 | 872 | 2 | sentiment | positive, negative |
| CoLA | 8551 | 521 | 522 | 2 | linguistic acceptability | acceptable, not acceptable | |
| Sentence-pair | MNLI | 391702 | 1000 | 19647 | 3 | NLI | entailment, neutral, contradiction |
| MRPC | 2668 | 1000 | 408 | 2 | paraphrase | equivalent, not equivalent | |
| QNLI | 103743 | 1000 | 5463 | 2 | NLI | entailment, not entailment | |
| QQP | 362846 | 1000 | 40430 | 2 | paraphrase | equivalent, not equivalent | |
| RTE | 1490 | 1000 | 277 | 2 | NLI | entailment, not entailment | |
| STS-B | 5750 | 750 | 751 | - | semantic similarity | - |
| Architecture | Test Error (%) | Params (M) | Search Cost (GPU days) | Search Framework |
| NASNet-A (Zoph et al., 2017) | 2.65 | 3.3 | 2000 | RL |
| AmoebaNet (Real et al., 2018) | 3.34 | 3.2 | 3150 | evolution |
| ENAS (Pham et al., 2018) | 2.89 | 4.6 | 0.5 | RL |
| DARTS (Liu et al., 2019a) | 3.02 | 3.3 | 0.4 | DNAS |
| SNAS (Xie et al., 2018) | 2.85 | 2.8 | 1.5 | DNAS |
| BayesNAS (Zhou et al., 2019) | 2.81 | 3.4 | 0.2 | DNAS |
| P-DARTS (Chen et al., 2021) | 2.50 | 3.4 | 0.3 | DNAS |
| R-DARTS (Zela et al., 2019) | 2.95 | - | 1.6 | DNAS |
| GDNAS (ours) | 2.52 | 3.3 | 0.6 | DNAS |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | FFN | elu | null_ACT | mha_8 |
| 2 | - | - | - | - |
| 3 | - | - | - | - |
| 4 | FFN | elu | null_ACT | mha_2 |
| 5 | Block | tanh | null_ACT | mha_2 |
| 6 | FFN | gelu_new | leaky_relu | conv_3 |
| 7 | FFN | null_ACT | tanh | mha_2 |
| 8 | - | - | - | - |
| 9 | Attn | elu | relu | conv_3 |
| 10 | - | - | - | - |
| 11 | Attn | gelu_new | relu | conv_1 |
| 12 | Attn | gelu_new | relu | conv_3 |
| 13 | Block | relu | leaky_relu | conv_1 |
| 14 | Attn | swish | relu | conv_3 |
| 15 | FFN | leaky_relu | relu | conv_3 |
| 16 | Block | leaky_relu | relu | conv_5 |
| 17 | FFN | leaky_relu | relu | conv_1 |
| 18 | Attn | leaky_relu | null_ACT | skip_connect |
| 19 | FFN | relu | relu | conv_3 |
| 20 | FFN | gelu_new | null_ACT | conv_3 |
| 21 | - | - | - | - |
| 22 | Block | tanh | null_ACT | mha_8 |
| 23 | Attn | tanh | null_ACT | conv_3 |
| 24 | FFN | tanh | null_ACT | mha_2 |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | Attn | null_ACT | null_ACT | conv_5 |
| 2 | - | - | - | - |
| 3 | FFN | elu | null_ACT | mha_2 |
| 4 | Attn | tanh | null_ACT | mha_2 |
| 5 | Attn | elu | null_ACT | skip_connect |
| 6 | FFN | gelu_new | relu | conv_5 |
| 7 | Block | leaky_relu | leaky_relu | conv_3 |
| 8 | - | - | - | - |
| 9 | Block | leaky_relu | leaky_relu | conv_5 |
| 10 | FFN | relu | relu | conv_3 |
| 11 | Block | tanh | null_ACT | mha_8 |
| 12 | - | - | - | - |
| 13 | Block | tanh | null_ACT | skip_connect |
| 14 | - | - | - | - |
| 15 | FFN | swish | swish | conv_1 |
| 16 | Block | relu | relu | conv_1 |
| 17 | FFN | relu | swish | conv_3 |
| 18 | FFN | leaky_relu | relu | conv_3 |
| 19 | Attn | gelu_new | relu | mha_8 |
| 20 | FFN | swish | relu | mha_8 |
| 21 | Attn | swish | null_ACT | conv_3 |
| 22 | Attn | elu | null_ACT | conv_3 |
| 23 | Attn | null_ACT | null_ACT | mha_8 |
| 24 | Attn | elu | null_ACT | mha_2 |
| Model | RTE | MRPC |
| M | 79.1 (0.5) | 89.3 (0.4) |
| Mgelu | 79.3 (0.2) | 89.5 (0.3) |
| Mswish | 79.6 (0.3) | 89.2 (0.4) |
| Mtanh | 79.2 (0.5) | 88.9 (0.6) |
| Msa | 79.5 (0.3) | 89.6 (0.2) |
| Mconv | 79.5 (0.6) | 89.4 (0.5) |
| M attenu | 79.0 (0.4) | 89.1 (0.3) |
| Mblock | 79.6 (0.5) | 89.4 (0.3) |
| Method | Tunable \nParams | CoLA \n(mcc) | SST-2 \n(acc) | MRPC \n(acc-f1) | QQP \n(acc-f1) | STS-B \n(corr) | MNLI \n(acc) | QNLI \n(acc) | RTE \n(acc) | Avg |
| Baselines | ||||||||||
| RoBERTa-large | 355M | 65.3 | 95.4 | 91.5 | 90.3 | 91.8 | 89.8 | 94.5 | 80.1 | 87.3 |
| Prompt Tuning | 21K | 54.6 (2.5) | 90.9 (0.5) | 74.8 (1.3) | 87.4 (0.5) | 90.1 (0.3) | 83.4 (0.2) | 92.4 (0.3) | 68.7 (1.9) | 80.3 |
| P-tuning v2 | 985K | 57.3 (2.1) | 92.3 (0.3) | 86.9 (1.2) | 88.1 (0.4) | 90.4 (0.1) | 87.3 (0.2) | 92.9 (0.4) | 77.2 (2.1) | 84.1 |
| BitFit | 273K | 59.2 (0.9) | 94.1 (0.3) | 88.6 (0.8) | 88.5 (0.6) | 91.1 (0.3) | 87.3 (0.1) | 93.2 (0.2) | 78.6 (1.4) | 85.1 |
| LoRA | 778K | 59.7 (1.4) | 93.6 (0.1) | 88.9 (0.7) | 88.3 (0.4) | 90.9 (0.2) | 87.9 (0.2) | 93.2 (0.1) | 78.8 (1.3) | 85.2 |
| UNIPELT | 1.4M | 61.5 (1.7) | 93.6 (0.2) | 89.2 (0.8) | 88.7 (0.4) | 91.0 (0.5) | 87.5 (0.2) | 92.9 (0.1) | 79.1 (0.9) | 85.4 |
| Adapter | 1.6M | 63.1 (1.4) | 93.8 (0.1) | 87.5 (0.8) | 88.7 (0.4) | 90.9 (0.4) | 88.5 (0.2) | 93.2 (0.1) | 78.9 (1.3) | 85.6 |
| LST | 1.7M | 63.6 (1.5) | 93.9 (0.2) | 89.2 (0.7) | 88.7 (0.2) | 91.2 (0.4) | 88.3 (0.1) | 93.0 (0.2) | 79.2 (0.9) | 85.9 |
| Parallel Adapters | 279K | 63.8 (1.5) | 94.2 (0.3) | 89.3 (0.5) | 88.9 (0.6) | 91.1 (0.4) | 88.3 (0.1) | 93.2 (0.1) | 79.1 (0.5) | 86.0 |
| Compacter | 279K | 63.7 (0.9) | 94.2 (0.4) | 89.1 (0.7) | 88.6 (0.3) | 90.8 (0.2) | 88.2 (0.2) | 92.8 (0.2) | 78.7 (0.9) | 85.8 |
| Our proposed methods | ||||||||||
| Learned Adapter | 294K | 64.3 (0.9) | 94.9 (0.3) | 89.8 (0.9) | 89.2 (0.3) | 91.3 (0.2) | 88.6 (0.2) | 93.5 (0.2) | 80.4 (1.3) | 86.5 |
| Source | Target | |
| RTE | SST-2 | |
| RTE | 80.4 (1.3) | 94.6 (0.1) |
| SST-2 | 80.3 (1.4) | 94.9 (0.3) |
| Search space | RTE (acc) | SST-2 (acc) |
| S | 80.4 (1.3) | 94.9 (0.3) |
| S1 | 80.1 (1.5) | 94.6 (0.3) |
| S2 | 79.8 (1.0) | 94.4 (0.4) |
| Parallel Adapter | 79.1 (0.5) | 94.2 (0.3) |
| Method | Tunable \nParams | RTE \n(acc) | SST-2 \n(acc) |
| DeBERTa-large backbone | |||
| Fine-tuning | 406M | 82.1 (1.1) | 95.7 (0.2) |
| Adapters | 1.6M | 80.7 (1.3) | 93.8 (0.1) |
| Parallel Adapters | 279K | 81.5 (1.2) | 93.9 (0.3) |
| Learned Adapter | 245K | 82.0 (0.9) | 94.5 (0.3) |
| GPT2-large backbone | |||
| Fine-tuning | 774M | 79.5 (0.8) | 95.5 (0.1) |
| Adapters | 2.7M | 78.1 (0.9) | 93.9 (0.1) |
| Parallel Adapters | 462K | 78.5 (1.2) | 93.8 (0.1) |
| Learned Adapter | 482K | 79.1 (0.6) | 94.2 (0.2) |
| Search method | RTE (acc) | SST-2 (acc) |
| GDNAS | 80.4 (1.3) | 94.9 (0.3) |
| GDNAS (single-level) | 80.2 (0.9) | 94.7 (0.2) |
| DARTS | 79.8 (1.2) | 94.5 (0.5) |
| ENAS | 80.3 (1.0) | 94.3 (0.4) |
| Category | Datasets | ltrainl | ldevl | ltestl | |y| | Type | Labels |
| Single-sentence | SST-2 | 66349 | 1000 | 872 | 2 | sentiment | positive, negative |
| CoLA | 8551 | 521 | 522 | 2 | linguistic acceptability | acceptable, not acceptable | |
| Sentence-pair | MNLI | 391702 | 1000 | 19647 | 3 | NLI | entailment, neutral, contradiction |
| MRPC | 2668 | 1000 | 408 | 2 | paraphrase | equivalent, not equivalent | |
| QNLI | 103743 | 1000 | 5463 | 2 | NLI | entailment, not entailment | |
| QQP | 362846 | 1000 | 40430 | 2 | paraphrase | equivalent, not equivalent | |
| RTE | 1490 | 1000 | 277 | 2 | NLI | entailment, not entailment | |
| STS-B | 5750 | 750 | 751 | - | semantic similarity | - |
| Architecture | Test Error (%) | Params (M) | Search Cost (GPU days) | Search Framework |
| NASNet-A (Zoph et al., 2017) | 2.65 | 3.3 | 2000 | RL |
| AmoebaNet (Real et al., 2018) | 3.34 | 3.2 | 3150 | evolution |
| ENAS (Pham et al., 2018) | 2.89 | 4.6 | 0.5 | RL |
| DARTS (Liu et al., 2019a) | 3.02 | 3.3 | 0.4 | DNAS |
| SNAS (Xie et al., 2018) | 2.85 | 2.8 | 1.5 | DNAS |
| BayesNAS (Zhou et al., 2019) | 2.81 | 3.4 | 0.2 | DNAS |
| P-DARTS (Chen et al., 2021) | 2.50 | 3.4 | 0.3 | DNAS |
| R-DARTS (Zela et al., 2019) | 2.95 | - | 1.6 | DNAS |
| GDNAS (ours) | 2.52 | 3.3 | 0.6 | DNAS |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | FFN | elu | null_ACT | mha_8 |
| 2 | - | - | - | - |
| 3 | - | - | - | - |
| 4 | FFN | elu | null_ACT | mha_2 |
| 5 | Block | tanh | null_ACT | mha_2 |
| 6 | FFN | gelu_new | leaky_relu | conv_3 |
| 7 | FFN | null_ACT | tanh | mha_2 |
| 8 | - | - | - | - |
| 9 | Attn | elu | relu | conv_3 |
| 10 | - | - | - | - |
| 11 | Attn | gelu_new | relu | conv_1 |
| 12 | Attn | gelu_new | relu | conv_3 |
| 13 | Block | relu | leaky_relu | conv_1 |
| 14 | Attn | swish | relu | conv_3 |
| 15 | FFN | leaky_relu | relu | conv_3 |
| 16 | Block | leaky_relu | relu | conv_5 |
| 17 | FFN | leaky_relu | relu | conv_1 |
| 18 | Attn | leaky_relu | null_ACT | skip_connect |
| 19 | FFN | relu | relu | conv_3 |
| 20 | FFN | gelu_new | null_ACT | conv_3 |
| 21 | - | - | - | - |
| 22 | Block | tanh | null_ACT | mha_8 |
| 23 | Attn | tanh | null_ACT | conv_3 |
| 24 | FFN | tanh | null_ACT | mha_2 |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | Attn | null_ACT | null_ACT | conv_5 |
| 2 | - | - | - | - |
| 3 | FFN | elu | null_ACT | mha_2 |
| 4 | Attn | tanh | null_ACT | mha_2 |
| 5 | Attn | elu | null_ACT | skip_connect |
| 6 | FFN | gelu_new | relu | conv_5 |
| 7 | Block | leaky_relu | leaky_relu | conv_3 |
| 8 | - | - | - | - |
| 9 | Block | leaky_relu | leaky_relu | conv_5 |
| 10 | FFN | relu | relu | conv_3 |
| 11 | Block | tanh | null_ACT | mha_8 |
| 12 | - | - | - | - |
| 13 | Block | tanh | null_ACT | skip_connect |
| 14 | - | - | - | - |
| 15 | FFN | swish | swish | conv_1 |
| 16 | Block | relu | relu | conv_1 |
| 17 | FFN | relu | swish | conv_3 |
| 18 | FFN | leaky_relu | relu | conv_3 |
| 19 | Attn | gelu_new | relu | mha_8 |
| 20 | FFN | swish | relu | mha_8 |
| 21 | Attn | swish | null_ACT | conv_3 |
| 22 | Attn | elu | null_ACT | conv_3 |
| 23 | Attn | null_ACT | null_ACT | mha_8 |
| 24 | Attn | elu | null_ACT | mha_2 |
| Model | RTE | MRPC |
| M | 79.1 (0.5) | 89.3 (0.4) |
| Mgelu | 79.3 (0.2) | 89.5 (0.3) |
| Mswish | 79.6 (0.3) | 89.2 (0.4) |
| Mtanh | 79.2 (0.5) | 88.9 (0.6) |
| Msa | 79.5 (0.3) | 89.6 (0.2) |
| Mconv | 79.5 (0.6) | 89.4 (0.5) |
| M attenu | 79.0 (0.4) | 89.1 (0.3) |
| Mblock | 79.6 (0.5) | 89.4 (0.3) |
| Method | Tunable +Params | CoLA +(mcc) | SST-2 +(acc) | MRPC +(acc-f1) | QQP +(acc-f1) | STS-B +(corr) | MNLI +(acc) | QNLI +(acc) | RTE +(acc) | Avg |
| Baselines | ||||||||||
| RoBERTa-large | 355M | 65.3 | 95.4 | 91.5 | 90.3 | 91.8 | 89.8 | 94.5 | 80.1 | 87.3 |
| Prompt Tuning | 21K | 54.6 (2.5) | 90.9 (0.5) | 74.8 (1.3) | 87.4 (0.5) | 90.1 (0.3) | 83.4 (0.2) | 92.4 (0.3) | 68.7 (1.9) | 80.3 |
| P-tuning v2 | 985K | 57.3 (2.1) | 92.3 (0.3) | 86.9 (1.2) | 88.1 (0.4) | 90.4 (0.1) | 87.3 (0.2) | 92.9 (0.4) | 77.2 (2.1) | 84.1 |
| BitFit | 273K | 59.2 (0.9) | 94.1 (0.3) | 88.6 (0.8) | 88.5 (0.6) | 91.1 (0.3) | 87.3 (0.1) | 93.2 (0.2) | 78.6 (1.4) | 85.1 |
| LoRA | 778K | 59.7 (1.4) | 93.6 (0.1) | 88.9 (0.7) | 88.3 (0.4) | 90.9 (0.2) | 87.9 (0.2) | 93.2 (0.1) | 78.8 (1.3) | 85.2 |
| UNIPELT | 1.4M | 61.5 (1.7) | 93.6 (0.2) | 89.2 (0.8) | 88.7 (0.4) | 91.0 (0.5) | 87.5 (0.2) | 92.9 (0.1) | 79.1 (0.9) | 85.4 |
| Adapter | 1.6M | 63.1 (1.4) | 93.8 (0.1) | 87.5 (0.8) | 88.7 (0.4) | 90.9 (0.4) | 88.5 (0.2) | 93.2 (0.1) | 78.9 (1.3) | 85.6 |
| LST | 1.7M | 63.6 (1.5) | 93.9 (0.2) | 89.2 (0.7) | 88.7 (0.2) | 91.2 (0.4) | 88.3 (0.1) | 93.0 (0.2) | 79.2 (0.9) | 85.9 |
| Parallel Adapters | 279K | 63.8 (1.5) | 94.2 (0.3) | 89.3 (0.5) | 88.9 (0.6) | 91.1 (0.4) | 88.3 (0.1) | 93.2 (0.1) | 79.1 (0.5) | 86.0 |
| Compacter | 279K | 63.7 (0.9) | 94.2 (0.4) | 89.1 (0.7) | 88.6 (0.3) | 90.8 (0.2) | 88.2 (0.2) | 92.8 (0.2) | 78.7 (0.9) | 85.8 |
| Our proposed methods | ||||||||||
| Learned Adapter | 294K | 64.3 (0.9) | 94.9 (0.3) | 89.8 (0.9) | 89.2 (0.3) | 91.3 (0.2) | 88.6 (0.2) | 93.5 (0.2) | 80.4 (1.3) | 86.5 |
| Source | Target | |
| RTE | SST-2 | |
| RTE | 80.4 (1.3) | 94.6 (0.1) |
| SST-2 | 80.3 (1.4) | 94.9 (0.3) |
| Search space | RTE (acc) | SST-2 (acc) |
| S | 80.4 (1.3) | 94.9 (0.3) |
| S1 | 80.1 (1.5) | 94.6 (0.3) |
| S2 | 79.8 (1.0) | 94.4 (0.4) |
| Parallel Adapter | 79.1 (0.5) | 94.2 (0.3) |
| Method | Tunable +Params | RTE +(acc) | SST-2 +(acc) |
| DeBERTa-large backbone | |||
| Fine-tuning | 406M | 82.1 (1.1) | 95.7 (0.2) |
| Adapters | 1.6M | 80.7 (1.3) | 93.8 (0.1) |
| Parallel Adapters | 279K | 81.5 (1.2) | 93.9 (0.3) |
| Learned Adapter | 245K | 82.0 (0.9) | 94.5 (0.3) |
| GPT2-large backbone | |||
| Fine-tuning | 774M | 79.5 (0.8) | 95.5 (0.1) |
| Adapters | 2.7M | 78.1 (0.9) | 93.9 (0.1) |
| Parallel Adapters | 462K | 78.5 (1.2) | 93.8 (0.1) |
| Learned Adapter | 482K | 79.1 (0.6) | 94.2 (0.2) |
| Search method | RTE (acc) | SST-2 (acc) |
| GDNAS | 80.4 (1.3) | 94.9 (0.3) |
| GDNAS (single-level) | 80.2 (0.9) | 94.7 (0.2) |
| DARTS | 79.8 (1.2) | 94.5 (0.5) |
| ENAS | 80.3 (1.0) | 94.3 (0.4) |
| Category | Datasets | ltrainl | ldevl | ltestl | |y| | Type | Labels |
| Single-sentence | SST-2 | 66349 | 1000 | 872 | 2 | sentiment | positive, negative |
| CoLA | 8551 | 521 | 522 | 2 | linguistic acceptability | acceptable, not acceptable | |
| Sentence-pair | MNLI | 391702 | 1000 | 19647 | 3 | NLI | entailment, neutral, contradiction |
| MRPC | 2668 | 1000 | 408 | 2 | paraphrase | equivalent, not equivalent | |
| QNLI | 103743 | 1000 | 5463 | 2 | NLI | entailment, not entailment | |
| QQP | 362846 | 1000 | 40430 | 2 | paraphrase | equivalent, not equivalent | |
| RTE | 1490 | 1000 | 277 | 2 | NLI | entailment, not entailment | |
| STS-B | 5750 | 750 | 751 | - | semantic similarity | - |
| Architecture | Test Error (%) | Params (M) | Search Cost (GPU days) | Search Framework |
| NASNet-A (Zoph et al., 2017) | 2.65 | 3.3 | 2000 | RL |
| AmoebaNet (Real et al., 2018) | 3.34 | 3.2 | 3150 | evolution |
| ENAS (Pham et al., 2018) | 2.89 | 4.6 | 0.5 | RL |
| DARTS (Liu et al., 2019a) | 3.02 | 3.3 | 0.4 | DNAS |
| SNAS (Xie et al., 2018) | 2.85 | 2.8 | 1.5 | DNAS |
| BayesNAS (Zhou et al., 2019) | 2.81 | 3.4 | 0.2 | DNAS |
| P-DARTS (Chen et al., 2021) | 2.50 | 3.4 | 0.3 | DNAS |
| R-DARTS (Zela et al., 2019) | 2.95 | - | 1.6 | DNAS |
| GDNAS (ours) | 2.52 | 3.3 | 0.6 | DNAS |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | FFN | elu | null_ACT | mha_8 |
| 2 | - | - | - | - |
| 3 | - | - | - | - |
| 4 | FFN | elu | null_ACT | mha_2 |
| 5 | Block | tanh | null_ACT | mha_2 |
| 6 | FFN | gelu_new | leaky_relu | conv_3 |
| 7 | FFN | null_ACT | tanh | mha_2 |
| 8 | - | - | - | - |
| 9 | Attn | elu | relu | conv_3 |
| 10 | - | - | - | - |
| 11 | Attn | gelu_new | relu | conv_1 |
| 12 | Attn | gelu_new | relu | conv_3 |
| 13 | Block | relu | leaky_relu | conv_1 |
| 14 | Attn | swish | relu | conv_3 |
| 15 | FFN | leaky_relu | relu | conv_3 |
| 16 | Block | leaky_relu | relu | conv_5 |
| 17 | FFN | leaky_relu | relu | conv_1 |
| 18 | Attn | leaky_relu | null_ACT | skip_connect |
| 19 | FFN | relu | relu | conv_3 |
| 20 | FFN | gelu_new | null_ACT | conv_3 |
| 21 | - | - | - | - |
| 22 | Block | tanh | null_ACT | mha_8 |
| 23 | Attn | tanh | null_ACT | conv_3 |
| 24 | FFN | tanh | null_ACT | mha_2 |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | Attn | null_ACT | null_ACT | conv_5 |
| 2 | - | - | - | - |
| 3 | FFN | elu | null_ACT | mha_2 |
| 4 | Attn | tanh | null_ACT | mha_2 |
| 5 | Attn | elu | null_ACT | skip_connect |
| 6 | FFN | gelu_new | relu | conv_5 |
| 7 | Block | leaky_relu | leaky_relu | conv_3 |
| 8 | - | - | - | - |
| 9 | Block | leaky_relu | leaky_relu | conv_5 |
| 10 | FFN | relu | relu | conv_3 |
| 11 | Block | tanh | null_ACT | mha_8 |
| 12 | - | - | - | - |
| 13 | Block | tanh | null_ACT | skip_connect |
| 14 | - | - | - | - |
| 15 | FFN | swish | swish | conv_1 |
| 16 | Block | relu | relu | conv_1 |
| 17 | FFN | relu | swish | conv_3 |
| 18 | FFN | leaky_relu | relu | conv_3 |
| 19 | Attn | gelu_new | relu | mha_8 |
| 20 | FFN | swish | relu | mha_8 |
| 21 | Attn | swish | null_ACT | conv_3 |
| 22 | Attn | elu | null_ACT | conv_3 |
| 23 | Attn | null_ACT | null_ACT | mha_8 |
| 24 | Attn | elu | null_ACT | mha_2 |
| Model | RTE | MRPC |
| M | 79.1 (0.5) | 89.3 (0.4) |
| Mgelu | 79.3 (0.2) | 89.5 (0.3) |
| Mswish | 79.6 (0.3) | 89.2 (0.4) |
| Mtanh | 79.2 (0.5) | 88.9 (0.6) |
| Msa | 79.5 (0.3) | 89.6 (0.2) |
| Mconv | 79.5 (0.6) | 89.4 (0.5) |
| M attenu | 79.0 (0.4) | 89.1 (0.3) |
| Mblock | 79.6 (0.5) | 89.4 (0.3) |
| Method | Tunable \nParams | CoLA \n(mcc) | SST-2 \n(acc) | MRPC \n(acc-f1) | QQP \n(acc-f1) | STS-B \n(corr) | MNLI \n(acc) | QNLI \n(acc) | RTE \n(acc) | Avg |
| Baselines | ||||||||||
| RoBERTa-large | 355M | 65.3 | 95.4 | 91.5 | 90.3 | 91.8 | 89.8 | 94.5 | 80.1 | 87.3 |
| Prompt Tuning | 21K | 54.6 (2.5) | 90.9 (0.5) | 74.8 (1.3) | 87.4 (0.5) | 90.1 (0.3) | 83.4 (0.2) | 92.4 (0.3) | 68.7 (1.9) | 80.3 |
| P-tuning v2 | 985K | 57.3 (2.1) | 92.3 (0.3) | 86.9 (1.2) | 88.1 (0.4) | 90.4 (0.1) | 87.3 (0.2) | 92.9 (0.4) | 77.2 (2.1) | 84.1 |
| BitFit | 273K | 59.2 (0.9) | 94.1 (0.3) | 88.6 (0.8) | 88.5 (0.6) | 91.1 (0.3) | 87.3 (0.1) | 93.2 (0.2) | 78.6 (1.4) | 85.1 |
| LoRA | 778K | 59.7 (1.4) | 93.6 (0.1) | 88.9 (0.7) | 88.3 (0.4) | 90.9 (0.2) | 87.9 (0.2) | 93.2 (0.1) | 78.8 (1.3) | 85.2 |
| UNIPELT | 1.4M | 61.5 (1.7) | 93.6 (0.2) | 89.2 (0.8) | 88.7 (0.4) | 91.0 (0.5) | 87.5 (0.2) | 92.9 (0.1) | 79.1 (0.9) | 85.4 |
| Adapter | 1.6M | 63.1 (1.4) | 93.8 (0.1) | 87.5 (0.8) | 88.7 (0.4) | 90.9 (0.4) | 88.5 (0.2) | 93.2 (0.1) | 78.9 (1.3) | 85.6 |
| LST | 1.7M | 63.6 (1.5) | 93.9 (0.2) | 89.2 (0.7) | 88.7 (0.2) | 91.2 (0.4) | 88.3 (0.1) | 93.0 (0.2) | 79.2 (0.9) | 85.9 |
| Parallel Adapters | 279K | 63.8 (1.5) | 94.2 (0.3) | 89.3 (0.5) | 88.9 (0.6) | 91.1 (0.4) | 88.3 (0.1) | 93.2 (0.1) | 79.1 (0.5) | 86.0 |
| Compacter | 279K | 63.7 (0.9) | 94.2 (0.4) | 89.1 (0.7) | 88.6 (0.3) | 90.8 (0.2) | 88.2 (0.2) | 92.8 (0.2) | 78.7 (0.9) | 85.8 |
| Our proposed methods | ||||||||||
| Learned Adapter | 294K | 64.3 (0.9) | 94.9 (0.3) | 89.8 (0.9) | 89.2 (0.3) | 91.3 (0.2) | 88.6 (0.2) | 93.5 (0.2) | 80.4 (1.3) | 86.5 |
| Source | Target | |
| RTE | SST-2 | |
| RTE | 80.4 (1.3) | 94.6 (0.1) |
| SST-2 | 80.3 (1.4) | 94.9 (0.3) |
| Search space | RTE (acc) | SST-2 (acc) |
| S | 80.4 (1.3) | 94.9 (0.3) |
| S1 | 80.1 (1.5) | 94.6 (0.3) |
| S2 | 79.8 (1.0) | 94.4 (0.4) |
| Parallel Adapter | 79.1 (0.5) | 94.2 (0.3) |
| Method | Tunable \nParams | RTE \n(acc) | SST-2 \n(acc) |
| DeBERTa-large backbone | |||
| Fine-tuning | 406M | 82.1 (1.1) | 95.7 (0.2) |
| Adapters | 1.6M | 80.7 (1.3) | 93.8 (0.1) |
| Parallel Adapters | 279K | 81.5 (1.2) | 93.9 (0.3) |
| Learned Adapter | 245K | 82.0 (0.9) | 94.5 (0.3) |
| GPT2-large backbone | |||
| Fine-tuning | 774M | 79.5 (0.8) | 95.5 (0.1) |
| Adapters | 2.7M | 78.1 (0.9) | 93.9 (0.1) |
| Parallel Adapters | 462K | 78.5 (1.2) | 93.8 (0.1) |
| Learned Adapter | 482K | 79.1 (0.6) | 94.2 (0.2) |
| Search method | RTE (acc) | SST-2 (acc) |
| GDNAS | 80.4 (1.3) | 94.9 (0.3) |
| GDNAS (single-level) | 80.2 (0.9) | 94.7 (0.2) |
| DARTS | 79.8 (1.2) | 94.5 (0.5) |
| ENAS | 80.3 (1.0) | 94.3 (0.4) |
| Category | Datasets | ltrainl | ldevl | ltestl | |y| | Type | Labels |
| Single-sentence | SST-2 | 66349 | 1000 | 872 | 2 | sentiment | positive, negative |
| CoLA | 8551 | 521 | 522 | 2 | linguistic acceptability | acceptable, not acceptable | |
| Sentence-pair | MNLI | 391702 | 1000 | 19647 | 3 | NLI | entailment, neutral, contradiction |
| MRPC | 2668 | 1000 | 408 | 2 | paraphrase | equivalent, not equivalent | |
| QNLI | 103743 | 1000 | 5463 | 2 | NLI | entailment, not entailment | |
| QQP | 362846 | 1000 | 40430 | 2 | paraphrase | equivalent, not equivalent | |
| RTE | 1490 | 1000 | 277 | 2 | NLI | entailment, not entailment | |
| STS-B | 5750 | 750 | 751 | - | semantic similarity | - |
| Architecture | Test Error (%) | Params (M) | Search Cost (GPU days) | Search Framework |
| NASNet-A (Zoph et al., 2017) | 2.65 | 3.3 | 2000 | RL |
| AmoebaNet (Real et al., 2018) | 3.34 | 3.2 | 3150 | evolution |
| ENAS (Pham et al., 2018) | 2.89 | 4.6 | 0.5 | RL |
| DARTS (Liu et al., 2019a) | 3.02 | 3.3 | 0.4 | DNAS |
| SNAS (Xie et al., 2018) | 2.85 | 2.8 | 1.5 | DNAS |
| BayesNAS (Zhou et al., 2019) | 2.81 | 3.4 | 0.2 | DNAS |
| P-DARTS (Chen et al., 2021) | 2.50 | 3.4 | 0.3 | DNAS |
| R-DARTS (Zela et al., 2019) | 2.95 | - | 1.6 | DNAS |
| GDNAS (ours) | 2.52 | 3.3 | 0.6 | DNAS |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | FFN | elu | null_ACT | mha_8 |
| 2 | - | - | - | - |
| 3 | - | - | - | - |
| 4 | FFN | elu | null_ACT | mha_2 |
| 5 | Block | tanh | null_ACT | mha_2 |
| 6 | FFN | gelu_new | leaky_relu | conv_3 |
| 7 | FFN | null_ACT | tanh | mha_2 |
| 8 | - | - | - | - |
| 9 | Attn | elu | relu | conv_3 |
| 10 | - | - | - | - |
| 11 | Attn | gelu_new | relu | conv_1 |
| 12 | Attn | gelu_new | relu | conv_3 |
| 13 | Block | relu | leaky_relu | conv_1 |
| 14 | Attn | swish | relu | conv_3 |
| 15 | FFN | leaky_relu | relu | conv_3 |
| 16 | Block | leaky_relu | relu | conv_5 |
| 17 | FFN | leaky_relu | relu | conv_1 |
| 18 | Attn | leaky_relu | null_ACT | skip_connect |
| 19 | FFN | relu | relu | conv_3 |
| 20 | FFN | gelu_new | null_ACT | conv_3 |
| 21 | - | - | - | - |
| 22 | Block | tanh | null_ACT | mha_8 |
| 23 | Attn | tanh | null_ACT | conv_3 |
| 24 | FFN | tanh | null_ACT | mha_2 |
| Layer index | Adapter placement | Activation g1 | Activation g2 | Encoder operation 1 |
| 1 | Attn | null_ACT | null_ACT | conv_5 |
| 2 | - | - | - | - |
| 3 | FFN | elu | null_ACT | mha_2 |
| 4 | Attn | tanh | null_ACT | mha_2 |
| 5 | Attn | elu | null_ACT | skip_connect |
| 6 | FFN | gelu_new | relu | conv_5 |
| 7 | Block | leaky_relu | leaky_relu | conv_3 |
| 8 | - | - | - | - |
| 9 | Block | leaky_relu | leaky_relu | conv_5 |
| 10 | FFN | relu | relu | conv_3 |
| 11 | Block | tanh | null_ACT | mha_8 |
| 12 | - | - | - | - |
| 13 | Block | tanh | null_ACT | skip_connect |
| 14 | - | - | - | - |
| 15 | FFN | swish | swish | conv_1 |
| 16 | Block | relu | relu | conv_1 |
| 17 | FFN | relu | swish | conv_3 |
| 18 | FFN | leaky_relu | relu | conv_3 |
| 19 | Attn | gelu_new | relu | mha_8 |
| 20 | FFN | swish | relu | mha_8 |
| 21 | Attn | swish | null_ACT | conv_3 |
| 22 | Attn | elu | null_ACT | conv_3 |
| 23 | Attn | null_ACT | null_ACT | mha_8 |
| 24 | Attn | elu | null_ACT | mha_2 |
| ... attend his daughter 's wedding(event1) ceremony ... \n... to be taken(event2) home later in the afternoon \nto marry(event1) his eldest daughter ... | |||||
| Event | Type | Polarity | Modality | Genericity | Tense |
| event1 | Life:Marry | Positive | Asserted | Specific | Future |
| event2 | Movement:Transport | Positive | Other | Specific | Future |
| Train | Dev | Test | |
| #Docs | 529 | 30 | 40 |
| #Sent | 19204 | 901 | 676 |
| #Event | 3342 | 327 | 293 |
| Parameters | Value |
| Epochs | 150 |
| Batch size | 16 |
| Task learning rate | 0.0001 |
| Transformer learning rate | 5e-5 |
| Hidden size | 500 |
| Event Alone Features (k) | Noisy Probability (pk) |
| Type | 0.00 |
| Polarity | 0.00 |
| Modality | 0.15 |
| Genericity | 0.15 |
| Tense | 0.25 |
| Models | Event Features | |||||
| Type | Polarity | Modality | Genericity | Tense | ||
| CDGM(2021) | train | 99.9 | 99.9 | 99.9 | 99.9 | 98.4 |
| test | 95.3 | 98.8 | 88.4 | 87.2 | 76.3 | |
| Ours | train | 99.9 | 99.9 | 99.4 | 99.9 | 99.7 |
| test | 95.3 | 98.6 | 89.2 | 90.3 | 77.6 |
| Models | CoNLL | AVG | |
| cross-validation | |||
| SSED(2016) | +Supervised | 55.23 | 52.50 |
| +MSEP | 53.80 | 51.40 | |
| test data | |||
| CDGM(2021) | +Simple | 57.55 | 54.79 |
| (all features) | +CDGM | 58.99 | 56.32 |
| +Simple +Noise | 60.43 | 57.85 | |
| +CDGM +Noise | 62.07 | 59.76 | |
| Ours | +Noise | 64.56 | 62.11 |
| (all features) |
| Models | CoNLL | AVG | |
| PAIREDRL(2020b) | 84.65 | - | |
| MSEP(2016) | 80.37 | 82.90 | |
| CDGM(2021) | +Simple | 75.32 | 74.94 |
| (all features) | +CDGM +Noise | 84.76 | 83.95 |
| Ours | +Noise | 86.78 | 86.63 |
| (all features) |
| Models | CoNLL | AVG | |
| CDGM(2021) | +Simple | 85.75 | 85.40 |
| (all features) | +CDGM | 87.90 | 88.30 |
| +CDGM +Noise | 85.40 | 85.38 | |
| Ours | +Noise | 87.10 | 86.83 |
| (all features) | - | 91.97 | 92.32 |
| Settings | AVG | Δ | |
| CDGM | Ours | ||
| End-to-end | 61.38 | 62.11 | +0.73 |
| Gold triggers | 85.72 | 86.63 | +0.91 |
| All gold | 92.05 | 92.32 | +0.27 |
| Models | AVG | |
| baseline | ||
| CDGM(2021) | + Average + Simple | 85.40 |
| Ours | + Attention + Simple | 90.45 |
| Models | AVG | Δ |
| -wo fusion layer | ||
| Baselineelement-wise similarity) | 90.50 | - |
| -wo fusion layer | ||
| Element-wise similarity | 90.92 | +0.42 |
| Cosine Similarity | 89.20 | -1.30 |
| Linear Similarity | 90.72 | +0.22 |
| Element-wise + Cosine Similarity | 91.07 | +0.57 |
| Element-wise + Linear Similarity | 91.73 | +1.23 |
| Multi-similarity | 92.32 | +1.82 |
| Settings | AVG | Δ | |
| Clean | Noisy | ||
| End-to-end | 57.00 | 62.11 | +5.11 |
| Gold triggers | 78.87 | 86.63 | +7.76 |
| All gold | 92.32 | 86.83 | -5.49 |
| Models | #Training Parameters |
| Baseline | 110.51M |
| CDGM(2021) | 108.57M |
| Ours | 110.52M |
| Settings | Models | MUC | B3 | CEAFe | BLANC | CoNLL | AVG |
| End-to-end | SSED + Supervised(2016) | 47.10 | 59.90 | 58.70 | 44.40 | 55.23 | 52.53 |
| CDGM(2021) | - | - | - | - | 62.07 | 59.76 | |
| Ours | 59.83 | 68.17 | 65.67 | 54.76 | 64.56 | 62.11 | |
| Gold Triggers | MSEP(2016) | 68.00 | 92.90 | 87.40 | 83.20 | 82.77 | 82.88 |
| PAIREDRL(2020b) | 76.10 | 90.70 | 87.20 | - | 84.65 | - | |
| CDGM(2021) | - | - | - | - | 84.76 | 83.95 | |
| Ours | 79.18 | 92.37 | 88.79 | 86.18 | 86.78 | 86.63 | |
| All Gold | CDGM(2021) | - | - | - | - | 85.40 | 85.38 |
| Ours | 87.44 | 95.48 | 92.99 | 93.38 | 91.97 | 92.32 |
| ... attend his daughter 's wedding(event1) ceremony ... \n... to be taken(event2) home later in the afternoon \nto marry(event1) his eldest daughter ... | |||||
| Event | Type | Polarity | Modality | Genericity | Tense |
| event1 | Life:Marry | Positive | Asserted | Specific | Future |
| event2 | Movement:Transport | Positive | Other | Specific | Future |
| Train | Dev | Test | |
| #Docs | 529 | 30 | 40 |
| #Sent | 19204 | 901 | 676 |
| #Event | 3342 | 327 | 293 |
| Parameters | Value |
| Epochs | 150 |
| Batch size | 16 |
| Task learning rate | 0.0001 |
| Transformer learning rate | 5e-5 |
| Hidden size | 500 |
| Event Alone Features (k) | Noisy Probability (pk) |
| Type | 0.00 |
| Polarity | 0.00 |
| Modality | 0.15 |
| Genericity | 0.15 |
| Tense | 0.25 |
| Models | Event Features | |||||
| Type | Polarity | Modality | Genericity | Tense | ||
| CDGM(2021) | train | 99.9 | 99.9 | 99.9 | 99.9 | 98.4 |
| test | 95.3 | 98.8 | 88.4 | 87.2 | 76.3 | |
| Ours | train | 99.9 | 99.9 | 99.4 | 99.9 | 99.7 |
| test | 95.3 | 98.6 | 89.2 | 90.3 | 77.6 |
| Models | CoNLL | AVG | |
| cross-validation | |||
| SSED(2016) | +Supervised | 55.23 | 52.50 |
| +MSEP | 53.80 | 51.40 | |
| test data | |||
| CDGM(2021) | +Simple | 57.55 | 54.79 |
| (all features) | +CDGM | 58.99 | 56.32 |
| +Simple +Noise | 60.43 | 57.85 | |
| +CDGM +Noise | 62.07 | 59.76 | |
| Ours | +Noise | 64.56 | 62.11 |
| (all features) |
| Models | CoNLL | AVG | |
| PAIREDRL(2020b) | 84.65 | - | |
| MSEP(2016) | 80.37 | 82.90 | |
| CDGM(2021) | +Simple | 75.32 | 74.94 |
| (all features) | +CDGM +Noise | 84.76 | 83.95 |
| Ours | +Noise | 86.78 | 86.63 |
| (all features) |
| Models | CoNLL | AVG | |
| CDGM(2021) | +Simple | 85.75 | 85.40 |
| (all features) | +CDGM | 87.90 | 88.30 |
| +CDGM +Noise | 85.40 | 85.38 | |
| Ours | +Noise | 87.10 | 86.83 |
| (all features) | - | 91.97 | 92.32 |
| Settings | AVG | Δ | |
| CDGM | Ours | ||
| End-to-end | 61.38 | 62.11 | +0.73 |
| Gold triggers | 85.72 | 86.63 | +0.91 |
| All gold | 92.05 | 92.32 | +0.27 |
| Models | AVG | |
| baseline | ||
| CDGM(2021) | + Average + Simple | 85.40 |
| Ours | + Attention + Simple | 90.45 |
| Models | AVG | Δ |
| -wo fusion layer | ||
| Baselineelement-wise similarity) | 90.50 | - |
| -wo fusion layer | ||
| Element-wise similarity | 90.92 | +0.42 |
| Cosine Similarity | 89.20 | -1.30 |
| Linear Similarity | 90.72 | +0.22 |
| Element-wise + Cosine Similarity | 91.07 | +0.57 |
| Element-wise + Linear Similarity | 91.73 | +1.23 |
| Multi-similarity | 92.32 | +1.82 |
| Settings | AVG | Δ | |
| Clean | Noisy | ||
| End-to-end | 57.00 | 62.11 | +5.11 |
| Gold triggers | 78.87 | 86.63 | +7.76 |
| All gold | 92.32 | 86.83 | -5.49 |
| Models | #Training Parameters |
| Baseline | 110.51M |
| CDGM(2021) | 108.57M |
| Ours | 110.52M |
| Settings | Models | MUC | B3 | CEAFe | BLANC | CoNLL | AVG |
| End-to-end | SSED + Supervised(2016) | 47.10 | 59.90 | 58.70 | 44.40 | 55.23 | 52.53 |
| CDGM(2021) | - | - | - | - | 62.07 | 59.76 | |
| Ours | 59.83 | 68.17 | 65.67 | 54.76 | 64.56 | 62.11 | |
| Gold Triggers | MSEP(2016) | 68.00 | 92.90 | 87.40 | 83.20 | 82.77 | 82.88 |
| PAIREDRL(2020b) | 76.10 | 90.70 | 87.20 | - | 84.65 | - | |
| CDGM(2021) | - | - | - | - | 84.76 | 83.95 | |
| Ours | 79.18 | 92.37 | 88.79 | 86.18 | 86.78 | 86.63 | |
| All Gold | CDGM(2021) | - | - | - | - | 85.40 | 85.38 |
| Ours | 87.44 | 95.48 | 92.99 | 93.38 | 91.97 | 92.32 |
| ... attend his daughter 's wedding(event1) ceremony ... +... to be taken(event2) home later in the afternoon +to marry(event1) his eldest daughter ... | |||||
| Event | Type | Polarity | Modality | Genericity | Tense |
| event1 | Life:Marry | Positive | Asserted | Specific | Future |
| event2 | Movement:Transport | Positive | Other | Specific | Future |
| Train | Dev | Test | |
| #Docs | 529 | 30 | 40 |
| #Sent | 19204 | 901 | 676 |
| #Event | 3342 | 327 | 293 |
| Parameters | Value |
| Epochs | 150 |
| Batch size | 16 |
| Task learning rate | 0.0001 |
| Transformer learning rate | 5e-5 |
| Hidden size | 500 |
| Event Alone Features (k) | Noisy Probability (pk) |
| Type | 0.00 |
| Polarity | 0.00 |
| Modality | 0.15 |
| Genericity | 0.15 |
| Tense | 0.25 |
| Models | Event Features | |||||
| Type | Polarity | Modality | Genericity | Tense | ||
| CDGM(2021) | train | 99.9 | 99.9 | 99.9 | 99.9 | 98.4 |
| test | 95.3 | 98.8 | 88.4 | 87.2 | 76.3 | |
| Ours | train | 99.9 | 99.9 | 99.4 | 99.9 | 99.7 |
| test | 95.3 | 98.6 | 89.2 | 90.3 | 77.6 | |
| Models | CoNLL | AVG | |
| cross-validation | |||
| SSED(2016) | +Supervised | 55.23 | 52.50 |
| +MSEP | 53.80 | 51.40 | |
| test data | |||
| CDGM(2021) | +Simple | 57.55 | 54.79 |
| (all features) | +CDGM | 58.99 | 56.32 |
| +Simple +Noise | 60.43 | 57.85 | |
| +CDGM +Noise | 62.07 | 59.76 | |
| Ours | +Noise | 64.56 | 62.11 |
| (all features) | |||
| Models | CoNLL | AVG | |
| PAIREDRL(2020b) | 84.65 | - | |
| MSEP(2016) | 80.37 | 82.90 | |
| CDGM(2021) | +Simple | 75.32 | 74.94 |
| (all features) | +CDGM +Noise | 84.76 | 83.95 |
| Ours | +Noise | 86.78 | 86.63 |
| (all features) | |||
| Models | CoNLL | AVG | |
| CDGM(2021) | +Simple | 85.75 | 85.40 |
| (all features) | +CDGM | 87.90 | 88.30 |
| +CDGM +Noise | 85.40 | 85.38 | |
| Ours | +Noise | 87.10 | 86.83 |
| (all features) | - | 91.97 | 92.32 |
| Settings | AVG | Δ | |
| CDGM | Ours | ||
| End-to-end | 61.38 | 62.11 | +0.73 |
| Gold triggers | 85.72 | 86.63 | +0.91 |
| All gold | 92.05 | 92.32 | +0.27 |
| Models | AVG | |
| baseline | ||
| CDGM(2021) | + Average + Simple | 85.40 |
| Ours | + Attention + Simple | 90.45 |
| Models | AVG | Δ |
| -wo fusion layer | ||
| Baselineelement-wise similarity) | 90.50 | - |
| -wo fusion layer | ||
| Element-wise similarity | 90.92 | +0.42 |
| Cosine Similarity | 89.20 | -1.30 |
| Linear Similarity | 90.72 | +0.22 |
| Element-wise + Cosine Similarity | 91.07 | +0.57 |
| Element-wise + Linear Similarity | 91.73 | +1.23 |
| Multi-similarity | 92.32 | +1.82 |
| Settings | AVG | Δ | |
| Clean | Noisy | ||
| End-to-end | 57.00 | 62.11 | +5.11 |
| Gold triggers | 78.87 | 86.63 | +7.76 |
| All gold | 92.32 | 86.83 | -5.49 |
| Models | #Training Parameters |
| Baseline | 110.51M |
| CDGM(2021) | 108.57M |
| Ours | 110.52M |
| Settings | Models | MUC | B3 | CEAFe | BLANC | CoNLL | AVG |
| End-to-end | SSED + Supervised(2016) | 47.10 | 59.90 | 58.70 | 44.40 | 55.23 | 52.53 |
| CDGM(2021) | - | - | - | - | 62.07 | 59.76 | |
| Ours | 59.83 | 68.17 | 65.67 | 54.76 | 64.56 | 62.11 | |
| Gold Triggers | MSEP(2016) | 68.00 | 92.90 | 87.40 | 83.20 | 82.77 | 82.88 |
| PAIREDRL(2020b) | 76.10 | 90.70 | 87.20 | - | 84.65 | - | |
| CDGM(2021) | - | - | - | - | 84.76 | 83.95 | |
| Ours | 79.18 | 92.37 | 88.79 | 86.18 | 86.78 | 86.63 | |
| All Gold | CDGM(2021) | - | - | - | - | 85.40 | 85.38 |
| Ours | 87.44 | 95.48 | 92.99 | 93.38 | 91.97 | 92.32 |
| ... attend his daughter 's wedding(event1) ceremony ... \n... to be taken(event2) home later in the afternoon \nto marry(event1) his eldest daughter ... | |||||
| Event | Type | Polarity | Modality | Genericity | Tense |
| event1 | Life:Marry | Positive | Asserted | Specific | Future |
| event2 | Movement:Transport | Positive | Other | Specific | Future |
| Train | Dev | Test | |
| #Docs | 529 | 30 | 40 |
| #Sent | 19204 | 901 | 676 |
| #Event | 3342 | 327 | 293 |
| Parameters | Value |
| Epochs | 150 |
| Batch size | 16 |
| Task learning rate | 0.0001 |
| Transformer learning rate | 5e-5 |
| Hidden size | 500 |
| Event Alone Features (k) | Noisy Probability (pk) |
| Type | 0.00 |
| Polarity | 0.00 |
| Modality | 0.15 |
| Genericity | 0.15 |
| Tense | 0.25 |
| Models | Event Features | |||||
| Type | Polarity | Modality | Genericity | Tense | ||
| CDGM(2021) | train | 99.9 | 99.9 | 99.9 | 99.9 | 98.4 |
| test | 95.3 | 98.8 | 88.4 | 87.2 | 76.3 | |
| Ours | train | 99.9 | 99.9 | 99.4 | 99.9 | 99.7 |
| test | 95.3 | 98.6 | 89.2 | 90.3 | 77.6 |
| Models | CoNLL | AVG | |
| cross-validation | |||
| SSED(2016) | +Supervised | 55.23 | 52.50 |
| +MSEP | 53.80 | 51.40 | |
| test data | |||
| CDGM(2021) | +Simple | 57.55 | 54.79 |
| (all features) | +CDGM | 58.99 | 56.32 |
| +Simple +Noise | 60.43 | 57.85 | |
| +CDGM +Noise | 62.07 | 59.76 | |
| Ours | +Noise | 64.56 | 62.11 |
| (all features) |
| Models | CoNLL | AVG | |
| PAIREDRL(2020b) | 84.65 | - | |
| MSEP(2016) | 80.37 | 82.90 | |
| CDGM(2021) | +Simple | 75.32 | 74.94 |
| (all features) | +CDGM +Noise | 84.76 | 83.95 |
| Ours | +Noise | 86.78 | 86.63 |
| (all features) |
| Models | CoNLL | AVG | |
| CDGM(2021) | +Simple | 85.75 | 85.40 |
| (all features) | +CDGM | 87.90 | 88.30 |
| +CDGM +Noise | 85.40 | 85.38 | |
| Ours | +Noise | 87.10 | 86.83 |
| (all features) | - | 91.97 | 92.32 |
| Settings | AVG | Δ | |
| CDGM | Ours | ||
| End-to-end | 61.38 | 62.11 | +0.73 |
| Gold triggers | 85.72 | 86.63 | +0.91 |
| All gold | 92.05 | 92.32 | +0.27 |
| Models | AVG | |
| baseline | ||
| CDGM(2021) | + Average + Simple | 85.40 |
| Ours | + Attention + Simple | 90.45 |
| Models | AVG | Δ |
| -wo fusion layer | ||
| Baselineelement-wise similarity) | 90.50 | - |
| -wo fusion layer | ||
| Element-wise similarity | 90.92 | +0.42 |
| Cosine Similarity | 89.20 | -1.30 |
| Linear Similarity | 90.72 | +0.22 |
| Element-wise + Cosine Similarity | 91.07 | +0.57 |
| Element-wise + Linear Similarity | 91.73 | +1.23 |
| Multi-similarity | 92.32 | +1.82 |
| Settings | AVG | Δ | |
| Clean | Noisy | ||
| End-to-end | 57.00 | 62.11 | +5.11 |
| Gold triggers | 78.87 | 86.63 | +7.76 |
| All gold | 92.32 | 86.83 | -5.49 |
| Models | #Training Parameters |
| Baseline | 110.51M |
| CDGM(2021) | 108.57M |
| Ours | 110.52M |
| Settings | Models | MUC | B3 | CEAFe | BLANC | CoNLL | AVG |
| End-to-end | SSED + Supervised(2016) | 47.10 | 59.90 | 58.70 | 44.40 | 55.23 | 52.53 |
| CDGM(2021) | - | - | - | - | 62.07 | 59.76 | |
| Ours | 59.83 | 68.17 | 65.67 | 54.76 | 64.56 | 62.11 | |
| Gold Triggers | MSEP(2016) | 68.00 | 92.90 | 87.40 | 83.20 | 82.77 | 82.88 |
| PAIREDRL(2020b) | 76.10 | 90.70 | 87.20 | - | 84.65 | - | |
| CDGM(2021) | - | - | - | - | 84.76 | 83.95 | |
| Ours | 79.18 | 92.37 | 88.79 | 86.18 | 86.78 | 86.63 | |
| All Gold | CDGM(2021) | - | - | - | - | 85.40 | 85.38 |
| Ours | 87.44 | 95.48 | 92.99 | 93.38 | 91.97 | 92.32 |
| Datasets | |E| | |R| | Ntrain | Nvalid | Ntest | |T| |
| ICEWS14 | 7,128 | 230 | 72,826 | 8,941 | 8,963 | 365 |
| ICEWS05-15 | 10,488 | 251 | 368,962 | 46,275 | 46,092 | 4,017 |
| GDELT | 500 | 20 | 2,735,685 | 341,961 | 341,961 | 366 |
| Method | ICEWS14 | ICEWS05-15 | GDELT | |||||||||
| MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | |
| TransE (2013) | .280 | .094 | - | .637 | .294 | .090 | - | .663 | - | - | - | - |
| DisMult (2015) | .441 | .325 | .498 | .668 | .457 | .338 | .515 | .691 | .210 | .133 | .224 | .365 |
| ComplEx (2016) | .442 | .440 | .430 | .664 | .464 | .347 | .524 | .696 | .213 | .133 | .225 | .366 |
| Simple (2018) | .458 | .341 | .516 | .687 | .478 | .359 | .539 | .708 | .206 | .124 | .220 | .366 |
| TTransE (2016) | .255 | .074 | - | .601 | .271 | .084 | - | .616 | .115 | .000 | .160 | .318 |
| TeRo (2020) | .562 | .468 | .621 | .732 | .586 | .469 | .668 | .795 | .245 | .154 | .264 | .420 |
| ChronoR (2021) | .625 | .547 | .669 | .773 | .675 | .596 | .723 | .820 | - | - | - | - |
| RotateQVS (2022) | .591 | .507 | .642 | .754 | .633 | .529 | .709 | .813 | .270 | .175 | .293 | .458 |
| BoxTE (2022) | .613 | .528 | .664 | .763 | .667 | .582 | .719 | .820 | .352 | .269 | .377 | .511 |
| TA-DisMult (2018) | .477 | .363 | - | .686 | .474 | .346 | - | .728 | .206 | .124 | .219 | .365 |
| DE-SimplE (2020) | .526 | .418 | .592 | .725 | .513 | .392 | .578 | .748 | .230 | .141 | .248 | .403 |
| TNTComplEx (2020) | .620 | .520 | .660 | .760 | .670 | .590 | .710 | .810 | .223 | .142 | .237 | .379 |
| TeLM (2021) | .625 | .545 | .673 | .774 | .678 | .599 | .728 | .823 | - | - | - | - |
| Time-LowFER (2022) | .623 | .549 | .671 | .757 | .638 | .555 | .690 | .791 | - | - | - | - |
| TeMP (2020) | .601 | .478 | .681 | .828 | .691 | .566 | .782 | .917 | .275 | .191 | .297 | .437 |
| T-GAP (2021) | .610 | .509 | .677 | .790 | .670 | .568 | .743 | .845 | - | - | - | - |
| SPA (concurrent work) | .658 | .544 | .737 | .857 | .713 | .580 | .820 | .933 | .360 | .282 | .384 | .510 |
| SToKE | .659 | .574 | .693 | .803 | .712 | .605 | .790 | .885 | .371 | .290 | .399 | .525 |
| Query: (A, Criticize_or_denounce, ?, 2014-1-29) Answer: B | ||||||||||
| Layer | 2014-1-27 | 2014-1-29 | ||||||||
| 12 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| - | - | - | - | - | .691 | .138 | .171 | - | - | |
| 11 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .256 | - | - | - | - | .097 | .065 | .083 | .021 | .034 | |
| 10 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .068 | .066 | .282 | .057 | .199 | .030 | - | - | - | - |
| Models | ICEWS14 | |||
| MRR | Hits@1 | Hits@3 | Hits@10 | |
| SToKE | .659 | .574 | .693 | .803 |
| - SC | .647 | .556 | .688 | .792 |
| - TC | .543 | .428 | .593 | .755 |
| +AFF | .641 | .550 | .678 | .796 |
| - TE | .641 | .557 | .675 | .791 |
| - SM | .634 | .547 | .669 | .778 |
| + BERT-large | .662 | .576 | .700 | .809 |
| GPU Resource | NVIDIA RTX A6000 | |
| Hyperparameter | Search Bound | Best Setup |
| length of time window m | choice{5, 10, 12, 14} | 10(ICEWS14), 14(ICEWS05-15), 5(GDELT) |
| number of concurrent facts n | choice{1, 2, 4} | 2 |
| length of input text l | choice{57, 107, 147} | 107(ICEWS14), 147(ICEWS05-15), 57(GDELT) |
| number of epochs | choice{30, 50} | 30 (ICEWS05-15, GDELT), 50 (ICEWS14) |
| batch size | choice{32, 64, 128} | 128 |
| learning rate | choice{2e-5, 5e-4} | 2e-5 |
| droupout rate | choice{0.1, 0.2} | 0.1 |
| number of model parameters | - | 98.6M(ICEWS14), 106.6M(ICEWS05-15), 88.4M(GDELT) |
| gpu hours | - | 12h(ICEWS14), 40h(ICEWS05-15), 60h(GDELT) |
| Datasets | |E| | |R| | Ntrain | Nvalid | Ntest | |T| |
| ICEWS14 | 7,128 | 230 | 72,826 | 8,941 | 8,963 | 365 |
| ICEWS05-15 | 10,488 | 251 | 368,962 | 46,275 | 46,092 | 4,017 |
| GDELT | 500 | 20 | 2,735,685 | 341,961 | 341,961 | 366 |
| Method | ICEWS14 | ICEWS05-15 | GDELT | |||||||||
| MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | |
| TransE (2013) | .280 | .094 | - | .637 | .294 | .090 | - | .663 | - | - | - | - |
| DisMult (2015) | .441 | .325 | .498 | .668 | .457 | .338 | .515 | .691 | .210 | .133 | .224 | .365 |
| ComplEx (2016) | .442 | .440 | .430 | .664 | .464 | .347 | .524 | .696 | .213 | .133 | .225 | .366 |
| Simple (2018) | .458 | .341 | .516 | .687 | .478 | .359 | .539 | .708 | .206 | .124 | .220 | .366 |
| TTransE (2016) | .255 | .074 | - | .601 | .271 | .084 | - | .616 | .115 | .000 | .160 | .318 |
| TeRo (2020) | .562 | .468 | .621 | .732 | .586 | .469 | .668 | .795 | .245 | .154 | .264 | .420 |
| ChronoR (2021) | .625 | .547 | .669 | .773 | .675 | .596 | .723 | .820 | - | - | - | - |
| RotateQVS (2022) | .591 | .507 | .642 | .754 | .633 | .529 | .709 | .813 | .270 | .175 | .293 | .458 |
| BoxTE (2022) | .613 | .528 | .664 | .763 | .667 | .582 | .719 | .820 | .352 | .269 | .377 | .511 |
| TA-DisMult (2018) | .477 | .363 | - | .686 | .474 | .346 | - | .728 | .206 | .124 | .219 | .365 |
| DE-SimplE (2020) | .526 | .418 | .592 | .725 | .513 | .392 | .578 | .748 | .230 | .141 | .248 | .403 |
| TNTComplEx (2020) | .620 | .520 | .660 | .760 | .670 | .590 | .710 | .810 | .223 | .142 | .237 | .379 |
| TeLM (2021) | .625 | .545 | .673 | .774 | .678 | .599 | .728 | .823 | - | - | - | - |
| Time-LowFER (2022) | .623 | .549 | .671 | .757 | .638 | .555 | .690 | .791 | - | - | - | - |
| TeMP (2020) | .601 | .478 | .681 | .828 | .691 | .566 | .782 | .917 | .275 | .191 | .297 | .437 |
| T-GAP (2021) | .610 | .509 | .677 | .790 | .670 | .568 | .743 | .845 | - | - | - | - |
| SPA (concurrent work) | .658 | .544 | .737 | .857 | .713 | .580 | .820 | .933 | .360 | .282 | .384 | .510 |
| SToKE | .659 | .574 | .693 | .803 | .712 | .605 | .790 | .885 | .371 | .290 | .399 | .525 |
| Query: (A, Criticize_or_denounce, ?, 2014-1-29) Answer: B | ||||||||||
| Layer | 2014-1-27 | 2014-1-29 | ||||||||
| 12 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| - | - | - | - | - | .691 | .138 | .171 | - | - | |
| 11 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .256 | - | - | - | - | .097 | .065 | .083 | .021 | .034 | |
| 10 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .068 | .066 | .282 | .057 | .199 | .030 | - | - | - | - |
| Models | ICEWS14 | |||
| MRR | Hits@1 | Hits@3 | Hits@10 | |
| SToKE | .659 | .574 | .693 | .803 |
| - SC | .647 | .556 | .688 | .792 |
| - TC | .543 | .428 | .593 | .755 |
| +AFF | .641 | .550 | .678 | .796 |
| - TE | .641 | .557 | .675 | .791 |
| - SM | .634 | .547 | .669 | .778 |
| + BERT-large | .662 | .576 | .700 | .809 |
| GPU Resource | NVIDIA RTX A6000 | |
| Hyperparameter | Search Bound | Best Setup |
| length of time window m | choice{5, 10, 12, 14} | 10(ICEWS14), 14(ICEWS05-15), 5(GDELT) |
| number of concurrent facts n | choice{1, 2, 4} | 2 |
| length of input text l | choice{57, 107, 147} | 107(ICEWS14), 147(ICEWS05-15), 57(GDELT) |
| number of epochs | choice{30, 50} | 30 (ICEWS05-15, GDELT), 50 (ICEWS14) |
| batch size | choice{32, 64, 128} | 128 |
| learning rate | choice{2e-5, 5e-4} | 2e-5 |
| droupout rate | choice{0.1, 0.2} | 0.1 |
| number of model parameters | - | 98.6M(ICEWS14), 106.6M(ICEWS05-15), 88.4M(GDELT) |
| gpu hours | - | 12h(ICEWS14), 40h(ICEWS05-15), 60h(GDELT) |
| Datasets | |E| | |R| | Ntrain | Nvalid | Ntest | |T| |
| ICEWS14 | 7,128 | 230 | 72,826 | 8,941 | 8,963 | 365 |
| ICEWS05-15 | 10,488 | 251 | 368,962 | 46,275 | 46,092 | 4,017 |
| GDELT | 500 | 20 | 2,735,685 | 341,961 | 341,961 | 366 |
| Method | ICEWS14 | ICEWS05-15 | GDELT | |||||||||
| MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | |
| TransE (2013) | .280 | .094 | - | .637 | .294 | .090 | - | .663 | - | - | - | - |
| DisMult (2015) | .441 | .325 | .498 | .668 | .457 | .338 | .515 | .691 | .210 | .133 | .224 | .365 |
| ComplEx (2016) | .442 | .440 | .430 | .664 | .464 | .347 | .524 | .696 | .213 | .133 | .225 | .366 |
| Simple (2018) | .458 | .341 | .516 | .687 | .478 | .359 | .539 | .708 | .206 | .124 | .220 | .366 |
| TTransE (2016) | .255 | .074 | - | .601 | .271 | .084 | - | .616 | .115 | .000 | .160 | .318 |
| TeRo (2020) | .562 | .468 | .621 | .732 | .586 | .469 | .668 | .795 | .245 | .154 | .264 | .420 |
| ChronoR (2021) | .625 | .547 | .669 | .773 | .675 | .596 | .723 | .820 | - | - | - | - |
| RotateQVS (2022) | .591 | .507 | .642 | .754 | .633 | .529 | .709 | .813 | .270 | .175 | .293 | .458 |
| BoxTE (2022) | .613 | .528 | .664 | .763 | .667 | .582 | .719 | .820 | .352 | .269 | .377 | .511 |
| TA-DisMult (2018) | .477 | .363 | - | .686 | .474 | .346 | - | .728 | .206 | .124 | .219 | .365 |
| DE-SimplE (2020) | .526 | .418 | .592 | .725 | .513 | .392 | .578 | .748 | .230 | .141 | .248 | .403 |
| TNTComplEx (2020) | .620 | .520 | .660 | .760 | .670 | .590 | .710 | .810 | .223 | .142 | .237 | .379 |
| TeLM (2021) | .625 | .545 | .673 | .774 | .678 | .599 | .728 | .823 | - | - | - | - |
| Time-LowFER (2022) | .623 | .549 | .671 | .757 | .638 | .555 | .690 | .791 | - | - | - | - |
| TeMP (2020) | .601 | .478 | .681 | .828 | .691 | .566 | .782 | .917 | .275 | .191 | .297 | .437 |
| T-GAP (2021) | .610 | .509 | .677 | .790 | .670 | .568 | .743 | .845 | - | - | - | - |
| SPA (concurrent work) | .658 | .544 | .737 | .857 | .713 | .580 | .820 | .933 | .360 | .282 | .384 | .510 |
| SToKE | .659 | .574 | .693 | .803 | .712 | .605 | .790 | .885 | .371 | .290 | .399 | .525 |
| Query: (A, Criticize_or_denounce, ?, 2014-1-29) Answer: B | ||||||||||
| Layer | 2014-1-27 | 2014-1-29 | ||||||||
| 12 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| - | - | - | - | - | .691 | .138 | .171 | - | - | |
| 11 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .256 | - | - | - | - | .097 | .065 | .083 | .021 | .034 | |
| 10 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .068 | .066 | .282 | .057 | .199 | .030 | - | - | - | - | |
| Models | ICEWS14 | |||
| MRR | Hits@1 | Hits@3 | Hits@10 | |
| SToKE | .659 | .574 | .693 | .803 |
| - SC | .647 | .556 | .688 | .792 |
| - TC | .543 | .428 | .593 | .755 |
| +AFF | .641 | .550 | .678 | .796 |
| - TE | .641 | .557 | .675 | .791 |
| - SM | .634 | .547 | .669 | .778 |
| + BERT-large | .662 | .576 | .700 | .809 |
| GPU Resource | NVIDIA RTX A6000 | |
| Hyperparameter | Search Bound | Best Setup |
| length of time window m | choice{5, 10, 12, 14} | 10(ICEWS14), 14(ICEWS05-15), 5(GDELT) |
| number of concurrent facts n | choice{1, 2, 4} | 2 |
| length of input text l | choice{57, 107, 147} | 107(ICEWS14), 147(ICEWS05-15), 57(GDELT) |
| number of epochs | choice{30, 50} | 30 (ICEWS05-15, GDELT), 50 (ICEWS14) |
| batch size | choice{32, 64, 128} | 128 |
| learning rate | choice{2e-5, 5e-4} | 2e-5 |
| droupout rate | choice{0.1, 0.2} | 0.1 |
| number of model parameters | - | 98.6M(ICEWS14), 106.6M(ICEWS05-15), 88.4M(GDELT) |
| gpu hours | - | 12h(ICEWS14), 40h(ICEWS05-15), 60h(GDELT) |
| Datasets | |E| | |R| | Ntrain | Nvalid | Ntest | |T| |
| ICEWS14 | 7,128 | 230 | 72,826 | 8,941 | 8,963 | 365 |
| ICEWS05-15 | 10,488 | 251 | 368,962 | 46,275 | 46,092 | 4,017 |
| GDELT | 500 | 20 | 2,735,685 | 341,961 | 341,961 | 366 |
| Method | ICEWS14 | ICEWS05-15 | GDELT | |||||||||
| MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | MRR | Hits@1 | Hits@3 | Hits@10 | |
| TransE (2013) | .280 | .094 | - | .637 | .294 | .090 | - | .663 | - | - | - | - |
| DisMult (2015) | .441 | .325 | .498 | .668 | .457 | .338 | .515 | .691 | .210 | .133 | .224 | .365 |
| ComplEx (2016) | .442 | .440 | .430 | .664 | .464 | .347 | .524 | .696 | .213 | .133 | .225 | .366 |
| Simple (2018) | .458 | .341 | .516 | .687 | .478 | .359 | .539 | .708 | .206 | .124 | .220 | .366 |
| TTransE (2016) | .255 | .074 | - | .601 | .271 | .084 | - | .616 | .115 | .000 | .160 | .318 |
| TeRo (2020) | .562 | .468 | .621 | .732 | .586 | .469 | .668 | .795 | .245 | .154 | .264 | .420 |
| ChronoR (2021) | .625 | .547 | .669 | .773 | .675 | .596 | .723 | .820 | - | - | - | - |
| RotateQVS (2022) | .591 | .507 | .642 | .754 | .633 | .529 | .709 | .813 | .270 | .175 | .293 | .458 |
| BoxTE (2022) | .613 | .528 | .664 | .763 | .667 | .582 | .719 | .820 | .352 | .269 | .377 | .511 |
| TA-DisMult (2018) | .477 | .363 | - | .686 | .474 | .346 | - | .728 | .206 | .124 | .219 | .365 |
| DE-SimplE (2020) | .526 | .418 | .592 | .725 | .513 | .392 | .578 | .748 | .230 | .141 | .248 | .403 |
| TNTComplEx (2020) | .620 | .520 | .660 | .760 | .670 | .590 | .710 | .810 | .223 | .142 | .237 | .379 |
| TeLM (2021) | .625 | .545 | .673 | .774 | .678 | .599 | .728 | .823 | - | - | - | - |
| Time-LowFER (2022) | .623 | .549 | .671 | .757 | .638 | .555 | .690 | .791 | - | - | - | - |
| TeMP (2020) | .601 | .478 | .681 | .828 | .691 | .566 | .782 | .917 | .275 | .191 | .297 | .437 |
| T-GAP (2021) | .610 | .509 | .677 | .790 | .670 | .568 | .743 | .845 | - | - | - | - |
| SPA (concurrent work) | .658 | .544 | .737 | .857 | .713 | .580 | .820 | .933 | .360 | .282 | .384 | .510 |
| SToKE | .659 | .574 | .693 | .803 | .712 | .605 | .790 | .885 | .371 | .290 | .399 | .525 |
| Query: (A, Criticize_or_denounce, ?, 2014-1-29) Answer: B | ||||||||||
| Layer | 2014-1-27 | 2014-1-29 | ||||||||
| 12 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| - | - | - | - | - | .691 | .138 | .171 | - | - | |
| 11 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .256 | - | - | - | - | .097 | .065 | .083 | .021 | .034 | |
| 10 | A | Criticize_or_denounce | B | Engage_in_negotiation | B | A | Criticize_or_denounce | [MASK] | Host_a.visit | C |
| .068 | .066 | .282 | .057 | .199 | .030 | - | - | - | - |
| Models | ICEWS14 | |||
| MRR | Hits@1 | Hits@3 | Hits@10 | |
| SToKE | .659 | .574 | .693 | .803 |
| - SC | .647 | .556 | .688 | .792 |
| - TC | .543 | .428 | .593 | .755 |
| +AFF | .641 | .550 | .678 | .796 |
| - TE | .641 | .557 | .675 | .791 |
| - SM | .634 | .547 | .669 | .778 |
| + BERT-large | .662 | .576 | .700 | .809 |
| GPU Resource | NVIDIA RTX A6000 | |
| Hyperparameter | Search Bound | Best Setup |
| length of time window m | choice{5, 10, 12, 14} | 10(ICEWS14), 14(ICEWS05-15), 5(GDELT) |
| number of concurrent facts n | choice{1, 2, 4} | 2 |
| length of input text l | choice{57, 107, 147} | 107(ICEWS14), 147(ICEWS05-15), 57(GDELT) |
| number of epochs | choice{30, 50} | 30 (ICEWS05-15, GDELT), 50 (ICEWS14) |
| batch size | choice{32, 64, 128} | 128 |
| learning rate | choice{2e-5, 5e-4} | 2e-5 |
| droupout rate | choice{0.1, 0.2} | 0.1 |
| number of model parameters | - | 98.6M(ICEWS14), 106.6M(ICEWS05-15), 88.4M(GDELT) |
| gpu hours | - | 12h(ICEWS14), 40h(ICEWS05-15), 60h(GDELT) |
| WN18RR | FB15k-237 | NELL-995 | ||||||||
| #Rel | #Ent | #Tri | #Rel | #Ent | #Tri | #R | #Ent | #Tri | ||
| V1 | train | 9 | 2746 | 6678 | 183 | 2000 | 5226 | 14 | 10915 | 5540 |
| test | 9 | 922 | 1991 | 146 | 1500 | 2404 | 14 | 225 | 1034 | |
| V2 | train | 10 | 6954 | 18968 | 203 | 3000 | 12085 | 88 | 2564 | 10109 |
| test | 10 | 2923 | 4863 | 176 | 2000 | 5092 | 79 | 4937 | 5521 | |
| V3 | train | 11 | 12078 | 32150 | 218 | 4000 | 22394 | 142 | 4647 | 20117 |
| test | 11 | 5084 | 7470 | 187 | 3000 | 9137 | 122 | 4921 | 9668 | |
| V3 | train | 9 | 3861 | 9842 | 222 | 5000 | 33916 | 77 | 2092 | 9289 |
| test | 9 | 7208 | 15157 | 204 | 3500 | 14554 | 61 | 3294 | 8520 |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.94 | 52.90 | 55.88 | 55.16 |
| DRUM† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.73 | 52.90 | 55.88 | 55.10 |
| RuleN† | 80.85 | 78.23 | 53.39 | 71.59 | 71.01 | 49.76 | 77.82 | 87.69 | 85.60 | 75.21 |
| Meta-iKG | - | - | - | - | - | 66.96 | 74.08 | 71.89 | 72.28 | 71.30 |
| MorsE | 84.14 | 81.50 | 70.92 | 79.61 | 79.04 | 83.17 | 95.67 | 95.69 | 95.89 | 92.61 |
| GraIL† | 82.45 | 78.68 | 58.43 | 73.41 | 73.24 | 64.15 | 81.80 | 82.83 | 89.29 | 79.51 |
| CoMPILE† | 83.60 | 79.82 | 60.69 | 75.49 | 74.90 | 67.64 | 82.98 | 84.67 | 87.44 | 80.68 |
| TACT† | 84.04 | 81.63 | 67.97 | 76.56 | 77.55 | 65.76 | 83.56 | 85.20 | 88.69 | 80.80 |
| ConGLR† | 85.64 | 92.93 | 70.74 | 92.90 | 85.55 | 68.29 | 85.98 | 88.61 | 89.31 | 82.93 |
| NodePiece | 83.00 | 88.60 | 78.50 | 80.70 | 82.70 | 87.30 | 93.90 | 94.40 | 94.90 | 92.35 |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 40.78 | 78.73 | 82.71 | 80.58 | 70.70 |
| DRUM† | 19.42 | 78.55 | 82.71 | 80.58 | 65.31 |
| RuleN† | 53.50 | 81.75 | 77.26 | 61.35 | 68.46 |
| Meta-iKG | 64.20 | 77.91 | 77.41 | 73.12 | 73.16 |
| MorsE | 65.20 | 80.70 | 87.67 | 53.44 | 71.75 |
| GraIL† | 59.50 | 93.25 | 91.41 | 73.19 | 79.33 |
| CoMPILE† | 58.38 | 93.87 | 92.77 | 75.19 | 80.05 |
| TACT* | 79.80 | 88.91 | 94.02 | 73.78 | 84.12 |
| ConGLR* | 81.07 | 94.92 | 94.36 | 81.61 | 87.99 |
| NodePiece | 89.00 | 90.10 | 93.60 | 89.30 | 90.50 |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| QAAR w/o D | 87.23 | 81.74 | 86.44 | 86.91 | 85.58 | 85.85 | 90.79 | 93.29 | 94.18 | 91.03 |
| Δ | -1.59 | -5.13 | -1.74 | -2.24 | -2.67 | -4.63 | -5.02 | -2.58 | -2.03 | -3.56 |
| QAAR w/o R | 86.17 | 86.16 | 86.77 | 89.43 | 87.13 | 84.63 | 92.46 | 93.64 | 95.64 | 91.59 |
| Δ | -2.65 | -0.71 | -1.41 | +0.28 | -1.12 | -5.85 | -3.35 | -2.23 | -0.57 | -3.00 |
| QAAR w/o S | 85.63 | 79.59 | 86.61 | 88.17 | 85.00 | 87.31 | 94.35 | 94.21 | 95.29 | 92.79 |
| Δ | -3.19 | -7.28 | -1.57 | -0.98 | -3.25 | -3.17 | -1.46 | -1.66 | -0.92 | -1.80 |
| QAAR w/o A | 86.96 | 81.51 | 87.76 | 88.38 | 86.15 | 84.39 | 94.97 | 94.91 | 95.36 | 92.41 |
| Δ | -1.86 | -5.36 | -0.42 | -0.77 | -2.10 | -6.09 | -0.84 | -0.96 | -0.85 | -2.18 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| QAAR w/o D | 87.00 | 90.33 | 90.61 | 4.81 | 88.19 |
| Δ | -2.50 | -5.99 | -3.95 | -6.34 | -4.69 |
| QAAR w/o R | 85.50 | 93.27 | 92.83 | 89.19 | 90.20 |
| Δ | -4.00 | -3.05 | -1.73 | -1.96 | -2.68 |
| QAAR w/o S | 82.50 | 92.85 | 92.58 | 90.28 | 88.55 |
| Δ | -7.00 | -3.47 | -1.98 | -0.87 | -3.33 |
| QAAR w/o A | 86.00 | 91.38 | 92.09 | 90.15 | 89.91 |
| Δ | -3.50 | -4.94 | -2.47 | -1.00 | -2.97 |
| #Neg. | WN18RR | FB15k-237 | NELL-995 | |||
| GraIL | QAAR | GraIL | QAAR | GraIL | QAAR | |
| 20 | 35.82 | 5.78 | 147.66 | 11.63 | 63.87 | 15.85 |
| 50 | 52.38 | 5.92 | 244.78 | 12.08 | 79.19 | 15.98 |
| 80 | 66.67 | 5.98 | 324.33 | 12.28 | 91.55 | 16.21 |
| 120 | 81.52 | 6.23 | 427.29 | 12.46 | 107.02 | 16.73 |
| 150 | 93.07 | 6.25 | 498.32 | 12.53 | 136.73 | 16.85 |
| WN18RR | FB15k-237 | NELL-995 | ||||||||
| #Rel | #Ent | #Tri | #Rel | #Ent | #Tri | #R | #Ent | #Tri | ||
| V1 | train | 9 | 2746 | 6678 | 183 | 2000 | 5226 | 14 | 10915 | 5540 |
| test | 9 | 922 | 1991 | 146 | 1500 | 2404 | 14 | 225 | 1034 | |
| V2 | train | 10 | 6954 | 18968 | 203 | 3000 | 12085 | 88 | 2564 | 10109 |
| test | 10 | 2923 | 4863 | 176 | 2000 | 5092 | 79 | 4937 | 5521 | |
| V3 | train | 11 | 12078 | 32150 | 218 | 4000 | 22394 | 142 | 4647 | 20117 |
| test | 11 | 5084 | 7470 | 187 | 3000 | 9137 | 122 | 4921 | 9668 | |
| V3 | train | 9 | 3861 | 9842 | 222 | 5000 | 33916 | 77 | 2092 | 9289 |
| test | 9 | 7208 | 15157 | 204 | 3500 | 14554 | 61 | 3294 | 8520 |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.94 | 52.90 | 55.88 | 55.16 |
| DRUM† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.73 | 52.90 | 55.88 | 55.10 |
| RuleN† | 80.85 | 78.23 | 53.39 | 71.59 | 71.01 | 49.76 | 77.82 | 87.69 | 85.60 | 75.21 |
| Meta-iKG | - | - | - | - | - | 66.96 | 74.08 | 71.89 | 72.28 | 71.30 |
| MorsE | 84.14 | 81.50 | 70.92 | 79.61 | 79.04 | 83.17 | 95.67 | 95.69 | 95.89 | 92.61 |
| GraIL† | 82.45 | 78.68 | 58.43 | 73.41 | 73.24 | 64.15 | 81.80 | 82.83 | 89.29 | 79.51 |
| CoMPILE† | 83.60 | 79.82 | 60.69 | 75.49 | 74.90 | 67.64 | 82.98 | 84.67 | 87.44 | 80.68 |
| TACT† | 84.04 | 81.63 | 67.97 | 76.56 | 77.55 | 65.76 | 83.56 | 85.20 | 88.69 | 80.80 |
| ConGLR† | 85.64 | 92.93 | 70.74 | 92.90 | 85.55 | 68.29 | 85.98 | 88.61 | 89.31 | 82.93 |
| NodePiece | 83.00 | 88.60 | 78.50 | 80.70 | 82.70 | 87.30 | 93.90 | 94.40 | 94.90 | 92.35 |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 40.78 | 78.73 | 82.71 | 80.58 | 70.70 |
| DRUM† | 19.42 | 78.55 | 82.71 | 80.58 | 65.31 |
| RuleN† | 53.50 | 81.75 | 77.26 | 61.35 | 68.46 |
| Meta-iKG | 64.20 | 77.91 | 77.41 | 73.12 | 73.16 |
| MorsE | 65.20 | 80.70 | 87.67 | 53.44 | 71.75 |
| GraIL† | 59.50 | 93.25 | 91.41 | 73.19 | 79.33 |
| CoMPILE† | 58.38 | 93.87 | 92.77 | 75.19 | 80.05 |
| TACT* | 79.80 | 88.91 | 94.02 | 73.78 | 84.12 |
| ConGLR* | 81.07 | 94.92 | 94.36 | 81.61 | 87.99 |
| NodePiece | 89.00 | 90.10 | 93.60 | 89.30 | 90.50 |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| QAAR w/o D | 87.23 | 81.74 | 86.44 | 86.91 | 85.58 | 85.85 | 90.79 | 93.29 | 94.18 | 91.03 |
| Δ | -1.59 | -5.13 | -1.74 | -2.24 | -2.67 | -4.63 | -5.02 | -2.58 | -2.03 | -3.56 |
| QAAR w/o R | 86.17 | 86.16 | 86.77 | 89.43 | 87.13 | 84.63 | 92.46 | 93.64 | 95.64 | 91.59 |
| Δ | -2.65 | -0.71 | -1.41 | +0.28 | -1.12 | -5.85 | -3.35 | -2.23 | -0.57 | -3.00 |
| QAAR w/o S | 85.63 | 79.59 | 86.61 | 88.17 | 85.00 | 87.31 | 94.35 | 94.21 | 95.29 | 92.79 |
| Δ | -3.19 | -7.28 | -1.57 | -0.98 | -3.25 | -3.17 | -1.46 | -1.66 | -0.92 | -1.80 |
| QAAR w/o A | 86.96 | 81.51 | 87.76 | 88.38 | 86.15 | 84.39 | 94.97 | 94.91 | 95.36 | 92.41 |
| Δ | -1.86 | -5.36 | -0.42 | -0.77 | -2.10 | -6.09 | -0.84 | -0.96 | -0.85 | -2.18 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| QAAR w/o D | 87.00 | 90.33 | 90.61 | 4.81 | 88.19 |
| Δ | -2.50 | -5.99 | -3.95 | -6.34 | -4.69 |
| QAAR w/o R | 85.50 | 93.27 | 92.83 | 89.19 | 90.20 |
| Δ | -4.00 | -3.05 | -1.73 | -1.96 | -2.68 |
| QAAR w/o S | 82.50 | 92.85 | 92.58 | 90.28 | 88.55 |
| Δ | -7.00 | -3.47 | -1.98 | -0.87 | -3.33 |
| QAAR w/o A | 86.00 | 91.38 | 92.09 | 90.15 | 89.91 |
| Δ | -3.50 | -4.94 | -2.47 | -1.00 | -2.97 |
| #Neg. | WN18RR | FB15k-237 | NELL-995 | |||
| GraIL | QAAR | GraIL | QAAR | GraIL | QAAR | |
| 20 | 35.82 | 5.78 | 147.66 | 11.63 | 63.87 | 15.85 |
| 50 | 52.38 | 5.92 | 244.78 | 12.08 | 79.19 | 15.98 |
| 80 | 66.67 | 5.98 | 324.33 | 12.28 | 91.55 | 16.21 |
| 120 | 81.52 | 6.23 | 427.29 | 12.46 | 107.02 | 16.73 |
| 150 | 93.07 | 6.25 | 498.32 | 12.53 | 136.73 | 16.85 |
| WN18RR | FB15k-237 | NELL-995 | ||||||||
| #Rel | #Ent | #Tri | #Rel | #Ent | #Tri | #R | #Ent | #Tri | ||
| V1 | train | 9 | 2746 | 6678 | 183 | 2000 | 5226 | 14 | 10915 | 5540 |
| test | 9 | 922 | 1991 | 146 | 1500 | 2404 | 14 | 225 | 1034 | |
| V2 | train | 10 | 6954 | 18968 | 203 | 3000 | 12085 | 88 | 2564 | 10109 |
| test | 10 | 2923 | 4863 | 176 | 2000 | 5092 | 79 | 4937 | 5521 | |
| V3 | train | 11 | 12078 | 32150 | 218 | 4000 | 22394 | 142 | 4647 | 20117 |
| test | 11 | 5084 | 7470 | 187 | 3000 | 9137 | 122 | 4921 | 9668 | |
| V3 | train | 9 | 3861 | 9842 | 222 | 5000 | 33916 | 77 | 2092 | 9289 |
| test | 9 | 7208 | 15157 | 204 | 3500 | 14554 | 61 | 3294 | 8520 | |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.94 | 52.90 | 55.88 | 55.16 |
| DRUM† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.73 | 52.90 | 55.88 | 55.10 |
| RuleN† | 80.85 | 78.23 | 53.39 | 71.59 | 71.01 | 49.76 | 77.82 | 87.69 | 85.60 | 75.21 |
| Meta-iKG | - | - | - | - | - | 66.96 | 74.08 | 71.89 | 72.28 | 71.30 |
| MorsE | 84.14 | 81.50 | 70.92 | 79.61 | 79.04 | 83.17 | 95.67 | 95.69 | 95.89 | 92.61 |
| GraIL† | 82.45 | 78.68 | 58.43 | 73.41 | 73.24 | 64.15 | 81.80 | 82.83 | 89.29 | 79.51 |
| CoMPILE† | 83.60 | 79.82 | 60.69 | 75.49 | 74.90 | 67.64 | 82.98 | 84.67 | 87.44 | 80.68 |
| TACT† | 84.04 | 81.63 | 67.97 | 76.56 | 77.55 | 65.76 | 83.56 | 85.20 | 88.69 | 80.80 |
| ConGLR† | 85.64 | 92.93 | 70.74 | 92.90 | 85.55 | 68.29 | 85.98 | 88.61 | 89.31 | 82.93 |
| NodePiece | 83.00 | 88.60 | 78.50 | 80.70 | 82.70 | 87.30 | 93.90 | 94.40 | 94.90 | 92.35 |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 40.78 | 78.73 | 82.71 | 80.58 | 70.70 |
| DRUM† | 19.42 | 78.55 | 82.71 | 80.58 | 65.31 |
| RuleN† | 53.50 | 81.75 | 77.26 | 61.35 | 68.46 |
| Meta-iKG | 64.20 | 77.91 | 77.41 | 73.12 | 73.16 |
| MorsE | 65.20 | 80.70 | 87.67 | 53.44 | 71.75 |
| GraIL† | 59.50 | 93.25 | 91.41 | 73.19 | 79.33 |
| CoMPILE† | 58.38 | 93.87 | 92.77 | 75.19 | 80.05 |
| TACT* | 79.80 | 88.91 | 94.02 | 73.78 | 84.12 |
| ConGLR* | 81.07 | 94.92 | 94.36 | 81.61 | 87.99 |
| NodePiece | 89.00 | 90.10 | 93.60 | 89.30 | 90.50 |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| QAAR w/o D | 87.23 | 81.74 | 86.44 | 86.91 | 85.58 | 85.85 | 90.79 | 93.29 | 94.18 | 91.03 |
| Δ | -1.59 | -5.13 | -1.74 | -2.24 | -2.67 | -4.63 | -5.02 | -2.58 | -2.03 | -3.56 |
| QAAR w/o R | 86.17 | 86.16 | 86.77 | 89.43 | 87.13 | 84.63 | 92.46 | 93.64 | 95.64 | 91.59 |
| Δ | -2.65 | -0.71 | -1.41 | +0.28 | -1.12 | -5.85 | -3.35 | -2.23 | -0.57 | -3.00 |
| QAAR w/o S | 85.63 | 79.59 | 86.61 | 88.17 | 85.00 | 87.31 | 94.35 | 94.21 | 95.29 | 92.79 |
| Δ | -3.19 | -7.28 | -1.57 | -0.98 | -3.25 | -3.17 | -1.46 | -1.66 | -0.92 | -1.80 |
| QAAR w/o A | 86.96 | 81.51 | 87.76 | 88.38 | 86.15 | 84.39 | 94.97 | 94.91 | 95.36 | 92.41 |
| Δ | -1.86 | -5.36 | -0.42 | -0.77 | -2.10 | -6.09 | -0.84 | -0.96 | -0.85 | -2.18 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| QAAR w/o D | 87.00 | 90.33 | 90.61 | 4.81 | 88.19 |
| Δ | -2.50 | -5.99 | -3.95 | -6.34 | -4.69 |
| QAAR w/o R | 85.50 | 93.27 | 92.83 | 89.19 | 90.20 |
| Δ | -4.00 | -3.05 | -1.73 | -1.96 | -2.68 |
| QAAR w/o S | 82.50 | 92.85 | 92.58 | 90.28 | 88.55 |
| Δ | -7.00 | -3.47 | -1.98 | -0.87 | -3.33 |
| QAAR w/o A | 86.00 | 91.38 | 92.09 | 90.15 | 89.91 |
| Δ | -3.50 | -4.94 | -2.47 | -1.00 | -2.97 |
| #Neg. | WN18RR | FB15k-237 | NELL-995 | |||
| GraIL | QAAR | GraIL | QAAR | GraIL | QAAR | |
| 20 | 35.82 | 5.78 | 147.66 | 11.63 | 63.87 | 15.85 |
| 50 | 52.38 | 5.92 | 244.78 | 12.08 | 79.19 | 15.98 |
| 80 | 66.67 | 5.98 | 324.33 | 12.28 | 91.55 | 16.21 |
| 120 | 81.52 | 6.23 | 427.29 | 12.46 | 107.02 | 16.73 |
| 150 | 93.07 | 6.25 | 498.32 | 12.53 | 136.73 | 16.85 |
| WN18RR | FB15k-237 | NELL-995 | ||||||||
| #Rel | #Ent | #Tri | #Rel | #Ent | #Tri | #R | #Ent | #Tri | ||
| V1 | train | 9 | 2746 | 6678 | 183 | 2000 | 5226 | 14 | 10915 | 5540 |
| test | 9 | 922 | 1991 | 146 | 1500 | 2404 | 14 | 225 | 1034 | |
| V2 | train | 10 | 6954 | 18968 | 203 | 3000 | 12085 | 88 | 2564 | 10109 |
| test | 10 | 2923 | 4863 | 176 | 2000 | 5092 | 79 | 4937 | 5521 | |
| V3 | train | 11 | 12078 | 32150 | 218 | 4000 | 22394 | 142 | 4647 | 20117 |
| test | 11 | 5084 | 7470 | 187 | 3000 | 9137 | 122 | 4921 | 9668 | |
| V3 | train | 9 | 3861 | 9842 | 222 | 5000 | 33916 | 77 | 2092 | 9289 |
| test | 9 | 7208 | 15157 | 204 | 3500 | 14554 | 61 | 3294 | 8520 |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.94 | 52.90 | 55.88 | 55.16 |
| DRUM† | 74.37 | 68.93 | 46.18 | 67.13 | 64.15 | 52.92 | 58.73 | 52.90 | 55.88 | 55.10 |
| RuleN† | 80.85 | 78.23 | 53.39 | 71.59 | 71.01 | 49.76 | 77.82 | 87.69 | 85.60 | 75.21 |
| Meta-iKG | - | - | - | - | - | 66.96 | 74.08 | 71.89 | 72.28 | 71.30 |
| MorsE | 84.14 | 81.50 | 70.92 | 79.61 | 79.04 | 83.17 | 95.67 | 95.69 | 95.89 | 92.61 |
| GraIL† | 82.45 | 78.68 | 58.43 | 73.41 | 73.24 | 64.15 | 81.80 | 82.83 | 89.29 | 79.51 |
| CoMPILE† | 83.60 | 79.82 | 60.69 | 75.49 | 74.90 | 67.64 | 82.98 | 84.67 | 87.44 | 80.68 |
| TACT† | 84.04 | 81.63 | 67.97 | 76.56 | 77.55 | 65.76 | 83.56 | 85.20 | 88.69 | 80.80 |
| ConGLR† | 85.64 | 92.93 | 70.74 | 92.90 | 85.55 | 68.29 | 85.98 | 88.61 | 89.31 | 82.93 |
| NodePiece | 83.00 | 88.60 | 78.50 | 80.70 | 82.70 | 87.30 | 93.90 | 94.40 | 94.90 | 92.35 |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| Neural-LP† | 40.78 | 78.73 | 82.71 | 80.58 | 70.70 |
| DRUM† | 19.42 | 78.55 | 82.71 | 80.58 | 65.31 |
| RuleN† | 53.50 | 81.75 | 77.26 | 61.35 | 68.46 |
| Meta-iKG | 64.20 | 77.91 | 77.41 | 73.12 | 73.16 |
| MorsE | 65.20 | 80.70 | 87.67 | 53.44 | 71.75 |
| GraIL† | 59.50 | 93.25 | 91.41 | 73.19 | 79.33 |
| CoMPILE† | 58.38 | 93.87 | 92.77 | 75.19 | 80.05 |
| TACT* | 79.80 | 88.91 | 94.02 | 73.78 | 84.12 |
| ConGLR* | 81.07 | 94.92 | 94.36 | 81.61 | 87.99 |
| NodePiece | 89.00 | 90.10 | 93.60 | 89.30 | 90.50 |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| Models | WN18RR | FB15k-237 | ||||||||
| v1 | v2 | v3 | v4 | Avg. | v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 88.82 | 86.87 | 88.18 | 89.15 | 88.25 | 90.48 | 95.81 | 95.87 | 96.21 | 94.59 |
| QAAR w/o D | 87.23 | 81.74 | 86.44 | 86.91 | 85.58 | 85.85 | 90.79 | 93.29 | 94.18 | 91.03 |
| Δ | -1.59 | -5.13 | -1.74 | -2.24 | -2.67 | -4.63 | -5.02 | -2.58 | -2.03 | -3.56 |
| QAAR w/o R | 86.17 | 86.16 | 86.77 | 89.43 | 87.13 | 84.63 | 92.46 | 93.64 | 95.64 | 91.59 |
| Δ | -2.65 | -0.71 | -1.41 | +0.28 | -1.12 | -5.85 | -3.35 | -2.23 | -0.57 | -3.00 |
| QAAR w/o S | 85.63 | 79.59 | 86.61 | 88.17 | 85.00 | 87.31 | 94.35 | 94.21 | 95.29 | 92.79 |
| Δ | -3.19 | -7.28 | -1.57 | -0.98 | -3.25 | -3.17 | -1.46 | -1.66 | -0.92 | -1.80 |
| QAAR w/o A | 86.96 | 81.51 | 87.76 | 88.38 | 86.15 | 84.39 | 94.97 | 94.91 | 95.36 | 92.41 |
| Δ | -1.86 | -5.36 | -0.42 | -0.77 | -2.10 | -6.09 | -0.84 | -0.96 | -0.85 | -2.18 |
| Models | NELL-995 | ||||
| v1 | v2 | v3 | v4 | Avg. | |
| QAAR | 89.50 | 96.32 | 94.56 | 91.15 | 92.88 |
| QAAR w/o D | 87.00 | 90.33 | 90.61 | 4.81 | 88.19 |
| Δ | -2.50 | -5.99 | -3.95 | -6.34 | -4.69 |
| QAAR w/o R | 85.50 | 93.27 | 92.83 | 89.19 | 90.20 |
| Δ | -4.00 | -3.05 | -1.73 | -1.96 | -2.68 |
| QAAR w/o S | 82.50 | 92.85 | 92.58 | 90.28 | 88.55 |
| Δ | -7.00 | -3.47 | -1.98 | -0.87 | -3.33 |
| QAAR w/o A | 86.00 | 91.38 | 92.09 | 90.15 | 89.91 |
| Δ | -3.50 | -4.94 | -2.47 | -1.00 | -2.97 |
| #Neg. | WN18RR | FB15k-237 | NELL-995 | |||
| GraIL | QAAR | GraIL | QAAR | GraIL | QAAR | |
| 20 | 35.82 | 5.78 | 147.66 | 11.63 | 63.87 | 15.85 |
| 50 | 52.38 | 5.92 | 244.78 | 12.08 | 79.19 | 15.98 |
| 80 | 66.67 | 5.98 | 324.33 | 12.28 | 91.55 | 16.21 |
| 120 | 81.52 | 6.23 | 427.29 | 12.46 | 107.02 | 16.73 |
| 150 | 93.07 | 6.25 | 498.32 | 12.53 | 136.73 | 16.85 |
| Data | Accuracy | Group-Accuracy |
| Unbiased-source | 34.9 | 29.5 |
| Unbiased-DQGF | 62.7 | 42.0 |
| Unbiased-GT | 78.4 | 64.0 |
| Strategy | Accuracy |
| All | 62.7 |
| (w/o)Sub-equations | 58.5 |
| (w/o)Same unit | 47.3 |
| (w/o)Different units | 60.4 |
| Methods | BLEU | BERT F1 | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Baseline | 52.3 | 87.4 | 77.2 | 59.4 | 70.6 |
| EQG(w/o)IM | 54.2 | 87.9 | 78.4 | 61.4 | 72.0 |
| EQG | 60.5 | 89.7 | 81.4 | 66.7 | 77.4 |
| Data | Accuracy | Group-Accuracy | Deq-Accuracy |
| Ori | 63.6 | 56.9 | 69.4 |
| Diverse | 68.4 | 60.2 | 48.1 |
| Scenario: A factory produce 3000 parts, 750 in the first 6 days and the rest in 15 days | Ori | Diverse |
| Question1: How many will be produced on average per day in the future? (original question) Equation1: x=(3000-750)/15 | True | True |
| Question2: How many more will be produced? Equation2: x=3000-750 | False | True |
| Question3: What is the average number of parts produced per day for the first 6 days? Equation3: x=750/6 | False | True |
| Original Data\nText: A pair of pants costs 58 dollars, and a jacket costs 4 times as much as a pair of pants.\nHow many dollars are spent on 5 sets of these clothes?\nEquation: x=5*(58+(58*4)) |
| Generated Data\nQuestion: How much do a pair of pants and a jacket cost in total?\nEquation: x=58+58*4 |
| Question: How much does a jacket cost?\nEquation: x=58*4 |
| Original Data\nText: Dingding has read 180 pages of a book and has 150 pages left to read.\nHow many pages are there in this book?\nEquation: x=180+150 |
| Generated Data\nQuestion: How many more pages have been read than have not been read?\nEquation: x=180-150 |
| Question: How many times more pages have been read than have not been read?\nEquation: x=180/150 |
| Original Data\nText: Qiangqiang's father and mother work outside. Father sends Qiangqiang 458 dollars a month and mother sends Qiangqiang 447 dollars a month. How much money do Qiangqiang's father and mother send to Qiangqiang each month?\nEquation: x=458+447 |
| Generated Data\nQuestion: How much more money does the mother send to Qiangqiang each month than the father?\nEquation: x=447-458 |
| Question: How many times more money does the mother send to Qiangqiang than the father each month?\nx=447/458 |
| Question: How much more money does the father send to Qiangqiang each month than the mother?\nEquation: x=458-447 |
| Question: How many times more money does the father send to Qiangqiang each month than the mother?\nEquation: x=458/447 |
| Original Data\nText: Mom bought a toothbrush for 3.6 dollars and a box of toothpaste for 9.5 dollars. How much is a toothbrush cheaper than a box of toothpaste?\nEquation: x=9.5-3.6 |
| Generated Data\nQuestion: What is the ratio of the price of a box of toothpaste to a toothbrush?\nEquation: x=9.5/3.6 |
| Question: How much do a toothbrush and a box of toothpaste cost in total?\nEquation: x=3.6+9.5 |
| Question: How much more expensive is a toothbrush than a box of toothpaste?\nEquation: x=3.6-9.5 |
| What is the ratio of the price of a toothbrush to a box of toothpaste?\nEquation: x=3.6/9.5 |
| Original Data\nText: A storybook has 438 pages and Xiao Liang has read 202 pages. How many pages does Xiao Liang have left to read?\nEquation: x=438-202 |
| Generated Data\nQuestion: What is the ratio of the number of pages Xiao Liang has read to the total number of pages in the storybook?\nEquation: x=202/438 |
| Question: How many times is the total number of pages in the storybook than the number of pages Xiao Liang has read?\nEquation: x=438/202 |
| Data | Accuracy | Group-Accuracy |
| Unbiased-source | 34.9 | 29.5 |
| Unbiased-DQGF | 62.7 | 42.0 |
| Unbiased-GT | 78.4 | 64.0 |
| Strategy | Accuracy |
| All | 62.7 |
| (w/o)Sub-equations | 58.5 |
| (w/o)Same unit | 47.3 |
| (w/o)Different units | 60.4 |
| Methods | BLEU | BERT F1 | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Baseline | 52.3 | 87.4 | 77.2 | 59.4 | 70.6 |
| EQG(w/o)IM | 54.2 | 87.9 | 78.4 | 61.4 | 72.0 |
| EQG | 60.5 | 89.7 | 81.4 | 66.7 | 77.4 |
| Data | Accuracy | Group-Accuracy | Deq-Accuracy |
| Ori | 63.6 | 56.9 | 69.4 |
| Diverse | 68.4 | 60.2 | 48.1 |
| Scenario: A factory produce 3000 parts, 750 in the first 6 days and the rest in 15 days | Ori | Diverse |
| Question1: How many will be produced on average per day in the future? (original question) Equation1: x=(3000-750)/15 | True | True |
| Question2: How many more will be produced? Equation2: x=3000-750 | False | True |
| Question3: What is the average number of parts produced per day for the first 6 days? Equation3: x=750/6 | False | True |
| Original Data\nText: A pair of pants costs 58 dollars, and a jacket costs 4 times as much as a pair of pants.\nHow many dollars are spent on 5 sets of these clothes?\nEquation: x=5*(58+(58*4)) |
| Generated Data\nQuestion: How much do a pair of pants and a jacket cost in total?\nEquation: x=58+58*4 |
| Question: How much does a jacket cost?\nEquation: x=58*4 |
| Original Data\nText: Dingding has read 180 pages of a book and has 150 pages left to read.\nHow many pages are there in this book?\nEquation: x=180+150 |
| Generated Data\nQuestion: How many more pages have been read than have not been read?\nEquation: x=180-150 |
| Question: How many times more pages have been read than have not been read?\nEquation: x=180/150 |
| Original Data\nText: Qiangqiang's father and mother work outside. Father sends Qiangqiang 458 dollars a month and mother sends Qiangqiang 447 dollars a month. How much money do Qiangqiang's father and mother send to Qiangqiang each month?\nEquation: x=458+447 |
| Generated Data\nQuestion: How much more money does the mother send to Qiangqiang each month than the father?\nEquation: x=447-458 |
| Question: How many times more money does the mother send to Qiangqiang than the father each month?\nx=447/458 |
| Question: How much more money does the father send to Qiangqiang each month than the mother?\nEquation: x=458-447 |
| Question: How many times more money does the father send to Qiangqiang each month than the mother?\nEquation: x=458/447 |
| Original Data\nText: Mom bought a toothbrush for 3.6 dollars and a box of toothpaste for 9.5 dollars. How much is a toothbrush cheaper than a box of toothpaste?\nEquation: x=9.5-3.6 |
| Generated Data\nQuestion: What is the ratio of the price of a box of toothpaste to a toothbrush?\nEquation: x=9.5/3.6 |
| Question: How much do a toothbrush and a box of toothpaste cost in total?\nEquation: x=3.6+9.5 |
| Question: How much more expensive is a toothbrush than a box of toothpaste?\nEquation: x=3.6-9.5 |
| What is the ratio of the price of a toothbrush to a box of toothpaste?\nEquation: x=3.6/9.5 |
| Original Data\nText: A storybook has 438 pages and Xiao Liang has read 202 pages. How many pages does Xiao Liang have left to read?\nEquation: x=438-202 |
| Generated Data\nQuestion: What is the ratio of the number of pages Xiao Liang has read to the total number of pages in the storybook?\nEquation: x=202/438 |
| Question: How many times is the total number of pages in the storybook than the number of pages Xiao Liang has read?\nEquation: x=438/202 |
| Data | Accuracy | Group-Accuracy |
| Unbiased-source | 34.9 | 29.5 |
| Unbiased-DQGF | 62.7 | 42.0 |
| Unbiased-GT | 78.4 | 64.0 |
| Strategy | Accuracy |
| All | 62.7 |
| (w/o)Sub-equations | 58.5 |
| (w/o)Same unit | 47.3 |
| (w/o)Different units | 60.4 |
| Methods | BLEU | BERT F1 | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Baseline | 52.3 | 87.4 | 77.2 | 59.4 | 70.6 |
| EQG(w/o)IM | 54.2 | 87.9 | 78.4 | 61.4 | 72.0 |
| EQG | 60.5 | 89.7 | 81.4 | 66.7 | 77.4 |
| Data | Accuracy | Group-Accuracy | Deq-Accuracy |
| Ori | 63.6 | 56.9 | 69.4 |
| Diverse | 68.4 | 60.2 | 48.1 |
| Scenario: A factory produce 3000 parts, 750 in the first 6 days and the rest in 15 days | Ori | Diverse |
| Question1: How many will be produced on average per day in the future? (original question) Equation1: x=(3000-750)/15 | True | True |
| Question2: How many more will be produced? Equation2: x=3000-750 | False | True |
| Question3: What is the average number of parts produced per day for the first 6 days? Equation3: x=750/6 | False | True |
| Original Data +Text: A pair of pants costs 58 dollars, and a jacket costs 4 times as much as a pair of pants. +How many dollars are spent on 5 sets of these clothes? +Equation: x=5*(58+(58*4)) |
| Generated Data +Question: How much do a pair of pants and a jacket cost in total? +Equation: x=58+58*4 |
| Question: How much does a jacket cost? +Equation: x=58*4 |
| Original Data +Text: Dingding has read 180 pages of a book and has 150 pages left to read. +How many pages are there in this book? +Equation: x=180+150 |
| Generated Data +Question: How many more pages have been read than have not been read? +Equation: x=180-150 |
| Question: How many times more pages have been read than have not been read? +Equation: x=180/150 |
| Original Data +Text: Qiangqiang's father and mother work outside. Father sends Qiangqiang 458 dollars a month and mother sends Qiangqiang 447 dollars a month. How much money do Qiangqiang's father and mother send to Qiangqiang each month? +Equation: x=458+447 |
| Generated Data +Question: How much more money does the mother send to Qiangqiang each month than the father? +Equation: x=447-458 |
| Question: How many times more money does the mother send to Qiangqiang than the father each month? +x=447/458 |
| Question: How much more money does the father send to Qiangqiang each month than the mother? +Equation: x=458-447 |
| Question: How many times more money does the father send to Qiangqiang each month than the mother? +Equation: x=458/447 |
| Original Data +Text: Mom bought a toothbrush for 3.6 dollars and a box of toothpaste for 9.5 dollars. How much is a toothbrush cheaper than a box of toothpaste? +Equation: x=9.5-3.6 |
| Generated Data +Question: What is the ratio of the price of a box of toothpaste to a toothbrush? +Equation: x=9.5/3.6 |
| Question: How much do a toothbrush and a box of toothpaste cost in total? +Equation: x=3.6+9.5 |
| Question: How much more expensive is a toothbrush than a box of toothpaste? +Equation: x=3.6-9.5 |
| What is the ratio of the price of a toothbrush to a box of toothpaste? +Equation: x=3.6/9.5 |
| Original Data +Text: A storybook has 438 pages and Xiao Liang has read 202 pages. How many pages does Xiao Liang have left to read? +Equation: x=438-202 |
| Generated Data +Question: What is the ratio of the number of pages Xiao Liang has read to the total number of pages in the storybook? +Equation: x=202/438 |
| Question: How many times is the total number of pages in the storybook than the number of pages Xiao Liang has read? +Equation: x=438/202 |
| Data | Accuracy | Group-Accuracy |
| Unbiased-source | 34.9 | 29.5 |
| Unbiased-DQGF | 62.7 | 42.0 |
| Unbiased-GT | 78.4 | 64.0 |
| Strategy | Accuracy |
| All | 62.7 |
| (w/o)Sub-equations | 58.5 |
| (w/o)Same unit | 47.3 |
| (w/o)Different units | 60.4 |
| Methods | BLEU | BERT F1 | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Baseline | 52.3 | 87.4 | 77.2 | 59.4 | 70.6 |
| EQG(w/o)IM | 54.2 | 87.9 | 78.4 | 61.4 | 72.0 |
| EQG | 60.5 | 89.7 | 81.4 | 66.7 | 77.4 |
| Data | Accuracy | Group-Accuracy | Deq-Accuracy |
| Ori | 63.6 | 56.9 | 69.4 |
| Diverse | 68.4 | 60.2 | 48.1 |
| Scenario: A factory produce 3000 parts, 750 in the first 6 days and the rest in 15 days | Ori | Diverse |
| Question1: How many will be produced on average per day in the future? (original question) Equation1: x=(3000-750)/15 | True | True |
| Question2: How many more will be produced? Equation2: x=3000-750 | False | True |
| Question3: What is the average number of parts produced per day for the first 6 days? Equation3: x=750/6 | False | True |
| Original Data\nText: A pair of pants costs 58 dollars, and a jacket costs 4 times as much as a pair of pants.\nHow many dollars are spent on 5 sets of these clothes?\nEquation: x=5*(58+(58*4)) |
| Generated Data\nQuestion: How much do a pair of pants and a jacket cost in total?\nEquation: x=58+58*4 |
| Question: How much does a jacket cost?\nEquation: x=58*4 |
| Original Data\nText: Dingding has read 180 pages of a book and has 150 pages left to read.\nHow many pages are there in this book?\nEquation: x=180+150 |
| Generated Data\nQuestion: How many more pages have been read than have not been read?\nEquation: x=180-150 |
| Question: How many times more pages have been read than have not been read?\nEquation: x=180/150 |
| Original Data\nText: Qiangqiang's father and mother work outside. Father sends Qiangqiang 458 dollars a month and mother sends Qiangqiang 447 dollars a month. How much money do Qiangqiang's father and mother send to Qiangqiang each month?\nEquation: x=458+447 |
| Generated Data\nQuestion: How much more money does the mother send to Qiangqiang each month than the father?\nEquation: x=447-458 |
| Question: How many times more money does the mother send to Qiangqiang than the father each month?\nx=447/458 |
| Question: How much more money does the father send to Qiangqiang each month than the mother?\nEquation: x=458-447 |
| Question: How many times more money does the father send to Qiangqiang each month than the mother?\nEquation: x=458/447 |
| Original Data\nText: Mom bought a toothbrush for 3.6 dollars and a box of toothpaste for 9.5 dollars. How much is a toothbrush cheaper than a box of toothpaste?\nEquation: x=9.5-3.6 |
| Generated Data\nQuestion: What is the ratio of the price of a box of toothpaste to a toothbrush?\nEquation: x=9.5/3.6 |
| Question: How much do a toothbrush and a box of toothpaste cost in total?\nEquation: x=3.6+9.5 |
| Question: How much more expensive is a toothbrush than a box of toothpaste?\nEquation: x=3.6-9.5 |
| What is the ratio of the price of a toothbrush to a box of toothpaste?\nEquation: x=3.6/9.5 |
| Original Data\nText: A storybook has 438 pages and Xiao Liang has read 202 pages. How many pages does Xiao Liang have left to read?\nEquation: x=438-202 |
| Generated Data\nQuestion: What is the ratio of the number of pages Xiao Liang has read to the total number of pages in the storybook?\nEquation: x=202/438 |
| Question: How many times is the total number of pages in the storybook than the number of pages Xiao Liang has read?\nEquation: x=438/202 |
| Model | Pretrained Visual BB | Language Backbone | Accuracy |
| OVR-CNN0 | X | GloVE | 20.62±0.86 |
| Ours | X | GloVE | 21.64±1.02 |
| OVR-CNN0 | X | BERT | 22.73±0.06 |
| Ours | X | BERT | 23.74±0.48 |
| OVR-CNN0 | ✓ | BERT | 34.46±0.11 |
| Ours | ✓ | BERT | 35.49±0.21 |
| Model | P1 Obj. | P2 Obj. |
| OVR-CNN0 | 49.36 | 16.61 |
| Ours | 50.9 | 26.1 |
| Model | Accuracy |
| OVR-CNN0 | 13.10 |
| Ours | 17.45 |
| Model | Cur. | Loss | Accuracy |
| OVR-CNN0 | L | 19.73 | |
| OVR-CNN0 | LCR | 20.65 | |
| Ours | ✓ | L | 21.02 |
| Ours | ✓ | LCR | 21.58 |
| Ours | ✓ | LCP | 22.57 |
| Model | Base | Target | Generalized | ||
| Base | Target | All | |||
| OVR-CNN | 46.8 | 27.5 | 46.0 | 22.8 | 39.9 |
| Ours | 48.89 | 26.65 | 48.18 | 21.99 | 41.33 |
| Model | Accuracy |
| OVR-CNN | 23.45 |
| Ours | 23.33 |
| Model | Pretrained Visual BB | Language Backbone | Accuracy |
| OVR-CNN0 | X | GloVE | 20.62±0.86 |
| Ours | X | GloVE | 21.64±1.02 |
| OVR-CNN0 | X | BERT | 22.73±0.06 |
| Ours | X | BERT | 23.74±0.48 |
| OVR-CNN0 | ✓ | BERT | 34.46±0.11 |
| Ours | ✓ | BERT | 35.49±0.21 |
| Model | P1 Obj. | P2 Obj. |
| OVR-CNN0 | 49.36 | 16.61 |
| Ours | 50.9 | 26.1 |
| Model | Accuracy |
| OVR-CNN0 | 13.10 |
| Ours | 17.45 |
| Model | Cur. | Loss | Accuracy |
| OVR-CNN0 | L | 19.73 | |
| OVR-CNN0 | LCR | 20.65 | |
| Ours | ✓ | L | 21.02 |
| Ours | ✓ | LCR | 21.58 |
| Ours | ✓ | LCP | 22.57 |
| Model | Base | Target | Generalized | ||
| Base | Target | All | |||
| OVR-CNN | 46.8 | 27.5 | 46.0 | 22.8 | 39.9 |
| Ours | 48.89 | 26.65 | 48.18 | 21.99 | 41.33 |
| Model | Accuracy |
| OVR-CNN | 23.45 |
| Ours | 23.33 |
| Model | Pretrained Visual BB | Language Backbone | Accuracy |
| OVR-CNN0 | X | GloVE | 20.62±0.86 |
| Ours | X | GloVE | 21.64±1.02 |
| OVR-CNN0 | X | BERT | 22.73±0.06 |
| Ours | X | BERT | 23.74±0.48 |
| OVR-CNN0 | ✓ | BERT | 34.46±0.11 |
| Ours | ✓ | BERT | 35.49±0.21 |
| Model | P1 Obj. | P2 Obj. |
| OVR-CNN0 | 49.36 | 16.61 |
| Ours | 50.9 | 26.1 |
| Model | Accuracy |
| OVR-CNN0 | 13.10 |
| Ours | 17.45 |
| Model | Cur. | Loss | Accuracy |
| OVR-CNN0 | L | 19.73 | |
| OVR-CNN0 | LCR | 20.65 | |
| Ours | ✓ | L | 21.02 |
| Ours | ✓ | LCR | 21.58 |
| Ours | ✓ | LCP | 22.57 |
| Model | Base | Target | Generalized | ||
| Base | Target | All | |||
| OVR-CNN | 46.8 | 27.5 | 46.0 | 22.8 | 39.9 |
| Ours | 48.89 | 26.65 | 48.18 | 21.99 | 41.33 |
| Model | Accuracy |
| OVR-CNN | 23.45 |
| Ours | 23.33 |
| Model | Pretrained Visual BB | Language Backbone | Accuracy |
| OVR-CNN0 | X | GloVE | 20.62±0.86 |
| Ours | X | GloVE | 21.64±1.02 |
| OVR-CNN0 | X | BERT | 22.73±0.06 |
| Ours | X | BERT | 23.74±0.48 |
| OVR-CNN0 | ✓ | BERT | 34.46±0.11 |
| Ours | ✓ | BERT | 35.49±0.21 |
| Model | P1 Obj. | P2 Obj. |
| OVR-CNN0 | 49.36 | 16.61 |
| Ours | 50.9 | 26.1 |
| Model | Accuracy |
| OVR-CNN0 | 13.10 |
| Ours | 17.45 |
| Model | Cur. | Loss | Accuracy |
| OVR-CNN0 | L | 19.73 | |
| OVR-CNN0 | LCR | 20.65 | |
| Ours | ✓ | L | 21.02 |
| Ours | ✓ | LCR | 21.58 |
| Ours | ✓ | LCP | 22.57 |
| Model | Base | Target | Generalized | ||
| Base | Target | All | |||
| OVR-CNN | 46.8 | 27.5 | 46.0 | 22.8 | 39.9 |
| Ours | 48.89 | 26.65 | 48.18 | 21.99 | 41.33 |
| Model | Accuracy |
| OVR-CNN | 23.45 |
| Ours | 23.33 |
| Tell that to the family of Margaret Hassan, the school teacher who was brutally tortured and then slaughtered by these same guys, they aren't so bad are they Chris Matthews? | AMR Path: [slaughtered, ARG1, teacher, MODIFIER, Margaret Hassan] \nPred Arg: O; Compatibility Score: -0.99 \nGold Arg: Victim; Compatibility Score: 0.99 |
| It is irritating enough to get sued by Sam Sloan; imagine how irritating it would be to get BEATEN by him because you have done something so egregious that a court is forced to agree with him. | AMR Path: [sued, ARG0, Sam Sloan] \nPred Arg: Adjudicator; Compatibility Score: -0.99 \nGold Arg: Plaintiff; Compatibility Score: 0.67 |
| Protests against the action aimed at toppling Iraqi President Saddam Hussein were held in cities across Libya, Egypt and Lebanon, as well as in Amman, Damascus and the Gaza Strip. | AMR Path: [Protests, ARG0, held, ARG1, were, PLACE, Amman] \nPred Arg: O; Compatibility Score: -0.52 \nGold Arg: Place; Compatibility Score: 0.74 |
| Meanwhile Blair arrived in Washington late Wednesday for two days of talks with Bush at the Camp David presidential retreat. | AMR Path: [arrived, OTHER, talks, PLACE, retreat] \nPred Arg: Destination; Compatibility Score: -0.94 \nGold Arg: O; Compatibility Score: 0.31 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Scoring w/o AMR | 87.4 | 85.9 | 87.9 | 86.9 | 82.8 | 83.1 |
| Scoring w/ AMR | 88.2 | 87.4 | 88.8 | 88.6 | 84.4 | 84.5 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Tri-C | Arg-C | Tri-C | Arg-C | Tri-C | Arg-C | |
| DyGIE (Wadden et al., 2019) | 69.7 | 48.8 | 67.3 | 42.7 | - | - |
| BERT_QA(Arg (Du and Cardie, 2020) | 72.4 | 53.3 | 70.6 | 48.3 | 57.0 | 39.2 |
| FourIE (Nguyen et al., 2021) | 75.4 | 58.0 | 73.3 | 57.5 | 57.9 | 48.6 |
| Text2Event (Lu et al., 2021) | 71.9 | 53.8 | 71.8 | 54.4 | 59.4 | 48.3 |
| DEGREE (Hsu et al., 2022) | 73.3 | 55.8 | 70.9 | 56.3 | 57.1 | 49.6 |
| Query Extract (Wang et al., 2022) | - | - | 73.6 | 55.1 | 60.4 | 50.4 |
| UIE (Lu et al., 2022) | 73.4 | 54.8 | - | - | - | - |
| Base OneIE (Lin et al., 2020) | 74.0 | 57.4 | 73.4 | 57.2 | 60.2 | 49.8 |
| +self-training* (Rosenberg et al., 2005) | 74.0 | 57.2 | 73.8 | 57.3 | 60.1 | 49.4 |
| +GradLRE* (Hu et al., 2021b) | 74.6 | 57.4 | 73.5 | 57.4 | 60.5 | 50.3 |
| +STFW/O_AMR | 74.4 | 57.9 | 73.8 | 57.6 | 60.4 | 51.0 |
| +STFAMR (ours) | 75.0 | 58.9 | 73.4 | 59.0 | 60.6 | 52.0 |
| Base AMR-IE (Zhang and Ji, 2021) | 74.4 | 57.7 | 73.4 | 57.2 | 60.4 | 50.5 |
| +self-training* (Rosenberg et al., 2005) | 74.2 | 57.4 | 73.4 | 57.1 | 60.1 | 50.2 |
| +GradLRE* (Hu et al., 2021b) | 74.4 | 57.8 | 73.3 | 57.4 | 60.3 | 50.5 |
| +STFW/O_AMR | 74.3 | 58.0 | 73.5 | 57.6 | 60.5 | 51.1 |
| +STFAMR (ours) | 74.5 | 58.5 | 73.6 | 58.2 | 60.4 | 51.7 |
| Sentence & Gold Event Mentions | AMR Path | Base OneIE | STF |
| With marathon talks at the top world body failing late Thursday to reconcile French and Russian opposition to US-British war plans, the United States upped its military presence, deploying more missile-firing warships to the Red Sea. (Movement:Transport) deploying -> (Agent) -> United States | ARG0\ndeploying United States | Arg Role:Other\nCompatibility\nScore: -1.0 | Arg Role:Agent\nCompatibility\nScore: 1.0 |
| In Paris, the French media group said parent company chairman Jean - Rene Fourtou will replace Diller as chairman and chief executive of US unit. (Personnel:Start-Position) replace -> (Entity) -> unit | and ARG2 Op Medium replace chairman | Arg Role:Other\nCompatibility\nScore: 0.22 | Arg Role:Entity\nCompatibility\nScore: 0.93 |
| In a verdict handed down on Saturday, the judge also ordered Ranjha to pay a fine of 50,000 rupees (about 870 US dollars), they said. (Justice:Fine) fine -> (Adjudicator) -> judge | pay ARG1 ARG2 arg0 fine ordered | Arg Role:Entity\nCompatibility\nScore: -0.98 | Arg Role:Judiciary\nCompatibility\nScore: 1.0 |
| The Daily Planet raised 3.5 million dollars (2.2 million US) in its initial public offering with one of the new 600 shareholders acquiring 1.0 million dollars worth of shares. (Transaction:Transfer-Money) acquiring -> (Other) -> shareholders | Other\nacquiring shareholders | Arg Role:Buyer\nCompatibility\nScore: -1.0 | Arg Role:Other\nCompatibility\nScore: 0.72 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Base OneIE | 70.1 | 68.4 | 76.9 | 61.9 | 76.4 | 69.2 |
| + STFAMR | 72.2 | 70.8 | 80.2 | 64.0 | 78.0 | 75.1 |
| ACE-E | ACE-E+ | ERE-EN | |
| Base OneIE | 57.4 | 57.2 | 49.8 |
| + STF w/o AMR | 57.9 | 57.6 | 51.0 |
| + STF w/ sys_AMR | 58.2 | 58.1 | 51.4 |
| + STF w/ gold_AMR | 58.9 | 59.0 | 52.0 |
| Group Label | AMR Relations |
| ARG0 | ARG0 |
| ARG1 | ARG1 |
| ARG2 | ARG2 |
| ARG3 | ARG3 |
| ARG4 | ARG4 |
| Destination | destination |
| Source | source |
| Instrument | instrument |
| Beneficiary | beneficiary |
| Prep roles | role starts with prep |
| Op roles | role start with op |
| Entity role | wiki, name |
| Arg-X role | ARG5, ARG6, ARG7 ARG8, ARG9 |
| Place role | location, path, direction |
| Medium role | manner, poss, medium, topic |
| Modifier role | domain, mod, example |
| Part-whole role | part, consist, subevent, subset |
| Time role | calendar, century, day, dayperiod, decade, era, month, quarter, season,timezone, weekday, year, year2, time |
| Others | purpose, li, quant, polarity, condition, extent, degree, cnt1, cnt2, ARG5, cnt3, concession, ord, unit, mode, value, frequency, polite, age, accompanier, cnt4, cnt10, cnt5, cnt6, cnt7, cnt8, cnt9, cnt11, scale, conj-as-if, rel |
| ACE05-E | ACE05-E+ | ERE-EN | |||||||
| Train | Dev | Test | Train | Dev | Test | Train | Dev | Test | |
| # Sent | 17,172 | 923 | 832 | 19,240 | 902 | 676 | 14,736 | 1,209 | 1,163 |
| # Entities | 29,006 | 2,451 | 3,017 | 47,525 | 3,422 | 3,673 | 38,864 | 3,320 | 3,291 |
| # Events | 4,202 | 450 | 403 | 4,419 | 468 | 424 | 6,208 | 525 | 551 |
| ACE05-E | ACE05-E+ | ERE-EN | |
| Arg-C | Arg-C | Arg-C | |
| Base OneIE | 57.4 ±1.23 | 57.2 ±0.32 | 49.8 ±0.45 |
| +STFAMR (ours) | 58.9 ±1.28 | 59.0 ±1.03 | 52.0 ±0.40 |
| Tell that to the family of Margaret Hassan, the school teacher who was brutally tortured and then slaughtered by these same guys, they aren't so bad are they Chris Matthews? | AMR Path: [slaughtered, ARG1, teacher, MODIFIER, Margaret Hassan] \nPred Arg: O; Compatibility Score: -0.99 \nGold Arg: Victim; Compatibility Score: 0.99 |
| It is irritating enough to get sued by Sam Sloan; imagine how irritating it would be to get BEATEN by him because you have done something so egregious that a court is forced to agree with him. | AMR Path: [sued, ARG0, Sam Sloan] \nPred Arg: Adjudicator; Compatibility Score: -0.99 \nGold Arg: Plaintiff; Compatibility Score: 0.67 |
| Protests against the action aimed at toppling Iraqi President Saddam Hussein were held in cities across Libya, Egypt and Lebanon, as well as in Amman, Damascus and the Gaza Strip. | AMR Path: [Protests, ARG0, held, ARG1, were, PLACE, Amman] \nPred Arg: O; Compatibility Score: -0.52 \nGold Arg: Place; Compatibility Score: 0.74 |
| Meanwhile Blair arrived in Washington late Wednesday for two days of talks with Bush at the Camp David presidential retreat. | AMR Path: [arrived, OTHER, talks, PLACE, retreat] \nPred Arg: Destination; Compatibility Score: -0.94 \nGold Arg: O; Compatibility Score: 0.31 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Scoring w/o AMR | 87.4 | 85.9 | 87.9 | 86.9 | 82.8 | 83.1 |
| Scoring w/ AMR | 88.2 | 87.4 | 88.8 | 88.6 | 84.4 | 84.5 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Tri-C | Arg-C | Tri-C | Arg-C | Tri-C | Arg-C | |
| DyGIE (Wadden et al., 2019) | 69.7 | 48.8 | 67.3 | 42.7 | - | - |
| BERT_QA(Arg (Du and Cardie, 2020) | 72.4 | 53.3 | 70.6 | 48.3 | 57.0 | 39.2 |
| FourIE (Nguyen et al., 2021) | 75.4 | 58.0 | 73.3 | 57.5 | 57.9 | 48.6 |
| Text2Event (Lu et al., 2021) | 71.9 | 53.8 | 71.8 | 54.4 | 59.4 | 48.3 |
| DEGREE (Hsu et al., 2022) | 73.3 | 55.8 | 70.9 | 56.3 | 57.1 | 49.6 |
| Query Extract (Wang et al., 2022) | - | - | 73.6 | 55.1 | 60.4 | 50.4 |
| UIE (Lu et al., 2022) | 73.4 | 54.8 | - | - | - | - |
| Base OneIE (Lin et al., 2020) | 74.0 | 57.4 | 73.4 | 57.2 | 60.2 | 49.8 |
| +self-training* (Rosenberg et al., 2005) | 74.0 | 57.2 | 73.8 | 57.3 | 60.1 | 49.4 |
| +GradLRE* (Hu et al., 2021b) | 74.6 | 57.4 | 73.5 | 57.4 | 60.5 | 50.3 |
| +STFW/O_AMR | 74.4 | 57.9 | 73.8 | 57.6 | 60.4 | 51.0 |
| +STFAMR (ours) | 75.0 | 58.9 | 73.4 | 59.0 | 60.6 | 52.0 |
| Base AMR-IE (Zhang and Ji, 2021) | 74.4 | 57.7 | 73.4 | 57.2 | 60.4 | 50.5 |
| +self-training* (Rosenberg et al., 2005) | 74.2 | 57.4 | 73.4 | 57.1 | 60.1 | 50.2 |
| +GradLRE* (Hu et al., 2021b) | 74.4 | 57.8 | 73.3 | 57.4 | 60.3 | 50.5 |
| +STFW/O_AMR | 74.3 | 58.0 | 73.5 | 57.6 | 60.5 | 51.1 |
| +STFAMR (ours) | 74.5 | 58.5 | 73.6 | 58.2 | 60.4 | 51.7 |
| Sentence & Gold Event Mentions | AMR Path | Base OneIE | STF |
| With marathon talks at the top world body failing late Thursday to reconcile French and Russian opposition to US-British war plans, the United States upped its military presence, deploying more missile-firing warships to the Red Sea. (Movement:Transport) deploying -> (Agent) -> United States | ARG0\ndeploying United States | Arg Role:Other\nCompatibility\nScore: -1.0 | Arg Role:Agent\nCompatibility\nScore: 1.0 |
| In Paris, the French media group said parent company chairman Jean - Rene Fourtou will replace Diller as chairman and chief executive of US unit. (Personnel:Start-Position) replace -> (Entity) -> unit | and ARG2 Op Medium replace chairman | Arg Role:Other\nCompatibility\nScore: 0.22 | Arg Role:Entity\nCompatibility\nScore: 0.93 |
| In a verdict handed down on Saturday, the judge also ordered Ranjha to pay a fine of 50,000 rupees (about 870 US dollars), they said. (Justice:Fine) fine -> (Adjudicator) -> judge | pay ARG1 ARG2 arg0 fine ordered | Arg Role:Entity\nCompatibility\nScore: -0.98 | Arg Role:Judiciary\nCompatibility\nScore: 1.0 |
| The Daily Planet raised 3.5 million dollars (2.2 million US) in its initial public offering with one of the new 600 shareholders acquiring 1.0 million dollars worth of shares. (Transaction:Transfer-Money) acquiring -> (Other) -> shareholders | Other\nacquiring shareholders | Arg Role:Buyer\nCompatibility\nScore: -1.0 | Arg Role:Other\nCompatibility\nScore: 0.72 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Base OneIE | 70.1 | 68.4 | 76.9 | 61.9 | 76.4 | 69.2 |
| + STFAMR | 72.2 | 70.8 | 80.2 | 64.0 | 78.0 | 75.1 |
| ACE-E | ACE-E+ | ERE-EN | |
| Base OneIE | 57.4 | 57.2 | 49.8 |
| + STF w/o AMR | 57.9 | 57.6 | 51.0 |
| + STF w/ sys_AMR | 58.2 | 58.1 | 51.4 |
| + STF w/ gold_AMR | 58.9 | 59.0 | 52.0 |
| Group Label | AMR Relations |
| ARG0 | ARG0 |
| ARG1 | ARG1 |
| ARG2 | ARG2 |
| ARG3 | ARG3 |
| ARG4 | ARG4 |
| Destination | destination |
| Source | source |
| Instrument | instrument |
| Beneficiary | beneficiary |
| Prep roles | role starts with prep |
| Op roles | role start with op |
| Entity role | wiki, name |
| Arg-X role | ARG5, ARG6, ARG7 ARG8, ARG9 |
| Place role | location, path, direction |
| Medium role | manner, poss, medium, topic |
| Modifier role | domain, mod, example |
| Part-whole role | part, consist, subevent, subset |
| Time role | calendar, century, day, dayperiod, decade, era, month, quarter, season,timezone, weekday, year, year2, time |
| Others | purpose, li, quant, polarity, condition, extent, degree, cnt1, cnt2, ARG5, cnt3, concession, ord, unit, mode, value, frequency, polite, age, accompanier, cnt4, cnt10, cnt5, cnt6, cnt7, cnt8, cnt9, cnt11, scale, conj-as-if, rel |
| ACE05-E | ACE05-E+ | ERE-EN | |||||||
| Train | Dev | Test | Train | Dev | Test | Train | Dev | Test | |
| # Sent | 17,172 | 923 | 832 | 19,240 | 902 | 676 | 14,736 | 1,209 | 1,163 |
| # Entities | 29,006 | 2,451 | 3,017 | 47,525 | 3,422 | 3,673 | 38,864 | 3,320 | 3,291 |
| # Events | 4,202 | 450 | 403 | 4,419 | 468 | 424 | 6,208 | 525 | 551 |
| ACE05-E | ACE05-E+ | ERE-EN | |
| Arg-C | Arg-C | Arg-C | |
| Base OneIE | 57.4 ±1.23 | 57.2 ±0.32 | 49.8 ±0.45 |
| +STFAMR (ours) | 58.9 ±1.28 | 59.0 ±1.03 | 52.0 ±0.40 |
| Tell that to the family of Margaret Hassan, the school teacher who was brutally tortured and then slaughtered by these same guys, they aren't so bad are they Chris Matthews? | AMR Path: [slaughtered, ARG1, teacher, MODIFIER, Margaret Hassan] +Pred Arg: O; Compatibility Score: -0.99 +Gold Arg: Victim; Compatibility Score: 0.99 |
| It is irritating enough to get sued by Sam Sloan; imagine how irritating it would be to get BEATEN by him because you have done something so egregious that a court is forced to agree with him. | AMR Path: [sued, ARG0, Sam Sloan] +Pred Arg: Adjudicator; Compatibility Score: -0.99 +Gold Arg: Plaintiff; Compatibility Score: 0.67 |
| Protests against the action aimed at toppling Iraqi President Saddam Hussein were held in cities across Libya, Egypt and Lebanon, as well as in Amman, Damascus and the Gaza Strip. | AMR Path: [Protests, ARG0, held, ARG1, were, PLACE, Amman] +Pred Arg: O; Compatibility Score: -0.52 +Gold Arg: Place; Compatibility Score: 0.74 |
| Meanwhile Blair arrived in Washington late Wednesday for two days of talks with Bush at the Camp David presidential retreat. | AMR Path: [arrived, OTHER, talks, PLACE, retreat] +Pred Arg: Destination; Compatibility Score: -0.94 +Gold Arg: O; Compatibility Score: 0.31 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Scoring w/o AMR | 87.4 | 85.9 | 87.9 | 86.9 | 82.8 | 83.1 |
| Scoring w/ AMR | 88.2 | 87.4 | 88.8 | 88.6 | 84.4 | 84.5 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Tri-C | Arg-C | Tri-C | Arg-C | Tri-C | Arg-C | |
| DyGIE (Wadden et al., 2019) | 69.7 | 48.8 | 67.3 | 42.7 | - | - |
| BERT_QA(Arg (Du and Cardie, 2020) | 72.4 | 53.3 | 70.6 | 48.3 | 57.0 | 39.2 |
| FourIE (Nguyen et al., 2021) | 75.4 | 58.0 | 73.3 | 57.5 | 57.9 | 48.6 |
| Text2Event (Lu et al., 2021) | 71.9 | 53.8 | 71.8 | 54.4 | 59.4 | 48.3 |
| DEGREE (Hsu et al., 2022) | 73.3 | 55.8 | 70.9 | 56.3 | 57.1 | 49.6 |
| Query Extract (Wang et al., 2022) | - | - | 73.6 | 55.1 | 60.4 | 50.4 |
| UIE (Lu et al., 2022) | 73.4 | 54.8 | - | - | - | - |
| Base OneIE (Lin et al., 2020) | 74.0 | 57.4 | 73.4 | 57.2 | 60.2 | 49.8 |
| +self-training* (Rosenberg et al., 2005) | 74.0 | 57.2 | 73.8 | 57.3 | 60.1 | 49.4 |
| +GradLRE* (Hu et al., 2021b) | 74.6 | 57.4 | 73.5 | 57.4 | 60.5 | 50.3 |
| +STFW/O_AMR | 74.4 | 57.9 | 73.8 | 57.6 | 60.4 | 51.0 |
| +STFAMR (ours) | 75.0 | 58.9 | 73.4 | 59.0 | 60.6 | 52.0 |
| Base AMR-IE (Zhang and Ji, 2021) | 74.4 | 57.7 | 73.4 | 57.2 | 60.4 | 50.5 |
| +self-training* (Rosenberg et al., 2005) | 74.2 | 57.4 | 73.4 | 57.1 | 60.1 | 50.2 |
| +GradLRE* (Hu et al., 2021b) | 74.4 | 57.8 | 73.3 | 57.4 | 60.3 | 50.5 |
| +STFW/O_AMR | 74.3 | 58.0 | 73.5 | 57.6 | 60.5 | 51.1 |
| +STFAMR (ours) | 74.5 | 58.5 | 73.6 | 58.2 | 60.4 | 51.7 |
| Sentence & Gold Event Mentions | AMR Path | Base OneIE | STF |
| With marathon talks at the top world body failing late Thursday to reconcile French and Russian opposition to US-British war plans, the United States upped its military presence, deploying more missile-firing warships to the Red Sea. (Movement:Transport) deploying -> (Agent) -> United States | ARG0 +deploying United States | Arg Role:Other +Compatibility +Score: -1.0 | Arg Role:Agent +Compatibility +Score: 1.0 |
| In Paris, the French media group said parent company chairman Jean - Rene Fourtou will replace Diller as chairman and chief executive of US unit. (Personnel:Start-Position) replace -> (Entity) -> unit | and ARG2 Op Medium replace chairman | Arg Role:Other +Compatibility +Score: 0.22 | Arg Role:Entity +Compatibility +Score: 0.93 |
| In a verdict handed down on Saturday, the judge also ordered Ranjha to pay a fine of 50,000 rupees (about 870 US dollars), they said. (Justice:Fine) fine -> (Adjudicator) -> judge | pay ARG1 ARG2 arg0 fine ordered | Arg Role:Entity +Compatibility +Score: -0.98 | Arg Role:Judiciary +Compatibility +Score: 1.0 |
| The Daily Planet raised 3.5 million dollars (2.2 million US) in its initial public offering with one of the new 600 shareholders acquiring 1.0 million dollars worth of shares. (Transaction:Transfer-Money) acquiring -> (Other) -> shareholders | Other +acquiring shareholders | Arg Role:Buyer +Compatibility +Score: -1.0 | Arg Role:Other +Compatibility +Score: 0.72 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Base OneIE | 70.1 | 68.4 | 76.9 | 61.9 | 76.4 | 69.2 |
| + STFAMR | 72.2 | 70.8 | 80.2 | 64.0 | 78.0 | 75.1 |
| ACE-E | ACE-E+ | ERE-EN | |
| Base OneIE | 57.4 | 57.2 | 49.8 |
| + STF w/o AMR | 57.9 | 57.6 | 51.0 |
| + STF w/ sys_AMR | 58.2 | 58.1 | 51.4 |
| + STF w/ gold_AMR | 58.9 | 59.0 | 52.0 |
| Group Label | AMR Relations |
| ARG0 | ARG0 |
| ARG1 | ARG1 |
| ARG2 | ARG2 |
| ARG3 | ARG3 |
| ARG4 | ARG4 |
| Destination | destination |
| Source | source |
| Instrument | instrument |
| Beneficiary | beneficiary |
| Prep roles | role starts with prep |
| Op roles | role start with op |
| Entity role | wiki, name |
| Arg-X role | ARG5, ARG6, ARG7 ARG8, ARG9 |
| Place role | location, path, direction |
| Medium role | manner, poss, medium, topic |
| Modifier role | domain, mod, example |
| Part-whole role | part, consist, subevent, subset |
| Time role | calendar, century, day, dayperiod, decade, era, month, quarter, season,timezone, weekday, year, year2, time |
| Others | purpose, li, quant, polarity, condition, extent, degree, cnt1, cnt2, ARG5, cnt3, concession, ord, unit, mode, value, frequency, polite, age, accompanier, cnt4, cnt10, cnt5, cnt6, cnt7, cnt8, cnt9, cnt11, scale, conj-as-if, rel |
| ACE05-E | ACE05-E+ | ERE-EN | |||||||
| Train | Dev | Test | Train | Dev | Test | Train | Dev | Test | |
| # Sent | 17,172 | 923 | 832 | 19,240 | 902 | 676 | 14,736 | 1,209 | 1,163 |
| # Entities | 29,006 | 2,451 | 3,017 | 47,525 | 3,422 | 3,673 | 38,864 | 3,320 | 3,291 |
| # Events | 4,202 | 450 | 403 | 4,419 | 468 | 424 | 6,208 | 525 | 551 |
| ACE05-E | ACE05-E+ | ERE-EN | |
| Arg-C | Arg-C | Arg-C | |
| Base OneIE | 57.4 ±1.23 | 57.2 ±0.32 | 49.8 ±0.45 |
| +STFAMR (ours) | 58.9 ±1.28 | 59.0 ±1.03 | 52.0 ±0.40 |
| Tell that to the family of Margaret Hassan, the school teacher who was brutally tortured and then slaughtered by these same guys, they aren't so bad are they Chris Matthews? | AMR Path: [slaughtered, ARG1, teacher, MODIFIER, Margaret Hassan] \nPred Arg: O; Compatibility Score: -0.99 \nGold Arg: Victim; Compatibility Score: 0.99 |
| It is irritating enough to get sued by Sam Sloan; imagine how irritating it would be to get BEATEN by him because you have done something so egregious that a court is forced to agree with him. | AMR Path: [sued, ARG0, Sam Sloan] \nPred Arg: Adjudicator; Compatibility Score: -0.99 \nGold Arg: Plaintiff; Compatibility Score: 0.67 |
| Protests against the action aimed at toppling Iraqi President Saddam Hussein were held in cities across Libya, Egypt and Lebanon, as well as in Amman, Damascus and the Gaza Strip. | AMR Path: [Protests, ARG0, held, ARG1, were, PLACE, Amman] \nPred Arg: O; Compatibility Score: -0.52 \nGold Arg: Place; Compatibility Score: 0.74 |
| Meanwhile Blair arrived in Washington late Wednesday for two days of talks with Bush at the Camp David presidential retreat. | AMR Path: [arrived, OTHER, talks, PLACE, retreat] \nPred Arg: Destination; Compatibility Score: -0.94 \nGold Arg: O; Compatibility Score: 0.31 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Scoring w/o AMR | 87.4 | 85.9 | 87.9 | 86.9 | 82.8 | 83.1 |
| Scoring w/ AMR | 88.2 | 87.4 | 88.8 | 88.6 | 84.4 | 84.5 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Tri-C | Arg-C | Tri-C | Arg-C | Tri-C | Arg-C | |
| DyGIE (Wadden et al., 2019) | 69.7 | 48.8 | 67.3 | 42.7 | - | - |
| BERT_QA(Arg (Du and Cardie, 2020) | 72.4 | 53.3 | 70.6 | 48.3 | 57.0 | 39.2 |
| FourIE (Nguyen et al., 2021) | 75.4 | 58.0 | 73.3 | 57.5 | 57.9 | 48.6 |
| Text2Event (Lu et al., 2021) | 71.9 | 53.8 | 71.8 | 54.4 | 59.4 | 48.3 |
| DEGREE (Hsu et al., 2022) | 73.3 | 55.8 | 70.9 | 56.3 | 57.1 | 49.6 |
| Query Extract (Wang et al., 2022) | - | - | 73.6 | 55.1 | 60.4 | 50.4 |
| UIE (Lu et al., 2022) | 73.4 | 54.8 | - | - | - | - |
| Base OneIE (Lin et al., 2020) | 74.0 | 57.4 | 73.4 | 57.2 | 60.2 | 49.8 |
| +self-training* (Rosenberg et al., 2005) | 74.0 | 57.2 | 73.8 | 57.3 | 60.1 | 49.4 |
| +GradLRE* (Hu et al., 2021b) | 74.6 | 57.4 | 73.5 | 57.4 | 60.5 | 50.3 |
| +STFW/O_AMR | 74.4 | 57.9 | 73.8 | 57.6 | 60.4 | 51.0 |
| +STFAMR (ours) | 75.0 | 58.9 | 73.4 | 59.0 | 60.6 | 52.0 |
| Base AMR-IE (Zhang and Ji, 2021) | 74.4 | 57.7 | 73.4 | 57.2 | 60.4 | 50.5 |
| +self-training* (Rosenberg et al., 2005) | 74.2 | 57.4 | 73.4 | 57.1 | 60.1 | 50.2 |
| +GradLRE* (Hu et al., 2021b) | 74.4 | 57.8 | 73.3 | 57.4 | 60.3 | 50.5 |
| +STFW/O_AMR | 74.3 | 58.0 | 73.5 | 57.6 | 60.5 | 51.1 |
| +STFAMR (ours) | 74.5 | 58.5 | 73.6 | 58.2 | 60.4 | 51.7 |
| Sentence & Gold Event Mentions | AMR Path | Base OneIE | STF |
| With marathon talks at the top world body failing late Thursday to reconcile French and Russian opposition to US-British war plans, the United States upped its military presence, deploying more missile-firing warships to the Red Sea. (Movement:Transport) deploying -> (Agent) -> United States | ARG0\ndeploying United States | Arg Role:Other\nCompatibility\nScore: -1.0 | Arg Role:Agent\nCompatibility\nScore: 1.0 |
| In Paris, the French media group said parent company chairman Jean - Rene Fourtou will replace Diller as chairman and chief executive of US unit. (Personnel:Start-Position) replace -> (Entity) -> unit | and ARG2 Op Medium replace chairman | Arg Role:Other\nCompatibility\nScore: 0.22 | Arg Role:Entity\nCompatibility\nScore: 0.93 |
| In a verdict handed down on Saturday, the judge also ordered Ranjha to pay a fine of 50,000 rupees (about 870 US dollars), they said. (Justice:Fine) fine -> (Adjudicator) -> judge | pay ARG1 ARG2 arg0 fine ordered | Arg Role:Entity\nCompatibility\nScore: -0.98 | Arg Role:Judiciary\nCompatibility\nScore: 1.0 |
| The Daily Planet raised 3.5 million dollars (2.2 million US) in its initial public offering with one of the new 600 shareholders acquiring 1.0 million dollars worth of shares. (Transaction:Transfer-Money) acquiring -> (Other) -> shareholders | Other\nacquiring shareholders | Arg Role:Buyer\nCompatibility\nScore: -1.0 | Arg Role:Other\nCompatibility\nScore: 0.72 |
| ACE05-E | ACE05-E+ | ERE-EN | ||||
| Dev | Test | Dev | Test | Dev | Test | |
| Base OneIE | 70.1 | 68.4 | 76.9 | 61.9 | 76.4 | 69.2 |
| + STFAMR | 72.2 | 70.8 | 80.2 | 64.0 | 78.0 | 75.1 |
| ACE-E | ACE-E+ | ERE-EN | |
| Base OneIE | 57.4 | 57.2 | 49.8 |
| + STF w/o AMR | 57.9 | 57.6 | 51.0 |
| + STF w/ sys_AMR | 58.2 | 58.1 | 51.4 |
| + STF w/ gold_AMR | 58.9 | 59.0 | 52.0 |
| Group Label | AMR Relations |
| ARG0 | ARG0 |
| ARG1 | ARG1 |
| ARG2 | ARG2 |
| ARG3 | ARG3 |
| ARG4 | ARG4 |
| Destination | destination |
| Source | source |
| Instrument | instrument |
| Beneficiary | beneficiary |
| Prep roles | role starts with prep |
| Op roles | role start with op |
| Entity role | wiki, name |
| Arg-X role | ARG5, ARG6, ARG7 ARG8, ARG9 |
| Place role | location, path, direction |
| Medium role | manner, poss, medium, topic |
| Modifier role | domain, mod, example |
| Part-whole role | part, consist, subevent, subset |
| Time role | calendar, century, day, dayperiod, decade, era, month, quarter, season,timezone, weekday, year, year2, time |
| Others | purpose, li, quant, polarity, condition, extent, degree, cnt1, cnt2, ARG5, cnt3, concession, ord, unit, mode, value, frequency, polite, age, accompanier, cnt4, cnt10, cnt5, cnt6, cnt7, cnt8, cnt9, cnt11, scale, conj-as-if, rel |
| ACE05-E | ACE05-E+ | ERE-EN | |||||||
| Train | Dev | Test | Train | Dev | Test | Train | Dev | Test | |
| # Sent | 17,172 | 923 | 832 | 19,240 | 902 | 676 | 14,736 | 1,209 | 1,163 |
| # Entities | 29,006 | 2,451 | 3,017 | 47,525 | 3,422 | 3,673 | 38,864 | 3,320 | 3,291 |
| # Events | 4,202 | 450 | 403 | 4,419 | 468 | 424 | 6,208 | 525 | 551 |
| ACE05-E | ACE05-E+ | ERE-EN | |
| Arg-C | Arg-C | Arg-C | |
| Base OneIE | 57.4 ±1.23 | 57.2 ±0.32 | 49.8 ±0.45 |
| +STFAMR (ours) | 58.9 ±1.28 | 59.0 ±1.03 | 52.0 ±0.40 |
| BART | RoBERTa | ||||||||||||||
| Train/Test | Consumer | Industrial | Technology | Train/Test | Consumer | Industrial | Technology | ||||||||
| F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | ||||
| Con | Ori | 70.29(+0.36) | 24.53(+0.47) | 68.44(+1.7) | 24.07(+2.78) | 68.61(+3.39) | 23.61(+4.63) | Con | Ori | 78.20(+0.19) | 51.38(+1.86) | 72.49(+2.58) | 47.68(+0.93) | 74.63(+2.83) | 49.07(+0.47) |
| Ours | Ours | ||||||||||||||
| Ind | Ori | 70.11(+3.95) | 31.31(+4.75) | 72.53(+4.05) | 32.41(+7.48) | 69.63(+3.84) | 27.27(+9.93) | Ind | Ori | 77.81(+2.74) | 49.45(+9.57) | 80.05(+0.65) | 58.46(+2.20) | 77.91(+1.09) | 48.35(+9.02) |
| Ours | Ours | ||||||||||||||
| Tech | Ori | 69.89(+2.78) | 30.30(+3.03) | 69.53(+2.98) | 27.77(+3.45) | 71.79(+2.23) | 33.83(+0.51) | Tech | Ori | 75.54(+2.39) | 55.05(+0.00) | 73.99(+2.71) | 48.98(+7.08) | 76.49(+0.27) | 54.04(-0.50) |
| Ours | Ours |
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| BART | 74.16 | 36.50 | 73.26 | 34.49 |
| BART + LP-FT | 74.06 | 35.03 | 73.83 | 34/00 |
| BART + P | 75.60 | 37.47 | 75.33 | 37.66 |
| BART + P + LP-FT | 77.60 | 41.22 | 76.90 | 39.44 |
| RoBERTa | 83.97 | 61.82 | 83.45 | 61.28 |
| RoBERTa + LP-FT | 84.80 | 62.15 | 83.49 | 61.18 |
| RoBERTa + P | 84.55 | 62.20 | 83.61 | 61.34 |
| RoBERTa + P + LP-FT | 84.56 | 62.15 | 83.87 | 61.42 |
| S->N | S->C | N->S | N->C | C->S | C->N | |
| RoBERTa | 37.60 | 66.58 | 49.87 | 44.22 | 19.44 | 7.45 |
| RoBERTa+DA (Yue et al., 2022) | 38.26 | 66.14 | 50.31 | 43.05 | 22.74 | 7.15 |
| RoBERTa+P | 38.17* | 66.84 | 50.97* | 48.37* | 21.41* | 8.64* |
| RoBERTa+LPFT | 37.95* | 66.60 | 50.28* | 45.86* | 20.92* | 7.5 |
| RoBERTa+P+LPFT | 38.76* | 66.86* | 52.41* | 51.64* | 23.02* | 7.73* |
| BART | 33.71 | 46.97 | 43.49 | 31.78 | 26.14 | 8.69 |
| BART+DA (Yue et al., 2022) | 35.09 | 55.65 | 44.05 | 33.47 | 26.98 | 9.02 |
| BART+P | 36.81* | 56.22* | 43.61* | 31.91 | 25.96 | 9.26* |
| BART+LPFT | 33.29 | 53.29* | 44.05* | 31.95 | 26.87* | 9.49* |
| BART+P+LPFT | 35.23* | 55.86* | 44.36* | 33.79* | 27.61* | 9.54* |
| Context & Prompt | Question | Gold Answer | Baseline Output | Our Output |
| Predictive Model: SpanBERT-base | ||||
| As Terex has expanded its MP product line, it has captured a larger global market share of the industry, allowing it to gain greater insight into customer demand. This may provide revenue synergies in the future......\nPrompt: To answer a "Why useful" question, you need to look for "allowing" | Why Terex has expanded its MP product line is useful? | gain greater insight into customer | it has captured a larger global market share of | it has captured a larger global market share of the industry, allowing it to gain greater insight into customer |
| However, Avnet's management conceded...... Suppliers get access to a broad range of customers by using Avnet without having to make significant investment in sales and engineering teams. In exchange for these services, Avnet can generate ...... Prompt: The sentiment of this question is positive, you need to look for "significant" | What will be prevented if suppliers get access to a broad range of customers? | investment in sales and engineering | using Avnet | using Avnet without having to make significant investment in sales and engineering |
| Generative Model: BART-base | ||||
| At the end of 2020, the store base had grown about 29% over the prior five-year period, to about 1,920 locations (around 2,100 including Petsense), driving sales and EPS compound annual growth rates over the past three years of 14% and 27%, respectively. We forecast that the firm will grow to around ...... Prompt: The entity "EPS" is mentioned in the question. This timing "annual" is mentioned in the question. | Why sales and EPS compound annual growth rates increase? | the store base had grown about 29% over the prior five-year period | 14% and 27% | the store base had grown about 29% over the prior five years |
| Finally, we view the likelihood of sustained economic value creation as quite high for the restaurant brand, which finds itself on the leading edge of most of the changes we expect to impact the restaurant industry over the medium to long term. Though Chipotle saw economic value destruction in 201 ...... Prompt: "restaurant industry" is an important phrase. And also pay attention to these words:"edge", "changes" | What will happen if on the leading edge of most of the changes we expect to impact the restaurant industry? | the likelihood of sustained economic value creation as quite high for the restaurant brand | over the medium to long term | we view the likelihood of sustained economic value creation as quite high for the restaurant brand |
| Template | F1 | |
| Baseline | None | 70.29 |
| Experiment1 | "There is no important phrase in this query.\nBut also pay attention to these words: _” | 69.57 |
| Experiment2 | "There is no important phrase in this query.\nAnd also pay attention to these words: _” | 70.84 |
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| SpanBERT-base | 84.77 | 62.62 | 84.85 | 64.04 |
| SpanBERT-large | 85.40 | 61.41 | 85.53 | 62.26 |
| Domain | Con | Ind | Tech | |||
| F1 | EM | F1 | EM | F1 | EM | |
| Con | 85.84 | 60.64 | 84.80 | 56.01 | 85.54 | 55.09 |
| Ind | 85.76 | 66.66 | 85.84 | 67.21 | 85.34 | 65.57 |
| Tech | 80.24 | 58.08 | 81.51 | 58.58 | 81.98 | 58.08 |
| Train/Test | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Con | Ori | 59.30 | 15.74 | 56.69 | 18.60 | 56.13 | 17.67 |
| Ours | (+2.19) | (+3.87) | (+1.79) | (+0.84) | (+2.24) | (+1.77) | |
| Ind | Ori | 59.85 | 24.72 | 61.41 | 24.03 | 59.94 | 25.27 |
| Ours | (+2.52) | (-0.68) | (+0.06) | (-3.81) | (+1.64) | (-4.51) | |
| Tech | Ori | 55.30 | 17.85 | 54.41 | 16.83 | 58.76 | 20.20 |
| Ours | (+3.70) | (+5.38) | (+4.83) | (+5.39) | (+0.53) | (+2.02) |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: RoBERTa | Con | 78.20 | 51.38 | 72.49 | 47.68 | 74.63 | 49.07 |
| Ind | 77.81 | 49.45 | 80.05 | 58.46 | 77.91 | 48.35 | |
| Tec | 75.54 | 55.05 | 73.99 | 48.98 | 76.49 | 54.04 | |
| RoBERTa + LP-FT | Con | 75.09 | 47.68 | 73.82 | 46.75 | 75.72 | 50.00 |
| Ind | 78.81 | 48.90 | 81.01* | 51.10 | 76.71 | 50.00 | |
| Tec | 79.23 | 56.06* | 77.50 | 55.05 | 78.22 | 53.53 | |
| RoBERTa + Prompt | Con | 78.97* | 53.24 | 77.09* | 50.00* | 77.93* | 51.85* |
| Ind | 78.45 | 56.28 | 80.05 | 61.74* | 80.05* | 57.92* | |
| Tec | 77.46 | 54.54 | 78.50* | 57.58* | 80.62* | 58.08* | |
| RoBERTa + LP-FT + Prompt | Con | 78.39 | 53.24* | 75.07 | 48.61 | 77.46 | 49.54 |
| Ind | 80.55* | 59.02* | 80.70 | 60.66 | 79.00 | 57.37 | |
| Tec | 77.93* | 55.05 | 76.70 | 56.06 | 76.76 | 53.54 |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: BART | Con | 70.29 | 24.53 | 68.44 | 24.07 | 68.61 | 23.61 |
| Ind | 70.11 | 31.31 | 72.53 | 32.41 | 69.53 | 27.77 | |
| Tec | 69.89 | 30.30 | 69.53 | 27.77 | 71.79 | 33.83 | |
| BART + LP-FT | Con | 70.81* | 24.07 | 70.22* | 24.53 | 68.26 | 23.61 |
| Ind | 70.61 | 30.77 | 73.39 | 31.87 | 70.91 | 28.57 | |
| Tec | 71.08 | 35.35 | 72.31 | 30.30 | 72.37 | 33.33 | |
| BART + Prompt | Con | 70.41 | 27.31* | 69.10 | 23.15 | 69.20 | 23.61 |
| Ind | 74.42* | 42.07* | 76.27 | 37.16 | 73.00 | 34.43 | |
| Tec | 72.20 | 32.32 | 72.51 | 30.3 | 73.32 | 35.35* | |
| BART + LP-FT + Prompt | Con | 70.65 | 25.01 | 70.14 | 26.85* | 72.00* | 28.24* |
| Ind | 74.06 | 36.06 | 76.58* | 39.89* | 73.37* | 37.70* | |
| Tec | 72.67* | 33.33 | 72.96* | 31.31* | 74.02* | 34.34 |
| BART | RoBERTa | ||||||||||||||
| Train/Test | Consumer | Industrial | Technology | Train/Test | Consumer | Industrial | Technology | ||||||||
| F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | ||||
| Con | Ori | 70.29(+0.36) | 24.53(+0.47) | 68.44(+1.7) | 24.07(+2.78) | 68.61(+3.39) | 23.61(+4.63) | Con | Ori | 78.20(+0.19) | 51.38(+1.86) | 72.49(+2.58) | 47.68(+0.93) | 74.63(+2.83) | 49.07(+0.47) |
| Ours | Ours | ||||||||||||||
| Ind | Ori | 70.11(+3.95) | 31.31(+4.75) | 72.53(+4.05) | 32.41(+7.48) | 69.63(+3.84) | 27.27(+9.93) | Ind | Ori | 77.81(+2.74) | 49.45(+9.57) | 80.05(+0.65) | 58.46(+2.20) | 77.91(+1.09) | 48.35(+9.02) |
| Ours | Ours | ||||||||||||||
| Tech | Ori | 69.89(+2.78) | 30.30(+3.03) | 69.53(+2.98) | 27.77(+3.45) | 71.79(+2.23) | 33.83(+0.51) | Tech | Ori | 75.54(+2.39) | 55.05(+0.00) | 73.99(+2.71) | 48.98(+7.08) | 76.49(+0.27) | 54.04(-0.50) |
| Ours | Ours |
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| BART | 74.16 | 36.50 | 73.26 | 34.49 |
| BART + LP-FT | 74.06 | 35.03 | 73.83 | 34/00 |
| BART + P | 75.60 | 37.47 | 75.33 | 37.66 |
| BART + P + LP-FT | 77.60 | 41.22 | 76.90 | 39.44 |
| RoBERTa | 83.97 | 61.82 | 83.45 | 61.28 |
| RoBERTa + LP-FT | 84.80 | 62.15 | 83.49 | 61.18 |
| RoBERTa + P | 84.55 | 62.20 | 83.61 | 61.34 |
| RoBERTa + P + LP-FT | 84.56 | 62.15 | 83.87 | 61.42 |
| S->N | S->C | N->S | N->C | C->S | C->N | |
| RoBERTa | 37.60 | 66.58 | 49.87 | 44.22 | 19.44 | 7.45 |
| RoBERTa+DA (Yue et al., 2022) | 38.26 | 66.14 | 50.31 | 43.05 | 22.74 | 7.15 |
| RoBERTa+P | 38.17* | 66.84 | 50.97* | 48.37* | 21.41* | 8.64* |
| RoBERTa+LPFT | 37.95* | 66.60 | 50.28* | 45.86* | 20.92* | 7.5 |
| RoBERTa+P+LPFT | 38.76* | 66.86* | 52.41* | 51.64* | 23.02* | 7.73* |
| BART | 33.71 | 46.97 | 43.49 | 31.78 | 26.14 | 8.69 |
| BART+DA (Yue et al., 2022) | 35.09 | 55.65 | 44.05 | 33.47 | 26.98 | 9.02 |
| BART+P | 36.81* | 56.22* | 43.61* | 31.91 | 25.96 | 9.26* |
| BART+LPFT | 33.29 | 53.29* | 44.05* | 31.95 | 26.87* | 9.49* |
| BART+P+LPFT | 35.23* | 55.86* | 44.36* | 33.79* | 27.61* | 9.54* |
| Context & Prompt | Question | Gold Answer | Baseline Output | Our Output |
| Predictive Model: SpanBERT-base | ||||
| As Terex has expanded its MP product line, it has captured a larger global market share of the industry, allowing it to gain greater insight into customer demand. This may provide revenue synergies in the future......\nPrompt: To answer a "Why useful" question, you need to look for "allowing" | Why Terex has expanded its MP product line is useful? | gain greater insight into customer | it has captured a larger global market share of | it has captured a larger global market share of the industry, allowing it to gain greater insight into customer |
| However, Avnet's management conceded...... Suppliers get access to a broad range of customers by using Avnet without having to make significant investment in sales and engineering teams. In exchange for these services, Avnet can generate ...... Prompt: The sentiment of this question is positive, you need to look for "significant" | What will be prevented if suppliers get access to a broad range of customers? | investment in sales and engineering | using Avnet | using Avnet without having to make significant investment in sales and engineering |
| Generative Model: BART-base | ||||
| At the end of 2020, the store base had grown about 29% over the prior five-year period, to about 1,920 locations (around 2,100 including Petsense), driving sales and EPS compound annual growth rates over the past three years of 14% and 27%, respectively. We forecast that the firm will grow to around ...... Prompt: The entity "EPS" is mentioned in the question. This timing "annual" is mentioned in the question. | Why sales and EPS compound annual growth rates increase? | the store base had grown about 29% over the prior five-year period | 14% and 27% | the store base had grown about 29% over the prior five years |
| Finally, we view the likelihood of sustained economic value creation as quite high for the restaurant brand, which finds itself on the leading edge of most of the changes we expect to impact the restaurant industry over the medium to long term. Though Chipotle saw economic value destruction in 201 ...... Prompt: "restaurant industry" is an important phrase. And also pay attention to these words:"edge", "changes" | What will happen if on the leading edge of most of the changes we expect to impact the restaurant industry? | the likelihood of sustained economic value creation as quite high for the restaurant brand | over the medium to long term | we view the likelihood of sustained economic value creation as quite high for the restaurant brand |
| Template | F1 | |
| Baseline | None | 70.29 |
| Experiment1 | "There is no important phrase in this query.\nBut also pay attention to these words: _” | 69.57 |
| Experiment2 | "There is no important phrase in this query.\nAnd also pay attention to these words: _” | 70.84 |
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| SpanBERT-base | 84.77 | 62.62 | 84.85 | 64.04 |
| SpanBERT-large | 85.40 | 61.41 | 85.53 | 62.26 |
| Domain | Con | Ind | Tech | |||
| F1 | EM | F1 | EM | F1 | EM | |
| Con | 85.84 | 60.64 | 84.80 | 56.01 | 85.54 | 55.09 |
| Ind | 85.76 | 66.66 | 85.84 | 67.21 | 85.34 | 65.57 |
| Tech | 80.24 | 58.08 | 81.51 | 58.58 | 81.98 | 58.08 |
| Train/Test | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Con | Ori | 59.30 | 15.74 | 56.69 | 18.60 | 56.13 | 17.67 |
| Ours | (+2.19) | (+3.87) | (+1.79) | (+0.84) | (+2.24) | (+1.77) | |
| Ind | Ori | 59.85 | 24.72 | 61.41 | 24.03 | 59.94 | 25.27 |
| Ours | (+2.52) | (-0.68) | (+0.06) | (-3.81) | (+1.64) | (-4.51) | |
| Tech | Ori | 55.30 | 17.85 | 54.41 | 16.83 | 58.76 | 20.20 |
| Ours | (+3.70) | (+5.38) | (+4.83) | (+5.39) | (+0.53) | (+2.02) |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: RoBERTa | Con | 78.20 | 51.38 | 72.49 | 47.68 | 74.63 | 49.07 |
| Ind | 77.81 | 49.45 | 80.05 | 58.46 | 77.91 | 48.35 | |
| Tec | 75.54 | 55.05 | 73.99 | 48.98 | 76.49 | 54.04 | |
| RoBERTa + LP-FT | Con | 75.09 | 47.68 | 73.82 | 46.75 | 75.72 | 50.00 |
| Ind | 78.81 | 48.90 | 81.01* | 51.10 | 76.71 | 50.00 | |
| Tec | 79.23 | 56.06* | 77.50 | 55.05 | 78.22 | 53.53 | |
| RoBERTa + Prompt | Con | 78.97* | 53.24 | 77.09* | 50.00* | 77.93* | 51.85* |
| Ind | 78.45 | 56.28 | 80.05 | 61.74* | 80.05* | 57.92* | |
| Tec | 77.46 | 54.54 | 78.50* | 57.58* | 80.62* | 58.08* | |
| RoBERTa + LP-FT + Prompt | Con | 78.39 | 53.24* | 75.07 | 48.61 | 77.46 | 49.54 |
| Ind | 80.55* | 59.02* | 80.70 | 60.66 | 79.00 | 57.37 | |
| Tec | 77.93* | 55.05 | 76.70 | 56.06 | 76.76 | 53.54 |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: BART | Con | 70.29 | 24.53 | 68.44 | 24.07 | 68.61 | 23.61 |
| Ind | 70.11 | 31.31 | 72.53 | 32.41 | 69.53 | 27.77 | |
| Tec | 69.89 | 30.30 | 69.53 | 27.77 | 71.79 | 33.83 | |
| BART + LP-FT | Con | 70.81* | 24.07 | 70.22* | 24.53 | 68.26 | 23.61 |
| Ind | 70.61 | 30.77 | 73.39 | 31.87 | 70.91 | 28.57 | |
| Tec | 71.08 | 35.35 | 72.31 | 30.30 | 72.37 | 33.33 | |
| BART + Prompt | Con | 70.41 | 27.31* | 69.10 | 23.15 | 69.20 | 23.61 |
| Ind | 74.42* | 42.07* | 76.27 | 37.16 | 73.00 | 34.43 | |
| Tec | 72.20 | 32.32 | 72.51 | 30.3 | 73.32 | 35.35* | |
| BART + LP-FT + Prompt | Con | 70.65 | 25.01 | 70.14 | 26.85* | 72.00* | 28.24* |
| Ind | 74.06 | 36.06 | 76.58* | 39.89* | 73.37* | 37.70* | |
| Tec | 72.67* | 33.33 | 72.96* | 31.31* | 74.02* | 34.34 |
| BART | RoBERTa | ||||||||||||||
| Train/Test | Consumer | Industrial | Technology | Train/Test | Consumer | Industrial | Technology | ||||||||
| F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | ||||
| Con | Ori | 70.29(+0.36) | 24.53(+0.47) | 68.44(+1.7) | 24.07(+2.78) | 68.61(+3.39) | 23.61(+4.63) | Con | Ori | 78.20(+0.19) | 51.38(+1.86) | 72.49(+2.58) | 47.68(+0.93) | 74.63(+2.83) | 49.07(+0.47) |
| Ours | Ours | ||||||||||||||
| Ind | Ori | 70.11(+3.95) | 31.31(+4.75) | 72.53(+4.05) | 32.41(+7.48) | 69.63(+3.84) | 27.27(+9.93) | Ind | Ori | 77.81(+2.74) | 49.45(+9.57) | 80.05(+0.65) | 58.46(+2.20) | 77.91(+1.09) | 48.35(+9.02) |
| Ours | Ours | ||||||||||||||
| Tech | Ori | 69.89(+2.78) | 30.30(+3.03) | 69.53(+2.98) | 27.77(+3.45) | 71.79(+2.23) | 33.83(+0.51) | Tech | Ori | 75.54(+2.39) | 55.05(+0.00) | 73.99(+2.71) | 48.98(+7.08) | 76.49(+0.27) | 54.04(-0.50) |
| Ours | Ours | ||||||||||||||
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| BART | 74.16 | 36.50 | 73.26 | 34.49 |
| BART + LP-FT | 74.06 | 35.03 | 73.83 | 34/00 |
| BART + P | 75.60 | 37.47 | 75.33 | 37.66 |
| BART + P + LP-FT | 77.60 | 41.22 | 76.90 | 39.44 |
| RoBERTa | 83.97 | 61.82 | 83.45 | 61.28 |
| RoBERTa + LP-FT | 84.80 | 62.15 | 83.49 | 61.18 |
| RoBERTa + P | 84.55 | 62.20 | 83.61 | 61.34 |
| RoBERTa + P + LP-FT | 84.56 | 62.15 | 83.87 | 61.42 |
| S->N | S->C | N->S | N->C | C->S | C->N | |
| RoBERTa | 37.60 | 66.58 | 49.87 | 44.22 | 19.44 | 7.45 |
| RoBERTa+DA (Yue et al., 2022) | 38.26 | 66.14 | 50.31 | 43.05 | 22.74 | 7.15 |
| RoBERTa+P | 38.17* | 66.84 | 50.97* | 48.37* | 21.41* | 8.64* |
| RoBERTa+LPFT | 37.95* | 66.60 | 50.28* | 45.86* | 20.92* | 7.5 |
| RoBERTa+P+LPFT | 38.76* | 66.86* | 52.41* | 51.64* | 23.02* | 7.73* |
| BART | 33.71 | 46.97 | 43.49 | 31.78 | 26.14 | 8.69 |
| BART+DA (Yue et al., 2022) | 35.09 | 55.65 | 44.05 | 33.47 | 26.98 | 9.02 |
| BART+P | 36.81* | 56.22* | 43.61* | 31.91 | 25.96 | 9.26* |
| BART+LPFT | 33.29 | 53.29* | 44.05* | 31.95 | 26.87* | 9.49* |
| BART+P+LPFT | 35.23* | 55.86* | 44.36* | 33.79* | 27.61* | 9.54* |
| Context & Prompt | Question | Gold Answer | Baseline Output | Our Output |
| Predictive Model: SpanBERT-base | ||||
| As Terex has expanded its MP product line, it has captured a larger global market share of the industry, allowing it to gain greater insight into customer demand. This may provide revenue synergies in the future...... +Prompt: To answer a "Why useful" question, you need to look for "allowing" | Why Terex has expanded its MP product line is useful? | gain greater insight into customer | it has captured a larger global market share of | it has captured a larger global market share of the industry, allowing it to gain greater insight into customer |
| However, Avnet's management conceded...... Suppliers get access to a broad range of customers by using Avnet without having to make significant investment in sales and engineering teams. In exchange for these services, Avnet can generate ...... Prompt: The sentiment of this question is positive, you need to look for "significant" | What will be prevented if suppliers get access to a broad range of customers? | investment in sales and engineering | using Avnet | using Avnet without having to make significant investment in sales and engineering |
| Generative Model: BART-base | ||||
| At the end of 2020, the store base had grown about 29% over the prior five-year period, to about 1,920 locations (around 2,100 including Petsense), driving sales and EPS compound annual growth rates over the past three years of 14% and 27%, respectively. We forecast that the firm will grow to around ...... Prompt: The entity "EPS" is mentioned in the question. This timing "annual" is mentioned in the question. | Why sales and EPS compound annual growth rates increase? | the store base had grown about 29% over the prior five-year period | 14% and 27% | the store base had grown about 29% over the prior five years |
| Finally, we view the likelihood of sustained economic value creation as quite high for the restaurant brand, which finds itself on the leading edge of most of the changes we expect to impact the restaurant industry over the medium to long term. Though Chipotle saw economic value destruction in 201 ...... Prompt: "restaurant industry" is an important phrase. And also pay attention to these words:"edge", "changes" | What will happen if on the leading edge of most of the changes we expect to impact the restaurant industry? | the likelihood of sustained economic value creation as quite high for the restaurant brand | over the medium to long term | we view the likelihood of sustained economic value creation as quite high for the restaurant brand |
| Template | F1 | |
| Baseline | None | 70.29 |
| Experiment1 | "There is no important phrase in this query. +But also pay attention to these words: _” | 69.57 |
| Experiment2 | "There is no important phrase in this query. +And also pay attention to these words: _” | 70.84 |
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| SpanBERT-base | 84.77 | 62.62 | 84.85 | 64.04 |
| SpanBERT-large | 85.40 | 61.41 | 85.53 | 62.26 |
| Domain | Con | Ind | Tech | |||
| F1 | EM | F1 | EM | F1 | EM | |
| Con | 85.84 | 60.64 | 84.80 | 56.01 | 85.54 | 55.09 |
| Ind | 85.76 | 66.66 | 85.84 | 67.21 | 85.34 | 65.57 |
| Tech | 80.24 | 58.08 | 81.51 | 58.58 | 81.98 | 58.08 |
| Train/Test | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Con | Ori | 59.30 | 15.74 | 56.69 | 18.60 | 56.13 | 17.67 |
| Ours | (+2.19) | (+3.87) | (+1.79) | (+0.84) | (+2.24) | (+1.77) | |
| Ind | Ori | 59.85 | 24.72 | 61.41 | 24.03 | 59.94 | 25.27 |
| Ours | (+2.52) | (-0.68) | (+0.06) | (-3.81) | (+1.64) | (-4.51) | |
| Tech | Ori | 55.30 | 17.85 | 54.41 | 16.83 | 58.76 | 20.20 |
| Ours | (+3.70) | (+5.38) | (+4.83) | (+5.39) | (+0.53) | (+2.02) | |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: RoBERTa | Con | 78.20 | 51.38 | 72.49 | 47.68 | 74.63 | 49.07 |
| Ind | 77.81 | 49.45 | 80.05 | 58.46 | 77.91 | 48.35 | |
| Tec | 75.54 | 55.05 | 73.99 | 48.98 | 76.49 | 54.04 | |
| RoBERTa + LP-FT | Con | 75.09 | 47.68 | 73.82 | 46.75 | 75.72 | 50.00 |
| Ind | 78.81 | 48.90 | 81.01* | 51.10 | 76.71 | 50.00 | |
| Tec | 79.23 | 56.06* | 77.50 | 55.05 | 78.22 | 53.53 | |
| RoBERTa + Prompt | Con | 78.97* | 53.24 | 77.09* | 50.00* | 77.93* | 51.85* |
| Ind | 78.45 | 56.28 | 80.05 | 61.74* | 80.05* | 57.92* | |
| Tec | 77.46 | 54.54 | 78.50* | 57.58* | 80.62* | 58.08* | |
| RoBERTa + LP-FT + Prompt | Con | 78.39 | 53.24* | 75.07 | 48.61 | 77.46 | 49.54 |
| Ind | 80.55* | 59.02* | 80.70 | 60.66 | 79.00 | 57.37 | |
| Tec | 77.93* | 55.05 | 76.70 | 56.06 | 76.76 | 53.54 | |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: BART | Con | 70.29 | 24.53 | 68.44 | 24.07 | 68.61 | 23.61 |
| Ind | 70.11 | 31.31 | 72.53 | 32.41 | 69.53 | 27.77 | |
| Tec | 69.89 | 30.30 | 69.53 | 27.77 | 71.79 | 33.83 | |
| BART + LP-FT | Con | 70.81* | 24.07 | 70.22* | 24.53 | 68.26 | 23.61 |
| Ind | 70.61 | 30.77 | 73.39 | 31.87 | 70.91 | 28.57 | |
| Tec | 71.08 | 35.35 | 72.31 | 30.30 | 72.37 | 33.33 | |
| BART + Prompt | Con | 70.41 | 27.31* | 69.10 | 23.15 | 69.20 | 23.61 |
| Ind | 74.42* | 42.07* | 76.27 | 37.16 | 73.00 | 34.43 | |
| Tec | 72.20 | 32.32 | 72.51 | 30.3 | 73.32 | 35.35* | |
| BART + LP-FT + Prompt | Con | 70.65 | 25.01 | 70.14 | 26.85* | 72.00* | 28.24* |
| Ind | 74.06 | 36.06 | 76.58* | 39.89* | 73.37* | 37.70* | |
| Tec | 72.67* | 33.33 | 72.96* | 31.31* | 74.02* | 34.34 | |
| BART | RoBERTa | ||||||||||||||
| Train/Test | Consumer | Industrial | Technology | Train/Test | Consumer | Industrial | Technology | ||||||||
| F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | ||||
| Con | Ori | 70.29(+0.36) | 24.53(+0.47) | 68.44(+1.7) | 24.07(+2.78) | 68.61(+3.39) | 23.61(+4.63) | Con | Ori | 78.20(+0.19) | 51.38(+1.86) | 72.49(+2.58) | 47.68(+0.93) | 74.63(+2.83) | 49.07(+0.47) |
| Ours | Ours | ||||||||||||||
| Ind | Ori | 70.11(+3.95) | 31.31(+4.75) | 72.53(+4.05) | 32.41(+7.48) | 69.63(+3.84) | 27.27(+9.93) | Ind | Ori | 77.81(+2.74) | 49.45(+9.57) | 80.05(+0.65) | 58.46(+2.20) | 77.91(+1.09) | 48.35(+9.02) |
| Ours | Ours | ||||||||||||||
| Tech | Ori | 69.89(+2.78) | 30.30(+3.03) | 69.53(+2.98) | 27.77(+3.45) | 71.79(+2.23) | 33.83(+0.51) | Tech | Ori | 75.54(+2.39) | 55.05(+0.00) | 73.99(+2.71) | 48.98(+7.08) | 76.49(+0.27) | 54.04(-0.50) |
| Ours | Ours |
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| BART | 74.16 | 36.50 | 73.26 | 34.49 |
| BART + LP-FT | 74.06 | 35.03 | 73.83 | 34/00 |
| BART + P | 75.60 | 37.47 | 75.33 | 37.66 |
| BART + P + LP-FT | 77.60 | 41.22 | 76.90 | 39.44 |
| RoBERTa | 83.97 | 61.82 | 83.45 | 61.28 |
| RoBERTa + LP-FT | 84.80 | 62.15 | 83.49 | 61.18 |
| RoBERTa + P | 84.55 | 62.20 | 83.61 | 61.34 |
| RoBERTa + P + LP-FT | 84.56 | 62.15 | 83.87 | 61.42 |
| S->N | S->C | N->S | N->C | C->S | C->N | |
| RoBERTa | 37.60 | 66.58 | 49.87 | 44.22 | 19.44 | 7.45 |
| RoBERTa+DA (Yue et al., 2022) | 38.26 | 66.14 | 50.31 | 43.05 | 22.74 | 7.15 |
| RoBERTa+P | 38.17* | 66.84 | 50.97* | 48.37* | 21.41* | 8.64* |
| RoBERTa+LPFT | 37.95* | 66.60 | 50.28* | 45.86* | 20.92* | 7.5 |
| RoBERTa+P+LPFT | 38.76* | 66.86* | 52.41* | 51.64* | 23.02* | 7.73* |
| BART | 33.71 | 46.97 | 43.49 | 31.78 | 26.14 | 8.69 |
| BART+DA (Yue et al., 2022) | 35.09 | 55.65 | 44.05 | 33.47 | 26.98 | 9.02 |
| BART+P | 36.81* | 56.22* | 43.61* | 31.91 | 25.96 | 9.26* |
| BART+LPFT | 33.29 | 53.29* | 44.05* | 31.95 | 26.87* | 9.49* |
| BART+P+LPFT | 35.23* | 55.86* | 44.36* | 33.79* | 27.61* | 9.54* |
| Context & Prompt | Question | Gold Answer | Baseline Output | Our Output |
| Predictive Model: SpanBERT-base | ||||
| As Terex has expanded its MP product line, it has captured a larger global market share of the industry, allowing it to gain greater insight into customer demand. This may provide revenue synergies in the future......\nPrompt: To answer a "Why useful" question, you need to look for "allowing" | Why Terex has expanded its MP product line is useful? | gain greater insight into customer | it has captured a larger global market share of | it has captured a larger global market share of the industry, allowing it to gain greater insight into customer |
| However, Avnet's management conceded...... Suppliers get access to a broad range of customers by using Avnet without having to make significant investment in sales and engineering teams. In exchange for these services, Avnet can generate ...... Prompt: The sentiment of this question is positive, you need to look for "significant" | What will be prevented if suppliers get access to a broad range of customers? | investment in sales and engineering | using Avnet | using Avnet without having to make significant investment in sales and engineering |
| Generative Model: BART-base | ||||
| At the end of 2020, the store base had grown about 29% over the prior five-year period, to about 1,920 locations (around 2,100 including Petsense), driving sales and EPS compound annual growth rates over the past three years of 14% and 27%, respectively. We forecast that the firm will grow to around ...... Prompt: The entity "EPS" is mentioned in the question. This timing "annual" is mentioned in the question. | Why sales and EPS compound annual growth rates increase? | the store base had grown about 29% over the prior five-year period | 14% and 27% | the store base had grown about 29% over the prior five years |
| Finally, we view the likelihood of sustained economic value creation as quite high for the restaurant brand, which finds itself on the leading edge of most of the changes we expect to impact the restaurant industry over the medium to long term. Though Chipotle saw economic value destruction in 201 ...... Prompt: "restaurant industry" is an important phrase. And also pay attention to these words:"edge", "changes" | What will happen if on the leading edge of most of the changes we expect to impact the restaurant industry? | the likelihood of sustained economic value creation as quite high for the restaurant brand | over the medium to long term | we view the likelihood of sustained economic value creation as quite high for the restaurant brand |
| Template | F1 | |
| Baseline | None | 70.29 |
| Experiment1 | "There is no important phrase in this query.\nBut also pay attention to these words: _” | 69.57 |
| Experiment2 | "There is no important phrase in this query.\nAnd also pay attention to these words: _” | 70.84 |
| Methods | Dev | Test | ||
| F1 | EM | F1 | EM | |
| SpanBERT-base | 84.77 | 62.62 | 84.85 | 64.04 |
| SpanBERT-large | 85.40 | 61.41 | 85.53 | 62.26 |
| Domain | Con | Ind | Tech | |||
| F1 | EM | F1 | EM | F1 | EM | |
| Con | 85.84 | 60.64 | 84.80 | 56.01 | 85.54 | 55.09 |
| Ind | 85.76 | 66.66 | 85.84 | 67.21 | 85.34 | 65.57 |
| Tech | 80.24 | 58.08 | 81.51 | 58.58 | 81.98 | 58.08 |
| Train/Test | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Con | Ori | 59.30 | 15.74 | 56.69 | 18.60 | 56.13 | 17.67 |
| Ours | (+2.19) | (+3.87) | (+1.79) | (+0.84) | (+2.24) | (+1.77) | |
| Ind | Ori | 59.85 | 24.72 | 61.41 | 24.03 | 59.94 | 25.27 |
| Ours | (+2.52) | (-0.68) | (+0.06) | (-3.81) | (+1.64) | (-4.51) | |
| Tech | Ori | 55.30 | 17.85 | 54.41 | 16.83 | 58.76 | 20.20 |
| Ours | (+3.70) | (+5.38) | (+4.83) | (+5.39) | (+0.53) | (+2.02) |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: RoBERTa | Con | 78.20 | 51.38 | 72.49 | 47.68 | 74.63 | 49.07 |
| Ind | 77.81 | 49.45 | 80.05 | 58.46 | 77.91 | 48.35 | |
| Tec | 75.54 | 55.05 | 73.99 | 48.98 | 76.49 | 54.04 | |
| RoBERTa + LP-FT | Con | 75.09 | 47.68 | 73.82 | 46.75 | 75.72 | 50.00 |
| Ind | 78.81 | 48.90 | 81.01* | 51.10 | 76.71 | 50.00 | |
| Tec | 79.23 | 56.06* | 77.50 | 55.05 | 78.22 | 53.53 | |
| RoBERTa + Prompt | Con | 78.97* | 53.24 | 77.09* | 50.00* | 77.93* | 51.85* |
| Ind | 78.45 | 56.28 | 80.05 | 61.74* | 80.05* | 57.92* | |
| Tec | 77.46 | 54.54 | 78.50* | 57.58* | 80.62* | 58.08* | |
| RoBERTa + LP-FT + Prompt | Con | 78.39 | 53.24* | 75.07 | 48.61 | 77.46 | 49.54 |
| Ind | 80.55* | 59.02* | 80.70 | 60.66 | 79.00 | 57.37 | |
| Tec | 77.93* | 55.05 | 76.70 | 56.06 | 76.76 | 53.54 |
| Methods | Consumer | Industrial | Technology | ||||
| F1 | EM | F1 | EM | F1 | EM | ||
| Baseline: BART | Con | 70.29 | 24.53 | 68.44 | 24.07 | 68.61 | 23.61 |
| Ind | 70.11 | 31.31 | 72.53 | 32.41 | 69.53 | 27.77 | |
| Tec | 69.89 | 30.30 | 69.53 | 27.77 | 71.79 | 33.83 | |
| BART + LP-FT | Con | 70.81* | 24.07 | 70.22* | 24.53 | 68.26 | 23.61 |
| Ind | 70.61 | 30.77 | 73.39 | 31.87 | 70.91 | 28.57 | |
| Tec | 71.08 | 35.35 | 72.31 | 30.30 | 72.37 | 33.33 | |
| BART + Prompt | Con | 70.41 | 27.31* | 69.10 | 23.15 | 69.20 | 23.61 |
| Ind | 74.42* | 42.07* | 76.27 | 37.16 | 73.00 | 34.43 | |
| Tec | 72.20 | 32.32 | 72.51 | 30.3 | 73.32 | 35.35* | |
| BART + LP-FT + Prompt | Con | 70.65 | 25.01 | 70.14 | 26.85* | 72.00* | 28.24* |
| Ind | 74.06 | 36.06 | 76.58* | 39.89* | 73.37* | 37.70* | |
| Tec | 72.67* | 33.33 | 72.96* | 31.31* | 74.02* | 34.34 |
| Model | CHIP-2020 | CHIP-2022 | DIAKG | ||||||
| Entity | Relation | Overall | Entity | Relation | Overall | Entity | Relation | Overall | |
| PFN (Yan et al., 2021) | 72.2 | 52.7 | 62.5 | 59.3 | 42.0 | 50.7 | 63.0 | 44.5 | 53.8 |
| CASREL (Wei et al., 2020) | 66.1 | 49.5 | 57.8 | 56.5 | 39.1 | 47.8 | 50.4 | 31.0 | 40.7 |
| SPERT (Eberts and Ulges, 2020) | 73.5 | 50.1 | 61.8 | 57.8 | 38.7 | 48.3 | 65.5 | 29.1 | 47.3 |
| KB-graph (Verlinden et al., 2021) | 73.9 | 61.8 | 67.9 | 65.0 | 32.5 | 48.8 | 69.3 | 51.1 | 60.2 |
| KECI (Lai et al., 2021) | 74.3 | 61.1 | 67.7 | 63.4 | 32.6 | 48.0 | 72.6 | 55.1 | 63.8 |
| KECI_nconv (Li et al., 2017) | 74.2 | 61.7 | 68.0 | 63.5 | 32.4 | 48.0 | 73.2 | 54.1 | 63.7 |
| KECI_gea (Huang et al., 2020) | 75.6 | 61.0 | 68.3 | 61.8 | 32.5 | 47.2 | 73.1 | 54.9 | 64.0 |
| KECI gn (Battaglia et al., 2018) | 74.7 | 60.5 | 67.6 | 64.0 | 32.4 | 48.2 | 73.9 | 54.5 | 64.2 |
| Our Model | 76.5 | 61.8 | 69.2 | 66.7 | 32.5 | 49.6 | 74.2 | 54.4 | 64.3 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| surgery | 55.78 | 54.30 | 55.03 | 47.34 | 64.90 | 54.75 |
| inspection | 65.75 | 61.03 | 63.30 | 58.07 | 72.31 | 64.41 |
| epidemiology | 60.76 | 58.22 | 59.46 | 52.92 | 73.26 | 61.45 |
| disease | 70.04 | 84.52 | 76.60 | 79.95 | 80.24 | 80.09 |
| symptom | 71.70 | 63.12 | 67.14 | 68.00 | 72.74 | 70.29 |
| sociology | 66.83 | 46.64 | 54.94 | 60.14 | 57.79 | 58.94 |
| medicine | 61.27 | 75.68 | 67.72 | 66.84 | 75.78 | 71.03 |
| part | 72.93 | 46.41 | 56.72 | 60.35 | 66.27 | 63.17 |
| prognosis | - | - | - | - | - | - |
| other-therapy | 47.70 | 33.24 | 39.18 | 47.67 | 38.78 | 42.77 |
| others | - | - | - | 100.00 | 5.83 | 11.02 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| ADE | 68.42 | 54.17 | 60.47 | 77.55 | 52.78 | 62.81 |
| Amount | 20.00 | 17.65 | 18.75 | 26.67 | 23.53 | 25.00 |
| Anatomy | 79.31 | 92.00 | 85.18 | 82.51 | 92.00 | 87.00 |
| Class | 87.13 | 60.69 | 71.55 | 86.27 | 60.69 | 71.25 |
| Disease | 78.22 | 77.49 | 77.85 | 73.40 | 82.60 | 77.73 |
| Drug | 65.02 | 75.24 | 69.76 | 66.80 | 77.62 | 71.80 |
| Method | - | - | - | 36.84 | 50.00 | 42.42 |
| Pathogenesis | - | - | - | 22.22 | 25.00 | 23.53 |
| Reason | 30.00 | 33.33 | 31.58 | - | - | - |
| Symptom | 48.15 | 40.62 | 44.07 | 52.17 | 37.50 | 43.63 |
| Test | 43.33 | 61.90 | 50.98 | 50.00 | 66.67 | 57.14 |
| Test_items | 59.54 | 60.94 | 60.23 | 53.99 | 68.75 | 60.48 |
| Treatment | 42.86 | 57.69 | 49.18 | 38.24 | 50.00 | 43.34 |
| Operation | - | - | - | - | - | - |
| Model | CHIP-2020 | CHIP-2022 | DIAKG | ||||||
| Entity | Relation | Overall | Entity | Relation | Overall | Entity | Relation | Overall | |
| PFN (Yan et al., 2021) | 72.2 | 52.7 | 62.5 | 59.3 | 42.0 | 50.7 | 63.0 | 44.5 | 53.8 |
| CASREL (Wei et al., 2020) | 66.1 | 49.5 | 57.8 | 56.5 | 39.1 | 47.8 | 50.4 | 31.0 | 40.7 |
| SPERT (Eberts and Ulges, 2020) | 73.5 | 50.1 | 61.8 | 57.8 | 38.7 | 48.3 | 65.5 | 29.1 | 47.3 |
| KB-graph (Verlinden et al., 2021) | 73.9 | 61.8 | 67.9 | 65.0 | 32.5 | 48.8 | 69.3 | 51.1 | 60.2 |
| KECI (Lai et al., 2021) | 74.3 | 61.1 | 67.7 | 63.4 | 32.6 | 48.0 | 72.6 | 55.1 | 63.8 |
| KECI_nconv (Li et al., 2017) | 74.2 | 61.7 | 68.0 | 63.5 | 32.4 | 48.0 | 73.2 | 54.1 | 63.7 |
| KECI_gea (Huang et al., 2020) | 75.6 | 61.0 | 68.3 | 61.8 | 32.5 | 47.2 | 73.1 | 54.9 | 64.0 |
| KECI gn (Battaglia et al., 2018) | 74.7 | 60.5 | 67.6 | 64.0 | 32.4 | 48.2 | 73.9 | 54.5 | 64.2 |
| Our Model | 76.5 | 61.8 | 69.2 | 66.7 | 32.5 | 49.6 | 74.2 | 54.4 | 64.3 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| surgery | 55.78 | 54.30 | 55.03 | 47.34 | 64.90 | 54.75 |
| inspection | 65.75 | 61.03 | 63.30 | 58.07 | 72.31 | 64.41 |
| epidemiology | 60.76 | 58.22 | 59.46 | 52.92 | 73.26 | 61.45 |
| disease | 70.04 | 84.52 | 76.60 | 79.95 | 80.24 | 80.09 |
| symptom | 71.70 | 63.12 | 67.14 | 68.00 | 72.74 | 70.29 |
| sociology | 66.83 | 46.64 | 54.94 | 60.14 | 57.79 | 58.94 |
| medicine | 61.27 | 75.68 | 67.72 | 66.84 | 75.78 | 71.03 |
| part | 72.93 | 46.41 | 56.72 | 60.35 | 66.27 | 63.17 |
| prognosis | - | - | - | - | - | - |
| other-therapy | 47.70 | 33.24 | 39.18 | 47.67 | 38.78 | 42.77 |
| others | - | - | - | 100.00 | 5.83 | 11.02 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| ADE | 68.42 | 54.17 | 60.47 | 77.55 | 52.78 | 62.81 |
| Amount | 20.00 | 17.65 | 18.75 | 26.67 | 23.53 | 25.00 |
| Anatomy | 79.31 | 92.00 | 85.18 | 82.51 | 92.00 | 87.00 |
| Class | 87.13 | 60.69 | 71.55 | 86.27 | 60.69 | 71.25 |
| Disease | 78.22 | 77.49 | 77.85 | 73.40 | 82.60 | 77.73 |
| Drug | 65.02 | 75.24 | 69.76 | 66.80 | 77.62 | 71.80 |
| Method | - | - | - | 36.84 | 50.00 | 42.42 |
| Pathogenesis | - | - | - | 22.22 | 25.00 | 23.53 |
| Reason | 30.00 | 33.33 | 31.58 | - | - | - |
| Symptom | 48.15 | 40.62 | 44.07 | 52.17 | 37.50 | 43.63 |
| Test | 43.33 | 61.90 | 50.98 | 50.00 | 66.67 | 57.14 |
| Test_items | 59.54 | 60.94 | 60.23 | 53.99 | 68.75 | 60.48 |
| Treatment | 42.86 | 57.69 | 49.18 | 38.24 | 50.00 | 43.34 |
| Operation | - | - | - | - | - | - |
| Model | CHIP-2020 | CHIP-2022 | DIAKG | ||||||
| Entity | Relation | Overall | Entity | Relation | Overall | Entity | Relation | Overall | |
| PFN (Yan et al., 2021) | 72.2 | 52.7 | 62.5 | 59.3 | 42.0 | 50.7 | 63.0 | 44.5 | 53.8 |
| CASREL (Wei et al., 2020) | 66.1 | 49.5 | 57.8 | 56.5 | 39.1 | 47.8 | 50.4 | 31.0 | 40.7 |
| SPERT (Eberts and Ulges, 2020) | 73.5 | 50.1 | 61.8 | 57.8 | 38.7 | 48.3 | 65.5 | 29.1 | 47.3 |
| KB-graph (Verlinden et al., 2021) | 73.9 | 61.8 | 67.9 | 65.0 | 32.5 | 48.8 | 69.3 | 51.1 | 60.2 |
| KECI (Lai et al., 2021) | 74.3 | 61.1 | 67.7 | 63.4 | 32.6 | 48.0 | 72.6 | 55.1 | 63.8 |
| KECI_nconv (Li et al., 2017) | 74.2 | 61.7 | 68.0 | 63.5 | 32.4 | 48.0 | 73.2 | 54.1 | 63.7 |
| KECI_gea (Huang et al., 2020) | 75.6 | 61.0 | 68.3 | 61.8 | 32.5 | 47.2 | 73.1 | 54.9 | 64.0 |
| KECI gn (Battaglia et al., 2018) | 74.7 | 60.5 | 67.6 | 64.0 | 32.4 | 48.2 | 73.9 | 54.5 | 64.2 |
| Our Model | 76.5 | 61.8 | 69.2 | 66.7 | 32.5 | 49.6 | 74.2 | 54.4 | 64.3 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| surgery | 55.78 | 54.30 | 55.03 | 47.34 | 64.90 | 54.75 |
| inspection | 65.75 | 61.03 | 63.30 | 58.07 | 72.31 | 64.41 |
| epidemiology | 60.76 | 58.22 | 59.46 | 52.92 | 73.26 | 61.45 |
| disease | 70.04 | 84.52 | 76.60 | 79.95 | 80.24 | 80.09 |
| symptom | 71.70 | 63.12 | 67.14 | 68.00 | 72.74 | 70.29 |
| sociology | 66.83 | 46.64 | 54.94 | 60.14 | 57.79 | 58.94 |
| medicine | 61.27 | 75.68 | 67.72 | 66.84 | 75.78 | 71.03 |
| part | 72.93 | 46.41 | 56.72 | 60.35 | 66.27 | 63.17 |
| prognosis | - | - | - | - | - | - |
| other-therapy | 47.70 | 33.24 | 39.18 | 47.67 | 38.78 | 42.77 |
| others | - | - | - | 100.00 | 5.83 | 11.02 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| ADE | 68.42 | 54.17 | 60.47 | 77.55 | 52.78 | 62.81 |
| Amount | 20.00 | 17.65 | 18.75 | 26.67 | 23.53 | 25.00 |
| Anatomy | 79.31 | 92.00 | 85.18 | 82.51 | 92.00 | 87.00 |
| Class | 87.13 | 60.69 | 71.55 | 86.27 | 60.69 | 71.25 |
| Disease | 78.22 | 77.49 | 77.85 | 73.40 | 82.60 | 77.73 |
| Drug | 65.02 | 75.24 | 69.76 | 66.80 | 77.62 | 71.80 |
| Method | - | - | - | 36.84 | 50.00 | 42.42 |
| Pathogenesis | - | - | - | 22.22 | 25.00 | 23.53 |
| Reason | 30.00 | 33.33 | 31.58 | - | - | - |
| Symptom | 48.15 | 40.62 | 44.07 | 52.17 | 37.50 | 43.63 |
| Test | 43.33 | 61.90 | 50.98 | 50.00 | 66.67 | 57.14 |
| Test_items | 59.54 | 60.94 | 60.23 | 53.99 | 68.75 | 60.48 |
| Treatment | 42.86 | 57.69 | 49.18 | 38.24 | 50.00 | 43.34 |
| Operation | - | - | - | - | - | - |
| Model | CHIP-2020 | CHIP-2022 | DIAKG | ||||||
| Entity | Relation | Overall | Entity | Relation | Overall | Entity | Relation | Overall | |
| PFN (Yan et al., 2021) | 72.2 | 52.7 | 62.5 | 59.3 | 42.0 | 50.7 | 63.0 | 44.5 | 53.8 |
| CASREL (Wei et al., 2020) | 66.1 | 49.5 | 57.8 | 56.5 | 39.1 | 47.8 | 50.4 | 31.0 | 40.7 |
| SPERT (Eberts and Ulges, 2020) | 73.5 | 50.1 | 61.8 | 57.8 | 38.7 | 48.3 | 65.5 | 29.1 | 47.3 |
| KB-graph (Verlinden et al., 2021) | 73.9 | 61.8 | 67.9 | 65.0 | 32.5 | 48.8 | 69.3 | 51.1 | 60.2 |
| KECI (Lai et al., 2021) | 74.3 | 61.1 | 67.7 | 63.4 | 32.6 | 48.0 | 72.6 | 55.1 | 63.8 |
| KECI_nconv (Li et al., 2017) | 74.2 | 61.7 | 68.0 | 63.5 | 32.4 | 48.0 | 73.2 | 54.1 | 63.7 |
| KECI_gea (Huang et al., 2020) | 75.6 | 61.0 | 68.3 | 61.8 | 32.5 | 47.2 | 73.1 | 54.9 | 64.0 |
| KECI gn (Battaglia et al., 2018) | 74.7 | 60.5 | 67.6 | 64.0 | 32.4 | 48.2 | 73.9 | 54.5 | 64.2 |
| Our Model | 76.5 | 61.8 | 69.2 | 66.7 | 32.5 | 49.6 | 74.2 | 54.4 | 64.3 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| surgery | 55.78 | 54.30 | 55.03 | 47.34 | 64.90 | 54.75 |
| inspection | 65.75 | 61.03 | 63.30 | 58.07 | 72.31 | 64.41 |
| epidemiology | 60.76 | 58.22 | 59.46 | 52.92 | 73.26 | 61.45 |
| disease | 70.04 | 84.52 | 76.60 | 79.95 | 80.24 | 80.09 |
| symptom | 71.70 | 63.12 | 67.14 | 68.00 | 72.74 | 70.29 |
| sociology | 66.83 | 46.64 | 54.94 | 60.14 | 57.79 | 58.94 |
| medicine | 61.27 | 75.68 | 67.72 | 66.84 | 75.78 | 71.03 |
| part | 72.93 | 46.41 | 56.72 | 60.35 | 66.27 | 63.17 |
| prognosis | - | - | - | - | - | - |
| other-therapy | 47.70 | 33.24 | 39.18 | 47.67 | 38.78 | 42.77 |
| others | - | - | - | 100.00 | 5.83 | 11.02 |
| Entity Type | KECI(Lai et al., 2021) | Our Model | ||||
| P | R | F1 | P | R | F1 | |
| ADE | 68.42 | 54.17 | 60.47 | 77.55 | 52.78 | 62.81 |
| Amount | 20.00 | 17.65 | 18.75 | 26.67 | 23.53 | 25.00 |
| Anatomy | 79.31 | 92.00 | 85.18 | 82.51 | 92.00 | 87.00 |
| Class | 87.13 | 60.69 | 71.55 | 86.27 | 60.69 | 71.25 |
| Disease | 78.22 | 77.49 | 77.85 | 73.40 | 82.60 | 77.73 |
| Drug | 65.02 | 75.24 | 69.76 | 66.80 | 77.62 | 71.80 |
| Method | - | - | - | 36.84 | 50.00 | 42.42 |
| Pathogenesis | - | - | - | 22.22 | 25.00 | 23.53 |
| Reason | 30.00 | 33.33 | 31.58 | - | - | - |
| Symptom | 48.15 | 40.62 | 44.07 | 52.17 | 37.50 | 43.63 |
| Test | 43.33 | 61.90 | 50.98 | 50.00 | 66.67 | 57.14 |
| Test_items | 59.54 | 60.94 | 60.23 | 53.99 | 68.75 | 60.48 |
| Treatment | 42.86 | 57.69 | 49.18 | 38.24 | 50.00 | 43.34 |
| Operation | - | - | - | - | - | - |
| Method | Self Persona | Their Persona | ||||||||
| ppt | Dst | Hits@1 | Cons | P-Cover | ppt | Dst | Hits@1 | Cons | P-Cover | |
| GPMN | 36.11 | 13.5 | 54.9 | 0.15 | .018 | 36.45 | 14.8 | 51.4 | 0.10 | .021 |
| DialogGPT | 13.62 | 23.1 | 83.2 | 0.35 | .052 | 14.03 | 23.9 | 78.9 | 0.27 | .059 |
| DialogGPT+PE | 13.57 | 24.8 | 84.5 | 0.38 | .055 | 13.90 | 25.1 | 79.3 | 0.28 | .063 |
| PersonaCVAE | 14.83 | 25.7* | 84.0 | 0.37 | .061 | 14.88 | 25.6 | 78.3 | 0.24 | .066 |
| DialogGPT w/o persona | 15.49 | 19.6 | 72.9 | 0.13 | .012 | - | - | - | - | |
| Persona Approximator | 14.42 | 24.2 | 83.3 | 0.33 | .038 | 14.63 | 24.9 | 78.4 | 0.24 | .040 |
| Persona Generator | 13.39* | 25.2 | 84.2 | 0.38 | .049 | 13.82 | 25.8 | 79.1 | 0.29 | .057 |
| Model | Persona | Fluency | Engagingness | Consistency | Persona Detection |
| DialogGPT | self | 3.56 | 3.57 | 3.63 | 0.88 |
| DialogGPT | their | 3.49 | 3.59 | 3.47 | 0.80 |
| DialogGPT | both | 3.63 | 3.69 | 3.60 | 0.88 |
| DialogGPT+PE | self | 3.62 | 3.49 | 3.61 | 0.87 |
| DialogGPT+PE | their | 3.57 | 3.51 | 3.52 | 0.82 |
| DialogGPT+PE | both | 3.69 | 3.65 | 3.68* | 0.90 |
| PersonaCVAE | self | 3.51 | 3.55 | 3.53 | 0.85 |
| PersonaCVAE | their | 3.50 | 3.52 | 3.42 | 0.77 |
| PersonaCVAE | both | 3.57 | 3.59 | 3.51 | 0.83 |
| DialogGPT w/o persona | - | 3.39 | 3.28 | 3.30 | 0.69 |
| Persona Approximator | self | 3.45 | 3.40 | 3.35 | 0.78 |
| Persona Approximator | their | 3.36 | 3.43 | 3.27 | 0.73 |
| Persona Generator | self | 3.67 | 3.61 | 3.58 | 0.89 |
| Persona Generator | their | 3.61 | 3.69 | 3.52 | 0.84 |
| Persona Generator | both | 3.72* | 3.74* | 3.63 | 0.90 |
| Model | Per | Fluen | Engag | Consis |
| DialogGPT w/o persona | - | 3.42 | 3.41 | 3.48 |
| w/ Persona Generator | self | 3.53 | 3.52 | 3.58 |
| w/ Persona Generator | their | 3.48 | 3.57 | 3.56 |
| Model | perplexity | Dst | Hits@1 | Cons |
| DialogGPT w/o Persona | 15.49 | 19.6 | 72.9 | 0.13 |
| - first half | 18.31 | 15.7 | 66.5 | 0.05 |
| - second half | 13.24 | 23.8 | 79.3 | 0.19 |
| w/ Persona Generator | 13.39 | 25.2 | 84.2 | 0.38 |
| - first half | 16.24 | 23.5 | 79.8 | 0.32 |
| - second half | 12.01 | 26.9 | 88.6 | 0.44 |
| - w/o shared layers | 13.92 | 24.9 | 83.5 | 0.35 |
| - w/o joint training | 14.05 | 24.7 | 83.8 | 0.36 |
| No persona | I don’t know what you could not do? |
| PE w/ self | I am going to the club now. |
| PE w/ their | Do you want to play frisbee or something? |
| PG w/ self | okay I am going to make a cake. |
| - Generated Persona: | ... I craving eating cake... |
| PG w/ their | I prefer that let’s watch tv together. |
| - Generated Persona: | ... I like TV show... |
| Method | Self Persona | Their Persona | ||||||||
| ppt | Dst | Hits@1 | Cons | P-Cover | ppt | Dst | Hits@1 | Cons | P-Cover | |
| GPMN | 36.11 | 13.5 | 54.9 | 0.15 | .018 | 36.45 | 14.8 | 51.4 | 0.10 | .021 |
| DialogGPT | 13.62 | 23.1 | 83.2 | 0.35 | .052 | 14.03 | 23.9 | 78.9 | 0.27 | .059 |
| DialogGPT+PE | 13.57 | 24.8 | 84.5 | 0.38 | .055 | 13.90 | 25.1 | 79.3 | 0.28 | .063 |
| PersonaCVAE | 14.83 | 25.7* | 84.0 | 0.37 | .061 | 14.88 | 25.6 | 78.3 | 0.24 | .066 |
| DialogGPT w/o persona | 15.49 | 19.6 | 72.9 | 0.13 | .012 | - | - | - | - | |
| Persona Approximator | 14.42 | 24.2 | 83.3 | 0.33 | .038 | 14.63 | 24.9 | 78.4 | 0.24 | .040 |
| Persona Generator | 13.39* | 25.2 | 84.2 | 0.38 | .049 | 13.82 | 25.8 | 79.1 | 0.29 | .057 |
| Model | Persona | Fluency | Engagingness | Consistency | Persona Detection |
| DialogGPT | self | 3.56 | 3.57 | 3.63 | 0.88 |
| DialogGPT | their | 3.49 | 3.59 | 3.47 | 0.80 |
| DialogGPT | both | 3.63 | 3.69 | 3.60 | 0.88 |
| DialogGPT+PE | self | 3.62 | 3.49 | 3.61 | 0.87 |
| DialogGPT+PE | their | 3.57 | 3.51 | 3.52 | 0.82 |
| DialogGPT+PE | both | 3.69 | 3.65 | 3.68* | 0.90 |
| PersonaCVAE | self | 3.51 | 3.55 | 3.53 | 0.85 |
| PersonaCVAE | their | 3.50 | 3.52 | 3.42 | 0.77 |
| PersonaCVAE | both | 3.57 | 3.59 | 3.51 | 0.83 |
| DialogGPT w/o persona | - | 3.39 | 3.28 | 3.30 | 0.69 |
| Persona Approximator | self | 3.45 | 3.40 | 3.35 | 0.78 |
| Persona Approximator | their | 3.36 | 3.43 | 3.27 | 0.73 |
| Persona Generator | self | 3.67 | 3.61 | 3.58 | 0.89 |
| Persona Generator | their | 3.61 | 3.69 | 3.52 | 0.84 |
| Persona Generator | both | 3.72* | 3.74* | 3.63 | 0.90 |
| Model | Per | Fluen | Engag | Consis |
| DialogGPT w/o persona | - | 3.42 | 3.41 | 3.48 |
| w/ Persona Generator | self | 3.53 | 3.52 | 3.58 |
| w/ Persona Generator | their | 3.48 | 3.57 | 3.56 |
| Model | perplexity | Dst | Hits@1 | Cons |
| DialogGPT w/o Persona | 15.49 | 19.6 | 72.9 | 0.13 |
| - first half | 18.31 | 15.7 | 66.5 | 0.05 |
| - second half | 13.24 | 23.8 | 79.3 | 0.19 |
| w/ Persona Generator | 13.39 | 25.2 | 84.2 | 0.38 |
| - first half | 16.24 | 23.5 | 79.8 | 0.32 |
| - second half | 12.01 | 26.9 | 88.6 | 0.44 |
| - w/o shared layers | 13.92 | 24.9 | 83.5 | 0.35 |
| - w/o joint training | 14.05 | 24.7 | 83.8 | 0.36 |
| No persona | I don’t know what you could not do? |
| PE w/ self | I am going to the club now. |
| PE w/ their | Do you want to play frisbee or something? |
| PG w/ self | okay I am going to make a cake. |
| - Generated Persona: | ... I craving eating cake... |
| PG w/ their | I prefer that let’s watch tv together. |
| - Generated Persona: | ... I like TV show... |
| Method | Self Persona | Their Persona | ||||||||
| ppt | Dst | Hits@1 | Cons | P-Cover | ppt | Dst | Hits@1 | Cons | P-Cover | |
| GPMN | 36.11 | 13.5 | 54.9 | 0.15 | .018 | 36.45 | 14.8 | 51.4 | 0.10 | .021 |
| DialogGPT | 13.62 | 23.1 | 83.2 | 0.35 | .052 | 14.03 | 23.9 | 78.9 | 0.27 | .059 |
| DialogGPT+PE | 13.57 | 24.8 | 84.5 | 0.38 | .055 | 13.90 | 25.1 | 79.3 | 0.28 | .063 |
| PersonaCVAE | 14.83 | 25.7* | 84.0 | 0.37 | .061 | 14.88 | 25.6 | 78.3 | 0.24 | .066 |
| DialogGPT w/o persona | 15.49 | 19.6 | 72.9 | 0.13 | .012 | - | - | - | - | |
| Persona Approximator | 14.42 | 24.2 | 83.3 | 0.33 | .038 | 14.63 | 24.9 | 78.4 | 0.24 | .040 |
| Persona Generator | 13.39* | 25.2 | 84.2 | 0.38 | .049 | 13.82 | 25.8 | 79.1 | 0.29 | .057 |
| Model | Persona | Fluency | Engagingness | Consistency | Persona Detection |
| DialogGPT | self | 3.56 | 3.57 | 3.63 | 0.88 |
| DialogGPT | their | 3.49 | 3.59 | 3.47 | 0.80 |
| DialogGPT | both | 3.63 | 3.69 | 3.60 | 0.88 |
| DialogGPT+PE | self | 3.62 | 3.49 | 3.61 | 0.87 |
| DialogGPT+PE | their | 3.57 | 3.51 | 3.52 | 0.82 |
| DialogGPT+PE | both | 3.69 | 3.65 | 3.68* | 0.90 |
| PersonaCVAE | self | 3.51 | 3.55 | 3.53 | 0.85 |
| PersonaCVAE | their | 3.50 | 3.52 | 3.42 | 0.77 |
| PersonaCVAE | both | 3.57 | 3.59 | 3.51 | 0.83 |
| DialogGPT w/o persona | - | 3.39 | 3.28 | 3.30 | 0.69 |
| Persona Approximator | self | 3.45 | 3.40 | 3.35 | 0.78 |
| Persona Approximator | their | 3.36 | 3.43 | 3.27 | 0.73 |
| Persona Generator | self | 3.67 | 3.61 | 3.58 | 0.89 |
| Persona Generator | their | 3.61 | 3.69 | 3.52 | 0.84 |
| Persona Generator | both | 3.72* | 3.74* | 3.63 | 0.90 |
| Model | Per | Fluen | Engag | Consis |
| DialogGPT w/o persona | - | 3.42 | 3.41 | 3.48 |
| w/ Persona Generator | self | 3.53 | 3.52 | 3.58 |
| w/ Persona Generator | their | 3.48 | 3.57 | 3.56 |
| Model | perplexity | Dst | Hits@1 | Cons |
| DialogGPT w/o Persona | 15.49 | 19.6 | 72.9 | 0.13 |
| - first half | 18.31 | 15.7 | 66.5 | 0.05 |
| - second half | 13.24 | 23.8 | 79.3 | 0.19 |
| w/ Persona Generator | 13.39 | 25.2 | 84.2 | 0.38 |
| - first half | 16.24 | 23.5 | 79.8 | 0.32 |
| - second half | 12.01 | 26.9 | 88.6 | 0.44 |
| - w/o shared layers | 13.92 | 24.9 | 83.5 | 0.35 |
| - w/o joint training | 14.05 | 24.7 | 83.8 | 0.36 |
| No persona | I don’t know what you could not do? |
| PE w/ self | I am going to the club now. |
| PE w/ their | Do you want to play frisbee or something? |
| PG w/ self | okay I am going to make a cake. |
| - Generated Persona: | ... I craving eating cake... |
| PG w/ their | I prefer that let’s watch tv together. |
| - Generated Persona: | ... I like TV show... |
| Method | Self Persona | Their Persona | ||||||||
| ppt | Dst | Hits@1 | Cons | P-Cover | ppt | Dst | Hits@1 | Cons | P-Cover | |
| GPMN | 36.11 | 13.5 | 54.9 | 0.15 | .018 | 36.45 | 14.8 | 51.4 | 0.10 | .021 |
| DialogGPT | 13.62 | 23.1 | 83.2 | 0.35 | .052 | 14.03 | 23.9 | 78.9 | 0.27 | .059 |
| DialogGPT+PE | 13.57 | 24.8 | 84.5 | 0.38 | .055 | 13.90 | 25.1 | 79.3 | 0.28 | .063 |
| PersonaCVAE | 14.83 | 25.7* | 84.0 | 0.37 | .061 | 14.88 | 25.6 | 78.3 | 0.24 | .066 |
| DialogGPT w/o persona | 15.49 | 19.6 | 72.9 | 0.13 | .012 | - | - | - | - | |
| Persona Approximator | 14.42 | 24.2 | 83.3 | 0.33 | .038 | 14.63 | 24.9 | 78.4 | 0.24 | .040 |
| Persona Generator | 13.39* | 25.2 | 84.2 | 0.38 | .049 | 13.82 | 25.8 | 79.1 | 0.29 | .057 |
| Model | Persona | Fluency | Engagingness | Consistency | Persona Detection |
| DialogGPT | self | 3.56 | 3.57 | 3.63 | 0.88 |
| DialogGPT | their | 3.49 | 3.59 | 3.47 | 0.80 |
| DialogGPT | both | 3.63 | 3.69 | 3.60 | 0.88 |
| DialogGPT+PE | self | 3.62 | 3.49 | 3.61 | 0.87 |
| DialogGPT+PE | their | 3.57 | 3.51 | 3.52 | 0.82 |
| DialogGPT+PE | both | 3.69 | 3.65 | 3.68* | 0.90 |
| PersonaCVAE | self | 3.51 | 3.55 | 3.53 | 0.85 |
| PersonaCVAE | their | 3.50 | 3.52 | 3.42 | 0.77 |
| PersonaCVAE | both | 3.57 | 3.59 | 3.51 | 0.83 |
| DialogGPT w/o persona | - | 3.39 | 3.28 | 3.30 | 0.69 |
| Persona Approximator | self | 3.45 | 3.40 | 3.35 | 0.78 |
| Persona Approximator | their | 3.36 | 3.43 | 3.27 | 0.73 |
| Persona Generator | self | 3.67 | 3.61 | 3.58 | 0.89 |
| Persona Generator | their | 3.61 | 3.69 | 3.52 | 0.84 |
| Persona Generator | both | 3.72* | 3.74* | 3.63 | 0.90 |
| Model | Per | Fluen | Engag | Consis |
| DialogGPT w/o persona | - | 3.42 | 3.41 | 3.48 |
| w/ Persona Generator | self | 3.53 | 3.52 | 3.58 |
| w/ Persona Generator | their | 3.48 | 3.57 | 3.56 |
| Model | perplexity | Dst | Hits@1 | Cons |
| DialogGPT w/o Persona | 15.49 | 19.6 | 72.9 | 0.13 |
| - first half | 18.31 | 15.7 | 66.5 | 0.05 |
| - second half | 13.24 | 23.8 | 79.3 | 0.19 |
| w/ Persona Generator | 13.39 | 25.2 | 84.2 | 0.38 |
| - first half | 16.24 | 23.5 | 79.8 | 0.32 |
| - second half | 12.01 | 26.9 | 88.6 | 0.44 |
| - w/o shared layers | 13.92 | 24.9 | 83.5 | 0.35 |
| - w/o joint training | 14.05 | 24.7 | 83.8 | 0.36 |
| No persona | I don’t know what you could not do? |
| PE w/ self | I am going to the club now. |
| PE w/ their | Do you want to play frisbee or something? |
| PG w/ self | okay I am going to make a cake. |
| - Generated Persona: | ... I craving eating cake... |
| PG w/ their | I prefer that let’s watch tv together. |
| - Generated Persona: | ... I like TV show... |
| Models | ROUGE-1 | ROUGE-2 | ROUGE-L | Extractor Size(M) |
| TextRank (Mihalcea and Tarau, 2004) | 16.27 | 2.69 | 15.41 | - |
| PGNet (See et al., 2017) | 28.74 | 5.98 | 25.13 | - |
| BART (Lewis et al., 2020) | 29.20 | 6.37 | 25.49 | - |
| LEAD + BART | 32.06 | 9.67 | 27.93 | - |
| HMNet (Zhu et al., 2020) | 32.29 | 8.67 | 28.17 | - |
| Longformer (Beltagy et al., 2020) | 34.18 | 10.32 | 29.95 | - |
| DialogLM (Zhong et al., 2021a) | 33.69 | 9.32 | 30.01 | - |
| SUMMN (Zhang et al., 2022) | 34.03 | 9.28 | 29.48 | - |
| DYLE (Mao et al., 2022) | 34.42 | 9.71 | 30.10 | 501 |
| Pointer Network + PGNet (Zhong et al., 2021b) | 31.37 | 8.47 | 27.08 | 440 |
| Pointer Network + BART (Zhong et al., 2021b) | 31.74 | 8.53 | 28.21 | 440 |
| RELREG-TT (Vig et al., 2022) | 33.02 | 10.17 | 28.90 | 329 |
| RELREG (Vig et al., 2022) | 34.91 | 11.91 | 30.73 | 1372 |
| Oracle | 43.80 | 19.63 | 39.10 | |
| Locator-Generator | 31.47(-3.77) | 8.53(-3.70) | 28.21(-3.07) | 134 |
| Simulator-Generator | 32.92(-2.59) | 9.46(-2.77) | 28.93(-2.35) | 134 |
| Ranker-Generator | 35.51 | 12.23 | 31.28 | 134 |
| RankSUM(w/o re-ranking) | 33.02(-2.49) | 9.73(-2.50) | 29.15(-2.13) | 134 |
| Models | Top 5 | Top 10 | ||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Gold | 26.32 | 7.58 | 24.43 | 20.55 | 5.15 | 19.29 |
| LEAD | 11.15 | 0.99 | 10.17 | 12.11 | 1.11 | 11.10 |
| RELREG | 18.02 | 2.46 | 15.30 | 15.02 | 2.35 | 13.23 |
| Locator | 16.89 | 2.24 | 13.97 | 14.10 | 1.97 | 12.75 |
| Simulator | 17.06 | 2.36 | 14.88 | 14.44 | 2.14 | 13.06 |
| Ours | 20.07 | 3.69 | 17.78 | 17.08 | 3.01 | 15.48 |
| Models | Flu. | QR. | FC. |
| Gold | 4.88 | 4.90 | 4.92 |
| BART | 4.48 | 3.78 | 3.64 |
| RELREG | 4.51 | 4.12 | 4.07 |
| Locator-Generator | 4.45 | 3.90 | 3.83 |
| Simulator-Generator | 4.48 | 4.01 | 4.02 |
| Ours | 4.52 | 4.40 | 4.21 |
| Query: What did Barry Hughes think about the targets regarding performance measure? |
| Gold: Barry Hughes stated that they did not have targets and only prosecuted with reasonable prospects. As for the police, there were no targets for the police anymore, except for some individuals. |
| Previous: Barry Hughes thought that there was a shortage of registered intermediaries in Wales, and the Ministry of Justice had taken action to deal with that, and they had had a number of people who were now in a position to act as intermediaries. Now, of course, if they were to decide not to do that anymore, they may have a problem, but, in turn, they would be looking to recruit more people into those positions. So, it had the potential to serve as a barrier, but in practice, he didn't think it would be a barrier. |
| Ours: Barry Hughes thought that the police were much more sensible these days than they may have been 20 years ago, in terms of trying to get cases charged in order to meet some notional target. In the CPS, they did not have targets for securing convictions or not. Clearly, they prosecuted if they thought that there was a reasonable prospect, a realistic prospect, of conviction and it was in the public interest. |
| Models | ROUGE-1 | ROUGE-2 | ROUGE-L | Extractor Size(M) |
| TextRank (Mihalcea and Tarau, 2004) | 16.27 | 2.69 | 15.41 | - |
| PGNet (See et al., 2017) | 28.74 | 5.98 | 25.13 | - |
| BART (Lewis et al., 2020) | 29.20 | 6.37 | 25.49 | - |
| LEAD + BART | 32.06 | 9.67 | 27.93 | - |
| HMNet (Zhu et al., 2020) | 32.29 | 8.67 | 28.17 | - |
| Longformer (Beltagy et al., 2020) | 34.18 | 10.32 | 29.95 | - |
| DialogLM (Zhong et al., 2021a) | 33.69 | 9.32 | 30.01 | - |
| SUMMN (Zhang et al., 2022) | 34.03 | 9.28 | 29.48 | - |
| DYLE (Mao et al., 2022) | 34.42 | 9.71 | 30.10 | 501 |
| Pointer Network + PGNet (Zhong et al., 2021b) | 31.37 | 8.47 | 27.08 | 440 |
| Pointer Network + BART (Zhong et al., 2021b) | 31.74 | 8.53 | 28.21 | 440 |
| RELREG-TT (Vig et al., 2022) | 33.02 | 10.17 | 28.90 | 329 |
| RELREG (Vig et al., 2022) | 34.91 | 11.91 | 30.73 | 1372 |
| Oracle | 43.80 | 19.63 | 39.10 | |
| Locator-Generator | 31.47(-3.77) | 8.53(-3.70) | 28.21(-3.07) | 134 |
| Simulator-Generator | 32.92(-2.59) | 9.46(-2.77) | 28.93(-2.35) | 134 |
| Ranker-Generator | 35.51 | 12.23 | 31.28 | 134 |
| RankSUM(w/o re-ranking) | 33.02(-2.49) | 9.73(-2.50) | 29.15(-2.13) | 134 |
| Models | Top 5 | Top 10 | ||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Gold | 26.32 | 7.58 | 24.43 | 20.55 | 5.15 | 19.29 |
| LEAD | 11.15 | 0.99 | 10.17 | 12.11 | 1.11 | 11.10 |
| RELREG | 18.02 | 2.46 | 15.30 | 15.02 | 2.35 | 13.23 |
| Locator | 16.89 | 2.24 | 13.97 | 14.10 | 1.97 | 12.75 |
| Simulator | 17.06 | 2.36 | 14.88 | 14.44 | 2.14 | 13.06 |
| Ours | 20.07 | 3.69 | 17.78 | 17.08 | 3.01 | 15.48 |
| Models | Flu. | QR. | FC. |
| Gold | 4.88 | 4.90 | 4.92 |
| BART | 4.48 | 3.78 | 3.64 |
| RELREG | 4.51 | 4.12 | 4.07 |
| Locator-Generator | 4.45 | 3.90 | 3.83 |
| Simulator-Generator | 4.48 | 4.01 | 4.02 |
| Ours | 4.52 | 4.40 | 4.21 |
| Query: What did Barry Hughes think about the targets regarding performance measure? |
| Gold: Barry Hughes stated that they did not have targets and only prosecuted with reasonable prospects. As for the police, there were no targets for the police anymore, except for some individuals. |
| Previous: Barry Hughes thought that there was a shortage of registered intermediaries in Wales, and the Ministry of Justice had taken action to deal with that, and they had had a number of people who were now in a position to act as intermediaries. Now, of course, if they were to decide not to do that anymore, they may have a problem, but, in turn, they would be looking to recruit more people into those positions. So, it had the potential to serve as a barrier, but in practice, he didn't think it would be a barrier. |
| Ours: Barry Hughes thought that the police were much more sensible these days than they may have been 20 years ago, in terms of trying to get cases charged in order to meet some notional target. In the CPS, they did not have targets for securing convictions or not. Clearly, they prosecuted if they thought that there was a reasonable prospect, a realistic prospect, of conviction and it was in the public interest. |
| Models | ROUGE-1 | ROUGE-2 | ROUGE-L | Extractor Size(M) |
| TextRank (Mihalcea and Tarau, 2004) | 16.27 | 2.69 | 15.41 | - |
| PGNet (See et al., 2017) | 28.74 | 5.98 | 25.13 | - |
| BART (Lewis et al., 2020) | 29.20 | 6.37 | 25.49 | - |
| LEAD + BART | 32.06 | 9.67 | 27.93 | - |
| HMNet (Zhu et al., 2020) | 32.29 | 8.67 | 28.17 | - |
| Longformer (Beltagy et al., 2020) | 34.18 | 10.32 | 29.95 | - |
| DialogLM (Zhong et al., 2021a) | 33.69 | 9.32 | 30.01 | - |
| SUMMN (Zhang et al., 2022) | 34.03 | 9.28 | 29.48 | - |
| DYLE (Mao et al., 2022) | 34.42 | 9.71 | 30.10 | 501 |
| Pointer Network + PGNet (Zhong et al., 2021b) | 31.37 | 8.47 | 27.08 | 440 |
| Pointer Network + BART (Zhong et al., 2021b) | 31.74 | 8.53 | 28.21 | 440 |
| RELREG-TT (Vig et al., 2022) | 33.02 | 10.17 | 28.90 | 329 |
| RELREG (Vig et al., 2022) | 34.91 | 11.91 | 30.73 | 1372 |
| Oracle | 43.80 | 19.63 | 39.10 | |
| Locator-Generator | 31.47(-3.77) | 8.53(-3.70) | 28.21(-3.07) | 134 |
| Simulator-Generator | 32.92(-2.59) | 9.46(-2.77) | 28.93(-2.35) | 134 |
| Ranker-Generator | 35.51 | 12.23 | 31.28 | 134 |
| RankSUM(w/o re-ranking) | 33.02(-2.49) | 9.73(-2.50) | 29.15(-2.13) | 134 |
| Models | Top 5 | Top 10 | ||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Gold | 26.32 | 7.58 | 24.43 | 20.55 | 5.15 | 19.29 |
| LEAD | 11.15 | 0.99 | 10.17 | 12.11 | 1.11 | 11.10 |
| RELREG | 18.02 | 2.46 | 15.30 | 15.02 | 2.35 | 13.23 |
| Locator | 16.89 | 2.24 | 13.97 | 14.10 | 1.97 | 12.75 |
| Simulator | 17.06 | 2.36 | 14.88 | 14.44 | 2.14 | 13.06 |
| Ours | 20.07 | 3.69 | 17.78 | 17.08 | 3.01 | 15.48 |
| Models | Flu. | QR. | FC. |
| Gold | 4.88 | 4.90 | 4.92 |
| BART | 4.48 | 3.78 | 3.64 |
| RELREG | 4.51 | 4.12 | 4.07 |
| Locator-Generator | 4.45 | 3.90 | 3.83 |
| Simulator-Generator | 4.48 | 4.01 | 4.02 |
| Ours | 4.52 | 4.40 | 4.21 |
| Query: What did Barry Hughes think about the targets regarding performance measure? |
| Gold: Barry Hughes stated that they did not have targets and only prosecuted with reasonable prospects. As for the police, there were no targets for the police anymore, except for some individuals. |
| Previous: Barry Hughes thought that there was a shortage of registered intermediaries in Wales, and the Ministry of Justice had taken action to deal with that, and they had had a number of people who were now in a position to act as intermediaries. Now, of course, if they were to decide not to do that anymore, they may have a problem, but, in turn, they would be looking to recruit more people into those positions. So, it had the potential to serve as a barrier, but in practice, he didn't think it would be a barrier. |
| Ours: Barry Hughes thought that the police were much more sensible these days than they may have been 20 years ago, in terms of trying to get cases charged in order to meet some notional target. In the CPS, they did not have targets for securing convictions or not. Clearly, they prosecuted if they thought that there was a reasonable prospect, a realistic prospect, of conviction and it was in the public interest. |
| Models | ROUGE-1 | ROUGE-2 | ROUGE-L | Extractor Size(M) |
| TextRank (Mihalcea and Tarau, 2004) | 16.27 | 2.69 | 15.41 | - |
| PGNet (See et al., 2017) | 28.74 | 5.98 | 25.13 | - |
| BART (Lewis et al., 2020) | 29.20 | 6.37 | 25.49 | - |
| LEAD + BART | 32.06 | 9.67 | 27.93 | - |
| HMNet (Zhu et al., 2020) | 32.29 | 8.67 | 28.17 | - |
| Longformer (Beltagy et al., 2020) | 34.18 | 10.32 | 29.95 | - |
| DialogLM (Zhong et al., 2021a) | 33.69 | 9.32 | 30.01 | - |
| SUMMN (Zhang et al., 2022) | 34.03 | 9.28 | 29.48 | - |
| DYLE (Mao et al., 2022) | 34.42 | 9.71 | 30.10 | 501 |
| Pointer Network + PGNet (Zhong et al., 2021b) | 31.37 | 8.47 | 27.08 | 440 |
| Pointer Network + BART (Zhong et al., 2021b) | 31.74 | 8.53 | 28.21 | 440 |
| RELREG-TT (Vig et al., 2022) | 33.02 | 10.17 | 28.90 | 329 |
| RELREG (Vig et al., 2022) | 34.91 | 11.91 | 30.73 | 1372 |
| Oracle | 43.80 | 19.63 | 39.10 | |
| Locator-Generator | 31.47(-3.77) | 8.53(-3.70) | 28.21(-3.07) | 134 |
| Simulator-Generator | 32.92(-2.59) | 9.46(-2.77) | 28.93(-2.35) | 134 |
| Ranker-Generator | 35.51 | 12.23 | 31.28 | 134 |
| RankSUM(w/o re-ranking) | 33.02(-2.49) | 9.73(-2.50) | 29.15(-2.13) | 134 |
| Models | Top 5 | Top 10 | ||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Gold | 26.32 | 7.58 | 24.43 | 20.55 | 5.15 | 19.29 |
| LEAD | 11.15 | 0.99 | 10.17 | 12.11 | 1.11 | 11.10 |
| RELREG | 18.02 | 2.46 | 15.30 | 15.02 | 2.35 | 13.23 |
| Locator | 16.89 | 2.24 | 13.97 | 14.10 | 1.97 | 12.75 |
| Simulator | 17.06 | 2.36 | 14.88 | 14.44 | 2.14 | 13.06 |
| Ours | 20.07 | 3.69 | 17.78 | 17.08 | 3.01 | 15.48 |
| Models | Flu. | QR. | FC. |
| Gold | 4.88 | 4.90 | 4.92 |
| BART | 4.48 | 3.78 | 3.64 |
| RELREG | 4.51 | 4.12 | 4.07 |
| Locator-Generator | 4.45 | 3.90 | 3.83 |
| Simulator-Generator | 4.48 | 4.01 | 4.02 |
| Ours | 4.52 | 4.40 | 4.21 |
| Query: What did Barry Hughes think about the targets regarding performance measure? |
| Gold: Barry Hughes stated that they did not have targets and only prosecuted with reasonable prospects. As for the police, there were no targets for the police anymore, except for some individuals. |
| Previous: Barry Hughes thought that there was a shortage of registered intermediaries in Wales, and the Ministry of Justice had taken action to deal with that, and they had had a number of people who were now in a position to act as intermediaries. Now, of course, if they were to decide not to do that anymore, they may have a problem, but, in turn, they would be looking to recruit more people into those positions. So, it had the potential to serve as a barrier, but in practice, he didn't think it would be a barrier. |
| Ours: Barry Hughes thought that the police were much more sensible these days than they may have been 20 years ago, in terms of trying to get cases charged in order to meet some notional target. In the CPS, they did not have targets for securing convictions or not. Clearly, they prosecuted if they thought that there was a reasonable prospect, a realistic prospect, of conviction and it was in the public interest. |
| Model | Param. | X → En | X → Zh | X → De | X → Ar | X → Ne | X → Az | X → Ceb | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 15.7 | 7.2 | 11.2 | 4.6 | 0.6 | 3.0 | 3.1 | 6.5 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 21.2 | 10.3 | 14.2 | 11.5 | 1.3 | 2.4 | 4.9 | 9.4 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 21.6 | 11.0 | 16.1 | 8.8 | 1.0 | 4.7 | 5.3 | 9.8 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 26.3 | 12.9 | 19.3 | 8.1 | 1.4 | 4.6 | 6.8 | 11.3 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 28.0 | 13.3 | 21.3 | 15.1 | 2.9 | 6.4 | 8.8 | 13.7 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 27.9 | 13.0 | 19.5 | 17.2 | 5.5 | 4.2 | 0.5 | 12.5 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 27.4 | 13.9 | 20.9 | 15.2 | 12.1 | 9.4 | 10.3 | 15.6 |
| 8: Lego-MT | × 1.0 | 30.7 | 16.4 | 23.8 | 18.2 | 15.0 | 11.9 | 15.1 | 18.7 |
| Model | Param. | En → X | Zh → X | De → X | Ar → X | Ne → X | Az → X | Ceb → X | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 12.7 | 7.8 | 11.6 | 6.9 | 2.2 | 2.8 | 5.4 | 7.1 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 17.3 | 10.1 | 14.1 | 11.5 | 4.0 | 4.2 | 6.1 | 9.6 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 18.0 | 11.1 | 15.6 | 11.2 | 5.2 | 4.3 | 7.9 | 10.5 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 21.5 | 13.1 | 17.7 | 12.6 | 7.1 | 6.1 | 9.5 | 12.5 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 24.7 | 14.9 | 20.3 | 16.4 | 9.7 | 6.2 | 12.5 | 15.0 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 21.3 | 10.9 | 15.8 | 14.9 | 3.9 | 3.0 | 1.5 | 10.2 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 21.8 | 13.5 | 18.4 | 14.7 | 13.4 | 11.1 | 12.4 | 15.0 |
| 8: Lego-MT | × 1.0 | 25.0 | 16.3 | 21.4 | 18.4 | 17.0 | 13.5 | 16.8 | 18.3 |
| Method | #Tokens | Size (GB) | Time (Hour) |
| Multi-Way Training | 3,000 | 60.9 | 16.9 |
| Lego-MT Training | 3,000 | 10.3 | 0.6 |
| Lang | Mix-Flow | Dec-Flow + Mix-Flow | Enc-Flow + Mix-Flow | |||
| x→zh | zh→x | x→zh | zh→x | x→zh | zh→x | |
| Be | 8.1 | 3.5 | 7.5 | 3.3 | 13.1 | 6.1 |
| De | 16.0 | 14.7 | 14.0 | 13.4 | 22.1 | 19.2 |
| Fa | 13.1 | 11.9 | 11.6 | 11.3 | 17.7 | 15.1 |
| Jv | 6.3 | 3.1 | 6.0 | 3.1 | 8.8 | 3.3 |
| Ne | 9.6 | 5.6 | 8.7 | 4.2 | 7.2 | 4.0 |
| Pa | 1.4 | 1.1 | 1.2 | 0.5 | 1.8 | 0.5 |
| Sw | 8.5 | 8.6 | 7.5 | 6.6 | 12.4 | 12.4 |
| AVG. | 9.0 | 6.9 | 8.1 | 6.1 | 11.9 | 8.7 |
| Model | Ceb→Ha | Ceb→Ig | Ceb→Ln | Ceb→Yo | AVG. |
| M-1.2B | 5.5 | 5.9 | 0.9 | 2.4 | 3.7 |
| M-12B | 8.7 | 10.8 | 0.9 | 2.9 | 5.8 |
| M-FT | 6.4 | 6.6 | 0.8 | 2.1 | 4.0 |
| Lego-MT | 12.5 | 13.9 | 2.3 | 3.2 | 8.0 |
| Model | Ha→Ceb | Ig→Ceb | Ln→Ceb | Yo→Ceb | AVG. |
| M-1.2B | 7.4 | 7.5 | 4.2 | 3.4 | 5.6 |
| M-12B | 8.8 | 8.8 | 3.8 | 4.1 | 6.4 |
| M-FT | 7.5 | 8.6 | 3.7 | 4.8 | 6.2 |
| Lego-MT | 12.3 | 12.3 | 6.2 | 6.7 | 9.4 |
| Model | X→Ast | X→Da | X→Hu | X→Lo | AVG. |
| M-1.2B | 16.7 | 22.0 | 17.7 | 4.8 | 15.3 |
| M-12B | 13.0 | 23.3 | 19.1 | 9.0 | 16.1 |
| M-FT | 13.8 | 23.2 | 17.8 | 0.9 | 13.9 |
| Lego-MT | 15.4 | 25.4 | 20.1 | 5.6 | 16.6 |
| Model | Ast→X | Da→X | Hu→X | Lo→X | AVG. |
| M-1.2B | 13.2 | 18.3 | 16.0 | 6.6 | 13.5 |
| M-12B | 15.2 | 20.9 | 18.2 | 8.8 | 15.8 |
| M-FT | 14.4 | 17.9 | 15.5 | 5.8 | 13.4 |
| Lego-MT | 15.5 | 20.8 | 18.1 | 8.9 | 15.8 |
| Model | Ast→X | Hu→X | Da→X | Lo→X | En→X | De→X | Ar→X | Az→X | Ceb→X | Ne→X | Zh→X | AVG. |
| Multilingual FT | 3.8 | 2.6 | 2.7 | 6.3 | 7.6 | 4.2 | 1.7 | 2.6 | 0.9 | 6.2 | 3.7 | 3.9 |
| Lego-MT | 14.2 | 10.1 | 11.9 | 17.5 | 20.6 | 14.1 | 6.9 | 12.2 | 5.8 | 17.7 | 11.7 | 13.1 |
| Model | X→Ast | X→Hu | X→Da | X→Lo | X→En | X→De | X→Ar | X→Az | X→Ceb | X→Ne | X→Zh | AVG. |
| Multilingual FT | 4.9 | 2.2 | 1.6 | 6.9 | 10.8 | 3.0 | 0.5 | 2.7 | 0.7 | 7.8 | 4.3 | 4.1 |
| Lego-MT | 14.1 | 8.8 | 10.5 | 17.8 | 24.1 | 13.7 | 3.6 | 11.0 | 4.0 | 21.2 | 11.8 | 12.9 |
| Model | X→En | En→X | AVG. |
| ChatGPT zero-shot | 27.9 | 23.9 | 25.9 |
| ChatGPT eight-shot | 31.9 | 24.7 | 28.3 |
| Lego-MT | 30.2 | 25.7 | 28.0 |
| Original | Replaced | Original | Replaced | Original | Replaced |
| ak | aka | es | es_HN | pt | pt_BR |
| am | amh | es | es_EC | pt | pt_br |
| ar | ara | es | es_CO | pt | pt_PT |
| ar | ar_SY | fa | fa_IR | rn | run |
| ar | ar_TN | fa | fa_AF | rw | kin |
| ay | aym | ff | ful | sn | sna |
| az | az_IR | fr | fr_FR | so | som |
| bg | bg_BG | fr | fr_CA | sr | srp |
| bm | bam | fr | fr_BE | sr | sr_ME |
| bn | bn_IN | fr | fr_ca | st | sot |
| ca | cat | ha | hau | sw | swa |
| da | da_DK | hi | hi_IN | ta | ta_LK |
| de | de_CH | ig | ibo | tg | tg_TJ |
| de | de_AT | it | it_IT | ti | tir |
| de | de_DE | jp | jap | tl | tl_PH |
| es | es_CL | kr | kau | tr | tr_TR |
| es | es_SV | kv | kpv | ur | ur_PK |
| es | es_NI | ln | lin | vi | vi_VN |
| es | es_UY | mg | mlg | wo | wol |
| es | es_PE | ms | ms_MY | xh | xho |
| es | es_VE | nb | nb_NO | yo | yor |
| es | es_AR | nds | nds_nl | ze | ze_zh |
| es | es_MX | nl | nl_NL | ze | ze_en |
| es | es_PA | nl | nl_BE | zh | zh_cn |
| es | es_CR | nn | nn_NO | zh | zh_CN |
| es | es_PR | no | no_nb | zthrad | zh_HK |
| es | es_ES | ny | nya | zthrad | zh_TW |
| es | es_GT | om | orm | zthrad | zh_tw |
| es | es_DO | pa | pan | zu | zul |
| Code | En | De | Ar | Zh | Ne | Az | Ceb |
| #Sentence Pairs | 811,238,712 | 360,369,144 | 152,457,830 | 92,763,445 | 6,654,270 | 4,208,025 | 1,683,531 |
| Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset |
| crp | bible-uedin | cb | MultiCCAligned | sz | MultiCCAligned | sgn | QED | cycl | Tatoeba |
| tc | EUbookshop | cx | MultiCCAligned | zz | MultiCCAligned | iro | QED | nah | Tatoeba |
| zhs | GlobalVoices | ns | MultiCCAligned | ze | OpenSubtitles | mo | QED,Ubuntu | ||
| zht | GlobalVoices | qd | MultiCCAligned | bh | QED | ber | QED,Ubuntu | ||
| tmp | GNOME | qa | MultiCCAligned | bnt | QED | toki | Tatoeba | ||
| gr | GNOME | tz | MultiCCAligned | ry | QED | kzj | Tatoeba |
| Language | Code | Language | Code | Language | Code | Language | Code | Language | Code | Language | Code |
| Abkhazian | ab | Corsican | co | Iban | ibla | Lower Sorbian | dsb | Ossetian | os | Swahili (macrolanguage) | sw |
| Achinese | ace | Cree | cr | Icelandic | is | Lukpa | dop | Ottoman Turkish (1500-1928) | ota | Swti | ss |
| Achuar-Shiwiar | acu | Creek | mus | Ido | io | Luo (Kenya and Tanzania) | luo | Paite Chin | pck | Swedish | sv |
| Adyghe | ady | Crimean Tatar | chr | Igbo | ig | Lushootseed | lut | Palaun | pau | Swiss German | gsw |
| Afar | aa | Croatian | hr | Ilok | ilo | Luxembourgish | lb | Pali | pi | Syriac | syr |
| Afrihili | afh | Cusco Quechua | quz | Indonesian | id | Luyia | luy | Pampanga | pam | Tachawit | shy |
| Afrikaans | af | Czech | cs | Ingrian | izh | Macedonian | mk | Pangasinan | pag | Tachelih | shi |
| Aguaruna | agr | Danish | da | Inghush | inh | Macedo-Romanian | rup | Panjabi | pa | Tagal Murut | mvv |
| Ainu (Japan) | ain | Dari | prs | Interlingua | ia | Madurese | mad | Papiamento | pap | Tagalog | tl |
| Akan | ak | Dinka | din | Interlingue | ie | Maithili | mai | Papuan Malay | pmy | Tahaggart Tamahaq | thv |
| Akawaio | ake | Drents | drt | Inuktitut | iu | Malagasy | mg | Pedi | nso | Tahitian | ty |
| Akanlon | akl | Dungan | dng | Inupiaq | ik | Malay (individual language) | zlm | Pennsylvania German | pdc | Tajik | tg |
| Albanian | sq | Dutch | nl | Iranian Persian | pes | Malay (macrolanguage) | ms | Persian | fa | Talusssan | tzl |
| Algerian Arabic | arq | Dutton World Speedwords | dws | Irish | ga | Malayalam | ml | Phoenician | phn | Talysh | tly |
| American Sign Language | ase | Dzongkha | dz | Italian | it | Maltese | mt | Picard | ped | Tamashek | tmh |
| Amharic | am | Eastern Canadian Inuktitut | ike | Jakun | jak | Mam | mam | Piemontese | pms | Tamil | ta |
| Ancient Greek (to 1453) | grc | Eastern Mari | mhr | Jamaican Creole English | jam | Mambae | mgn | Pipil | ppl | Tarifit | rif |
| Ancient Hebrew | hbo | Eastern Maroon Creole | djk | Japanese | ja | Mandarin Chinese | cmn | Plateau Malagasy | plt | Tase Naga | nst |
| Arabic | ar | Efik | efi | Javanese | jv | Manx | gv | Polish | pl | Tatar | tt |
| Argonesse | an | Egyptian Arabic | artz | Jewish Babylonian Aramaic | tmr | Maori | mi | Portuguese | pt | Telugu | te |
| Armenian | hy | Emilian | egl | Kabyle | kab | Marathi | mr | Potawatomi | pot | Tena Lowland Quichua | qw |
| Arpitan | frp | English | en | Kadazan Dusun | dtp | Marshallese | mh | Prussian | prg | Tetelingino Nahuatl | nhg |
| Asháninka | cni | Erzya | myv | Kalaallusit | kl | Mesopotamian Arabic | acm | Pushto | ps | Tetum | tet |
| Assamese | as | Esperanto | eo | Kalmyk | xal | Miahuatlán Zapotec | zam | Quechua | qu | Thai | th |
| Asturian | ast | Estonian | et | Kamba (Kenya) | kam | Middle English (1100-1500) | emm | Quenya | qya | Tibetan | bo |
| Avaric | av | Evenki | evn | Kamada | kn | Middle French (ca. 1400-1600) | frm | Quitopepec Chinantech | chq | Tigrinya | ti |
| Avestan | ae | Ewe | ee | Kanuri | kr | Mikasuki | mik | Rapanui | rap | Tohono O'odham | ood |
| Awadiya | awa | Extremaduran | ext | Kaçchikel | cak | Mi'kmaq | mic | Romanian | ro | Tok Pisin | tpi |
| Aymara | ay | Faroeose | fo | Karalian | krl | Min Dong Chinese | cdo | Romansh | rn | Tonga (Tonga Islands) | to |
| Azerbaijan | az | Fiji Hindi | hif | Kashmiri | ks | Min Nan Chinese | nan | Romany | rom | Traditional Chinese | zhtrad |
| Baluchi | bal | Fijian | fj | Kashubian | csb | Minangkabau | min | Rundi | rn | Tsonga | ts |
| Bambara | bm | Filipino | fil | Kazakh | kk | Mingrelian | xmf | Russian | ru | Tswana | tn |
| Banjar | bjn | Finnish | fi | Kekchi | kek | Mirandese | mwI | Rusyn | rue | Tupi | tpw |
| Barasana-Eduria | bsn | French | fr | Khakasi | kjh | Miskito | miq | Samoan | sm | Turkish | tr |
| Bashkir | ba | Friulian | fur | Khasi | kha | Modern Greek (1453-) | el | Samogtian | sgs | Turkmen | tk |
| Basque | eu | Fulah | ff | Khemer | km | Mohawk | moh | Sango | sg | Tuvalu | tvl |
| Bavarian | bar | Galela | gbi | K'iche' | quc | Mongolian | mn | Sanskrit | sa | Twi | tw |
| Baybayanon | bvy | Galician | gl | Kikuyu | kik | Morisieny | mfe | Santali | sat | Uab Meto | aoz |
| Belarusian | be | Gan Chinese | gan | Kinyarwanda | rw | Moroccan Arabic | ary | Sardinian | sc | Udmurt | udm |
| Bemba (Zambia) | bem | Ganda | lg | Kirghiz | ky | Mossi | mos | Satterffriesisch | stq | Uighur | ug |
| Bengali | bn | Garhwali | gbm | Klingon | thh | Nauru | na | Scots | sco | Ukrainian | uk |
| Berom | born | Georgian | ka | Koaati | ku | Navajo | nv | Scottish Gaelic | gd | Uma | ppk |
| Bhojpuri | bho | German | de | Kölsch | ksh | Neapolitan | nap | Sediq | trv | Umbundu | umb |
| Bislama | bi | Gheg Albanian | aln | Komii | kv | Nepali (individual language) | npi | Serbian | sr | Upper Sorbian | hsb |
| Bodo (India) | brx | Gilbertese | gil | Komii-Permyak | koi | Nepali (macrolanguage) | ne | Serbo-Croatian | sh | Urdu | ur |
| Bosnian | bs | Goan Konkani | gom | Kongo | kg | Nigerian Fulfulde | fuv | Shan | shn | Uspantecco | usp |
| Breton | br | Gothic | got | Korean | ko | Niuean | niu | Shona | sn | Uzbek | uz |
| Brithening | bzt | Gronings | gos | Kotava | avk | Nogai | nog | Shuar | jv | Venda | ve |
| Buginese | bug | Guadeloupean Creole French | gcf | Kriang | ngt | North Levantine Arabic | apc | Schuswap | shs | Venetian | vec |
| Bulgarian | bg | Guarani | gn | Kuanyama | kj | North Molucan Malay | max | Sicilian | scn | Vietnamese | vi |
| Buraiti | bua | Guerrero Amuzgo | amu | Kurdish | ku | Northern Frisian | frr | Silesian | szl | Vlaams | vls |
| Burmese | my | Guerrero Nahuatl | ngu | Kven Finnish | fkv | Northern Kurdish | kmr | Sindarin | sjn | Volapuk | vo |
| Cabecar | cjp | Gujarat | gu | Laidan | ldn | Northern Sami | se | Sindi | sd | Walloon | wa |
| Camsé | kbh | Gulf Arabic | afb | Ladin | lld | Northwestern Ojibwa | obj | Sinhala | si | Walser | wae |
| Catalan | ca | Haida | hai | Ladino | lad | Norwegian | no | Slovak | sk | Waray (Philippines) | war |
| Cebuano | ceb | Haitian | ht | Lakota | lkt | Norwegian Bokmål | nb | Slovenian | sl | Welsh | cy |
| Central Huasteca Nahuatl | nch | Hakka Chin | cnh | Lao | lo | Norwegian Nynorsk | nn | Somali | so | Western Frisan | fy |
| Central Kurdish | ckb | Hakka Chinese | hak | Latgalian | ltg | Novial | nov | South Azerbaijani | azb | Western Panjabi | pnb |
| Central Sama | sml | Hausa | ha | Latin | la | Nuer | nus | South Ndebele | nr | Wolayta | wal |
| Chamorro | ch | Hawaiian | haw | Latvian | lv | Nyanja | ny | Southern Kurdish | sdh | Wolof | wo |
| Chavacano | cbk | Hebrew | he | Ligurian | lij | Occitan (post 1500) | oc | Southern Sami | sma | Wu Chinese | wuu |
| Chechen | ce | Hilgaynon | hil | Limburgan | li | Old English (ca. 450-1100) | ang | Southern Sotho | st | Xhosa | xh |
| Cherokee | chr | Hindi | hi | Lingala | ln | Old French (842-ca. 1400) | fro | Southwestern Dinka | dik | Yakut | sah |
| Chhattisgarhi | hne | Hiri Motu | ho | Lingua Franca Nova | lfn | Old Frisan | ofs | Spanish | es | Yaqui | yqa |
| Chinese | zh | Hmong Daw | mww | Literary Chinese | lzh | Old Norse | non | Standard Malay | zsm | Yiddish | yi |
| Choctaw | cho | Ho | hoc | Lithuanian | lt | Old Russian | ovr | Standard Moroccan Tamazight | zgh | Yoruba | yo |
| Church Slavic | cu | Huastec | hus | Liv | liv | Old Spanish | osp | Sumerian | sux | Zarma | dje |
| Chuvash | cv | Hungarian | hu | Lojban | jbo | Oriya (macrolanguage) | or | Sundanese | su | Zaza | za |
| Coptic | cop | Hunsrik | hrx | Lomburd | lmo | Orizaba Nahuatl | nlv | Swaiban | swg | Zulu | zu |
| Cornish | kw | Hupa | hup | Low German | nds | Oromo | om | Swahili (individual language) | swh |
| Model | Param. | X → En | X → Zh | X → De | X → Ar | X → Ne | X → Az | X → Ceb | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 15.7 | 7.2 | 11.2 | 4.6 | 0.6 | 3.0 | 3.1 | 6.5 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 21.2 | 10.3 | 14.2 | 11.5 | 1.3 | 2.4 | 4.9 | 9.4 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 21.6 | 11.0 | 16.1 | 8.8 | 1.0 | 4.7 | 5.3 | 9.8 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 26.3 | 12.9 | 19.3 | 8.1 | 1.4 | 4.6 | 6.8 | 11.3 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 28.0 | 13.3 | 21.3 | 15.1 | 2.9 | 6.4 | 8.8 | 13.7 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 27.9 | 13.0 | 19.5 | 17.2 | 5.5 | 4.2 | 0.5 | 12.5 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 27.4 | 13.9 | 20.9 | 15.2 | 12.1 | 9.4 | 10.3 | 15.6 |
| 8: Lego-MT | × 1.0 | 30.7 | 16.4 | 23.8 | 18.2 | 15.0 | 11.9 | 15.1 | 18.7 |
| Model | Param. | En → X | Zh → X | De → X | Ar → X | Ne → X | Az → X | Ceb → X | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 12.7 | 7.8 | 11.6 | 6.9 | 2.2 | 2.8 | 5.4 | 7.1 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 17.3 | 10.1 | 14.1 | 11.5 | 4.0 | 4.2 | 6.1 | 9.6 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 18.0 | 11.1 | 15.6 | 11.2 | 5.2 | 4.3 | 7.9 | 10.5 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 21.5 | 13.1 | 17.7 | 12.6 | 7.1 | 6.1 | 9.5 | 12.5 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 24.7 | 14.9 | 20.3 | 16.4 | 9.7 | 6.2 | 12.5 | 15.0 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 21.3 | 10.9 | 15.8 | 14.9 | 3.9 | 3.0 | 1.5 | 10.2 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 21.8 | 13.5 | 18.4 | 14.7 | 13.4 | 11.1 | 12.4 | 15.0 |
| 8: Lego-MT | × 1.0 | 25.0 | 16.3 | 21.4 | 18.4 | 17.0 | 13.5 | 16.8 | 18.3 |
| Method | #Tokens | Size (GB) | Time (Hour) |
| Multi-Way Training | 3,000 | 60.9 | 16.9 |
| Lego-MT Training | 3,000 | 10.3 | 0.6 |
| Lang | Mix-Flow | Dec-Flow + Mix-Flow | Enc-Flow + Mix-Flow | |||
| x→zh | zh→x | x→zh | zh→x | x→zh | zh→x | |
| Be | 8.1 | 3.5 | 7.5 | 3.3 | 13.1 | 6.1 |
| De | 16.0 | 14.7 | 14.0 | 13.4 | 22.1 | 19.2 |
| Fa | 13.1 | 11.9 | 11.6 | 11.3 | 17.7 | 15.1 |
| Jv | 6.3 | 3.1 | 6.0 | 3.1 | 8.8 | 3.3 |
| Ne | 9.6 | 5.6 | 8.7 | 4.2 | 7.2 | 4.0 |
| Pa | 1.4 | 1.1 | 1.2 | 0.5 | 1.8 | 0.5 |
| Sw | 8.5 | 8.6 | 7.5 | 6.6 | 12.4 | 12.4 |
| AVG. | 9.0 | 6.9 | 8.1 | 6.1 | 11.9 | 8.7 |
| Model | Ceb→Ha | Ceb→Ig | Ceb→Ln | Ceb→Yo | AVG. |
| M-1.2B | 5.5 | 5.9 | 0.9 | 2.4 | 3.7 |
| M-12B | 8.7 | 10.8 | 0.9 | 2.9 | 5.8 |
| M-FT | 6.4 | 6.6 | 0.8 | 2.1 | 4.0 |
| Lego-MT | 12.5 | 13.9 | 2.3 | 3.2 | 8.0 |
| Model | Ha→Ceb | Ig→Ceb | Ln→Ceb | Yo→Ceb | AVG. |
| M-1.2B | 7.4 | 7.5 | 4.2 | 3.4 | 5.6 |
| M-12B | 8.8 | 8.8 | 3.8 | 4.1 | 6.4 |
| M-FT | 7.5 | 8.6 | 3.7 | 4.8 | 6.2 |
| Lego-MT | 12.3 | 12.3 | 6.2 | 6.7 | 9.4 |
| Model | X→Ast | X→Da | X→Hu | X→Lo | AVG. |
| M-1.2B | 16.7 | 22.0 | 17.7 | 4.8 | 15.3 |
| M-12B | 13.0 | 23.3 | 19.1 | 9.0 | 16.1 |
| M-FT | 13.8 | 23.2 | 17.8 | 0.9 | 13.9 |
| Lego-MT | 15.4 | 25.4 | 20.1 | 5.6 | 16.6 |
| Model | Ast→X | Da→X | Hu→X | Lo→X | AVG. |
| M-1.2B | 13.2 | 18.3 | 16.0 | 6.6 | 13.5 |
| M-12B | 15.2 | 20.9 | 18.2 | 8.8 | 15.8 |
| M-FT | 14.4 | 17.9 | 15.5 | 5.8 | 13.4 |
| Lego-MT | 15.5 | 20.8 | 18.1 | 8.9 | 15.8 |
| Model | Ast→X | Hu→X | Da→X | Lo→X | En→X | De→X | Ar→X | Az→X | Ceb→X | Ne→X | Zh→X | AVG. |
| Multilingual FT | 3.8 | 2.6 | 2.7 | 6.3 | 7.6 | 4.2 | 1.7 | 2.6 | 0.9 | 6.2 | 3.7 | 3.9 |
| Lego-MT | 14.2 | 10.1 | 11.9 | 17.5 | 20.6 | 14.1 | 6.9 | 12.2 | 5.8 | 17.7 | 11.7 | 13.1 |
| Model | X→Ast | X→Hu | X→Da | X→Lo | X→En | X→De | X→Ar | X→Az | X→Ceb | X→Ne | X→Zh | AVG. |
| Multilingual FT | 4.9 | 2.2 | 1.6 | 6.9 | 10.8 | 3.0 | 0.5 | 2.7 | 0.7 | 7.8 | 4.3 | 4.1 |
| Lego-MT | 14.1 | 8.8 | 10.5 | 17.8 | 24.1 | 13.7 | 3.6 | 11.0 | 4.0 | 21.2 | 11.8 | 12.9 |
| Model | X→En | En→X | AVG. |
| ChatGPT zero-shot | 27.9 | 23.9 | 25.9 |
| ChatGPT eight-shot | 31.9 | 24.7 | 28.3 |
| Lego-MT | 30.2 | 25.7 | 28.0 |
| Original | Replaced | Original | Replaced | Original | Replaced |
| ak | aka | es | es_HN | pt | pt_BR |
| am | amh | es | es_EC | pt | pt_br |
| ar | ara | es | es_CO | pt | pt_PT |
| ar | ar_SY | fa | fa_IR | rn | run |
| ar | ar_TN | fa | fa_AF | rw | kin |
| ay | aym | ff | ful | sn | sna |
| az | az_IR | fr | fr_FR | so | som |
| bg | bg_BG | fr | fr_CA | sr | srp |
| bm | bam | fr | fr_BE | sr | sr_ME |
| bn | bn_IN | fr | fr_ca | st | sot |
| ca | cat | ha | hau | sw | swa |
| da | da_DK | hi | hi_IN | ta | ta_LK |
| de | de_CH | ig | ibo | tg | tg_TJ |
| de | de_AT | it | it_IT | ti | tir |
| de | de_DE | jp | jap | tl | tl_PH |
| es | es_CL | kr | kau | tr | tr_TR |
| es | es_SV | kv | kpv | ur | ur_PK |
| es | es_NI | ln | lin | vi | vi_VN |
| es | es_UY | mg | mlg | wo | wol |
| es | es_PE | ms | ms_MY | xh | xho |
| es | es_VE | nb | nb_NO | yo | yor |
| es | es_AR | nds | nds_nl | ze | ze_zh |
| es | es_MX | nl | nl_NL | ze | ze_en |
| es | es_PA | nl | nl_BE | zh | zh_cn |
| es | es_CR | nn | nn_NO | zh | zh_CN |
| es | es_PR | no | no_nb | zthrad | zh_HK |
| es | es_ES | ny | nya | zthrad | zh_TW |
| es | es_GT | om | orm | zthrad | zh_tw |
| es | es_DO | pa | pan | zu | zul |
| Code | En | De | Ar | Zh | Ne | Az | Ceb |
| #Sentence Pairs | 811,238,712 | 360,369,144 | 152,457,830 | 92,763,445 | 6,654,270 | 4,208,025 | 1,683,531 |
| Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset |
| crp | bible-uedin | cb | MultiCCAligned | sz | MultiCCAligned | sgn | QED | cycl | Tatoeba |
| tc | EUbookshop | cx | MultiCCAligned | zz | MultiCCAligned | iro | QED | nah | Tatoeba |
| zhs | GlobalVoices | ns | MultiCCAligned | ze | OpenSubtitles | mo | QED,Ubuntu | ||
| zht | GlobalVoices | qd | MultiCCAligned | bh | QED | ber | QED,Ubuntu | ||
| tmp | GNOME | qa | MultiCCAligned | bnt | QED | toki | Tatoeba | ||
| gr | GNOME | tz | MultiCCAligned | ry | QED | kzj | Tatoeba |
| Language | Code | Language | Code | Language | Code | Language | Code | Language | Code | Language | Code |
| Abkhazian | ab | Corsican | co | Iban | ibla | Lower Sorbian | dsb | Ossetian | os | Swahili (macrolanguage) | sw |
| Achinese | ace | Cree | cr | Icelandic | is | Lukpa | dop | Ottoman Turkish (1500-1928) | ota | Swti | ss |
| Achuar-Shiwiar | acu | Creek | mus | Ido | io | Luo (Kenya and Tanzania) | luo | Paite Chin | pck | Swedish | sv |
| Adyghe | ady | Crimean Tatar | chr | Igbo | ig | Lushootseed | lut | Palaun | pau | Swiss German | gsw |
| Afar | aa | Croatian | hr | Ilok | ilo | Luxembourgish | lb | Pali | pi | Syriac | syr |
| Afrihili | afh | Cusco Quechua | quz | Indonesian | id | Luyia | luy | Pampanga | pam | Tachawit | shy |
| Afrikaans | af | Czech | cs | Ingrian | izh | Macedonian | mk | Pangasinan | pag | Tachelih | shi |
| Aguaruna | agr | Danish | da | Inghush | inh | Macedo-Romanian | rup | Panjabi | pa | Tagal Murut | mvv |
| Ainu (Japan) | ain | Dari | prs | Interlingua | ia | Madurese | mad | Papiamento | pap | Tagalog | tl |
| Akan | ak | Dinka | din | Interlingue | ie | Maithili | mai | Papuan Malay | pmy | Tahaggart Tamahaq | thv |
| Akawaio | ake | Drents | drt | Inuktitut | iu | Malagasy | mg | Pedi | nso | Tahitian | ty |
| Akanlon | akl | Dungan | dng | Inupiaq | ik | Malay (individual language) | zlm | Pennsylvania German | pdc | Tajik | tg |
| Albanian | sq | Dutch | nl | Iranian Persian | pes | Malay (macrolanguage) | ms | Persian | fa | Talusssan | tzl |
| Algerian Arabic | arq | Dutton World Speedwords | dws | Irish | ga | Malayalam | ml | Phoenician | phn | Talysh | tly |
| American Sign Language | ase | Dzongkha | dz | Italian | it | Maltese | mt | Picard | ped | Tamashek | tmh |
| Amharic | am | Eastern Canadian Inuktitut | ike | Jakun | jak | Mam | mam | Piemontese | pms | Tamil | ta |
| Ancient Greek (to 1453) | grc | Eastern Mari | mhr | Jamaican Creole English | jam | Mambae | mgn | Pipil | ppl | Tarifit | rif |
| Ancient Hebrew | hbo | Eastern Maroon Creole | djk | Japanese | ja | Mandarin Chinese | cmn | Plateau Malagasy | plt | Tase Naga | nst |
| Arabic | ar | Efik | efi | Javanese | jv | Manx | gv | Polish | pl | Tatar | tt |
| Argonesse | an | Egyptian Arabic | artz | Jewish Babylonian Aramaic | tmr | Maori | mi | Portuguese | pt | Telugu | te |
| Armenian | hy | Emilian | egl | Kabyle | kab | Marathi | mr | Potawatomi | pot | Tena Lowland Quichua | qw |
| Arpitan | frp | English | en | Kadazan Dusun | dtp | Marshallese | mh | Prussian | prg | Tetelingino Nahuatl | nhg |
| Asháninka | cni | Erzya | myv | Kalaallusit | kl | Mesopotamian Arabic | acm | Pushto | ps | Tetum | tet |
| Assamese | as | Esperanto | eo | Kalmyk | xal | Miahuatlán Zapotec | zam | Quechua | qu | Thai | th |
| Asturian | ast | Estonian | et | Kamba (Kenya) | kam | Middle English (1100-1500) | emm | Quenya | qya | Tibetan | bo |
| Avaric | av | Evenki | evn | Kamada | kn | Middle French (ca. 1400-1600) | frm | Quitopepec Chinantech | chq | Tigrinya | ti |
| Avestan | ae | Ewe | ee | Kanuri | kr | Mikasuki | mik | Rapanui | rap | Tohono O'odham | ood |
| Awadiya | awa | Extremaduran | ext | Kaçchikel | cak | Mi'kmaq | mic | Romanian | ro | Tok Pisin | tpi |
| Aymara | ay | Faroeose | fo | Karalian | krl | Min Dong Chinese | cdo | Romansh | rn | Tonga (Tonga Islands) | to |
| Azerbaijan | az | Fiji Hindi | hif | Kashmiri | ks | Min Nan Chinese | nan | Romany | rom | Traditional Chinese | zhtrad |
| Baluchi | bal | Fijian | fj | Kashubian | csb | Minangkabau | min | Rundi | rn | Tsonga | ts |
| Bambara | bm | Filipino | fil | Kazakh | kk | Mingrelian | xmf | Russian | ru | Tswana | tn |
| Banjar | bjn | Finnish | fi | Kekchi | kek | Mirandese | mwI | Rusyn | rue | Tupi | tpw |
| Barasana-Eduria | bsn | French | fr | Khakasi | kjh | Miskito | miq | Samoan | sm | Turkish | tr |
| Bashkir | ba | Friulian | fur | Khasi | kha | Modern Greek (1453-) | el | Samogtian | sgs | Turkmen | tk |
| Basque | eu | Fulah | ff | Khemer | km | Mohawk | moh | Sango | sg | Tuvalu | tvl |
| Bavarian | bar | Galela | gbi | K'iche' | quc | Mongolian | mn | Sanskrit | sa | Twi | tw |
| Baybayanon | bvy | Galician | gl | Kikuyu | kik | Morisieny | mfe | Santali | sat | Uab Meto | aoz |
| Belarusian | be | Gan Chinese | gan | Kinyarwanda | rw | Moroccan Arabic | ary | Sardinian | sc | Udmurt | udm |
| Bemba (Zambia) | bem | Ganda | lg | Kirghiz | ky | Mossi | mos | Satterffriesisch | stq | Uighur | ug |
| Bengali | bn | Garhwali | gbm | Klingon | thh | Nauru | na | Scots | sco | Ukrainian | uk |
| Berom | born | Georgian | ka | Koaati | ku | Navajo | nv | Scottish Gaelic | gd | Uma | ppk |
| Bhojpuri | bho | German | de | Kölsch | ksh | Neapolitan | nap | Sediq | trv | Umbundu | umb |
| Bislama | bi | Gheg Albanian | aln | Komii | kv | Nepali (individual language) | npi | Serbian | sr | Upper Sorbian | hsb |
| Bodo (India) | brx | Gilbertese | gil | Komii-Permyak | koi | Nepali (macrolanguage) | ne | Serbo-Croatian | sh | Urdu | ur |
| Bosnian | bs | Goan Konkani | gom | Kongo | kg | Nigerian Fulfulde | fuv | Shan | shn | Uspantecco | usp |
| Breton | br | Gothic | got | Korean | ko | Niuean | niu | Shona | sn | Uzbek | uz |
| Brithening | bzt | Gronings | gos | Kotava | avk | Nogai | nog | Shuar | jv | Venda | ve |
| Buginese | bug | Guadeloupean Creole French | gcf | Kriang | ngt | North Levantine Arabic | apc | Schuswap | shs | Venetian | vec |
| Bulgarian | bg | Guarani | gn | Kuanyama | kj | North Molucan Malay | max | Sicilian | scn | Vietnamese | vi |
| Buraiti | bua | Guerrero Amuzgo | amu | Kurdish | ku | Northern Frisian | frr | Silesian | szl | Vlaams | vls |
| Burmese | my | Guerrero Nahuatl | ngu | Kven Finnish | fkv | Northern Kurdish | kmr | Sindarin | sjn | Volapuk | vo |
| Cabecar | cjp | Gujarat | gu | Laidan | ldn | Northern Sami | se | Sindi | sd | Walloon | wa |
| Camsé | kbh | Gulf Arabic | afb | Ladin | lld | Northwestern Ojibwa | obj | Sinhala | si | Walser | wae |
| Catalan | ca | Haida | hai | Ladino | lad | Norwegian | no | Slovak | sk | Waray (Philippines) | war |
| Cebuano | ceb | Haitian | ht | Lakota | lkt | Norwegian Bokmål | nb | Slovenian | sl | Welsh | cy |
| Central Huasteca Nahuatl | nch | Hakka Chin | cnh | Lao | lo | Norwegian Nynorsk | nn | Somali | so | Western Frisan | fy |
| Central Kurdish | ckb | Hakka Chinese | hak | Latgalian | ltg | Novial | nov | South Azerbaijani | azb | Western Panjabi | pnb |
| Central Sama | sml | Hausa | ha | Latin | la | Nuer | nus | South Ndebele | nr | Wolayta | wal |
| Chamorro | ch | Hawaiian | haw | Latvian | lv | Nyanja | ny | Southern Kurdish | sdh | Wolof | wo |
| Chavacano | cbk | Hebrew | he | Ligurian | lij | Occitan (post 1500) | oc | Southern Sami | sma | Wu Chinese | wuu |
| Chechen | ce | Hilgaynon | hil | Limburgan | li | Old English (ca. 450-1100) | ang | Southern Sotho | st | Xhosa | xh |
| Cherokee | chr | Hindi | hi | Lingala | ln | Old French (842-ca. 1400) | fro | Southwestern Dinka | dik | Yakut | sah |
| Chhattisgarhi | hne | Hiri Motu | ho | Lingua Franca Nova | lfn | Old Frisan | ofs | Spanish | es | Yaqui | yqa |
| Chinese | zh | Hmong Daw | mww | Literary Chinese | lzh | Old Norse | non | Standard Malay | zsm | Yiddish | yi |
| Choctaw | cho | Ho | hoc | Lithuanian | lt | Old Russian | ovr | Standard Moroccan Tamazight | zgh | Yoruba | yo |
| Church Slavic | cu | Huastec | hus | Liv | liv | Old Spanish | osp | Sumerian | sux | Zarma | dje |
| Chuvash | cv | Hungarian | hu | Lojban | jbo | Oriya (macrolanguage) | or | Sundanese | su | Zaza | za |
| Coptic | cop | Hunsrik | hrx | Lomburd | lmo | Orizaba Nahuatl | nlv | Swaiban | swg | Zulu | zu |
| Cornish | kw | Hupa | hup | Low German | nds | Oromo | om | Swahili (individual language) | swh |
| Model | Param. | X → En | X → Zh | X → De | X → Ar | X → Ne | X → Az | X → Ceb | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 15.7 | 7.2 | 11.2 | 4.6 | 0.6 | 3.0 | 3.1 | 6.5 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 21.2 | 10.3 | 14.2 | 11.5 | 1.3 | 2.4 | 4.9 | 9.4 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 21.6 | 11.0 | 16.1 | 8.8 | 1.0 | 4.7 | 5.3 | 9.8 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 26.3 | 12.9 | 19.3 | 8.1 | 1.4 | 4.6 | 6.8 | 11.3 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 28.0 | 13.3 | 21.3 | 15.1 | 2.9 | 6.4 | 8.8 | 13.7 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 27.9 | 13.0 | 19.5 | 17.2 | 5.5 | 4.2 | 0.5 | 12.5 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 27.4 | 13.9 | 20.9 | 15.2 | 12.1 | 9.4 | 10.3 | 15.6 |
| 8: Lego-MT | × 1.0 | 30.7 | 16.4 | 23.8 | 18.2 | 15.0 | 11.9 | 15.1 | 18.7 |
| Model | Param. | En → X | Zh → X | De → X | Ar → X | Ne → X | Az → X | Ceb → X | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 12.7 | 7.8 | 11.6 | 6.9 | 2.2 | 2.8 | 5.4 | 7.1 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 17.3 | 10.1 | 14.1 | 11.5 | 4.0 | 4.2 | 6.1 | 9.6 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 18.0 | 11.1 | 15.6 | 11.2 | 5.2 | 4.3 | 7.9 | 10.5 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 21.5 | 13.1 | 17.7 | 12.6 | 7.1 | 6.1 | 9.5 | 12.5 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 24.7 | 14.9 | 20.3 | 16.4 | 9.7 | 6.2 | 12.5 | 15.0 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 21.3 | 10.9 | 15.8 | 14.9 | 3.9 | 3.0 | 1.5 | 10.2 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 21.8 | 13.5 | 18.4 | 14.7 | 13.4 | 11.1 | 12.4 | 15.0 |
| 8: Lego-MT | × 1.0 | 25.0 | 16.3 | 21.4 | 18.4 | 17.0 | 13.5 | 16.8 | 18.3 |
| Method | #Tokens | Size (GB) | Time (Hour) |
| Multi-Way Training | 3,000 | 60.9 | 16.9 |
| Lego-MT Training | 3,000 | 10.3 | 0.6 |
| Lang | Mix-Flow | Dec-Flow + Mix-Flow | Enc-Flow + Mix-Flow | |||
| x→zh | zh→x | x→zh | zh→x | x→zh | zh→x | |
| Be | 8.1 | 3.5 | 7.5 | 3.3 | 13.1 | 6.1 |
| De | 16.0 | 14.7 | 14.0 | 13.4 | 22.1 | 19.2 |
| Fa | 13.1 | 11.9 | 11.6 | 11.3 | 17.7 | 15.1 |
| Jv | 6.3 | 3.1 | 6.0 | 3.1 | 8.8 | 3.3 |
| Ne | 9.6 | 5.6 | 8.7 | 4.2 | 7.2 | 4.0 |
| Pa | 1.4 | 1.1 | 1.2 | 0.5 | 1.8 | 0.5 |
| Sw | 8.5 | 8.6 | 7.5 | 6.6 | 12.4 | 12.4 |
| AVG. | 9.0 | 6.9 | 8.1 | 6.1 | 11.9 | 8.7 |
| Model | Ceb→Ha | Ceb→Ig | Ceb→Ln | Ceb→Yo | AVG. |
| M-1.2B | 5.5 | 5.9 | 0.9 | 2.4 | 3.7 |
| M-12B | 8.7 | 10.8 | 0.9 | 2.9 | 5.8 |
| M-FT | 6.4 | 6.6 | 0.8 | 2.1 | 4.0 |
| Lego-MT | 12.5 | 13.9 | 2.3 | 3.2 | 8.0 |
| Model | Ha→Ceb | Ig→Ceb | Ln→Ceb | Yo→Ceb | AVG. |
| M-1.2B | 7.4 | 7.5 | 4.2 | 3.4 | 5.6 |
| M-12B | 8.8 | 8.8 | 3.8 | 4.1 | 6.4 |
| M-FT | 7.5 | 8.6 | 3.7 | 4.8 | 6.2 |
| Lego-MT | 12.3 | 12.3 | 6.2 | 6.7 | 9.4 |
| Model | X→Ast | X→Da | X→Hu | X→Lo | AVG. |
| M-1.2B | 16.7 | 22.0 | 17.7 | 4.8 | 15.3 |
| M-12B | 13.0 | 23.3 | 19.1 | 9.0 | 16.1 |
| M-FT | 13.8 | 23.2 | 17.8 | 0.9 | 13.9 |
| Lego-MT | 15.4 | 25.4 | 20.1 | 5.6 | 16.6 |
| Model | Ast→X | Da→X | Hu→X | Lo→X | AVG. |
| M-1.2B | 13.2 | 18.3 | 16.0 | 6.6 | 13.5 |
| M-12B | 15.2 | 20.9 | 18.2 | 8.8 | 15.8 |
| M-FT | 14.4 | 17.9 | 15.5 | 5.8 | 13.4 |
| Lego-MT | 15.5 | 20.8 | 18.1 | 8.9 | 15.8 |
| Model | Ast→X | Hu→X | Da→X | Lo→X | En→X | De→X | Ar→X | Az→X | Ceb→X | Ne→X | Zh→X | AVG. |
| Multilingual FT | 3.8 | 2.6 | 2.7 | 6.3 | 7.6 | 4.2 | 1.7 | 2.6 | 0.9 | 6.2 | 3.7 | 3.9 |
| Lego-MT | 14.2 | 10.1 | 11.9 | 17.5 | 20.6 | 14.1 | 6.9 | 12.2 | 5.8 | 17.7 | 11.7 | 13.1 |
| Model | X→Ast | X→Hu | X→Da | X→Lo | X→En | X→De | X→Ar | X→Az | X→Ceb | X→Ne | X→Zh | AVG. |
| Multilingual FT | 4.9 | 2.2 | 1.6 | 6.9 | 10.8 | 3.0 | 0.5 | 2.7 | 0.7 | 7.8 | 4.3 | 4.1 |
| Lego-MT | 14.1 | 8.8 | 10.5 | 17.8 | 24.1 | 13.7 | 3.6 | 11.0 | 4.0 | 21.2 | 11.8 | 12.9 |
| Model | X→En | En→X | AVG. |
| ChatGPT zero-shot | 27.9 | 23.9 | 25.9 |
| ChatGPT eight-shot | 31.9 | 24.7 | 28.3 |
| Lego-MT | 30.2 | 25.7 | 28.0 |
| Original | Replaced | Original | Replaced | Original | Replaced |
| ak | aka | es | es_HN | pt | pt_BR |
| am | amh | es | es_EC | pt | pt_br |
| ar | ara | es | es_CO | pt | pt_PT |
| ar | ar_SY | fa | fa_IR | rn | run |
| ar | ar_TN | fa | fa_AF | rw | kin |
| ay | aym | ff | ful | sn | sna |
| az | az_IR | fr | fr_FR | so | som |
| bg | bg_BG | fr | fr_CA | sr | srp |
| bm | bam | fr | fr_BE | sr | sr_ME |
| bn | bn_IN | fr | fr_ca | st | sot |
| ca | cat | ha | hau | sw | swa |
| da | da_DK | hi | hi_IN | ta | ta_LK |
| de | de_CH | ig | ibo | tg | tg_TJ |
| de | de_AT | it | it_IT | ti | tir |
| de | de_DE | jp | jap | tl | tl_PH |
| es | es_CL | kr | kau | tr | tr_TR |
| es | es_SV | kv | kpv | ur | ur_PK |
| es | es_NI | ln | lin | vi | vi_VN |
| es | es_UY | mg | mlg | wo | wol |
| es | es_PE | ms | ms_MY | xh | xho |
| es | es_VE | nb | nb_NO | yo | yor |
| es | es_AR | nds | nds_nl | ze | ze_zh |
| es | es_MX | nl | nl_NL | ze | ze_en |
| es | es_PA | nl | nl_BE | zh | zh_cn |
| es | es_CR | nn | nn_NO | zh | zh_CN |
| es | es_PR | no | no_nb | zthrad | zh_HK |
| es | es_ES | ny | nya | zthrad | zh_TW |
| es | es_GT | om | orm | zthrad | zh_tw |
| es | es_DO | pa | pan | zu | zul |
| Code | En | De | Ar | Zh | Ne | Az | Ceb |
| #Sentence Pairs | 811,238,712 | 360,369,144 | 152,457,830 | 92,763,445 | 6,654,270 | 4,208,025 | 1,683,531 |
| Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset |
| crp | bible-uedin | cb | MultiCCAligned | sz | MultiCCAligned | sgn | QED | cycl | Tatoeba |
| tc | EUbookshop | cx | MultiCCAligned | zz | MultiCCAligned | iro | QED | nah | Tatoeba |
| zhs | GlobalVoices | ns | MultiCCAligned | ze | OpenSubtitles | mo | QED,Ubuntu | ||
| zht | GlobalVoices | qd | MultiCCAligned | bh | QED | ber | QED,Ubuntu | ||
| tmp | GNOME | qa | MultiCCAligned | bnt | QED | toki | Tatoeba | ||
| gr | GNOME | tz | MultiCCAligned | ry | QED | kzj | Tatoeba |
| Language | Code | Language | Code | Language | Code | Language | Code | Language | Code | Language | Code |
| Abkhazian | ab | Corsican | co | Iban | ibla | Lower Sorbian | dsb | Ossetian | os | Swahili (macrolanguage) | sw |
| Achinese | ace | Cree | cr | Icelandic | is | Lukpa | dop | Ottoman Turkish (1500-1928) | ota | Swti | ss |
| Achuar-Shiwiar | acu | Creek | mus | Ido | io | Luo (Kenya and Tanzania) | luo | Paite Chin | pck | Swedish | sv |
| Adyghe | ady | Crimean Tatar | chr | Igbo | ig | Lushootseed | lut | Palaun | pau | Swiss German | gsw |
| Afar | aa | Croatian | hr | Ilok | ilo | Luxembourgish | lb | Pali | pi | Syriac | syr |
| Afrihili | afh | Cusco Quechua | quz | Indonesian | id | Luyia | luy | Pampanga | pam | Tachawit | shy |
| Afrikaans | af | Czech | cs | Ingrian | izh | Macedonian | mk | Pangasinan | pag | Tachelih | shi |
| Aguaruna | agr | Danish | da | Inghush | inh | Macedo-Romanian | rup | Panjabi | pa | Tagal Murut | mvv |
| Ainu (Japan) | ain | Dari | prs | Interlingua | ia | Madurese | mad | Papiamento | pap | Tagalog | tl |
| Akan | ak | Dinka | din | Interlingue | ie | Maithili | mai | Papuan Malay | pmy | Tahaggart Tamahaq | thv |
| Akawaio | ake | Drents | drt | Inuktitut | iu | Malagasy | mg | Pedi | nso | Tahitian | ty |
| Akanlon | akl | Dungan | dng | Inupiaq | ik | Malay (individual language) | zlm | Pennsylvania German | pdc | Tajik | tg |
| Albanian | sq | Dutch | nl | Iranian Persian | pes | Malay (macrolanguage) | ms | Persian | fa | Talusssan | tzl |
| Algerian Arabic | arq | Dutton World Speedwords | dws | Irish | ga | Malayalam | ml | Phoenician | phn | Talysh | tly |
| American Sign Language | ase | Dzongkha | dz | Italian | it | Maltese | mt | Picard | ped | Tamashek | tmh |
| Amharic | am | Eastern Canadian Inuktitut | ike | Jakun | jak | Mam | mam | Piemontese | pms | Tamil | ta |
| Ancient Greek (to 1453) | grc | Eastern Mari | mhr | Jamaican Creole English | jam | Mambae | mgn | Pipil | ppl | Tarifit | rif |
| Ancient Hebrew | hbo | Eastern Maroon Creole | djk | Japanese | ja | Mandarin Chinese | cmn | Plateau Malagasy | plt | Tase Naga | nst |
| Arabic | ar | Efik | efi | Javanese | jv | Manx | gv | Polish | pl | Tatar | tt |
| Argonesse | an | Egyptian Arabic | artz | Jewish Babylonian Aramaic | tmr | Maori | mi | Portuguese | pt | Telugu | te |
| Armenian | hy | Emilian | egl | Kabyle | kab | Marathi | mr | Potawatomi | pot | Tena Lowland Quichua | qw |
| Arpitan | frp | English | en | Kadazan Dusun | dtp | Marshallese | mh | Prussian | prg | Tetelingino Nahuatl | nhg |
| Asháninka | cni | Erzya | myv | Kalaallusit | kl | Mesopotamian Arabic | acm | Pushto | ps | Tetum | tet |
| Assamese | as | Esperanto | eo | Kalmyk | xal | Miahuatlán Zapotec | zam | Quechua | qu | Thai | th |
| Asturian | ast | Estonian | et | Kamba (Kenya) | kam | Middle English (1100-1500) | emm | Quenya | qya | Tibetan | bo |
| Avaric | av | Evenki | evn | Kamada | kn | Middle French (ca. 1400-1600) | frm | Quitopepec Chinantech | chq | Tigrinya | ti |
| Avestan | ae | Ewe | ee | Kanuri | kr | Mikasuki | mik | Rapanui | rap | Tohono O'odham | ood |
| Awadiya | awa | Extremaduran | ext | Kaçchikel | cak | Mi'kmaq | mic | Romanian | ro | Tok Pisin | tpi |
| Aymara | ay | Faroeose | fo | Karalian | krl | Min Dong Chinese | cdo | Romansh | rn | Tonga (Tonga Islands) | to |
| Azerbaijan | az | Fiji Hindi | hif | Kashmiri | ks | Min Nan Chinese | nan | Romany | rom | Traditional Chinese | zhtrad |
| Baluchi | bal | Fijian | fj | Kashubian | csb | Minangkabau | min | Rundi | rn | Tsonga | ts |
| Bambara | bm | Filipino | fil | Kazakh | kk | Mingrelian | xmf | Russian | ru | Tswana | tn |
| Banjar | bjn | Finnish | fi | Kekchi | kek | Mirandese | mwI | Rusyn | rue | Tupi | tpw |
| Barasana-Eduria | bsn | French | fr | Khakasi | kjh | Miskito | miq | Samoan | sm | Turkish | tr |
| Bashkir | ba | Friulian | fur | Khasi | kha | Modern Greek (1453-) | el | Samogtian | sgs | Turkmen | tk |
| Basque | eu | Fulah | ff | Khemer | km | Mohawk | moh | Sango | sg | Tuvalu | tvl |
| Bavarian | bar | Galela | gbi | K'iche' | quc | Mongolian | mn | Sanskrit | sa | Twi | tw |
| Baybayanon | bvy | Galician | gl | Kikuyu | kik | Morisieny | mfe | Santali | sat | Uab Meto | aoz |
| Belarusian | be | Gan Chinese | gan | Kinyarwanda | rw | Moroccan Arabic | ary | Sardinian | sc | Udmurt | udm |
| Bemba (Zambia) | bem | Ganda | lg | Kirghiz | ky | Mossi | mos | Satterffriesisch | stq | Uighur | ug |
| Bengali | bn | Garhwali | gbm | Klingon | thh | Nauru | na | Scots | sco | Ukrainian | uk |
| Berom | born | Georgian | ka | Koaati | ku | Navajo | nv | Scottish Gaelic | gd | Uma | ppk |
| Bhojpuri | bho | German | de | Kölsch | ksh | Neapolitan | nap | Sediq | trv | Umbundu | umb |
| Bislama | bi | Gheg Albanian | aln | Komii | kv | Nepali (individual language) | npi | Serbian | sr | Upper Sorbian | hsb |
| Bodo (India) | brx | Gilbertese | gil | Komii-Permyak | koi | Nepali (macrolanguage) | ne | Serbo-Croatian | sh | Urdu | ur |
| Bosnian | bs | Goan Konkani | gom | Kongo | kg | Nigerian Fulfulde | fuv | Shan | shn | Uspantecco | usp |
| Breton | br | Gothic | got | Korean | ko | Niuean | niu | Shona | sn | Uzbek | uz |
| Brithening | bzt | Gronings | gos | Kotava | avk | Nogai | nog | Shuar | jv | Venda | ve |
| Buginese | bug | Guadeloupean Creole French | gcf | Kriang | ngt | North Levantine Arabic | apc | Schuswap | shs | Venetian | vec |
| Bulgarian | bg | Guarani | gn | Kuanyama | kj | North Molucan Malay | max | Sicilian | scn | Vietnamese | vi |
| Buraiti | bua | Guerrero Amuzgo | amu | Kurdish | ku | Northern Frisian | frr | Silesian | szl | Vlaams | vls |
| Burmese | my | Guerrero Nahuatl | ngu | Kven Finnish | fkv | Northern Kurdish | kmr | Sindarin | sjn | Volapuk | vo |
| Cabecar | cjp | Gujarat | gu | Laidan | ldn | Northern Sami | se | Sindi | sd | Walloon | wa |
| Camsé | kbh | Gulf Arabic | afb | Ladin | lld | Northwestern Ojibwa | obj | Sinhala | si | Walser | wae |
| Catalan | ca | Haida | hai | Ladino | lad | Norwegian | no | Slovak | sk | Waray (Philippines) | war |
| Cebuano | ceb | Haitian | ht | Lakota | lkt | Norwegian Bokmål | nb | Slovenian | sl | Welsh | cy |
| Central Huasteca Nahuatl | nch | Hakka Chin | cnh | Lao | lo | Norwegian Nynorsk | nn | Somali | so | Western Frisan | fy |
| Central Kurdish | ckb | Hakka Chinese | hak | Latgalian | ltg | Novial | nov | South Azerbaijani | azb | Western Panjabi | pnb |
| Central Sama | sml | Hausa | ha | Latin | la | Nuer | nus | South Ndebele | nr | Wolayta | wal |
| Chamorro | ch | Hawaiian | haw | Latvian | lv | Nyanja | ny | Southern Kurdish | sdh | Wolof | wo |
| Chavacano | cbk | Hebrew | he | Ligurian | lij | Occitan (post 1500) | oc | Southern Sami | sma | Wu Chinese | wuu |
| Chechen | ce | Hilgaynon | hil | Limburgan | li | Old English (ca. 450-1100) | ang | Southern Sotho | st | Xhosa | xh |
| Cherokee | chr | Hindi | hi | Lingala | ln | Old French (842-ca. 1400) | fro | Southwestern Dinka | dik | Yakut | sah |
| Chhattisgarhi | hne | Hiri Motu | ho | Lingua Franca Nova | lfn | Old Frisan | ofs | Spanish | es | Yaqui | yqa |
| Chinese | zh | Hmong Daw | mww | Literary Chinese | lzh | Old Norse | non | Standard Malay | zsm | Yiddish | yi |
| Choctaw | cho | Ho | hoc | Lithuanian | lt | Old Russian | ovr | Standard Moroccan Tamazight | zgh | Yoruba | yo |
| Church Slavic | cu | Huastec | hus | Liv | liv | Old Spanish | osp | Sumerian | sux | Zarma | dje |
| Chuvash | cv | Hungarian | hu | Lojban | jbo | Oriya (macrolanguage) | or | Sundanese | su | Zaza | za |
| Coptic | cop | Hunsrik | hrx | Lomburd | lmo | Orizaba Nahuatl | nlv | Swaiban | swg | Zulu | zu |
| Cornish | kw | Hupa | hup | Low German | nds | Oromo | om | Swahili (individual language) | swh |
| Model | Param. | X → En | X → Zh | X → De | X → Ar | X → Ne | X → Az | X → Ceb | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 15.7 | 7.2 | 11.2 | 4.6 | 0.6 | 3.0 | 3.1 | 6.5 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 21.2 | 10.3 | 14.2 | 11.5 | 1.3 | 2.4 | 4.9 | 9.4 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 21.6 | 11.0 | 16.1 | 8.8 | 1.0 | 4.7 | 5.3 | 9.8 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 26.3 | 12.9 | 19.3 | 8.1 | 1.4 | 4.6 | 6.8 | 11.3 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 28.0 | 13.3 | 21.3 | 15.1 | 2.9 | 6.4 | 8.8 | 13.7 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 27.9 | 13.0 | 19.5 | 17.2 | 5.5 | 4.2 | 0.5 | 12.5 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 27.4 | 13.9 | 20.9 | 15.2 | 12.1 | 9.4 | 10.3 | 15.6 |
| 8: Lego-MT | × 1.0 | 30.7 | 16.4 | 23.8 | 18.2 | 15.0 | 11.9 | 15.1 | 18.7 |
| Model | Param. | En → X | Zh → X | De → X | Ar → X | Ne → X | Az → X | Ceb → X | AVG. |
| 1: Flores-175M (Goyal et al., 2022) | × 0.1 | 12.7 | 7.8 | 11.6 | 6.9 | 2.2 | 2.8 | 5.4 | 7.1 |
| 2: M2M-100-418M (Fan et al., 2021) | × 0.3 | 17.3 | 10.1 | 14.1 | 11.5 | 4.0 | 4.2 | 6.1 | 9.6 |
| 3: Flores-615M (Goyal et al., 2022) | × 0.5 | 18.0 | 11.1 | 15.6 | 11.2 | 5.2 | 4.3 | 7.9 | 10.5 |
| 4: M2M-100-1.2B (Fan et al., 2021) | × 1.0 | 21.5 | 13.1 | 17.7 | 12.6 | 7.1 | 6.1 | 9.5 | 12.5 |
| 5: M2M-100-12B (Fan et al., 2021) | × 10.0 | 24.7 | 14.9 | 20.3 | 16.4 | 9.7 | 6.2 | 12.5 | 15.0 |
| 6: (4) + LG-Centric Fine-Tuning | × 1.0 | 21.3 | 10.9 | 15.8 | 14.9 | 3.9 | 3.0 | 1.5 | 10.2 |
| 7: (4) + Multilingual Fine-Tuning | × 1.0 | 21.8 | 13.5 | 18.4 | 14.7 | 13.4 | 11.1 | 12.4 | 15.0 |
| 8: Lego-MT | × 1.0 | 25.0 | 16.3 | 21.4 | 18.4 | 17.0 | 13.5 | 16.8 | 18.3 |
| Method | #Tokens | Size (GB) | Time (Hour) |
| Multi-Way Training | 3,000 | 60.9 | 16.9 |
| Lego-MT Training | 3,000 | 10.3 | 0.6 |
| Lang | Mix-Flow | Dec-Flow + Mix-Flow | Enc-Flow + Mix-Flow | |||
| x→zh | zh→x | x→zh | zh→x | x→zh | zh→x | |
| Be | 8.1 | 3.5 | 7.5 | 3.3 | 13.1 | 6.1 |
| De | 16.0 | 14.7 | 14.0 | 13.4 | 22.1 | 19.2 |
| Fa | 13.1 | 11.9 | 11.6 | 11.3 | 17.7 | 15.1 |
| Jv | 6.3 | 3.1 | 6.0 | 3.1 | 8.8 | 3.3 |
| Ne | 9.6 | 5.6 | 8.7 | 4.2 | 7.2 | 4.0 |
| Pa | 1.4 | 1.1 | 1.2 | 0.5 | 1.8 | 0.5 |
| Sw | 8.5 | 8.6 | 7.5 | 6.6 | 12.4 | 12.4 |
| AVG. | 9.0 | 6.9 | 8.1 | 6.1 | 11.9 | 8.7 |
| Model | Ceb→Ha | Ceb→Ig | Ceb→Ln | Ceb→Yo | AVG. |
| M-1.2B | 5.5 | 5.9 | 0.9 | 2.4 | 3.7 |
| M-12B | 8.7 | 10.8 | 0.9 | 2.9 | 5.8 |
| M-FT | 6.4 | 6.6 | 0.8 | 2.1 | 4.0 |
| Lego-MT | 12.5 | 13.9 | 2.3 | 3.2 | 8.0 |
| Model | Ha→Ceb | Ig→Ceb | Ln→Ceb | Yo→Ceb | AVG. |
| M-1.2B | 7.4 | 7.5 | 4.2 | 3.4 | 5.6 |
| M-12B | 8.8 | 8.8 | 3.8 | 4.1 | 6.4 |
| M-FT | 7.5 | 8.6 | 3.7 | 4.8 | 6.2 |
| Lego-MT | 12.3 | 12.3 | 6.2 | 6.7 | 9.4 |
| Model | X→Ast | X→Da | X→Hu | X→Lo | AVG. |
| M-1.2B | 16.7 | 22.0 | 17.7 | 4.8 | 15.3 |
| M-12B | 13.0 | 23.3 | 19.1 | 9.0 | 16.1 |
| M-FT | 13.8 | 23.2 | 17.8 | 0.9 | 13.9 |
| Lego-MT | 15.4 | 25.4 | 20.1 | 5.6 | 16.6 |
| Model | Ast→X | Da→X | Hu→X | Lo→X | AVG. |
| M-1.2B | 13.2 | 18.3 | 16.0 | 6.6 | 13.5 |
| M-12B | 15.2 | 20.9 | 18.2 | 8.8 | 15.8 |
| M-FT | 14.4 | 17.9 | 15.5 | 5.8 | 13.4 |
| Lego-MT | 15.5 | 20.8 | 18.1 | 8.9 | 15.8 |
| Model | Ast→X | Hu→X | Da→X | Lo→X | En→X | De→X | Ar→X | Az→X | Ceb→X | Ne→X | Zh→X | AVG. |
| Multilingual FT | 3.8 | 2.6 | 2.7 | 6.3 | 7.6 | 4.2 | 1.7 | 2.6 | 0.9 | 6.2 | 3.7 | 3.9 |
| Lego-MT | 14.2 | 10.1 | 11.9 | 17.5 | 20.6 | 14.1 | 6.9 | 12.2 | 5.8 | 17.7 | 11.7 | 13.1 |
| Model | X→Ast | X→Hu | X→Da | X→Lo | X→En | X→De | X→Ar | X→Az | X→Ceb | X→Ne | X→Zh | AVG. |
| Multilingual FT | 4.9 | 2.2 | 1.6 | 6.9 | 10.8 | 3.0 | 0.5 | 2.7 | 0.7 | 7.8 | 4.3 | 4.1 |
| Lego-MT | 14.1 | 8.8 | 10.5 | 17.8 | 24.1 | 13.7 | 3.6 | 11.0 | 4.0 | 21.2 | 11.8 | 12.9 |
| Model | X→En | En→X | AVG. |
| ChatGPT zero-shot | 27.9 | 23.9 | 25.9 |
| ChatGPT eight-shot | 31.9 | 24.7 | 28.3 |
| Lego-MT | 30.2 | 25.7 | 28.0 |
| Original | Replaced | Original | Replaced | Original | Replaced |
| ak | aka | es | es_HN | pt | pt_BR |
| am | amh | es | es_EC | pt | pt_br |
| ar | ara | es | es_CO | pt | pt_PT |
| ar | ar_SY | fa | fa_IR | rn | run |
| ar | ar_TN | fa | fa_AF | rw | kin |
| ay | aym | ff | ful | sn | sna |
| az | az_IR | fr | fr_FR | so | som |
| bg | bg_BG | fr | fr_CA | sr | srp |
| bm | bam | fr | fr_BE | sr | sr_ME |
| bn | bn_IN | fr | fr_ca | st | sot |
| ca | cat | ha | hau | sw | swa |
| da | da_DK | hi | hi_IN | ta | ta_LK |
| de | de_CH | ig | ibo | tg | tg_TJ |
| de | de_AT | it | it_IT | ti | tir |
| de | de_DE | jp | jap | tl | tl_PH |
| es | es_CL | kr | kau | tr | tr_TR |
| es | es_SV | kv | kpv | ur | ur_PK |
| es | es_NI | ln | lin | vi | vi_VN |
| es | es_UY | mg | mlg | wo | wol |
| es | es_PE | ms | ms_MY | xh | xho |
| es | es_VE | nb | nb_NO | yo | yor |
| es | es_AR | nds | nds_nl | ze | ze_zh |
| es | es_MX | nl | nl_NL | ze | ze_en |
| es | es_PA | nl | nl_BE | zh | zh_cn |
| es | es_CR | nn | nn_NO | zh | zh_CN |
| es | es_PR | no | no_nb | zthrad | zh_HK |
| es | es_ES | ny | nya | zthrad | zh_TW |
| es | es_GT | om | orm | zthrad | zh_tw |
| es | es_DO | pa | pan | zu | zul |
| Code | En | De | Ar | Zh | Ne | Az | Ceb |
| #Sentence Pairs | 811,238,712 | 360,369,144 | 152,457,830 | 92,763,445 | 6,654,270 | 4,208,025 | 1,683,531 |
| Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset | Code | Dataset |
| crp | bible-uedin | cb | MultiCCAligned | sz | MultiCCAligned | sgn | QED | cycl | Tatoeba |
| tc | EUbookshop | cx | MultiCCAligned | zz | MultiCCAligned | iro | QED | nah | Tatoeba |
| zhs | GlobalVoices | ns | MultiCCAligned | ze | OpenSubtitles | mo | QED,Ubuntu | ||
| zht | GlobalVoices | qd | MultiCCAligned | bh | QED | ber | QED,Ubuntu | ||
| tmp | GNOME | qa | MultiCCAligned | bnt | QED | toki | Tatoeba | ||
| gr | GNOME | tz | MultiCCAligned | ry | QED | kzj | Tatoeba |
| Language | Code | Language | Code | Language | Code | Language | Code | Language | Code | Language | Code |
| Abkhazian | ab | Corsican | co | Iban | ibla | Lower Sorbian | dsb | Ossetian | os | Swahili (macrolanguage) | sw |
| Achinese | ace | Cree | cr | Icelandic | is | Lukpa | dop | Ottoman Turkish (1500-1928) | ota | Swti | ss |
| Achuar-Shiwiar | acu | Creek | mus | Ido | io | Luo (Kenya and Tanzania) | luo | Paite Chin | pck | Swedish | sv |
| Adyghe | ady | Crimean Tatar | chr | Igbo | ig | Lushootseed | lut | Palaun | pau | Swiss German | gsw |
| Afar | aa | Croatian | hr | Ilok | ilo | Luxembourgish | lb | Pali | pi | Syriac | syr |
| Afrihili | afh | Cusco Quechua | quz | Indonesian | id | Luyia | luy | Pampanga | pam | Tachawit | shy |
| Afrikaans | af | Czech | cs | Ingrian | izh | Macedonian | mk | Pangasinan | pag | Tachelih | shi |
| Aguaruna | agr | Danish | da | Inghush | inh | Macedo-Romanian | rup | Panjabi | pa | Tagal Murut | mvv |
| Ainu (Japan) | ain | Dari | prs | Interlingua | ia | Madurese | mad | Papiamento | pap | Tagalog | tl |
| Akan | ak | Dinka | din | Interlingue | ie | Maithili | mai | Papuan Malay | pmy | Tahaggart Tamahaq | thv |
| Akawaio | ake | Drents | drt | Inuktitut | iu | Malagasy | mg | Pedi | nso | Tahitian | ty |
| Akanlon | akl | Dungan | dng | Inupiaq | ik | Malay (individual language) | zlm | Pennsylvania German | pdc | Tajik | tg |
| Albanian | sq | Dutch | nl | Iranian Persian | pes | Malay (macrolanguage) | ms | Persian | fa | Talusssan | tzl |
| Algerian Arabic | arq | Dutton World Speedwords | dws | Irish | ga | Malayalam | ml | Phoenician | phn | Talysh | tly |
| American Sign Language | ase | Dzongkha | dz | Italian | it | Maltese | mt | Picard | ped | Tamashek | tmh |
| Amharic | am | Eastern Canadian Inuktitut | ike | Jakun | jak | Mam | mam | Piemontese | pms | Tamil | ta |
| Ancient Greek (to 1453) | grc | Eastern Mari | mhr | Jamaican Creole English | jam | Mambae | mgn | Pipil | ppl | Tarifit | rif |
| Ancient Hebrew | hbo | Eastern Maroon Creole | djk | Japanese | ja | Mandarin Chinese | cmn | Plateau Malagasy | plt | Tase Naga | nst |
| Arabic | ar | Efik | efi | Javanese | jv | Manx | gv | Polish | pl | Tatar | tt |
| Argonesse | an | Egyptian Arabic | artz | Jewish Babylonian Aramaic | tmr | Maori | mi | Portuguese | pt | Telugu | te |
| Armenian | hy | Emilian | egl | Kabyle | kab | Marathi | mr | Potawatomi | pot | Tena Lowland Quichua | qw |
| Arpitan | frp | English | en | Kadazan Dusun | dtp | Marshallese | mh | Prussian | prg | Tetelingino Nahuatl | nhg |
| Asháninka | cni | Erzya | myv | Kalaallusit | kl | Mesopotamian Arabic | acm | Pushto | ps | Tetum | tet |
| Assamese | as | Esperanto | eo | Kalmyk | xal | Miahuatlán Zapotec | zam | Quechua | qu | Thai | th |
| Asturian | ast | Estonian | et | Kamba (Kenya) | kam | Middle English (1100-1500) | emm | Quenya | qya | Tibetan | bo |
| Avaric | av | Evenki | evn | Kamada | kn | Middle French (ca. 1400-1600) | frm | Quitopepec Chinantech | chq | Tigrinya | ti |
| Avestan | ae | Ewe | ee | Kanuri | kr | Mikasuki | mik | Rapanui | rap | Tohono O'odham | ood |
| Awadiya | awa | Extremaduran | ext | Kaçchikel | cak | Mi'kmaq | mic | Romanian | ro | Tok Pisin | tpi |
| Aymara | ay | Faroeose | fo | Karalian | krl | Min Dong Chinese | cdo | Romansh | rn | Tonga (Tonga Islands) | to |
| Azerbaijan | az | Fiji Hindi | hif | Kashmiri | ks | Min Nan Chinese | nan | Romany | rom | Traditional Chinese | zhtrad |
| Baluchi | bal | Fijian | fj | Kashubian | csb | Minangkabau | min | Rundi | rn | Tsonga | ts |
| Bambara | bm | Filipino | fil | Kazakh | kk | Mingrelian | xmf | Russian | ru | Tswana | tn |
| Banjar | bjn | Finnish | fi | Kekchi | kek | Mirandese | mwI | Rusyn | rue | Tupi | tpw |
| Barasana-Eduria | bsn | French | fr | Khakasi | kjh | Miskito | miq | Samoan | sm | Turkish | tr |
| Bashkir | ba | Friulian | fur | Khasi | kha | Modern Greek (1453-) | el | Samogtian | sgs | Turkmen | tk |
| Basque | eu | Fulah | ff | Khemer | km | Mohawk | moh | Sango | sg | Tuvalu | tvl |
| Bavarian | bar | Galela | gbi | K'iche' | quc | Mongolian | mn | Sanskrit | sa | Twi | tw |
| Baybayanon | bvy | Galician | gl | Kikuyu | kik | Morisieny | mfe | Santali | sat | Uab Meto | aoz |
| Belarusian | be | Gan Chinese | gan | Kinyarwanda | rw | Moroccan Arabic | ary | Sardinian | sc | Udmurt | udm |
| Bemba (Zambia) | bem | Ganda | lg | Kirghiz | ky | Mossi | mos | Satterffriesisch | stq | Uighur | ug |
| Bengali | bn | Garhwali | gbm | Klingon | thh | Nauru | na | Scots | sco | Ukrainian | uk |
| Berom | born | Georgian | ka | Koaati | ku | Navajo | nv | Scottish Gaelic | gd | Uma | ppk |
| Bhojpuri | bho | German | de | Kölsch | ksh | Neapolitan | nap | Sediq | trv | Umbundu | umb |
| Bislama | bi | Gheg Albanian | aln | Komii | kv | Nepali (individual language) | npi | Serbian | sr | Upper Sorbian | hsb |
| Bodo (India) | brx | Gilbertese | gil | Komii-Permyak | koi | Nepali (macrolanguage) | ne | Serbo-Croatian | sh | Urdu | ur |
| Bosnian | bs | Goan Konkani | gom | Kongo | kg | Nigerian Fulfulde | fuv | Shan | shn | Uspantecco | usp |
| Breton | br | Gothic | got | Korean | ko | Niuean | niu | Shona | sn | Uzbek | uz |
| Brithening | bzt | Gronings | gos | Kotava | avk | Nogai | nog | Shuar | jv | Venda | ve |
| Buginese | bug | Guadeloupean Creole French | gcf | Kriang | ngt | North Levantine Arabic | apc | Schuswap | shs | Venetian | vec |
| Bulgarian | bg | Guarani | gn | Kuanyama | kj | North Molucan Malay | max | Sicilian | scn | Vietnamese | vi |
| Buraiti | bua | Guerrero Amuzgo | amu | Kurdish | ku | Northern Frisian | frr | Silesian | szl | Vlaams | vls |
| Burmese | my | Guerrero Nahuatl | ngu | Kven Finnish | fkv | Northern Kurdish | kmr | Sindarin | sjn | Volapuk | vo |
| Cabecar | cjp | Gujarat | gu | Laidan | ldn | Northern Sami | se | Sindi | sd | Walloon | wa |
| Camsé | kbh | Gulf Arabic | afb | Ladin | lld | Northwestern Ojibwa | obj | Sinhala | si | Walser | wae |
| Catalan | ca | Haida | hai | Ladino | lad | Norwegian | no | Slovak | sk | Waray (Philippines) | war |
| Cebuano | ceb | Haitian | ht | Lakota | lkt | Norwegian Bokmål | nb | Slovenian | sl | Welsh | cy |
| Central Huasteca Nahuatl | nch | Hakka Chin | cnh | Lao | lo | Norwegian Nynorsk | nn | Somali | so | Western Frisan | fy |
| Central Kurdish | ckb | Hakka Chinese | hak | Latgalian | ltg | Novial | nov | South Azerbaijani | azb | Western Panjabi | pnb |
| Central Sama | sml | Hausa | ha | Latin | la | Nuer | nus | South Ndebele | nr | Wolayta | wal |
| Chamorro | ch | Hawaiian | haw | Latvian | lv | Nyanja | ny | Southern Kurdish | sdh | Wolof | wo |
| Chavacano | cbk | Hebrew | he | Ligurian | lij | Occitan (post 1500) | oc | Southern Sami | sma | Wu Chinese | wuu |
| Chechen | ce | Hilgaynon | hil | Limburgan | li | Old English (ca. 450-1100) | ang | Southern Sotho | st | Xhosa | xh |
| Cherokee | chr | Hindi | hi | Lingala | ln | Old French (842-ca. 1400) | fro | Southwestern Dinka | dik | Yakut | sah |
| Chhattisgarhi | hne | Hiri Motu | ho | Lingua Franca Nova | lfn | Old Frisan | ofs | Spanish | es | Yaqui | yqa |
| Chinese | zh | Hmong Daw | mww | Literary Chinese | lzh | Old Norse | non | Standard Malay | zsm | Yiddish | yi |
| Choctaw | cho | Ho | hoc | Lithuanian | lt | Old Russian | ovr | Standard Moroccan Tamazight | zgh | Yoruba | yo |
| Church Slavic | cu | Huastec | hus | Liv | liv | Old Spanish | osp | Sumerian | sux | Zarma | dje |
| Chuvash | cv | Hungarian | hu | Lojban | jbo | Oriya (macrolanguage) | or | Sundanese | su | Zaza | za |
| Coptic | cop | Hunsrik | hrx | Lomburd | lmo | Orizaba Nahuatl | nlv | Swaiban | swg | Zulu | zu |
| Cornish | kw | Hupa | hup | Low German | nds | Oromo | om | Swahili (individual language) | swh |
| Model | ART | E-CARE | ||
| \\(Frac(↑)\\) | \\(PPL(↓)\\) | \\(Frac(↑)\\) | \\(PPL(↓)\\) | |
| MLE | 54.1 | 42.6 | 54.5 | 80.4 |
| MLE-LL | 56.6 | 42.5 | 52.6 | 84.8 |
| + CD | 59.9 | 49.8 | 63.4 | 107.3 |
| + DEXPERTS | 56.2 | 51.7 | 57.2 | 108.3 |
| BRAINSTORM | 79.4 | 40.7 | 58.1 | 69.2 |
| + CD | 79.7 | 50.2 | 67.2 | 88.1 |
| + DEXPERTS | 79.0 | 51.5 | 58.1 | 89.3 |
| QUARK | 85.9 | 27.5 | 68.2 | 80.8 |
| BRAINSTORM | ||||
| -\\(\\mathcal{L}_{\\text{margin}}\\) | 69.3 | 44.9 | 54.6 | 73.2 |
| -\\(\\mathcal{L}_{\\text{sim}}\\) | 58.2 | 52.6 | 53.2 | 83.7 |
| BRAINSTORM' | 58.3 | 52.0 | 55.1 | 71.2 |
| Model | ART | E-CARE | ||||||||
| Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | |
| MLE-LL | 42.3 | 15.2 | 22.7 | 9.5 | 10.3 | 35.4 | 15.6 | 5.7 | 18.6 | 24.7 |
| Quark | 14.7 | 20.8 | 51.0 | 4.3 | 9.2 | 35.2 | 15.1 | 5.7 | 3.3 | 40.7 |
| BRAINSTORM | 20.9 | 20.2 | 41.3 | 4.8 | 12.8 | 37.1 | 20.1 | 4.7 | 12.7 | 25.4 |
| Model | Likely | Less likely | Irrel. | ||
| High | Med. | Low | |||
| MLE-LL | 6.7 | 40.7 | 21.2 | 14.7 | 16.7 |
| MLE | 7.3 | 50.0 | 22.1 | 13.3 | 7.3 |
| BRAINSTORM' | 6.7 | 42.0 | 32.6 | 8.7 | 10.0 |
| Reference | 3.3 | 76.7 | 13.4 | 3.3 | 3.3 |
| Dataset | Train | Val | Test | |||||
| Likely | Less Likely | Less Likely | Less Likely | |||||
| MRIINTERPRET | 10097 | 1005 | 121 | — | ||||
| ART | 50509 | 50509 | 1781 | 3562 | ||||
| E-CARE | cause | effect | cause | effect | cause | effect | cause | effect |
| 6855 | 6580 | 6855 | 6580 | 762 | 731 | 1088 | 1044 |
| Model | ART | E-CARE | ||
| Gram. Correct Fluent | Contain Flu. Errors | Gram. Correct Fluent | Contain Flu. Errors | |
| MLE-LL | 93.9 | 6.1 | 99.0 | 1.0 |
| QUARK | 94.6 | 5.4 | 98.0 | 2.0 |
| BRAINSTORM | 93.5 | 6.6 | 95.9 | 4.1 |
| Task | Examples | Output | Explanation |
| Brain MRI | ...... Absence of evidence of restricted diffusion makes it unlikely to be | ~ acute ischemia ~ infarct | In diffusion weighted imaging sequences on MRI of the brain, one of the most common causes of diffusion restriction finding is due to acute ischemic stroke, also known as an infarct. Thus, the absence of restricted diffusion within brain tissue makes an interpretation unlikely to be acute ischemia/infarct. |
| ANLG | O1: Lisa went laptop shopping yesterday. O2: She was thankful she bought it. | + The price raised on the next day. ~ Lisa decided to buy a car. | This scenario makes sense—Lisa was grateful that she had made her laptop purchase the day before, as the price un-expectedly increased the following day. The event focus on Lisa's purchase of a laptop. It is not likely that she would suddenly decide to buy a car without any prior indication of her interest in doing so. |
| E-CARE | Tom goes to the gym every day. | + Tom improves his physical fitness. ~ He gets a promotion from his manager who saw him in the gym. | It directly relates to Tom's daily gym attendance. Regular exercise is a common and effective method for improving physical fitness. It directly connects to the premise involving Tom's daily gym attendance. However, the connection between gym attendance and job promotion is indirect. |
| Model | Hallucination (%) |
| MLE-LL | 23.3 |
| MLE | 30.0 |
| BRAINSTORM | 33.3 |
| Reference | 6.6 |
| MLE-LL | Riley's mother bought her cowboy boots. |
| + CD (τCD = 0.5) | Riley had bought cowboy shoes that she had not worn before. |
| + CD (τCD = 1.0) | Her mother bought a new cowboy shirt for Riley. |
| + CD (τCD = 1.5) | Riiley got her new cowboy boots torn. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 0.5) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 1.0) | Riley didn't have cowboy boots. |
| BRAINSTORM | Riley bought a pair of cowboy boots for her mother. |
| + CD (τCD = 0.5) | Rice bought cowboy hats with her money and jeans. |
| + CD (τCD = 1.0) | Rice's mother bought her a pair of cowboy boots to wear to school next to her cowboy outfit. |
| + CD (τCD = 1.5) | Her mom bought her a new pair of cowboy boots. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her a pair of cowboy boots. Riley hated them. |
| + DEXPERTS (α = 0.5) | Riley bought a pair of cowboy pants. |
| + DEXPERTS (α = 1.0) | Riley bought a pair of cowboy pants. |
| QUARK | Riley's mother bought her cowboy boots at the store and Riley hated them. |
| MLE-LL | The whole kitchen was burnt down more quickly. |
| + CD (τCD = 0.5) | Tom poured water in the pan of acid. |
| + CD (τCD = 1.0) | There was a big thunderstorm yesterday. |
| + CD (τCD = 1.5) | There is a big storm outside. |
| + DEXPERTS (α = 0.1) | The whole kitchen was burnt down more quickly. |
| + DEXPERTS (α = 0.5) | The fire in the kitchen spread outwards. |
| + DEXPERTS (α = 1.0) | There are more and more fires in this place. |
| BRAINSTORM | Tom put a lot of fuel on the fire. |
| + CD (τCD = 0.5) | Tom poured a bucket of water to a sink which has a high temperature. |
| + CD (τCD = 1.0) | There was an accident at night. |
| + CD (τCD = 1.5) | Tom poured gasoline to the stove. |
| + DEXPERTS (α = 0.1) | There is a fire in the kitchen. |
| + DEXPERTS (α = 0.5) | The whole kitchen was filled with smoke. |
| + DEXPERTS (α = 1.0) | Tom's kitchen is leaking water. |
| QUARK | The fire in the kitchen was very hot. |
| Model | ART | E-CARE | ||
| \\(Frac(↑)\\) | \\(PPL(↓)\\) | \\(Frac(↑)\\) | \\(PPL(↓)\\) | |
| MLE | 54.1 | 42.6 | 54.5 | 80.4 |
| MLE-LL | 56.6 | 42.5 | 52.6 | 84.8 |
| + CD | 59.9 | 49.8 | 63.4 | 107.3 |
| + DEXPERTS | 56.2 | 51.7 | 57.2 | 108.3 |
| BRAINSTORM | 79.4 | 40.7 | 58.1 | 69.2 |
| + CD | 79.7 | 50.2 | 67.2 | 88.1 |
| + DEXPERTS | 79.0 | 51.5 | 58.1 | 89.3 |
| QUARK | 85.9 | 27.5 | 68.2 | 80.8 |
| BRAINSTORM | ||||
| -\\(\\mathcal{L}_{\\text{margin}}\\) | 69.3 | 44.9 | 54.6 | 73.2 |
| -\\(\\mathcal{L}_{\\text{sim}}\\) | 58.2 | 52.6 | 53.2 | 83.7 |
| BRAINSTORM' | 58.3 | 52.0 | 55.1 | 71.2 |
| Model | ART | E-CARE | ||||||||
| Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | |
| MLE-LL | 42.3 | 15.2 | 22.7 | 9.5 | 10.3 | 35.4 | 15.6 | 5.7 | 18.6 | 24.7 |
| Quark | 14.7 | 20.8 | 51.0 | 4.3 | 9.2 | 35.2 | 15.1 | 5.7 | 3.3 | 40.7 |
| BRAINSTORM | 20.9 | 20.2 | 41.3 | 4.8 | 12.8 | 37.1 | 20.1 | 4.7 | 12.7 | 25.4 |
| Model | Likely | Less likely | Irrel. | ||
| High | Med. | Low | |||
| MLE-LL | 6.7 | 40.7 | 21.2 | 14.7 | 16.7 |
| MLE | 7.3 | 50.0 | 22.1 | 13.3 | 7.3 |
| BRAINSTORM' | 6.7 | 42.0 | 32.6 | 8.7 | 10.0 |
| Reference | 3.3 | 76.7 | 13.4 | 3.3 | 3.3 |
| Dataset | Train | Val | Test | |||||
| Likely | Less Likely | Less Likely | Less Likely | |||||
| MRIINTERPRET | 10097 | 1005 | 121 | — | ||||
| ART | 50509 | 50509 | 1781 | 3562 | ||||
| E-CARE | cause | effect | cause | effect | cause | effect | cause | effect |
| 6855 | 6580 | 6855 | 6580 | 762 | 731 | 1088 | 1044 |
| Model | ART | E-CARE | ||
| Gram. Correct Fluent | Contain Flu. Errors | Gram. Correct Fluent | Contain Flu. Errors | |
| MLE-LL | 93.9 | 6.1 | 99.0 | 1.0 |
| QUARK | 94.6 | 5.4 | 98.0 | 2.0 |
| BRAINSTORM | 93.5 | 6.6 | 95.9 | 4.1 |
| Task | Examples | Output | Explanation |
| Brain MRI | ...... Absence of evidence of restricted diffusion makes it unlikely to be | ~ acute ischemia ~ infarct | In diffusion weighted imaging sequences on MRI of the brain, one of the most common causes of diffusion restriction finding is due to acute ischemic stroke, also known as an infarct. Thus, the absence of restricted diffusion within brain tissue makes an interpretation unlikely to be acute ischemia/infarct. |
| ANLG | O1: Lisa went laptop shopping yesterday. O2: She was thankful she bought it. | + The price raised on the next day. ~ Lisa decided to buy a car. | This scenario makes sense—Lisa was grateful that she had made her laptop purchase the day before, as the price un-expectedly increased the following day. The event focus on Lisa's purchase of a laptop. It is not likely that she would suddenly decide to buy a car without any prior indication of her interest in doing so. |
| E-CARE | Tom goes to the gym every day. | + Tom improves his physical fitness. ~ He gets a promotion from his manager who saw him in the gym. | It directly relates to Tom's daily gym attendance. Regular exercise is a common and effective method for improving physical fitness. It directly connects to the premise involving Tom's daily gym attendance. However, the connection between gym attendance and job promotion is indirect. |
| Model | Hallucination (%) |
| MLE-LL | 23.3 |
| MLE | 30.0 |
| BRAINSTORM | 33.3 |
| Reference | 6.6 |
| MLE-LL | Riley's mother bought her cowboy boots. |
| + CD (τCD = 0.5) | Riley had bought cowboy shoes that she had not worn before. |
| + CD (τCD = 1.0) | Her mother bought a new cowboy shirt for Riley. |
| + CD (τCD = 1.5) | Riiley got her new cowboy boots torn. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 0.5) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 1.0) | Riley didn't have cowboy boots. |
| BRAINSTORM | Riley bought a pair of cowboy boots for her mother. |
| + CD (τCD = 0.5) | Rice bought cowboy hats with her money and jeans. |
| + CD (τCD = 1.0) | Rice's mother bought her a pair of cowboy boots to wear to school next to her cowboy outfit. |
| + CD (τCD = 1.5) | Her mom bought her a new pair of cowboy boots. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her a pair of cowboy boots. Riley hated them. |
| + DEXPERTS (α = 0.5) | Riley bought a pair of cowboy pants. |
| + DEXPERTS (α = 1.0) | Riley bought a pair of cowboy pants. |
| QUARK | Riley's mother bought her cowboy boots at the store and Riley hated them. |
| MLE-LL | The whole kitchen was burnt down more quickly. |
| + CD (τCD = 0.5) | Tom poured water in the pan of acid. |
| + CD (τCD = 1.0) | There was a big thunderstorm yesterday. |
| + CD (τCD = 1.5) | There is a big storm outside. |
| + DEXPERTS (α = 0.1) | The whole kitchen was burnt down more quickly. |
| + DEXPERTS (α = 0.5) | The fire in the kitchen spread outwards. |
| + DEXPERTS (α = 1.0) | There are more and more fires in this place. |
| BRAINSTORM | Tom put a lot of fuel on the fire. |
| + CD (τCD = 0.5) | Tom poured a bucket of water to a sink which has a high temperature. |
| + CD (τCD = 1.0) | There was an accident at night. |
| + CD (τCD = 1.5) | Tom poured gasoline to the stove. |
| + DEXPERTS (α = 0.1) | There is a fire in the kitchen. |
| + DEXPERTS (α = 0.5) | The whole kitchen was filled with smoke. |
| + DEXPERTS (α = 1.0) | Tom's kitchen is leaking water. |
| QUARK | The fire in the kitchen was very hot. |
| Model | ART | E-CARE | ||
| \(Frac(↑)\) | \(PPL(↓)\) | \(Frac(↑)\) | \(PPL(↓)\) | |
| MLE | 54.1 | 42.6 | 54.5 | 80.4 |
| MLE-LL | 56.6 | 42.5 | 52.6 | 84.8 |
| + CD | 59.9 | 49.8 | 63.4 | 107.3 |
| + DEXPERTS | 56.2 | 51.7 | 57.2 | 108.3 |
| BRAINSTORM | 79.4 | 40.7 | 58.1 | 69.2 |
| + CD | 79.7 | 50.2 | 67.2 | 88.1 |
| + DEXPERTS | 79.0 | 51.5 | 58.1 | 89.3 |
| QUARK | 85.9 | 27.5 | 68.2 | 80.8 |
| BRAINSTORM | ||||
| -\(\mathcal{L}_{\text{margin}}\) | 69.3 | 44.9 | 54.6 | 73.2 |
| -\(\mathcal{L}_{\text{sim}}\) | 58.2 | 52.6 | 53.2 | 83.7 |
| BRAINSTORM' | 58.3 | 52.0 | 55.1 | 71.2 |
| Model | ART | E-CARE | ||||||||
| Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | |
| MLE-LL | 42.3 | 15.2 | 22.7 | 9.5 | 10.3 | 35.4 | 15.6 | 5.7 | 18.6 | 24.7 |
| Quark | 14.7 | 20.8 | 51.0 | 4.3 | 9.2 | 35.2 | 15.1 | 5.7 | 3.3 | 40.7 |
| BRAINSTORM | 20.9 | 20.2 | 41.3 | 4.8 | 12.8 | 37.1 | 20.1 | 4.7 | 12.7 | 25.4 |
| Model | Likely | Less likely | Irrel. | ||
| High | Med. | Low | |||
| MLE-LL | 6.7 | 40.7 | 21.2 | 14.7 | 16.7 |
| MLE | 7.3 | 50.0 | 22.1 | 13.3 | 7.3 |
| BRAINSTORM' | 6.7 | 42.0 | 32.6 | 8.7 | 10.0 |
| Reference | 3.3 | 76.7 | 13.4 | 3.3 | 3.3 |
| Dataset | Train | Val | Test | |||||
| Likely | Less Likely | Less Likely | Less Likely | |||||
| MRIINTERPRET | 10097 | 1005 | 121 | — | ||||
| ART | 50509 | 50509 | 1781 | 3562 | ||||
| E-CARE | cause | effect | cause | effect | cause | effect | cause | effect |
| 6855 | 6580 | 6855 | 6580 | 762 | 731 | 1088 | 1044 | |
| Model | ART | E-CARE | ||
| Gram. Correct Fluent | Contain Flu. Errors | Gram. Correct Fluent | Contain Flu. Errors | |
| MLE-LL | 93.9 | 6.1 | 99.0 | 1.0 |
| QUARK | 94.6 | 5.4 | 98.0 | 2.0 |
| BRAINSTORM | 93.5 | 6.6 | 95.9 | 4.1 |
| Task | Examples | Output | Explanation |
| Brain MRI | ...... Absence of evidence of restricted diffusion makes it unlikely to be | ~ acute ischemia ~ infarct | In diffusion weighted imaging sequences on MRI of the brain, one of the most common causes of diffusion restriction finding is due to acute ischemic stroke, also known as an infarct. Thus, the absence of restricted diffusion within brain tissue makes an interpretation unlikely to be acute ischemia/infarct. |
| ANLG | O1: Lisa went laptop shopping yesterday. O2: She was thankful she bought it. | + The price raised on the next day. ~ Lisa decided to buy a car. | This scenario makes sense—Lisa was grateful that she had made her laptop purchase the day before, as the price un-expectedly increased the following day. The event focus on Lisa's purchase of a laptop. It is not likely that she would suddenly decide to buy a car without any prior indication of her interest in doing so. |
| E-CARE | Tom goes to the gym every day. | + Tom improves his physical fitness. ~ He gets a promotion from his manager who saw him in the gym. | It directly relates to Tom's daily gym attendance. Regular exercise is a common and effective method for improving physical fitness. It directly connects to the premise involving Tom's daily gym attendance. However, the connection between gym attendance and job promotion is indirect. |
| Model | Hallucination (%) |
| MLE-LL | 23.3 |
| MLE | 30.0 |
| BRAINSTORM | 33.3 |
| Reference | 6.6 |
| MLE-LL | Riley's mother bought her cowboy boots. |
| + CD (τCD = 0.5) | Riley had bought cowboy shoes that she had not worn before. |
| + CD (τCD = 1.0) | Her mother bought a new cowboy shirt for Riley. |
| + CD (τCD = 1.5) | Riiley got her new cowboy boots torn. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 0.5) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 1.0) | Riley didn't have cowboy boots. |
| BRAINSTORM | Riley bought a pair of cowboy boots for her mother. |
| + CD (τCD = 0.5) | Rice bought cowboy hats with her money and jeans. |
| + CD (τCD = 1.0) | Rice's mother bought her a pair of cowboy boots to wear to school next to her cowboy outfit. |
| + CD (τCD = 1.5) | Her mom bought her a new pair of cowboy boots. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her a pair of cowboy boots. Riley hated them. |
| + DEXPERTS (α = 0.5) | Riley bought a pair of cowboy pants. |
| + DEXPERTS (α = 1.0) | Riley bought a pair of cowboy pants. |
| QUARK | Riley's mother bought her cowboy boots at the store and Riley hated them. |
| MLE-LL | The whole kitchen was burnt down more quickly. |
| + CD (τCD = 0.5) | Tom poured water in the pan of acid. |
| + CD (τCD = 1.0) | There was a big thunderstorm yesterday. |
| + CD (τCD = 1.5) | There is a big storm outside. |
| + DEXPERTS (α = 0.1) | The whole kitchen was burnt down more quickly. |
| + DEXPERTS (α = 0.5) | The fire in the kitchen spread outwards. |
| + DEXPERTS (α = 1.0) | There are more and more fires in this place. |
| BRAINSTORM | Tom put a lot of fuel on the fire. |
| + CD (τCD = 0.5) | Tom poured a bucket of water to a sink which has a high temperature. |
| + CD (τCD = 1.0) | There was an accident at night. |
| + CD (τCD = 1.5) | Tom poured gasoline to the stove. |
| + DEXPERTS (α = 0.1) | There is a fire in the kitchen. |
| + DEXPERTS (α = 0.5) | The whole kitchen was filled with smoke. |
| + DEXPERTS (α = 1.0) | Tom's kitchen is leaking water. |
| QUARK | The fire in the kitchen was very hot. |
| Model | ART | E-CARE | ||
| \\(Frac(↑)\\) | \\(PPL(↓)\\) | \\(Frac(↑)\\) | \\(PPL(↓)\\) | |
| MLE | 54.1 | 42.6 | 54.5 | 80.4 |
| MLE-LL | 56.6 | 42.5 | 52.6 | 84.8 |
| + CD | 59.9 | 49.8 | 63.4 | 107.3 |
| + DEXPERTS | 56.2 | 51.7 | 57.2 | 108.3 |
| BRAINSTORM | 79.4 | 40.7 | 58.1 | 69.2 |
| + CD | 79.7 | 50.2 | 67.2 | 88.1 |
| + DEXPERTS | 79.0 | 51.5 | 58.1 | 89.3 |
| QUARK | 85.9 | 27.5 | 68.2 | 80.8 |
| BRAINSTORM | ||||
| -\\(\\mathcal{L}_{\\text{margin}}\\) | 69.3 | 44.9 | 54.6 | 73.2 |
| -\\(\\mathcal{L}_{\\text{sim}}\\) | 58.2 | 52.6 | 53.2 | 83.7 |
| BRAINSTORM' | 58.3 | 52.0 | 55.1 | 71.2 |
| Model | ART | E-CARE | ||||||||
| Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | Likely (↓) | L-Likely (↑) | Contra. (?) | Rep. (↓) | Irrel. (↓) | |
| MLE-LL | 42.3 | 15.2 | 22.7 | 9.5 | 10.3 | 35.4 | 15.6 | 5.7 | 18.6 | 24.7 |
| Quark | 14.7 | 20.8 | 51.0 | 4.3 | 9.2 | 35.2 | 15.1 | 5.7 | 3.3 | 40.7 |
| BRAINSTORM | 20.9 | 20.2 | 41.3 | 4.8 | 12.8 | 37.1 | 20.1 | 4.7 | 12.7 | 25.4 |
| Model | Likely | Less likely | Irrel. | ||
| High | Med. | Low | |||
| MLE-LL | 6.7 | 40.7 | 21.2 | 14.7 | 16.7 |
| MLE | 7.3 | 50.0 | 22.1 | 13.3 | 7.3 |
| BRAINSTORM' | 6.7 | 42.0 | 32.6 | 8.7 | 10.0 |
| Reference | 3.3 | 76.7 | 13.4 | 3.3 | 3.3 |
| Dataset | Train | Val | Test | |||||
| Likely | Less Likely | Less Likely | Less Likely | |||||
| MRIINTERPRET | 10097 | 1005 | 121 | — | ||||
| ART | 50509 | 50509 | 1781 | 3562 | ||||
| E-CARE | cause | effect | cause | effect | cause | effect | cause | effect |
| 6855 | 6580 | 6855 | 6580 | 762 | 731 | 1088 | 1044 |
| Model | ART | E-CARE | ||
| Gram. Correct Fluent | Contain Flu. Errors | Gram. Correct Fluent | Contain Flu. Errors | |
| MLE-LL | 93.9 | 6.1 | 99.0 | 1.0 |
| QUARK | 94.6 | 5.4 | 98.0 | 2.0 |
| BRAINSTORM | 93.5 | 6.6 | 95.9 | 4.1 |
| Task | Examples | Output | Explanation |
| Brain MRI | ...... Absence of evidence of restricted diffusion makes it unlikely to be | ~ acute ischemia ~ infarct | In diffusion weighted imaging sequences on MRI of the brain, one of the most common causes of diffusion restriction finding is due to acute ischemic stroke, also known as an infarct. Thus, the absence of restricted diffusion within brain tissue makes an interpretation unlikely to be acute ischemia/infarct. |
| ANLG | O1: Lisa went laptop shopping yesterday. O2: She was thankful she bought it. | + The price raised on the next day. ~ Lisa decided to buy a car. | This scenario makes sense—Lisa was grateful that she had made her laptop purchase the day before, as the price un-expectedly increased the following day. The event focus on Lisa's purchase of a laptop. It is not likely that she would suddenly decide to buy a car without any prior indication of her interest in doing so. |
| E-CARE | Tom goes to the gym every day. | + Tom improves his physical fitness. ~ He gets a promotion from his manager who saw him in the gym. | It directly relates to Tom's daily gym attendance. Regular exercise is a common and effective method for improving physical fitness. It directly connects to the premise involving Tom's daily gym attendance. However, the connection between gym attendance and job promotion is indirect. |
| Model | Hallucination (%) |
| MLE-LL | 23.3 |
| MLE | 30.0 |
| BRAINSTORM | 33.3 |
| Reference | 6.6 |
| MLE-LL | Riley's mother bought her cowboy boots. |
| + CD (τCD = 0.5) | Riley had bought cowboy shoes that she had not worn before. |
| + CD (τCD = 1.0) | Her mother bought a new cowboy shirt for Riley. |
| + CD (τCD = 1.5) | Riiley got her new cowboy boots torn. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 0.5) | Riley's mother bought her cowboy boots. |
| + DEXPERTS (α = 1.0) | Riley didn't have cowboy boots. |
| BRAINSTORM | Riley bought a pair of cowboy boots for her mother. |
| + CD (τCD = 0.5) | Rice bought cowboy hats with her money and jeans. |
| + CD (τCD = 1.0) | Rice's mother bought her a pair of cowboy boots to wear to school next to her cowboy outfit. |
| + CD (τCD = 1.5) | Her mom bought her a new pair of cowboy boots. |
| + DEXPERTS (α = 0.1) | Riley's mother bought her a pair of cowboy boots. Riley hated them. |
| + DEXPERTS (α = 0.5) | Riley bought a pair of cowboy pants. |
| + DEXPERTS (α = 1.0) | Riley bought a pair of cowboy pants. |
| QUARK | Riley's mother bought her cowboy boots at the store and Riley hated them. |
| MLE-LL | The whole kitchen was burnt down more quickly. |
| + CD (τCD = 0.5) | Tom poured water in the pan of acid. |
| + CD (τCD = 1.0) | There was a big thunderstorm yesterday. |
| + CD (τCD = 1.5) | There is a big storm outside. |
| + DEXPERTS (α = 0.1) | The whole kitchen was burnt down more quickly. |
| + DEXPERTS (α = 0.5) | The fire in the kitchen spread outwards. |
| + DEXPERTS (α = 1.0) | There are more and more fires in this place. |
| BRAINSTORM | Tom put a lot of fuel on the fire. |
| + CD (τCD = 0.5) | Tom poured a bucket of water to a sink which has a high temperature. |
| + CD (τCD = 1.0) | There was an accident at night. |
| + CD (τCD = 1.5) | Tom poured gasoline to the stove. |
| + DEXPERTS (α = 0.1) | There is a fire in the kitchen. |
| + DEXPERTS (α = 0.5) | The whole kitchen was filled with smoke. |
| + DEXPERTS (α = 1.0) | Tom's kitchen is leaking water. |
| QUARK | The fire in the kitchen was very hot. |
| Models | Synthetic data | BEA-test | CoNLL-14 | ||||
| P | R | F0.5 | P | R | F0.5 | ||
| Katsumata and Komachi (2020) | - | 68.3 | 57.1 | 65.6 | 69.3 | 45.0 | 62.6 |
| Kiyono et al. (2019) | ✓ | 69.5 | 59.4 | 64.2 | 67.9 | 44.1 | 61.3 |
| Kaneko et al. (2020) | ✓ | 67.1 | 61.0 | 65.6 | 69.2 | 45.6 | 62.6 |
| Rothe et al. (2021) | ✓ | - | - | 67.1 | - | - | 65.1 |
| Omelianchuk et al. (2020) | ✓ | 79.2 | 53.9 | 72.4 | 77.5 | 40.1 | 65.3 |
| AMR-GEC | - | 71.5 | 58.3 | 68.4 | 70.2 | 48.3 | 64.4 |
| Katsumata and Komachi (2020) | - | 68.8 | 57.1 | 66.1 | 69.9 | 45.1 | 63.0 |
| Kiyono et al. (2019) | ✓ | 74.7 | 56.7 | 70.2 | 67.3 | 44.0 | 67.9 |
| Omelianchuk et al. (2020) | ✓ | 79.4 | 57.2 | 73.7 | 78.2 | 41.5 | 66.5 |
| AMR-GEC | - | 73.5 | 55.9 | 69.1 | 70.3 | 48.2 | 64.4 |
| Error Type | T5-GEC | AMR-GEC | ||||
| P | R | F0.5 | P | R | F0.5 | |
| PUNCT | 79.8 | 49.4 | 71.0 | 78.7 | 72.9 | 77.4 |
| DET | 78.6 | 64.8 | 75.4 | 78.6 | 65.8 | 75.7 |
| PREP | 72.9 | 48.0 | 66.0 | 73.1 | 61.5 | 70.4 |
| ORTH | 84.6 | 55.7 | 76.7 | 69.5 | 62.9 | 68.1 |
| SPELL | 83.0 | 58.3 | 76.5 | 80.9 | 61.9 | 76.2 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GAT | 68.26 | 63.41 | 67.23 |
| AMR-DeepGCN | 66.34 | 62.57 | 65.55 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GCN (node/edge) | 73.52 | 55.91 | 69.14 |
| AMR-GCN (subgraph) | 72.12 | 57.60 | 68.60 |
| Models | Synthetic data | BEA-test | CoNLL-14 | ||||
| P | R | F0.5 | P | R | F0.5 | ||
| Katsumata and Komachi (2020) | - | 68.3 | 57.1 | 65.6 | 69.3 | 45.0 | 62.6 |
| Kiyono et al. (2019) | ✓ | 69.5 | 59.4 | 64.2 | 67.9 | 44.1 | 61.3 |
| Kaneko et al. (2020) | ✓ | 67.1 | 61.0 | 65.6 | 69.2 | 45.6 | 62.6 |
| Rothe et al. (2021) | ✓ | - | - | 67.1 | - | - | 65.1 |
| Omelianchuk et al. (2020) | ✓ | 79.2 | 53.9 | 72.4 | 77.5 | 40.1 | 65.3 |
| AMR-GEC | - | 71.5 | 58.3 | 68.4 | 70.2 | 48.3 | 64.4 |
| Katsumata and Komachi (2020) | - | 68.8 | 57.1 | 66.1 | 69.9 | 45.1 | 63.0 |
| Kiyono et al. (2019) | ✓ | 74.7 | 56.7 | 70.2 | 67.3 | 44.0 | 67.9 |
| Omelianchuk et al. (2020) | ✓ | 79.4 | 57.2 | 73.7 | 78.2 | 41.5 | 66.5 |
| AMR-GEC | - | 73.5 | 55.9 | 69.1 | 70.3 | 48.2 | 64.4 |
| Error Type | T5-GEC | AMR-GEC | ||||
| P | R | F0.5 | P | R | F0.5 | |
| PUNCT | 79.8 | 49.4 | 71.0 | 78.7 | 72.9 | 77.4 |
| DET | 78.6 | 64.8 | 75.4 | 78.6 | 65.8 | 75.7 |
| PREP | 72.9 | 48.0 | 66.0 | 73.1 | 61.5 | 70.4 |
| ORTH | 84.6 | 55.7 | 76.7 | 69.5 | 62.9 | 68.1 |
| SPELL | 83.0 | 58.3 | 76.5 | 80.9 | 61.9 | 76.2 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GAT | 68.26 | 63.41 | 67.23 |
| AMR-DeepGCN | 66.34 | 62.57 | 65.55 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GCN (node/edge) | 73.52 | 55.91 | 69.14 |
| AMR-GCN (subgraph) | 72.12 | 57.60 | 68.60 |
| Models | Synthetic data | BEA-test | CoNLL-14 | ||||
| P | R | F0.5 | P | R | F0.5 | ||
| Katsumata and Komachi (2020) | - | 68.3 | 57.1 | 65.6 | 69.3 | 45.0 | 62.6 |
| Kiyono et al. (2019) | ✓ | 69.5 | 59.4 | 64.2 | 67.9 | 44.1 | 61.3 |
| Kaneko et al. (2020) | ✓ | 67.1 | 61.0 | 65.6 | 69.2 | 45.6 | 62.6 |
| Rothe et al. (2021) | ✓ | - | - | 67.1 | - | - | 65.1 |
| Omelianchuk et al. (2020) | ✓ | 79.2 | 53.9 | 72.4 | 77.5 | 40.1 | 65.3 |
| AMR-GEC | - | 71.5 | 58.3 | 68.4 | 70.2 | 48.3 | 64.4 |
| Katsumata and Komachi (2020) | - | 68.8 | 57.1 | 66.1 | 69.9 | 45.1 | 63.0 |
| Kiyono et al. (2019) | ✓ | 74.7 | 56.7 | 70.2 | 67.3 | 44.0 | 67.9 |
| Omelianchuk et al. (2020) | ✓ | 79.4 | 57.2 | 73.7 | 78.2 | 41.5 | 66.5 |
| AMR-GEC | - | 73.5 | 55.9 | 69.1 | 70.3 | 48.2 | 64.4 |
| Error Type | T5-GEC | AMR-GEC | ||||
| P | R | F0.5 | P | R | F0.5 | |
| PUNCT | 79.8 | 49.4 | 71.0 | 78.7 | 72.9 | 77.4 |
| DET | 78.6 | 64.8 | 75.4 | 78.6 | 65.8 | 75.7 |
| PREP | 72.9 | 48.0 | 66.0 | 73.1 | 61.5 | 70.4 |
| ORTH | 84.6 | 55.7 | 76.7 | 69.5 | 62.9 | 68.1 |
| SPELL | 83.0 | 58.3 | 76.5 | 80.9 | 61.9 | 76.2 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GAT | 68.26 | 63.41 | 67.23 |
| AMR-DeepGCN | 66.34 | 62.57 | 65.55 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GCN (node/edge) | 73.52 | 55.91 | 69.14 |
| AMR-GCN (subgraph) | 72.12 | 57.60 | 68.60 |
| Models | Synthetic data | BEA-test | CoNLL-14 | ||||
| P | R | F0.5 | P | R | F0.5 | ||
| Katsumata and Komachi (2020) | - | 68.3 | 57.1 | 65.6 | 69.3 | 45.0 | 62.6 |
| Kiyono et al. (2019) | ✓ | 69.5 | 59.4 | 64.2 | 67.9 | 44.1 | 61.3 |
| Kaneko et al. (2020) | ✓ | 67.1 | 61.0 | 65.6 | 69.2 | 45.6 | 62.6 |
| Rothe et al. (2021) | ✓ | - | - | 67.1 | - | - | 65.1 |
| Omelianchuk et al. (2020) | ✓ | 79.2 | 53.9 | 72.4 | 77.5 | 40.1 | 65.3 |
| AMR-GEC | - | 71.5 | 58.3 | 68.4 | 70.2 | 48.3 | 64.4 |
| Katsumata and Komachi (2020) | - | 68.8 | 57.1 | 66.1 | 69.9 | 45.1 | 63.0 |
| Kiyono et al. (2019) | ✓ | 74.7 | 56.7 | 70.2 | 67.3 | 44.0 | 67.9 |
| Omelianchuk et al. (2020) | ✓ | 79.4 | 57.2 | 73.7 | 78.2 | 41.5 | 66.5 |
| AMR-GEC | - | 73.5 | 55.9 | 69.1 | 70.3 | 48.2 | 64.4 |
| Error Type | T5-GEC | AMR-GEC | ||||
| P | R | F0.5 | P | R | F0.5 | |
| PUNCT | 79.8 | 49.4 | 71.0 | 78.7 | 72.9 | 77.4 |
| DET | 78.6 | 64.8 | 75.4 | 78.6 | 65.8 | 75.7 |
| PREP | 72.9 | 48.0 | 66.0 | 73.1 | 61.5 | 70.4 |
| ORTH | 84.6 | 55.7 | 76.7 | 69.5 | 62.9 | 68.1 |
| SPELL | 83.0 | 58.3 | 76.5 | 80.9 | 61.9 | 76.2 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GAT | 68.26 | 63.41 | 67.23 |
| AMR-DeepGCN | 66.34 | 62.57 | 65.55 |
| Model | P | R | F0.5 |
| T5-GEC | 71.47 | 53.46 | 66.96 |
| AMR-GCN | 72.95 | 52.17 | 67.57 |
| AMR-GCN (node/edge) | 73.52 | 55.91 | 69.14 |
| AMR-GCN (subgraph) | 72.12 | 57.60 | 68.60 |
| Model | AST | WD | ||||||
| ABCD | SGD | ABCD | SGD | |||||
| B-Slot | Value | B-Slot | Value | B-Slot | Value | B-Slot | Value | |
| T5 | 79.5 | 82.2 | 51.8 | 31.6 | 65.9 | 66.8 | 58.7 | 28.3 |
| T5 + guide | 81.3 | 82.5 | NA | NA | 56.8 | 58.4 | NA | NA |
| KADS | 85.2 | 83.1 | 63.2 | 39.5 | 72.5 | 73.0 | 53.1 | 23.8 |
| Model | ABCD | SGD | ||
| B-Slot | Value | B-Slot | Value | |
| T5 | 0.0 | 11.6 | 46.2 | 25.8 |
| T5 + guide | 2.7 | 17.9 | NA | NA |
| KADS | 11.6 | 21.4 | 49.8 | 32.2 |
| full data | 94.6 | 85.7 | 61.3 | 38.1 |
| Dataset | Accuracy | ||
| DDM | MLM | AST | |
| ABCD | 98.0 | 82.3 | 85.4 |
| SGD | 66.2 | 74.9 | 63.9 |
| Model | B-Slot | Value |
| none | 82.7 | 79.4 |
| MLM only | 81.5 | 79.0 |
| DDM only | 82.6 | 78.5 |
| full | 85.2 | 83.1 |
| Input | [agent] hello! how can i help you today? [customer] i’m thinking about buying an item but first i would like to get some more info on the product [agent] sure. i can help you with that. what item are you looking for more information on? [customer] the tommy hilifiger shirt [agent] and what would you like to know about it? [customer] i would like to know how long is the arm length [agent] sure give me one second and i can find that out for you [customer] ok [action] search faq |
| Output | search shirt |
| Document | get shirt info [SEP] search faq; search shirt; select faq |
| Input | [customer] i am interested to know how the weather is going to be on 7th of march in san diego. |
| Output | offer temperature; offer precipitation |
| Document | get the weather of a certain location on a date [SEP] [required] city [optional] date [result] precipitation; humidity; wind; temperature; city; date |
| Model | AST | WD | ||||||
| ABCD | SGD | ABCD | SGD | |||||
| B-Slot | Value | B-Slot | Value | B-Slot | Value | B-Slot | Value | |
| T5 | 79.5 | 82.2 | 51.8 | 31.6 | 65.9 | 66.8 | 58.7 | 28.3 |
| T5 + guide | 81.3 | 82.5 | NA | NA | 56.8 | 58.4 | NA | NA |
| KADS | 85.2 | 83.1 | 63.2 | 39.5 | 72.5 | 73.0 | 53.1 | 23.8 |
| Model | ABCD | SGD | ||
| B-Slot | Value | B-Slot | Value | |
| T5 | 0.0 | 11.6 | 46.2 | 25.8 |
| T5 + guide | 2.7 | 17.9 | NA | NA |
| KADS | 11.6 | 21.4 | 49.8 | 32.2 |
| full data | 94.6 | 85.7 | 61.3 | 38.1 |
| Dataset | Accuracy | ||
| DDM | MLM | AST | |
| ABCD | 98.0 | 82.3 | 85.4 |
| SGD | 66.2 | 74.9 | 63.9 |
| Model | B-Slot | Value |
| none | 82.7 | 79.4 |
| MLM only | 81.5 | 79.0 |
| DDM only | 82.6 | 78.5 |
| full | 85.2 | 83.1 |
| Input | [agent] hello! how can i help you today? [customer] i’m thinking about buying an item but first i would like to get some more info on the product [agent] sure. i can help you with that. what item are you looking for more information on? [customer] the tommy hilifiger shirt [agent] and what would you like to know about it? [customer] i would like to know how long is the arm length [agent] sure give me one second and i can find that out for you [customer] ok [action] search faq |
| Output | search shirt |
| Document | get shirt info [SEP] search faq; search shirt; select faq |
| Input | [customer] i am interested to know how the weather is going to be on 7th of march in san diego. |
| Output | offer temperature; offer precipitation |
| Document | get the weather of a certain location on a date [SEP] [required] city [optional] date [result] precipitation; humidity; wind; temperature; city; date |
| Model | AST | WD | ||||||
| ABCD | SGD | ABCD | SGD | |||||
| B-Slot | Value | B-Slot | Value | B-Slot | Value | B-Slot | Value | |
| T5 | 79.5 | 82.2 | 51.8 | 31.6 | 65.9 | 66.8 | 58.7 | 28.3 |
| T5 + guide | 81.3 | 82.5 | NA | NA | 56.8 | 58.4 | NA | NA |
| KADS | 85.2 | 83.1 | 63.2 | 39.5 | 72.5 | 73.0 | 53.1 | 23.8 |
| Model | ABCD | SGD | ||
| B-Slot | Value | B-Slot | Value | |
| T5 | 0.0 | 11.6 | 46.2 | 25.8 |
| T5 + guide | 2.7 | 17.9 | NA | NA |
| KADS | 11.6 | 21.4 | 49.8 | 32.2 |
| full data | 94.6 | 85.7 | 61.3 | 38.1 |
| Dataset | Accuracy | ||
| DDM | MLM | AST | |
| ABCD | 98.0 | 82.3 | 85.4 |
| SGD | 66.2 | 74.9 | 63.9 |
| Model | B-Slot | Value |
| none | 82.7 | 79.4 |
| MLM only | 81.5 | 79.0 |
| DDM only | 82.6 | 78.5 |
| full | 85.2 | 83.1 |
| Input | [agent] hello! how can i help you today? [customer] i’m thinking about buying an item but first i would like to get some more info on the product [agent] sure. i can help you with that. what item are you looking for more information on? [customer] the tommy hilifiger shirt [agent] and what would you like to know about it? [customer] i would like to know how long is the arm length [agent] sure give me one second and i can find that out for you [customer] ok [action] search faq |
| Output | search shirt |
| Document | get shirt info [SEP] search faq; search shirt; select faq |
| Input | [customer] i am interested to know how the weather is going to be on 7th of march in san diego. |
| Output | offer temperature; offer precipitation |
| Document | get the weather of a certain location on a date [SEP] [required] city [optional] date [result] precipitation; humidity; wind; temperature; city; date |
| Model | AST | WD | ||||||
| ABCD | SGD | ABCD | SGD | |||||
| B-Slot | Value | B-Slot | Value | B-Slot | Value | B-Slot | Value | |
| T5 | 79.5 | 82.2 | 51.8 | 31.6 | 65.9 | 66.8 | 58.7 | 28.3 |
| T5 + guide | 81.3 | 82.5 | NA | NA | 56.8 | 58.4 | NA | NA |
| KADS | 85.2 | 83.1 | 63.2 | 39.5 | 72.5 | 73.0 | 53.1 | 23.8 |
| Model | ABCD | SGD | ||
| B-Slot | Value | B-Slot | Value | |
| T5 | 0.0 | 11.6 | 46.2 | 25.8 |
| T5 + guide | 2.7 | 17.9 | NA | NA |
| KADS | 11.6 | 21.4 | 49.8 | 32.2 |
| full data | 94.6 | 85.7 | 61.3 | 38.1 |
| Dataset | Accuracy | ||
| DDM | MLM | AST | |
| ABCD | 98.0 | 82.3 | 85.4 |
| SGD | 66.2 | 74.9 | 63.9 |
| Model | B-Slot | Value |
| none | 82.7 | 79.4 |
| MLM only | 81.5 | 79.0 |
| DDM only | 82.6 | 78.5 |
| full | 85.2 | 83.1 |
| Input | [agent] hello! how can i help you today? [customer] i’m thinking about buying an item but first i would like to get some more info on the product [agent] sure. i can help you with that. what item are you looking for more information on? [customer] the tommy hilifiger shirt [agent] and what would you like to know about it? [customer] i would like to know how long is the arm length [agent] sure give me one second and i can find that out for you [customer] ok [action] search faq |
| Output | search shirt |
| Document | get shirt info [SEP] search faq; search shirt; select faq |
| Input | [customer] i am interested to know how the weather is going to be on 7th of march in san diego. |
| Output | offer temperature; offer precipitation |
| Document | get the weather of a certain location on a date [SEP] [required] city [optional] date [result] precipitation; humidity; wind; temperature; city; date |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 17.6 | 20.9 | 3.3 |
| 5M | 22.6 | 23.5 | 0.9 | |
| 25M | 23.1 | 23.9 | 0.8 | |
| 6 × 6 | 0.5M | 18.0 | 22.5 | 4.5 |
| 5M | 24.7 | 25.7 | 1.0 | |
| 25M | 25.3 | 25.9 | 0.6 | |
| 24 × 6 | 0.5M | 18.6 | 22.9 | 4.3 |
| 5M | 25.7 | 26.5 | 0.8 | |
| 25M | 26.4 | 26.4 | 0.0 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 18.6 | 22.8 | 4.2 |
| 5M | 23.8 | 26.0 | 2.2 | |
| 57M | 24.2 | 26.3 | 2.1 | |
| 6 × 6 | 0.5M | 18.7 | 23.9 | 5.2 |
| 5M | 25.1 | 27.8 | 2.7 | |
| 57M | 26.6 | 28.4 | 1.8 | |
| 24 × 6 | 0.5M | 19.2 | 24.1 | 4.9 |
| 5M | 25.9 | 28.4 | 2.5 | |
| 57M | 27.1 | 29.0 | 1.9 |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| en-de | en-es | en-fr | Avg. | en-de | en-es | en-fr | Avg. | |||
| 2 × 2 | 1.5M | 24.9 | 29.6 | 29.3 | 27.9 | 28.2 | 31.4 | 31.3 | 30.3 | 2.4 |
| 15M | 29.5 | 31.8 | 32.6 | 31.3 | 31.9 | 33.0 | 33.5 | 32.8 | 1.5 | |
| 300M | 30.1 | 32.5 | 32.8 | 31.8 | 32.3 | 33.2 | 33.7 | 33.0 | 1.2 | |
| 6 × 6 | 1.5M | 25.9 | 30.1 | 30.2 | 28.7 | 31.0 | 32.5 | 32.5 | 32.0 | 3.3 |
| 15M | 33.2 | 33.9 | 35.0 | 34.0 | 36.0 | 34.4 | 35.4 | 35.2 | 1.2 | |
| 300M | 34.5 | 34.4 | 35.6 | 34.8 | 36.1 | 34.5 | 35.8 | 35.4 | 0.6 | |
| 24 × 6 | 1.5M | 25.8 | 30.0 | 29.6 | 28.5 | 31.8 | 33.3 | 33.3 | 32.8 | 4.3 |
| 15M | 35.0 | 34.5 | 35.7 | 35.0 | 36.9 | 34.9 | 36.4 | 36.0 | 1.0 | |
| 300M | 36.1 | 34.9 | 36.8 | 35.9 | 37.9 | 35.2 | 37.0 | 36.7 | 0.8 |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| de-en | es-en | fr-en | Avg. | de-en | es-en | fr-en | Avg. | |||
| 2 × 2 | 1.5M | 28.4 | 29.1 | 32.2 | 29.9 | 31.9 | 30.8 | 34.1 | 32.3 | 2.4 |
| 15M | 34.6 | 32.8 | 36.2 | 34.5 | 37.1 | 33.6 | 37.5 | 36.1 | 1.6 | |
| 300M | 35.5 | 32.7 | 36.9 | 35.0 | 38.3 | 33.8 | 37.6 | 36.6 | 1.6 | |
| 6 × 6 | 1.5M | 30.8 | 30.1 | 33.2 | 31.4 | 35.3 | 31.8 | 35.3 | 34.1 | 2.7 |
| 15M | 38.1 | 34.0 | 38.3 | 36.8 | 40.2 | 35.0 | 39.1 | 38.1 | 1.3 | |
| 300M | 40.0 | 34.2 | 38.8 | 37.7 | 41.3 | 35.5 | 39.6 | 38.8 | 1.1 | |
| 24 × 6 | 1.5M | 31.4 | 30.5 | 33.8 | 31.9 | 36.0 | 32.8 | 36.2 | 35.0 | 3.1 |
| 15M | 38.8 | 34.6 | 39.6 | 37.6 | 40.8 | 35.6 | 40.1 | 38.8 | 1.2 | |
| 300M | 40.1 | 35.6 | 40.9 | 38.9 | 41.8 | 36.3 | 41.2 | 39.8 | 0.9 |
| German targets source | Size | Must-C v2 | IWSLT tst | ||||||||
| dev | tst | 2010 | 2013 | 2014 | 2015 | 2018 | 2019 | 2020 | Avg. | ||
| Real | 590K | 25.2 | 27.8 | 22.1 | 27.9 | 23.8 | 21.3 | 20.9 | 20.1 | 20.9 | 22.4 |
| Synthetic, WMT21 teacher + fine-tuned on IWSLT + extra ASR data | 590K | 28.3 | 28.9 | 24.5 | 28.0 | 23.9 | 23.0 | 22.6 | 21.7 | 23.3 | 23.9 |
| 590K | 29.2 | 30.2 | 25.0 | 29.5 | 24.8 | 24.9 | 24.2 | 23.4 | 25.3 | 25.3 | |
| 1.25M | 30.6 | 31.0 | 27.2 | 31.3 | 27.4 | 25.8 | 25.1 | 24.3 | 26.4 | 26.8 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | IWSLT | 27.9 | 30.4 | 2.5 |
| Medical | 27.3 | 30.6 | 3.3 | |
| Law | 36.5 | 39.0 | 2.5 | |
| 6 × 6 | IWSLT | 30.0 | 31.6 | 1.6 |
| Medical | 30.1 | 31.7 | 1.6 | |
| Law | 39.8 | 40.8 | 1.0 | |
| 24 × 6 | IWSLT | 31.2 | 33.1 | 1.9 |
| Medical | 31.3 | 33.1 | 1.8 | |
| Law | 40.4 | 42.1 | 1.7 |
| Fine-tuning targets | 2 × 2 | 6 × 6 | 24 × 6 |
| Training from scratch | |||
| Real | 16.9 | 17.0 | 19.2 |
| Synthetic, WMT Teacher | 20.6 | 21.9 | 22.7 |
| + fine-tuned on IWSLT | 21.8 | 22.9 | 25.9 |
| Pre-training on real data | |||
| Real | 30.9 | 33.6 | 34.5 |
| Synthetic, WMT Teacher | 30.9 | 32.6 | 33.7 |
| + fine-tuned on IWSLT | 32.5 | 34.1 | 35.5 |
| Pre-training on synthetic data | |||
| Real | 32.4 | 33.8 | 34.9 |
| Synthetic, WMT Teacher | 32.0 | 33.3 | 33.6 |
| + fine-tuned on IWSLT | 33.4 | 35.1 | 35.6 |
| Language | Targets | Predictive Entropy | ||
| 2 × 2 | 6 × 6 | 24 × 6 | ||
| English to German | Real | 2.4 | 2.4 | 2.5 |
| Synthetic | 1.8 | 1.8 | 1.7 | |
| English to Russian | Real | 2.6 | 2.4 | 2.3 |
| Synthetic | 1.8 | 1.8 | 1.9 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 17.6 | 20.9 | 3.3 |
| 5M | 22.6 | 23.5 | 0.9 | |
| 25M | 23.1 | 23.9 | 0.8 | |
| 6 × 6 | 0.5M | 18.0 | 22.5 | 4.5 |
| 5M | 24.7 | 25.7 | 1.0 | |
| 25M | 25.3 | 25.9 | 0.6 | |
| 24 × 6 | 0.5M | 18.6 | 22.9 | 4.3 |
| 5M | 25.7 | 26.5 | 0.8 | |
| 25M | 26.4 | 26.4 | 0.0 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 18.6 | 22.8 | 4.2 |
| 5M | 23.8 | 26.0 | 2.2 | |
| 57M | 24.2 | 26.3 | 2.1 | |
| 6 × 6 | 0.5M | 18.7 | 23.9 | 5.2 |
| 5M | 25.1 | 27.8 | 2.7 | |
| 57M | 26.6 | 28.4 | 1.8 | |
| 24 × 6 | 0.5M | 19.2 | 24.1 | 4.9 |
| 5M | 25.9 | 28.4 | 2.5 | |
| 57M | 27.1 | 29.0 | 1.9 |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| en-de | en-es | en-fr | Avg. | en-de | en-es | en-fr | Avg. | |||
| 2 × 2 | 1.5M | 24.9 | 29.6 | 29.3 | 27.9 | 28.2 | 31.4 | 31.3 | 30.3 | 2.4 |
| 15M | 29.5 | 31.8 | 32.6 | 31.3 | 31.9 | 33.0 | 33.5 | 32.8 | 1.5 | |
| 300M | 30.1 | 32.5 | 32.8 | 31.8 | 32.3 | 33.2 | 33.7 | 33.0 | 1.2 | |
| 6 × 6 | 1.5M | 25.9 | 30.1 | 30.2 | 28.7 | 31.0 | 32.5 | 32.5 | 32.0 | 3.3 |
| 15M | 33.2 | 33.9 | 35.0 | 34.0 | 36.0 | 34.4 | 35.4 | 35.2 | 1.2 | |
| 300M | 34.5 | 34.4 | 35.6 | 34.8 | 36.1 | 34.5 | 35.8 | 35.4 | 0.6 | |
| 24 × 6 | 1.5M | 25.8 | 30.0 | 29.6 | 28.5 | 31.8 | 33.3 | 33.3 | 32.8 | 4.3 |
| 15M | 35.0 | 34.5 | 35.7 | 35.0 | 36.9 | 34.9 | 36.4 | 36.0 | 1.0 | |
| 300M | 36.1 | 34.9 | 36.8 | 35.9 | 37.9 | 35.2 | 37.0 | 36.7 | 0.8 |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| de-en | es-en | fr-en | Avg. | de-en | es-en | fr-en | Avg. | |||
| 2 × 2 | 1.5M | 28.4 | 29.1 | 32.2 | 29.9 | 31.9 | 30.8 | 34.1 | 32.3 | 2.4 |
| 15M | 34.6 | 32.8 | 36.2 | 34.5 | 37.1 | 33.6 | 37.5 | 36.1 | 1.6 | |
| 300M | 35.5 | 32.7 | 36.9 | 35.0 | 38.3 | 33.8 | 37.6 | 36.6 | 1.6 | |
| 6 × 6 | 1.5M | 30.8 | 30.1 | 33.2 | 31.4 | 35.3 | 31.8 | 35.3 | 34.1 | 2.7 |
| 15M | 38.1 | 34.0 | 38.3 | 36.8 | 40.2 | 35.0 | 39.1 | 38.1 | 1.3 | |
| 300M | 40.0 | 34.2 | 38.8 | 37.7 | 41.3 | 35.5 | 39.6 | 38.8 | 1.1 | |
| 24 × 6 | 1.5M | 31.4 | 30.5 | 33.8 | 31.9 | 36.0 | 32.8 | 36.2 | 35.0 | 3.1 |
| 15M | 38.8 | 34.6 | 39.6 | 37.6 | 40.8 | 35.6 | 40.1 | 38.8 | 1.2 | |
| 300M | 40.1 | 35.6 | 40.9 | 38.9 | 41.8 | 36.3 | 41.2 | 39.8 | 0.9 |
| German targets source | Size | Must-C v2 | IWSLT tst | ||||||||
| dev | tst | 2010 | 2013 | 2014 | 2015 | 2018 | 2019 | 2020 | Avg. | ||
| Real | 590K | 25.2 | 27.8 | 22.1 | 27.9 | 23.8 | 21.3 | 20.9 | 20.1 | 20.9 | 22.4 |
| Synthetic, WMT21 teacher + fine-tuned on IWSLT + extra ASR data | 590K | 28.3 | 28.9 | 24.5 | 28.0 | 23.9 | 23.0 | 22.6 | 21.7 | 23.3 | 23.9 |
| 590K | 29.2 | 30.2 | 25.0 | 29.5 | 24.8 | 24.9 | 24.2 | 23.4 | 25.3 | 25.3 | |
| 1.25M | 30.6 | 31.0 | 27.2 | 31.3 | 27.4 | 25.8 | 25.1 | 24.3 | 26.4 | 26.8 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | IWSLT | 27.9 | 30.4 | 2.5 |
| Medical | 27.3 | 30.6 | 3.3 | |
| Law | 36.5 | 39.0 | 2.5 | |
| 6 × 6 | IWSLT | 30.0 | 31.6 | 1.6 |
| Medical | 30.1 | 31.7 | 1.6 | |
| Law | 39.8 | 40.8 | 1.0 | |
| 24 × 6 | IWSLT | 31.2 | 33.1 | 1.9 |
| Medical | 31.3 | 33.1 | 1.8 | |
| Law | 40.4 | 42.1 | 1.7 |
| Fine-tuning targets | 2 × 2 | 6 × 6 | 24 × 6 |
| Training from scratch | |||
| Real | 16.9 | 17.0 | 19.2 |
| Synthetic, WMT Teacher | 20.6 | 21.9 | 22.7 |
| + fine-tuned on IWSLT | 21.8 | 22.9 | 25.9 |
| Pre-training on real data | |||
| Real | 30.9 | 33.6 | 34.5 |
| Synthetic, WMT Teacher | 30.9 | 32.6 | 33.7 |
| + fine-tuned on IWSLT | 32.5 | 34.1 | 35.5 |
| Pre-training on synthetic data | |||
| Real | 32.4 | 33.8 | 34.9 |
| Synthetic, WMT Teacher | 32.0 | 33.3 | 33.6 |
| + fine-tuned on IWSLT | 33.4 | 35.1 | 35.6 |
| Language | Targets | Predictive Entropy | ||
| 2 × 2 | 6 × 6 | 24 × 6 | ||
| English to German | Real | 2.4 | 2.4 | 2.5 |
| Synthetic | 1.8 | 1.8 | 1.7 | |
| English to Russian | Real | 2.6 | 2.4 | 2.3 |
| Synthetic | 1.8 | 1.8 | 1.9 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 17.6 | 20.9 | 3.3 |
| 5M | 22.6 | 23.5 | 0.9 | |
| 25M | 23.1 | 23.9 | 0.8 | |
| 6 × 6 | 0.5M | 18.0 | 22.5 | 4.5 |
| 5M | 24.7 | 25.7 | 1.0 | |
| 25M | 25.3 | 25.9 | 0.6 | |
| 24 × 6 | 0.5M | 18.6 | 22.9 | 4.3 |
| 5M | 25.7 | 26.5 | 0.8 | |
| 25M | 26.4 | 26.4 | 0.0 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 18.6 | 22.8 | 4.2 |
| 5M | 23.8 | 26.0 | 2.2 | |
| 57M | 24.2 | 26.3 | 2.1 | |
| 6 × 6 | 0.5M | 18.7 | 23.9 | 5.2 |
| 5M | 25.1 | 27.8 | 2.7 | |
| 57M | 26.6 | 28.4 | 1.8 | |
| 24 × 6 | 0.5M | 19.2 | 24.1 | 4.9 |
| 5M | 25.9 | 28.4 | 2.5 | |
| 57M | 27.1 | 29.0 | 1.9 |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| en-de | en-es | en-fr | Avg. | en-de | en-es | en-fr | Avg. | |||
| 2 × 2 | 1.5M | 24.9 | 29.6 | 29.3 | 27.9 | 28.2 | 31.4 | 31.3 | 30.3 | 2.4 |
| 15M | 29.5 | 31.8 | 32.6 | 31.3 | 31.9 | 33.0 | 33.5 | 32.8 | 1.5 | |
| 300M | 30.1 | 32.5 | 32.8 | 31.8 | 32.3 | 33.2 | 33.7 | 33.0 | 1.2 | |
| 6 × 6 | 1.5M | 25.9 | 30.1 | 30.2 | 28.7 | 31.0 | 32.5 | 32.5 | 32.0 | 3.3 |
| 15M | 33.2 | 33.9 | 35.0 | 34.0 | 36.0 | 34.4 | 35.4 | 35.2 | 1.2 | |
| 300M | 34.5 | 34.4 | 35.6 | 34.8 | 36.1 | 34.5 | 35.8 | 35.4 | 0.6 | |
| 24 × 6 | 1.5M | 25.8 | 30.0 | 29.6 | 28.5 | 31.8 | 33.3 | 33.3 | 32.8 | 4.3 |
| 15M | 35.0 | 34.5 | 35.7 | 35.0 | 36.9 | 34.9 | 36.4 | 36.0 | 1.0 | |
| 300M | 36.1 | 34.9 | 36.8 | 35.9 | 37.9 | 35.2 | 37.0 | 36.7 | 0.8 | |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| de-en | es-en | fr-en | Avg. | de-en | es-en | fr-en | Avg. | |||
| 2 × 2 | 1.5M | 28.4 | 29.1 | 32.2 | 29.9 | 31.9 | 30.8 | 34.1 | 32.3 | 2.4 |
| 15M | 34.6 | 32.8 | 36.2 | 34.5 | 37.1 | 33.6 | 37.5 | 36.1 | 1.6 | |
| 300M | 35.5 | 32.7 | 36.9 | 35.0 | 38.3 | 33.8 | 37.6 | 36.6 | 1.6 | |
| 6 × 6 | 1.5M | 30.8 | 30.1 | 33.2 | 31.4 | 35.3 | 31.8 | 35.3 | 34.1 | 2.7 |
| 15M | 38.1 | 34.0 | 38.3 | 36.8 | 40.2 | 35.0 | 39.1 | 38.1 | 1.3 | |
| 300M | 40.0 | 34.2 | 38.8 | 37.7 | 41.3 | 35.5 | 39.6 | 38.8 | 1.1 | |
| 24 × 6 | 1.5M | 31.4 | 30.5 | 33.8 | 31.9 | 36.0 | 32.8 | 36.2 | 35.0 | 3.1 |
| 15M | 38.8 | 34.6 | 39.6 | 37.6 | 40.8 | 35.6 | 40.1 | 38.8 | 1.2 | |
| 300M | 40.1 | 35.6 | 40.9 | 38.9 | 41.8 | 36.3 | 41.2 | 39.8 | 0.9 | |
| German targets source | Size | Must-C v2 | IWSLT tst | ||||||||
| dev | tst | 2010 | 2013 | 2014 | 2015 | 2018 | 2019 | 2020 | Avg. | ||
| Real | 590K | 25.2 | 27.8 | 22.1 | 27.9 | 23.8 | 21.3 | 20.9 | 20.1 | 20.9 | 22.4 |
| Synthetic, WMT21 teacher + fine-tuned on IWSLT + extra ASR data | 590K | 28.3 | 28.9 | 24.5 | 28.0 | 23.9 | 23.0 | 22.6 | 21.7 | 23.3 | 23.9 |
| 590K | 29.2 | 30.2 | 25.0 | 29.5 | 24.8 | 24.9 | 24.2 | 23.4 | 25.3 | 25.3 | |
| 1.25M | 30.6 | 31.0 | 27.2 | 31.3 | 27.4 | 25.8 | 25.1 | 24.3 | 26.4 | 26.8 | |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | IWSLT | 27.9 | 30.4 | 2.5 |
| Medical | 27.3 | 30.6 | 3.3 | |
| Law | 36.5 | 39.0 | 2.5 | |
| 6 × 6 | IWSLT | 30.0 | 31.6 | 1.6 |
| Medical | 30.1 | 31.7 | 1.6 | |
| Law | 39.8 | 40.8 | 1.0 | |
| 24 × 6 | IWSLT | 31.2 | 33.1 | 1.9 |
| Medical | 31.3 | 33.1 | 1.8 | |
| Law | 40.4 | 42.1 | 1.7 |
| Fine-tuning targets | 2 × 2 | 6 × 6 | 24 × 6 |
| Training from scratch | |||
| Real | 16.9 | 17.0 | 19.2 |
| Synthetic, WMT Teacher | 20.6 | 21.9 | 22.7 |
| + fine-tuned on IWSLT | 21.8 | 22.9 | 25.9 |
| Pre-training on real data | |||
| Real | 30.9 | 33.6 | 34.5 |
| Synthetic, WMT Teacher | 30.9 | 32.6 | 33.7 |
| + fine-tuned on IWSLT | 32.5 | 34.1 | 35.5 |
| Pre-training on synthetic data | |||
| Real | 32.4 | 33.8 | 34.9 |
| Synthetic, WMT Teacher | 32.0 | 33.3 | 33.6 |
| + fine-tuned on IWSLT | 33.4 | 35.1 | 35.6 |
| Language | Targets | Predictive Entropy | ||
| 2 × 2 | 6 × 6 | 24 × 6 | ||
| English to German | Real | 2.4 | 2.4 | 2.5 |
| Synthetic | 1.8 | 1.8 | 1.7 | |
| English to Russian | Real | 2.6 | 2.4 | 2.3 |
| Synthetic | 1.8 | 1.8 | 1.9 | |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 17.6 | 20.9 | 3.3 |
| 5M | 22.6 | 23.5 | 0.9 | |
| 25M | 23.1 | 23.9 | 0.8 | |
| 6 × 6 | 0.5M | 18.0 | 22.5 | 4.5 |
| 5M | 24.7 | 25.7 | 1.0 | |
| 25M | 25.3 | 25.9 | 0.6 | |
| 24 × 6 | 0.5M | 18.6 | 22.9 | 4.3 |
| 5M | 25.7 | 26.5 | 0.8 | |
| 25M | 26.4 | 26.4 | 0.0 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | 0.5M | 18.6 | 22.8 | 4.2 |
| 5M | 23.8 | 26.0 | 2.2 | |
| 57M | 24.2 | 26.3 | 2.1 | |
| 6 × 6 | 0.5M | 18.7 | 23.9 | 5.2 |
| 5M | 25.1 | 27.8 | 2.7 | |
| 57M | 26.6 | 28.4 | 1.8 | |
| 24 × 6 | 0.5M | 19.2 | 24.1 | 4.9 |
| 5M | 25.9 | 28.4 | 2.5 | |
| 57M | 27.1 | 29.0 | 1.9 |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| en-de | en-es | en-fr | Avg. | en-de | en-es | en-fr | Avg. | |||
| 2 × 2 | 1.5M | 24.9 | 29.6 | 29.3 | 27.9 | 28.2 | 31.4 | 31.3 | 30.3 | 2.4 |
| 15M | 29.5 | 31.8 | 32.6 | 31.3 | 31.9 | 33.0 | 33.5 | 32.8 | 1.5 | |
| 300M | 30.1 | 32.5 | 32.8 | 31.8 | 32.3 | 33.2 | 33.7 | 33.0 | 1.2 | |
| 6 × 6 | 1.5M | 25.9 | 30.1 | 30.2 | 28.7 | 31.0 | 32.5 | 32.5 | 32.0 | 3.3 |
| 15M | 33.2 | 33.9 | 35.0 | 34.0 | 36.0 | 34.4 | 35.4 | 35.2 | 1.2 | |
| 300M | 34.5 | 34.4 | 35.6 | 34.8 | 36.1 | 34.5 | 35.8 | 35.4 | 0.6 | |
| 24 × 6 | 1.5M | 25.8 | 30.0 | 29.6 | 28.5 | 31.8 | 33.3 | 33.3 | 32.8 | 4.3 |
| 15M | 35.0 | 34.5 | 35.7 | 35.0 | 36.9 | 34.9 | 36.4 | 36.0 | 1.0 | |
| 300M | 36.1 | 34.9 | 36.8 | 35.9 | 37.9 | 35.2 | 37.0 | 36.7 | 0.8 |
| Model | Data | Real | Synthetic | Δ Avg. | ||||||
| de-en | es-en | fr-en | Avg. | de-en | es-en | fr-en | Avg. | |||
| 2 × 2 | 1.5M | 28.4 | 29.1 | 32.2 | 29.9 | 31.9 | 30.8 | 34.1 | 32.3 | 2.4 |
| 15M | 34.6 | 32.8 | 36.2 | 34.5 | 37.1 | 33.6 | 37.5 | 36.1 | 1.6 | |
| 300M | 35.5 | 32.7 | 36.9 | 35.0 | 38.3 | 33.8 | 37.6 | 36.6 | 1.6 | |
| 6 × 6 | 1.5M | 30.8 | 30.1 | 33.2 | 31.4 | 35.3 | 31.8 | 35.3 | 34.1 | 2.7 |
| 15M | 38.1 | 34.0 | 38.3 | 36.8 | 40.2 | 35.0 | 39.1 | 38.1 | 1.3 | |
| 300M | 40.0 | 34.2 | 38.8 | 37.7 | 41.3 | 35.5 | 39.6 | 38.8 | 1.1 | |
| 24 × 6 | 1.5M | 31.4 | 30.5 | 33.8 | 31.9 | 36.0 | 32.8 | 36.2 | 35.0 | 3.1 |
| 15M | 38.8 | 34.6 | 39.6 | 37.6 | 40.8 | 35.6 | 40.1 | 38.8 | 1.2 | |
| 300M | 40.1 | 35.6 | 40.9 | 38.9 | 41.8 | 36.3 | 41.2 | 39.8 | 0.9 |
| German targets source | Size | Must-C v2 | IWSLT tst | ||||||||
| dev | tst | 2010 | 2013 | 2014 | 2015 | 2018 | 2019 | 2020 | Avg. | ||
| Real | 590K | 25.2 | 27.8 | 22.1 | 27.9 | 23.8 | 21.3 | 20.9 | 20.1 | 20.9 | 22.4 |
| Synthetic, WMT21 teacher + fine-tuned on IWSLT + extra ASR data | 590K | 28.3 | 28.9 | 24.5 | 28.0 | 23.9 | 23.0 | 22.6 | 21.7 | 23.3 | 23.9 |
| 590K | 29.2 | 30.2 | 25.0 | 29.5 | 24.8 | 24.9 | 24.2 | 23.4 | 25.3 | 25.3 | |
| 1.25M | 30.6 | 31.0 | 27.2 | 31.3 | 27.4 | 25.8 | 25.1 | 24.3 | 26.4 | 26.8 |
| Model | Data | Real | Synthetic | Δ |
| 2 × 2 | IWSLT | 27.9 | 30.4 | 2.5 |
| Medical | 27.3 | 30.6 | 3.3 | |
| Law | 36.5 | 39.0 | 2.5 | |
| 6 × 6 | IWSLT | 30.0 | 31.6 | 1.6 |
| Medical | 30.1 | 31.7 | 1.6 | |
| Law | 39.8 | 40.8 | 1.0 | |
| 24 × 6 | IWSLT | 31.2 | 33.1 | 1.9 |
| Medical | 31.3 | 33.1 | 1.8 | |
| Law | 40.4 | 42.1 | 1.7 |
| Fine-tuning targets | 2 × 2 | 6 × 6 | 24 × 6 |
| Training from scratch | |||
| Real | 16.9 | 17.0 | 19.2 |
| Synthetic, WMT Teacher | 20.6 | 21.9 | 22.7 |
| + fine-tuned on IWSLT | 21.8 | 22.9 | 25.9 |
| Pre-training on real data | |||
| Real | 30.9 | 33.6 | 34.5 |
| Synthetic, WMT Teacher | 30.9 | 32.6 | 33.7 |
| + fine-tuned on IWSLT | 32.5 | 34.1 | 35.5 |
| Pre-training on synthetic data | |||
| Real | 32.4 | 33.8 | 34.9 |
| Synthetic, WMT Teacher | 32.0 | 33.3 | 33.6 |
| + fine-tuned on IWSLT | 33.4 | 35.1 | 35.6 |
| Language | Targets | Predictive Entropy | ||
| 2 × 2 | 6 × 6 | 24 × 6 | ||
| English to German | Real | 2.4 | 2.4 | 2.5 |
| Synthetic | 1.8 | 1.8 | 1.7 | |
| English to Russian | Real | 2.6 | 2.4 | 2.3 |
| Synthetic | 1.8 | 1.8 | 1.9 |
| Dataset | #Train | #Program | #Valid | #Test |
| GSM8K | 7,473 | 6,363 (85.1%) | - | 1,319 |
| SVAMP | 3,138 | 3,071 (97.9%) | - | 1,000 |
| MathQA† | 16,191 | 7,676 (47.4%) | 2,411 | 1,605 |
| Model | #Param | GSM8K | SVAMP | MathQA | |
| Prompting | LaMDA (Thoppilan et al., 2022) | 137B | 17.1 | - | - |
| PaLM (Chowdhery et al., 2022) | 540B | 58.1 | 79.0 | - | |
| GPT-3 CoT (text-davinci-002) | 175B | 48.1 | - | - | |
| Codex CoT (code-davinci-002) | 175B | 65.6 | 74.8 | 29.9 | |
| Complex CoT (Fu et al., 2022) | 175B | 55.4 | - | 36.0† | |
| PAL (Gao et al., 2022) | 175B | 72.0 | 79.4 | - | |
| PAL (reproduced) | 175B | 71.6 | 77.4 | 30.0 | |
| Our Dynamic Program Prompting | 175B | 76.6 | 80.3 | 61.7 | |
| Fine-tuning | GPT-3 | 175B | 33.1 | - | - |
| CoT Fine-tune (Magister et al., 2022) | 11B | 38.2 | - | - | |
| CoT Fine-tune (CodeGen) | 6B | 35.3 | 40.2 | 25.3 | |
| Our Program Distillation | 6B | 39.0 | 48.0 | 50.6 |
| GSM8K | SVAMP | MathQA | ||
| Most Similar M Exemplars | OpenAI | 76.6 | 80.3 | 61.7 |
| SimCSE (Gao et al., 2021) | 76.4 | 80.1 | 61.0 | |
| ST5 (Ni et al., 2022) | 76.6 | 79.9 | 61.6 | |
| Random | - | 74.4 | 78.1 | 34.0 |
| Least Similar M Exemplars | OpenAI | 73.5 | 78.2 | 34.1 |
| SimCSE (Gao et al., 2021) | 76.0 | 78.4 | 34.7 | |
| ST5 (Ni et al., 2022) | 74.2 | 77.9 | 34.3 |
| Predicted Program\ndef solution():\nstudents_total = 20\ncontemporary_students = students_total * 0.2\njazz_students = (students_total - contemporary學生) * 0.25\nhip_hoplicants = students_total - contemporarylicants - jazz学生们\nhip_hop_percentage = hip_hoplicants / students_total * 100\nresult = hip_hop_percentage\nreturn result |
| Dataset | #Train | #Program | #Valid | #Test |
| GSM8K | 7,473 | 6,363 (85.1%) | - | 1,319 |
| SVAMP | 3,138 | 3,071 (97.9%) | - | 1,000 |
| MathQA† | 16,191 | 7,676 (47.4%) | 2,411 | 1,605 |
| Model | #Param | GSM8K | SVAMP | MathQA | |
| Prompting | LaMDA (Thoppilan et al., 2022) | 137B | 17.1 | - | - |
| PaLM (Chowdhery et al., 2022) | 540B | 58.1 | 79.0 | - | |
| GPT-3 CoT (text-davinci-002) | 175B | 48.1 | - | - | |
| Codex CoT (code-davinci-002) | 175B | 65.6 | 74.8 | 29.9 | |
| Complex CoT (Fu et al., 2022) | 175B | 55.4 | - | 36.0† | |
| PAL (Gao et al., 2022) | 175B | 72.0 | 79.4 | - | |
| PAL (reproduced) | 175B | 71.6 | 77.4 | 30.0 | |
| Our Dynamic Program Prompting | 175B | 76.6 | 80.3 | 61.7 | |
| Fine-tuning | GPT-3 | 175B | 33.1 | - | - |
| CoT Fine-tune (Magister et al., 2022) | 11B | 38.2 | - | - | |
| CoT Fine-tune (CodeGen) | 6B | 35.3 | 40.2 | 25.3 | |
| Our Program Distillation | 6B | 39.0 | 48.0 | 50.6 |
| GSM8K | SVAMP | MathQA | ||
| Most Similar M Exemplars | OpenAI | 76.6 | 80.3 | 61.7 |
| SimCSE (Gao et al., 2021) | 76.4 | 80.1 | 61.0 | |
| ST5 (Ni et al., 2022) | 76.6 | 79.9 | 61.6 | |
| Random | - | 74.4 | 78.1 | 34.0 |
| Least Similar M Exemplars | OpenAI | 73.5 | 78.2 | 34.1 |
| SimCSE (Gao et al., 2021) | 76.0 | 78.4 | 34.7 | |
| ST5 (Ni et al., 2022) | 74.2 | 77.9 | 34.3 |
| Predicted Program\ndef solution():\nstudents_total = 20\ncontemporary_students = students_total * 0.2\njazz_students = (students_total - contemporary學生) * 0.25\nhip_hoplicants = students_total - contemporarylicants - jazz学生们\nhip_hop_percentage = hip_hoplicants / students_total * 100\nresult = hip_hop_percentage\nreturn result |
| Dataset | #Train | #Program | #Valid | #Test |
| GSM8K | 7,473 | 6,363 (85.1%) | - | 1,319 |
| SVAMP | 3,138 | 3,071 (97.9%) | - | 1,000 |
| MathQA† | 16,191 | 7,676 (47.4%) | 2,411 | 1,605 |
| Model | #Param | GSM8K | SVAMP | MathQA | |
| Prompting | LaMDA (Thoppilan et al., 2022) | 137B | 17.1 | - | - |
| PaLM (Chowdhery et al., 2022) | 540B | 58.1 | 79.0 | - | |
| GPT-3 CoT (text-davinci-002) | 175B | 48.1 | - | - | |
| Codex CoT (code-davinci-002) | 175B | 65.6 | 74.8 | 29.9 | |
| Complex CoT (Fu et al., 2022) | 175B | 55.4 | - | 36.0† | |
| PAL (Gao et al., 2022) | 175B | 72.0 | 79.4 | - | |
| PAL (reproduced) | 175B | 71.6 | 77.4 | 30.0 | |
| Our Dynamic Program Prompting | 175B | 76.6 | 80.3 | 61.7 | |
| Fine-tuning | GPT-3 | 175B | 33.1 | - | - |
| CoT Fine-tune (Magister et al., 2022) | 11B | 38.2 | - | - | |
| CoT Fine-tune (CodeGen) | 6B | 35.3 | 40.2 | 25.3 | |
| Our Program Distillation | 6B | 39.0 | 48.0 | 50.6 |
| GSM8K | SVAMP | MathQA | ||
| Most Similar M Exemplars | OpenAI | 76.6 | 80.3 | 61.7 |
| SimCSE (Gao et al., 2021) | 76.4 | 80.1 | 61.0 | |
| ST5 (Ni et al., 2022) | 76.6 | 79.9 | 61.6 | |
| Random | - | 74.4 | 78.1 | 34.0 |
| Least Similar M Exemplars | OpenAI | 73.5 | 78.2 | 34.1 |
| SimCSE (Gao et al., 2021) | 76.0 | 78.4 | 34.7 | |
| ST5 (Ni et al., 2022) | 74.2 | 77.9 | 34.3 |
| Predicted Program +def solution(): +students_total = 20 +contemporary_students = students_total * 0.2 +jazz_students = (students_total - contemporary學生) * 0.25 +hip_hoplicants = students_total - contemporarylicants - jazz学生们 +hip_hop_percentage = hip_hoplicants / students_total * 100 +result = hip_hop_percentage +return result |
| Dataset | #Train | #Program | #Valid | #Test |
| GSM8K | 7,473 | 6,363 (85.1%) | - | 1,319 |
| SVAMP | 3,138 | 3,071 (97.9%) | - | 1,000 |
| MathQA† | 16,191 | 7,676 (47.4%) | 2,411 | 1,605 |
| Model | #Param | GSM8K | SVAMP | MathQA | |
| Prompting | LaMDA (Thoppilan et al., 2022) | 137B | 17.1 | - | - |
| PaLM (Chowdhery et al., 2022) | 540B | 58.1 | 79.0 | - | |
| GPT-3 CoT (text-davinci-002) | 175B | 48.1 | - | - | |
| Codex CoT (code-davinci-002) | 175B | 65.6 | 74.8 | 29.9 | |
| Complex CoT (Fu et al., 2022) | 175B | 55.4 | - | 36.0† | |
| PAL (Gao et al., 2022) | 175B | 72.0 | 79.4 | - | |
| PAL (reproduced) | 175B | 71.6 | 77.4 | 30.0 | |
| Our Dynamic Program Prompting | 175B | 76.6 | 80.3 | 61.7 | |
| Fine-tuning | GPT-3 | 175B | 33.1 | - | - |
| CoT Fine-tune (Magister et al., 2022) | 11B | 38.2 | - | - | |
| CoT Fine-tune (CodeGen) | 6B | 35.3 | 40.2 | 25.3 | |
| Our Program Distillation | 6B | 39.0 | 48.0 | 50.6 |
| GSM8K | SVAMP | MathQA | ||
| Most Similar M Exemplars | OpenAI | 76.6 | 80.3 | 61.7 |
| SimCSE (Gao et al., 2021) | 76.4 | 80.1 | 61.0 | |
| ST5 (Ni et al., 2022) | 76.6 | 79.9 | 61.6 | |
| Random | - | 74.4 | 78.1 | 34.0 |
| Least Similar M Exemplars | OpenAI | 73.5 | 78.2 | 34.1 |
| SimCSE (Gao et al., 2021) | 76.0 | 78.4 | 34.7 | |
| ST5 (Ni et al., 2022) | 74.2 | 77.9 | 34.3 |
| Predicted Program\ndef solution():\nstudents_total = 20\ncontemporary_students = students_total * 0.2\njazz_students = (students_total - contemporary學生) * 0.25\nhip_hoplicants = students_total - contemporarylicants - jazz学生们\nhip_hop_percentage = hip_hoplicants / students_total * 100\nresult = hip_hop_percentage\nreturn result |
| Source\n#1: 孙燕姿 非洲 送 关爱 遇 沙暴 挨 石头 被 划伤\n#13: ... 孙燕姿 经历 当地 难得一见 的 沙尘暴 ...\n#17: 之后 听 当地人 解释 沙尘暴 是好运 的 象征 ...\n#20: ... 请 孙燕姿 坐 上 骆驼 接受 当地人 欢呼 |
| Sentence-level NMT\n#1: sun yen-tzu in africa ... and gets injured by stones ...\n#13: ... sun yanzi experienced rare sandstorms in the area ...\n#17: after hearing the locals explain that dust storms are a symbol of good luck ...\n#20: ... invite sun yanzi to sit on the camel and receive cheers from the local people ... |
| Document-level NMT\n#1: sun yen-tzu sent ... she was hit by sand and rocks ...\n#13: ... sun yanzi experienced the rare dust storms in the area ...\n#17: after hearing the locals explain n that sand storms were symbols of good fortune ...\n#20: ... asking sun yanzi to sit on the camel and receive a cheer from the local people ... |
| Reference\n#1: sun yanzi delivers love to africa ...\n#13: ... sun yanzi experienced a sandstorm that was rarely seen in the area\n#17: she later heard the locals explaining that the sandstorm is a sign of good luck ...\n#20: ... and invited sun yanzi to sit on the camel to receive the cheers of the locals |
| Model | NIST (ZH→EN) | NIST (EN→ZH) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 48.45 | 50.70 | 65.25 | 78.61 | 25.82 | 27.24 | 64.59 | 67.87 |
| SentRepair (Trans.) | 48.49 | 50.76 | 64.89 | 78.03 | 25.71 | 27.12 | 64.39 | 67.69 |
| DocRepair (Trans.) | - | 51.12 | - | - | - | 27.01 | - | - |
| DocRepair (G-Trans.) | 49.25 | 51.54 | 65.39 | 78.11 | 26.31 | 27.76 | 64.66 | 67.92 |
| DocRepair (Ours) | 50.28 | 52.28 | 69.51 | 80.74 | 26.66 | 28.11 | 67.11 | 70.37 |
| Model | PDC (ZH→EN) | Europarl (DE→EN) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 27.49 | 30.23 | 74.48 | 71.84 | 38.44 | 40.94 | 68.81 | 81.51 |
| SentRepair (Trans.) | 27.31 | 30.08 | 73.64 | 71.44 | 38.66 | 41.20 | 69.27 | 79.33 |
| DocRepair (Trans.) | - | 30.57 | - | - | - | 41.23 | - | |
| DocRepair (G-Trans.) | 27.94 | 30.82 | 72.68 | 70.45 | 38.79 | 41.30 | 69.57 | 81.71 |
| DocRepair (Ours) | 28.19 | 31.05 | 77.51 | 73.66 | 39.07 | 41.56 | 74.02 | 82.15 |
| Model | s-BLEU | d-BLEU | LTCR | ref-L. |
| NIST ZH→EN | ||||
| Doc-level NMT | 48.77 | 51.11 | 65.89 | 78.06 |
| DocRep. (Ours) | 48.86 | 51.00 | 69.75 | 80.52 |
| NIST EN→ZH | ||||
| Doc-level NMT | 26.19 | 27.61 | 64.39 | 72.45 |
| DocRep. (Ours) | 26.50 | 27.94 | 67.74 | 73.81 |
| PDC ZH→EN | ||||
| Doc-level NMT | 28.48 | 31.33 | 74.73 | 72.53 |
| DocRep. (Ours) | 28.68 | 31.54 | 79.92 | 74.30 |
| Europarl DE→EN | ||||
| Doc-level NMT | 39.64 | 42.16 | 74.47 | 82.80 |
| DocRep. (Ours) | 39.82 | 42.36 | 76.92 | 82.71 |
| Ablation | s-BLEU | Δ | LTCR | Δ | ref-L. | Δ |
| Lattice-Input | 50.28 | - | 69.51 | - | 80.74 | - |
| w/o lat. pos. | 49.04 | -1.24 | 67.94 | -1.57 | 79.13 | -1.61 |
| w/o tri. word | 49.70 | -0.58 | 68.68 | -0.83 | 79.98 | -0.76 |
| w/o weights | 49.84 | -0.45 | 69.41 | -0.10 | 80.40 | -0.34 |
| Number | Count | % | Number | Count | % |
| 2 | 567541 | 78.51 | 3 | 117061 | 16.19 |
| 4 | 28693 | 4.01 | 5 | 6945 | 0.96 |
| >6 | 2355 | 0.33 | All | 722865 | 100 |
| Model | s-BLEU | LTCR | ref-L. |
| Sent-NMT | 48.45 | 65.25 | 78.61 |
| DocRepair (w/o pre-training) | 47.94 | 69.19 | 79.74 |
| DocRepair (w/ 0.83M) | 49.06 | 69.24 | 79.93 |
| DocRepair (w/ 2.0M) | 50.28 | 69.51 | 80.74 |
| Annotator | Equal | Better | Worse |
| 1 | 44% | 36% | 20% |
| 2 | 49% | 33% | 18% |
| Average | 46% | 35% | 19% |
| Set | NIST | PDC | Europarl | |||
| #Doc | #Sent | #Doc | #Sent | #Doc | #Sent | |
| Pre-Training | - | 2M | - | 3.39M | - | 1.67M |
| Fine-Tuning | 66,396 | 0.83M | 59,384 | 1.39M | 117,855 | 1.67M |
| Dev | 100 | 1664 | 100 | 2320 | 240 | 3587 |
| Test | 580 | 5833 | 148 | 4858 | 360 | 5134 |
| Source | <#1> ... 瑞典大使 毅损 艺术品 引发 ... <#2> ... 大使 马兹尔 破坏而 引发的争议 ... <#3> 马兹尔 大使 ... 自杀 炸弹 客照片 的艺术品 时 <#8> ... 今天 出面 挺 马兹尔 ... <#9> ... 微笑 的照片 插 在 船头, 作为 船帆 |
| Sentence-Level NMT | <#1> ... triggered by destruction of art by ... <#2> ... was undermined by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top of the ship ... BLEU: 43.12 |
| DocRepair (G-Trans) | <#1> ... by israeli ambassador to sweden's destruction of art <#2> ... was destroyed by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top ... BLEU: 44.07 |
| Our Approach | <#1> ... triggered by destruction of artwork by ... <#2> ... was damaged by israeli ambassador to sweden marzir ... <#3> ... ambassador marzir saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support marzir ... <#9> ... with a smiling picture of ... hanging on the top ... BLEU: 44.30 |
| Reference | <#1> israel 's ambassador to sweden vandalizes artwork <#2> ... museum of national history by mazel , israeli ambassador to sweden ... <#3> ambassador mazel visited ... artwork featuring a photo of the suicide bomber ... <#8> ... expressed his support for mazel today ... <#9> ... with a photo of a smiling hanadi jaradat placed on the ... |
| Model | s-BLEU | #Params (M) |
| Lattice-Input | 50.28 | 74.77 |
| w/o lat. pos. | 49.04 | 74.77 |
| w/o tri. word | 49.70 | 70.34 |
| w/o weights | 49.84 | 74.77 |
| Source\n#1: 孙燕姿 非洲 送 关爱 遇 沙暴 挨 石头 被 划伤\n#13: ... 孙燕姿 经历 当地 难得一见 的 沙尘暴 ...\n#17: 之后 听 当地人 解释 沙尘暴 是好运 的 象征 ...\n#20: ... 请 孙燕姿 坐 上 骆驼 接受 当地人 欢呼 |
| Sentence-level NMT\n#1: sun yen-tzu in africa ... and gets injured by stones ...\n#13: ... sun yanzi experienced rare sandstorms in the area ...\n#17: after hearing the locals explain that dust storms are a symbol of good luck ...\n#20: ... invite sun yanzi to sit on the camel and receive cheers from the local people ... |
| Document-level NMT\n#1: sun yen-tzu sent ... she was hit by sand and rocks ...\n#13: ... sun yanzi experienced the rare dust storms in the area ...\n#17: after hearing the locals explain n that sand storms were symbols of good fortune ...\n#20: ... asking sun yanzi to sit on the camel and receive a cheer from the local people ... |
| Reference\n#1: sun yanzi delivers love to africa ...\n#13: ... sun yanzi experienced a sandstorm that was rarely seen in the area\n#17: she later heard the locals explaining that the sandstorm is a sign of good luck ...\n#20: ... and invited sun yanzi to sit on the camel to receive the cheers of the locals |
| Model | NIST (ZH→EN) | NIST (EN→ZH) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 48.45 | 50.70 | 65.25 | 78.61 | 25.82 | 27.24 | 64.59 | 67.87 |
| SentRepair (Trans.) | 48.49 | 50.76 | 64.89 | 78.03 | 25.71 | 27.12 | 64.39 | 67.69 |
| DocRepair (Trans.) | - | 51.12 | - | - | - | 27.01 | - | - |
| DocRepair (G-Trans.) | 49.25 | 51.54 | 65.39 | 78.11 | 26.31 | 27.76 | 64.66 | 67.92 |
| DocRepair (Ours) | 50.28 | 52.28 | 69.51 | 80.74 | 26.66 | 28.11 | 67.11 | 70.37 |
| Model | PDC (ZH→EN) | Europarl (DE→EN) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 27.49 | 30.23 | 74.48 | 71.84 | 38.44 | 40.94 | 68.81 | 81.51 |
| SentRepair (Trans.) | 27.31 | 30.08 | 73.64 | 71.44 | 38.66 | 41.20 | 69.27 | 79.33 |
| DocRepair (Trans.) | - | 30.57 | - | - | - | 41.23 | - | |
| DocRepair (G-Trans.) | 27.94 | 30.82 | 72.68 | 70.45 | 38.79 | 41.30 | 69.57 | 81.71 |
| DocRepair (Ours) | 28.19 | 31.05 | 77.51 | 73.66 | 39.07 | 41.56 | 74.02 | 82.15 |
| Model | s-BLEU | d-BLEU | LTCR | ref-L. |
| NIST ZH→EN | ||||
| Doc-level NMT | 48.77 | 51.11 | 65.89 | 78.06 |
| DocRep. (Ours) | 48.86 | 51.00 | 69.75 | 80.52 |
| NIST EN→ZH | ||||
| Doc-level NMT | 26.19 | 27.61 | 64.39 | 72.45 |
| DocRep. (Ours) | 26.50 | 27.94 | 67.74 | 73.81 |
| PDC ZH→EN | ||||
| Doc-level NMT | 28.48 | 31.33 | 74.73 | 72.53 |
| DocRep. (Ours) | 28.68 | 31.54 | 79.92 | 74.30 |
| Europarl DE→EN | ||||
| Doc-level NMT | 39.64 | 42.16 | 74.47 | 82.80 |
| DocRep. (Ours) | 39.82 | 42.36 | 76.92 | 82.71 |
| Ablation | s-BLEU | Δ | LTCR | Δ | ref-L. | Δ |
| Lattice-Input | 50.28 | - | 69.51 | - | 80.74 | - |
| w/o lat. pos. | 49.04 | -1.24 | 67.94 | -1.57 | 79.13 | -1.61 |
| w/o tri. word | 49.70 | -0.58 | 68.68 | -0.83 | 79.98 | -0.76 |
| w/o weights | 49.84 | -0.45 | 69.41 | -0.10 | 80.40 | -0.34 |
| Number | Count | % | Number | Count | % |
| 2 | 567541 | 78.51 | 3 | 117061 | 16.19 |
| 4 | 28693 | 4.01 | 5 | 6945 | 0.96 |
| >6 | 2355 | 0.33 | All | 722865 | 100 |
| Model | s-BLEU | LTCR | ref-L. |
| Sent-NMT | 48.45 | 65.25 | 78.61 |
| DocRepair (w/o pre-training) | 47.94 | 69.19 | 79.74 |
| DocRepair (w/ 0.83M) | 49.06 | 69.24 | 79.93 |
| DocRepair (w/ 2.0M) | 50.28 | 69.51 | 80.74 |
| Annotator | Equal | Better | Worse |
| 1 | 44% | 36% | 20% |
| 2 | 49% | 33% | 18% |
| Average | 46% | 35% | 19% |
| Set | NIST | PDC | Europarl | |||
| #Doc | #Sent | #Doc | #Sent | #Doc | #Sent | |
| Pre-Training | - | 2M | - | 3.39M | - | 1.67M |
| Fine-Tuning | 66,396 | 0.83M | 59,384 | 1.39M | 117,855 | 1.67M |
| Dev | 100 | 1664 | 100 | 2320 | 240 | 3587 |
| Test | 580 | 5833 | 148 | 4858 | 360 | 5134 |
| Source | <#1> ... 瑞典大使 毅损 艺术品 引发 ... <#2> ... 大使 马兹尔 破坏而 引发的争议 ... <#3> 马兹尔 大使 ... 自杀 炸弹 客照片 的艺术品 时 <#8> ... 今天 出面 挺 马兹尔 ... <#9> ... 微笑 的照片 插 在 船头, 作为 船帆 |
| Sentence-Level NMT | <#1> ... triggered by destruction of art by ... <#2> ... was undermined by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top of the ship ... BLEU: 43.12 |
| DocRepair (G-Trans) | <#1> ... by israeli ambassador to sweden's destruction of art <#2> ... was destroyed by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top ... BLEU: 44.07 |
| Our Approach | <#1> ... triggered by destruction of artwork by ... <#2> ... was damaged by israeli ambassador to sweden marzir ... <#3> ... ambassador marzir saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support marzir ... <#9> ... with a smiling picture of ... hanging on the top ... BLEU: 44.30 |
| Reference | <#1> israel 's ambassador to sweden vandalizes artwork <#2> ... museum of national history by mazel , israeli ambassador to sweden ... <#3> ambassador mazel visited ... artwork featuring a photo of the suicide bomber ... <#8> ... expressed his support for mazel today ... <#9> ... with a photo of a smiling hanadi jaradat placed on the ... |
| Model | s-BLEU | #Params (M) |
| Lattice-Input | 50.28 | 74.77 |
| w/o lat. pos. | 49.04 | 74.77 |
| w/o tri. word | 49.70 | 70.34 |
| w/o weights | 49.84 | 74.77 |
| Source +#1: 孙燕姿 非洲 送 关爱 遇 沙暴 挨 石头 被 划伤 +#13: ... 孙燕姿 经历 当地 难得一见 的 沙尘暴 ... +#17: 之后 听 当地人 解释 沙尘暴 是好运 的 象征 ... +#20: ... 请 孙燕姿 坐 上 骆驼 接受 当地人 欢呼 |
| Sentence-level NMT +#1: sun yen-tzu in africa ... and gets injured by stones ... +#13: ... sun yanzi experienced rare sandstorms in the area ... +#17: after hearing the locals explain that dust storms are a symbol of good luck ... +#20: ... invite sun yanzi to sit on the camel and receive cheers from the local people ... |
| Document-level NMT +#1: sun yen-tzu sent ... she was hit by sand and rocks ... +#13: ... sun yanzi experienced the rare dust storms in the area ... +#17: after hearing the locals explain n that sand storms were symbols of good fortune ... +#20: ... asking sun yanzi to sit on the camel and receive a cheer from the local people ... |
| Reference +#1: sun yanzi delivers love to africa ... +#13: ... sun yanzi experienced a sandstorm that was rarely seen in the area +#17: she later heard the locals explaining that the sandstorm is a sign of good luck ... +#20: ... and invited sun yanzi to sit on the camel to receive the cheers of the locals |
| Model | NIST (ZH→EN) | NIST (EN→ZH) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 48.45 | 50.70 | 65.25 | 78.61 | 25.82 | 27.24 | 64.59 | 67.87 |
| SentRepair (Trans.) | 48.49 | 50.76 | 64.89 | 78.03 | 25.71 | 27.12 | 64.39 | 67.69 |
| DocRepair (Trans.) | - | 51.12 | - | - | - | 27.01 | - | - |
| DocRepair (G-Trans.) | 49.25 | 51.54 | 65.39 | 78.11 | 26.31 | 27.76 | 64.66 | 67.92 |
| DocRepair (Ours) | 50.28 | 52.28 | 69.51 | 80.74 | 26.66 | 28.11 | 67.11 | 70.37 |
| Model | PDC (ZH→EN) | Europarl (DE→EN) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 27.49 | 30.23 | 74.48 | 71.84 | 38.44 | 40.94 | 68.81 | 81.51 |
| SentRepair (Trans.) | 27.31 | 30.08 | 73.64 | 71.44 | 38.66 | 41.20 | 69.27 | 79.33 |
| DocRepair (Trans.) | - | 30.57 | - | - | - | 41.23 | - | |
| DocRepair (G-Trans.) | 27.94 | 30.82 | 72.68 | 70.45 | 38.79 | 41.30 | 69.57 | 81.71 |
| DocRepair (Ours) | 28.19 | 31.05 | 77.51 | 73.66 | 39.07 | 41.56 | 74.02 | 82.15 |
| Model | s-BLEU | d-BLEU | LTCR | ref-L. |
| NIST ZH→EN | ||||
| Doc-level NMT | 48.77 | 51.11 | 65.89 | 78.06 |
| DocRep. (Ours) | 48.86 | 51.00 | 69.75 | 80.52 |
| NIST EN→ZH | ||||
| Doc-level NMT | 26.19 | 27.61 | 64.39 | 72.45 |
| DocRep. (Ours) | 26.50 | 27.94 | 67.74 | 73.81 |
| PDC ZH→EN | ||||
| Doc-level NMT | 28.48 | 31.33 | 74.73 | 72.53 |
| DocRep. (Ours) | 28.68 | 31.54 | 79.92 | 74.30 |
| Europarl DE→EN | ||||
| Doc-level NMT | 39.64 | 42.16 | 74.47 | 82.80 |
| DocRep. (Ours) | 39.82 | 42.36 | 76.92 | 82.71 |
| Ablation | s-BLEU | Δ | LTCR | Δ | ref-L. | Δ |
| Lattice-Input | 50.28 | - | 69.51 | - | 80.74 | - |
| w/o lat. pos. | 49.04 | -1.24 | 67.94 | -1.57 | 79.13 | -1.61 |
| w/o tri. word | 49.70 | -0.58 | 68.68 | -0.83 | 79.98 | -0.76 |
| w/o weights | 49.84 | -0.45 | 69.41 | -0.10 | 80.40 | -0.34 |
| Number | Count | % | Number | Count | % |
| 2 | 567541 | 78.51 | 3 | 117061 | 16.19 |
| 4 | 28693 | 4.01 | 5 | 6945 | 0.96 |
| >6 | 2355 | 0.33 | All | 722865 | 100 |
| Model | s-BLEU | LTCR | ref-L. |
| Sent-NMT | 48.45 | 65.25 | 78.61 |
| DocRepair (w/o pre-training) | 47.94 | 69.19 | 79.74 |
| DocRepair (w/ 0.83M) | 49.06 | 69.24 | 79.93 |
| DocRepair (w/ 2.0M) | 50.28 | 69.51 | 80.74 |
| Annotator | Equal | Better | Worse |
| 1 | 44% | 36% | 20% |
| 2 | 49% | 33% | 18% |
| Average | 46% | 35% | 19% |
| Set | NIST | PDC | Europarl | |||
| #Doc | #Sent | #Doc | #Sent | #Doc | #Sent | |
| Pre-Training | - | 2M | - | 3.39M | - | 1.67M |
| Fine-Tuning | 66,396 | 0.83M | 59,384 | 1.39M | 117,855 | 1.67M |
| Dev | 100 | 1664 | 100 | 2320 | 240 | 3587 |
| Test | 580 | 5833 | 148 | 4858 | 360 | 5134 |
| Source | <#1> ... 瑞典大使 毅损 艺术品 引发 ... <#2> ... 大使 马兹尔 破坏而 引发的争议 ... <#3> 马兹尔 大使 ... 自杀 炸弹 客照片 的艺术品 时 <#8> ... 今天 出面 挺 马兹尔 ... <#9> ... 微笑 的照片 插 在 船头, 作为 船帆 |
| Sentence-Level NMT | <#1> ... triggered by destruction of art by ... <#2> ... was undermined by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top of the ship ... BLEU: 43.12 |
| DocRepair (G-Trans) | <#1> ... by israeli ambassador to sweden's destruction of art <#2> ... was destroyed by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top ... BLEU: 44.07 |
| Our Approach | <#1> ... triggered by destruction of artwork by ... <#2> ... was damaged by israeli ambassador to sweden marzir ... <#3> ... ambassador marzir saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support marzir ... <#9> ... with a smiling picture of ... hanging on the top ... BLEU: 44.30 |
| Reference | <#1> israel 's ambassador to sweden vandalizes artwork <#2> ... museum of national history by mazel , israeli ambassador to sweden ... <#3> ambassador mazel visited ... artwork featuring a photo of the suicide bomber ... <#8> ... expressed his support for mazel today ... <#9> ... with a photo of a smiling hanadi jaradat placed on the ... |
| Model | s-BLEU | #Params (M) |
| Lattice-Input | 50.28 | 74.77 |
| w/o lat. pos. | 49.04 | 74.77 |
| w/o tri. word | 49.70 | 70.34 |
| w/o weights | 49.84 | 74.77 |
| Source\n#1: 孙燕姿 非洲 送 关爱 遇 沙暴 挨 石头 被 划伤\n#13: ... 孙燕姿 经历 当地 难得一见 的 沙尘暴 ...\n#17: 之后 听 当地人 解释 沙尘暴 是好运 的 象征 ...\n#20: ... 请 孙燕姿 坐 上 骆驼 接受 当地人 欢呼 |
| Sentence-level NMT\n#1: sun yen-tzu in africa ... and gets injured by stones ...\n#13: ... sun yanzi experienced rare sandstorms in the area ...\n#17: after hearing the locals explain that dust storms are a symbol of good luck ...\n#20: ... invite sun yanzi to sit on the camel and receive cheers from the local people ... |
| Document-level NMT\n#1: sun yen-tzu sent ... she was hit by sand and rocks ...\n#13: ... sun yanzi experienced the rare dust storms in the area ...\n#17: after hearing the locals explain n that sand storms were symbols of good fortune ...\n#20: ... asking sun yanzi to sit on the camel and receive a cheer from the local people ... |
| Reference\n#1: sun yanzi delivers love to africa ...\n#13: ... sun yanzi experienced a sandstorm that was rarely seen in the area\n#17: she later heard the locals explaining that the sandstorm is a sign of good luck ...\n#20: ... and invited sun yanzi to sit on the camel to receive the cheers of the locals |
| Model | NIST (ZH→EN) | NIST (EN→ZH) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 48.45 | 50.70 | 65.25 | 78.61 | 25.82 | 27.24 | 64.59 | 67.87 |
| SentRepair (Trans.) | 48.49 | 50.76 | 64.89 | 78.03 | 25.71 | 27.12 | 64.39 | 67.69 |
| DocRepair (Trans.) | - | 51.12 | - | - | - | 27.01 | - | - |
| DocRepair (G-Trans.) | 49.25 | 51.54 | 65.39 | 78.11 | 26.31 | 27.76 | 64.66 | 67.92 |
| DocRepair (Ours) | 50.28 | 52.28 | 69.51 | 80.74 | 26.66 | 28.11 | 67.11 | 70.37 |
| Model | PDC (ZH→EN) | Europarl (DE→EN) | ||||||
| s-BLEU | d-BLEU | LTCR | ref-LTCR | s-BLEU | d-BLEU | LTCR | ref-LTCR | |
| Sent-level NMT | 27.49 | 30.23 | 74.48 | 71.84 | 38.44 | 40.94 | 68.81 | 81.51 |
| SentRepair (Trans.) | 27.31 | 30.08 | 73.64 | 71.44 | 38.66 | 41.20 | 69.27 | 79.33 |
| DocRepair (Trans.) | - | 30.57 | - | - | - | 41.23 | - | |
| DocRepair (G-Trans.) | 27.94 | 30.82 | 72.68 | 70.45 | 38.79 | 41.30 | 69.57 | 81.71 |
| DocRepair (Ours) | 28.19 | 31.05 | 77.51 | 73.66 | 39.07 | 41.56 | 74.02 | 82.15 |
| Model | s-BLEU | d-BLEU | LTCR | ref-L. |
| NIST ZH→EN | ||||
| Doc-level NMT | 48.77 | 51.11 | 65.89 | 78.06 |
| DocRep. (Ours) | 48.86 | 51.00 | 69.75 | 80.52 |
| NIST EN→ZH | ||||
| Doc-level NMT | 26.19 | 27.61 | 64.39 | 72.45 |
| DocRep. (Ours) | 26.50 | 27.94 | 67.74 | 73.81 |
| PDC ZH→EN | ||||
| Doc-level NMT | 28.48 | 31.33 | 74.73 | 72.53 |
| DocRep. (Ours) | 28.68 | 31.54 | 79.92 | 74.30 |
| Europarl DE→EN | ||||
| Doc-level NMT | 39.64 | 42.16 | 74.47 | 82.80 |
| DocRep. (Ours) | 39.82 | 42.36 | 76.92 | 82.71 |
| Ablation | s-BLEU | Δ | LTCR | Δ | ref-L. | Δ |
| Lattice-Input | 50.28 | - | 69.51 | - | 80.74 | - |
| w/o lat. pos. | 49.04 | -1.24 | 67.94 | -1.57 | 79.13 | -1.61 |
| w/o tri. word | 49.70 | -0.58 | 68.68 | -0.83 | 79.98 | -0.76 |
| w/o weights | 49.84 | -0.45 | 69.41 | -0.10 | 80.40 | -0.34 |
| Number | Count | % | Number | Count | % |
| 2 | 567541 | 78.51 | 3 | 117061 | 16.19 |
| 4 | 28693 | 4.01 | 5 | 6945 | 0.96 |
| >6 | 2355 | 0.33 | All | 722865 | 100 |
| Model | s-BLEU | LTCR | ref-L. |
| Sent-NMT | 48.45 | 65.25 | 78.61 |
| DocRepair (w/o pre-training) | 47.94 | 69.19 | 79.74 |
| DocRepair (w/ 0.83M) | 49.06 | 69.24 | 79.93 |
| DocRepair (w/ 2.0M) | 50.28 | 69.51 | 80.74 |
| Annotator | Equal | Better | Worse |
| 1 | 44% | 36% | 20% |
| 2 | 49% | 33% | 18% |
| Average | 46% | 35% | 19% |
| Set | NIST | PDC | Europarl | |||
| #Doc | #Sent | #Doc | #Sent | #Doc | #Sent | |
| Pre-Training | - | 2M | - | 3.39M | - | 1.67M |
| Fine-Tuning | 66,396 | 0.83M | 59,384 | 1.39M | 117,855 | 1.67M |
| Dev | 100 | 1664 | 100 | 2320 | 240 | 3587 |
| Test | 580 | 5833 | 148 | 4858 | 360 | 5134 |
| Source | <#1> ... 瑞典大使 毅损 艺术品 引发 ... <#2> ... 大使 马兹尔 破坏而 引发的争议 ... <#3> 马兹尔 大使 ... 自杀 炸弹 客照片 的艺术品 时 <#8> ... 今天 出面 挺 马兹尔 ... <#9> ... 微笑 的照片 插 在 船头, 作为 船帆 |
| Sentence-Level NMT | <#1> ... triggered by destruction of art by ... <#2> ... was undermined by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top of the ship ... BLEU: 43.12 |
| DocRepair (G-Trans) | <#1> ... by israeli ambassador to sweden's destruction of art <#2> ... was destroyed by israeli ambassador to sweden marzir ... <#3> ... ambassador marshall saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support mahathir ... <#9> ... with a smiling photo of ... hanging on the top ... BLEU: 44.07 |
| Our Approach | <#1> ... triggered by destruction of artwork by ... <#2> ... was damaged by israeli ambassador to sweden marzir ... <#3> ... ambassador marzir saw an artwork showing pictures of a suicide bomber ... <#8> ... came out today to support marzir ... <#9> ... with a smiling picture of ... hanging on the top ... BLEU: 44.30 |
| Reference | <#1> israel 's ambassador to sweden vandalizes artwork <#2> ... museum of national history by mazel , israeli ambassador to sweden ... <#3> ambassador mazel visited ... artwork featuring a photo of the suicide bomber ... <#8> ... expressed his support for mazel today ... <#9> ... with a photo of a smiling hanadi jaradat placed on the ... |
| Model | s-BLEU | #Params (M) |
| Lattice-Input | 50.28 | 74.77 |
| w/o lat. pos. | 49.04 | 74.77 |
| w/o tri. word | 49.70 | 70.34 |
| w/o weights | 49.84 | 74.77 |
| Model | Rank | Params. | BLEU |
| Transformer | - | 36.8M | 34.6 |
| LMF Transformer | 256 | 21.9M | 34.3 |
| 112 | 10.4M | 34.1 | |
| 64 | 6.6M | 33.7 | |
| 32 | 4.0M | 31.7 | |
| 16 | 2.7M | 28.8 | |
| 4 | 1.8M | 17.4 |
| Model | IWSLT'14 De-En | WMT'14 En-De | ||||||
| Params. | Ratio | Speed | BLEU | Params. | Ratio | Speed | BLEU | |
| Transformer | 36.8M | 1.0× | 1.0× | 34.5 | 63.2M | 1.0× | 1.0× | 27.3 |
| Lite Transformer | 13.9M | 2.6× | 1.5× | 33.6 | 33.6M | 1.9× | 1.1× | 26.5 |
| HAT Transformer | 28.2M | 1.3× | 1.7× | 34.5 | 46.2M | 1.4× | 1.7× | 26.9 |
| DelighT | 19.9M | 1.8× | 0.8× | 34.4 | 23.3M | 2.7× | 1.2× | 26.7 |
| LightFormer | 7.7M | 4.8× | 2.1× | 34.6 | 22.5M | 2.8× | 1.5× | 27.1 |
| w/o SVD WT | 7.7M | 4.8× | 2.1× | 34.0 | 22.5M | 2.8× | 1.5× | 26.5 |
| Model | Params. | Ratio | Speed | BLEU |
| Transformer | 62M | 1.0× | 1.0× | 34.4 |
| Delight | 22.0M | 2.8× | 1.2× | 34.3 |
| Subformer | 20.0M | 3.1× | - | 34.1 |
| LightFormer | 15.3M | 4.1× | 1.8× | 34.3 |
| w/o SVD WT | 15.3M | 4.1× | 1.8× | 33.9 |
| Model | Params. | R-1 | R-2 | R-L |
| Transformer | 51.28M | 37.5 | 18.9 | 34.7 |
| LightFormer | 13.04M | 37.5 | 19.2 | 35.0 |
| w/o SVD WT | 13.04M | 37.1 | 18.7 | 34.7 |
| Model | Params. | CR | MR | SUBJ | MPQA | Average |
| Transformer | 12.8M | 86.2 | 80.1 | 95.4 | 90.0 | 87.93 |
| LightFormer | 0.5M | 87.3 | 80.6 | 95.4 | 90.1 | 88.35 |
| w/o SVD WT | 0.5M | 86.5 | 80.3 | 95.1 | 89.7 | 87.90 |
| Model | E-D | Params. | BLEU | |
| r1 | Transformer | 6-6 | 36.8M | 34.5 |
| r2 | 18-3 | 52.5M | 34.8 | |
| r3 | + Weight Sharing | 18-3 | 33.6M | 34.6 |
| r4 | + LMF | 18-3 | 14.6M | 33.9 |
| r5 | + LMF & WS | 18-3 | 7.7M | 34.0 |
| r6 | LightFormer | 18-3 | 7.7M | 34.6 |
| Model | E-D | Params. | Speed | BLEU |
| Transformer | 6-6 | 36.8M | 1.0× | 34.5 |
| LightFormer | 6-6 | 7.7M | 1.3× | 34.5 |
| 18-3 | 7.7M | 2.1× | 34.6 | |
| 18-2 | 6.7M | 2.4× | 34.2 | |
| 18-1 | 5.8M | 3.0× | 33.6 | |
| 12-3 | 7.7M | 2.0× | 34.4 | |
| 12-2 | 6.7M | 2.5× | 34.1 | |
| 24-3 | 7.7M | 1.8× | 34.6 |
| Model | Groups | Params. | BLEU |
| LightFormer 6-6 | 3 | 6.9M | 34.4 |
| 2 | 6.4M | 34.2 | |
| 1 | 5.8M | 34.2 | |
| LightFormer 18-3 | 9 | 9.4M | 34.7 |
| 6 | 7.7M | 34.6 | |
| 3 | 5.9M | 34.1 | |
| 1 | 4.8M | 34.0 |
| Dataset | Model | Params | S(Pi) | BLEU |
| De-En | Transformer | 36.8M | 1.0× | 34.5 / 33.4 |
| LightFormer | 7.7M | 2.5× | 34.6 / 33.6 | |
| En-De | Transformer | 63.2M | 1.0× | 27.3 / 26.3 |
| LightFormer | 22.5M | 1.4× | 27.1 / 26.2 | |
| En-Ro | Transformer | 62.0M | 1.0× | 34.4 / 33.5 |
| LightFormer | 15.3M | 1.9× | 34.3 / 33.6 |
| Model | Rank | Params. | BLEU |
| Transformer | - | 36.8M | 34.6 |
| LMF Transformer | 256 | 21.9M | 34.3 |
| 112 | 10.4M | 34.1 | |
| 64 | 6.6M | 33.7 | |
| 32 | 4.0M | 31.7 | |
| 16 | 2.7M | 28.8 | |
| 4 | 1.8M | 17.4 |
| Model | IWSLT'14 De-En | WMT'14 En-De | ||||||
| Params. | Ratio | Speed | BLEU | Params. | Ratio | Speed | BLEU | |
| Transformer | 36.8M | 1.0× | 1.0× | 34.5 | 63.2M | 1.0× | 1.0× | 27.3 |
| Lite Transformer | 13.9M | 2.6× | 1.5× | 33.6 | 33.6M | 1.9× | 1.1× | 26.5 |
| HAT Transformer | 28.2M | 1.3× | 1.7× | 34.5 | 46.2M | 1.4× | 1.7× | 26.9 |
| DelighT | 19.9M | 1.8× | 0.8× | 34.4 | 23.3M | 2.7× | 1.2× | 26.7 |
| LightFormer | 7.7M | 4.8× | 2.1× | 34.6 | 22.5M | 2.8× | 1.5× | 27.1 |
| w/o SVD WT | 7.7M | 4.8× | 2.1× | 34.0 | 22.5M | 2.8× | 1.5× | 26.5 |
| Model | Params. | Ratio | Speed | BLEU |
| Transformer | 62M | 1.0× | 1.0× | 34.4 |
| Delight | 22.0M | 2.8× | 1.2× | 34.3 |
| Subformer | 20.0M | 3.1× | - | 34.1 |
| LightFormer | 15.3M | 4.1× | 1.8× | 34.3 |
| w/o SVD WT | 15.3M | 4.1× | 1.8× | 33.9 |
| Model | Params. | R-1 | R-2 | R-L |
| Transformer | 51.28M | 37.5 | 18.9 | 34.7 |
| LightFormer | 13.04M | 37.5 | 19.2 | 35.0 |
| w/o SVD WT | 13.04M | 37.1 | 18.7 | 34.7 |
| Model | Params. | CR | MR | SUBJ | MPQA | Average |
| Transformer | 12.8M | 86.2 | 80.1 | 95.4 | 90.0 | 87.93 |
| LightFormer | 0.5M | 87.3 | 80.6 | 95.4 | 90.1 | 88.35 |
| w/o SVD WT | 0.5M | 86.5 | 80.3 | 95.1 | 89.7 | 87.90 |
| Model | E-D | Params. | BLEU | |
| r1 | Transformer | 6-6 | 36.8M | 34.5 |
| r2 | 18-3 | 52.5M | 34.8 | |
| r3 | + Weight Sharing | 18-3 | 33.6M | 34.6 |
| r4 | + LMF | 18-3 | 14.6M | 33.9 |
| r5 | + LMF & WS | 18-3 | 7.7M | 34.0 |
| r6 | LightFormer | 18-3 | 7.7M | 34.6 |
| Model | E-D | Params. | Speed | BLEU |
| Transformer | 6-6 | 36.8M | 1.0× | 34.5 |
| LightFormer | 6-6 | 7.7M | 1.3× | 34.5 |
| 18-3 | 7.7M | 2.1× | 34.6 | |
| 18-2 | 6.7M | 2.4× | 34.2 | |
| 18-1 | 5.8M | 3.0× | 33.6 | |
| 12-3 | 7.7M | 2.0× | 34.4 | |
| 12-2 | 6.7M | 2.5× | 34.1 | |
| 24-3 | 7.7M | 1.8× | 34.6 |
| Model | Groups | Params. | BLEU |
| LightFormer 6-6 | 3 | 6.9M | 34.4 |
| 2 | 6.4M | 34.2 | |
| 1 | 5.8M | 34.2 | |
| LightFormer 18-3 | 9 | 9.4M | 34.7 |
| 6 | 7.7M | 34.6 | |
| 3 | 5.9M | 34.1 | |
| 1 | 4.8M | 34.0 |
| Dataset | Model | Params | S(Pi) | BLEU |
| De-En | Transformer | 36.8M | 1.0× | 34.5 / 33.4 |
| LightFormer | 7.7M | 2.5× | 34.6 / 33.6 | |
| En-De | Transformer | 63.2M | 1.0× | 27.3 / 26.3 |
| LightFormer | 22.5M | 1.4× | 27.1 / 26.2 | |
| En-Ro | Transformer | 62.0M | 1.0× | 34.4 / 33.5 |
| LightFormer | 15.3M | 1.9× | 34.3 / 33.6 |
| Model | Rank | Params. | BLEU |
| Transformer | - | 36.8M | 34.6 |
| LMF Transformer | 256 | 21.9M | 34.3 |
| 112 | 10.4M | 34.1 | |
| 64 | 6.6M | 33.7 | |
| 32 | 4.0M | 31.7 | |
| 16 | 2.7M | 28.8 | |
| 4 | 1.8M | 17.4 |
| Model | IWSLT'14 De-En | WMT'14 En-De | ||||||
| Params. | Ratio | Speed | BLEU | Params. | Ratio | Speed | BLEU | |
| Transformer | 36.8M | 1.0× | 1.0× | 34.5 | 63.2M | 1.0× | 1.0× | 27.3 |
| Lite Transformer | 13.9M | 2.6× | 1.5× | 33.6 | 33.6M | 1.9× | 1.1× | 26.5 |
| HAT Transformer | 28.2M | 1.3× | 1.7× | 34.5 | 46.2M | 1.4× | 1.7× | 26.9 |
| DelighT | 19.9M | 1.8× | 0.8× | 34.4 | 23.3M | 2.7× | 1.2× | 26.7 |
| LightFormer | 7.7M | 4.8× | 2.1× | 34.6 | 22.5M | 2.8× | 1.5× | 27.1 |
| w/o SVD WT | 7.7M | 4.8× | 2.1× | 34.0 | 22.5M | 2.8× | 1.5× | 26.5 |
| Model | Params. | Ratio | Speed | BLEU |
| Transformer | 62M | 1.0× | 1.0× | 34.4 |
| Delight | 22.0M | 2.8× | 1.2× | 34.3 |
| Subformer | 20.0M | 3.1× | - | 34.1 |
| LightFormer | 15.3M | 4.1× | 1.8× | 34.3 |
| w/o SVD WT | 15.3M | 4.1× | 1.8× | 33.9 |
| Model | Params. | R-1 | R-2 | R-L |
| Transformer | 51.28M | 37.5 | 18.9 | 34.7 |
| LightFormer | 13.04M | 37.5 | 19.2 | 35.0 |
| w/o SVD WT | 13.04M | 37.1 | 18.7 | 34.7 |
| Model | Params. | CR | MR | SUBJ | MPQA | Average |
| Transformer | 12.8M | 86.2 | 80.1 | 95.4 | 90.0 | 87.93 |
| LightFormer | 0.5M | 87.3 | 80.6 | 95.4 | 90.1 | 88.35 |
| w/o SVD WT | 0.5M | 86.5 | 80.3 | 95.1 | 89.7 | 87.90 |
| Model | E-D | Params. | BLEU | |
| r1 | Transformer | 6-6 | 36.8M | 34.5 |
| r2 | 18-3 | 52.5M | 34.8 | |
| r3 | + Weight Sharing | 18-3 | 33.6M | 34.6 |
| r4 | + LMF | 18-3 | 14.6M | 33.9 |
| r5 | + LMF & WS | 18-3 | 7.7M | 34.0 |
| r6 | LightFormer | 18-3 | 7.7M | 34.6 |
| Model | E-D | Params. | Speed | BLEU |
| Transformer | 6-6 | 36.8M | 1.0× | 34.5 |
| LightFormer | 6-6 | 7.7M | 1.3× | 34.5 |
| 18-3 | 7.7M | 2.1× | 34.6 | |
| 18-2 | 6.7M | 2.4× | 34.2 | |
| 18-1 | 5.8M | 3.0× | 33.6 | |
| 12-3 | 7.7M | 2.0× | 34.4 | |
| 12-2 | 6.7M | 2.5× | 34.1 | |
| 24-3 | 7.7M | 1.8× | 34.6 |
| Model | Groups | Params. | BLEU |
| LightFormer 6-6 | 3 | 6.9M | 34.4 |
| 2 | 6.4M | 34.2 | |
| 1 | 5.8M | 34.2 | |
| LightFormer 18-3 | 9 | 9.4M | 34.7 |
| 6 | 7.7M | 34.6 | |
| 3 | 5.9M | 34.1 | |
| 1 | 4.8M | 34.0 |
| Dataset | Model | Params | S(Pi) | BLEU |
| De-En | Transformer | 36.8M | 1.0× | 34.5 / 33.4 |
| LightFormer | 7.7M | 2.5× | 34.6 / 33.6 | |
| En-De | Transformer | 63.2M | 1.0× | 27.3 / 26.3 |
| LightFormer | 22.5M | 1.4× | 27.1 / 26.2 | |
| En-Ro | Transformer | 62.0M | 1.0× | 34.4 / 33.5 |
| LightFormer | 15.3M | 1.9× | 34.3 / 33.6 |
| Model | Rank | Params. | BLEU |
| Transformer | - | 36.8M | 34.6 |
| LMF Transformer | 256 | 21.9M | 34.3 |
| 112 | 10.4M | 34.1 | |
| 64 | 6.6M | 33.7 | |
| 32 | 4.0M | 31.7 | |
| 16 | 2.7M | 28.8 | |
| 4 | 1.8M | 17.4 |
| Model | IWSLT'14 De-En | WMT'14 En-De | ||||||
| Params. | Ratio | Speed | BLEU | Params. | Ratio | Speed | BLEU | |
| Transformer | 36.8M | 1.0× | 1.0× | 34.5 | 63.2M | 1.0× | 1.0× | 27.3 |
| Lite Transformer | 13.9M | 2.6× | 1.5× | 33.6 | 33.6M | 1.9× | 1.1× | 26.5 |
| HAT Transformer | 28.2M | 1.3× | 1.7× | 34.5 | 46.2M | 1.4× | 1.7× | 26.9 |
| DelighT | 19.9M | 1.8× | 0.8× | 34.4 | 23.3M | 2.7× | 1.2× | 26.7 |
| LightFormer | 7.7M | 4.8× | 2.1× | 34.6 | 22.5M | 2.8× | 1.5× | 27.1 |
| w/o SVD WT | 7.7M | 4.8× | 2.1× | 34.0 | 22.5M | 2.8× | 1.5× | 26.5 |
| Model | Params. | Ratio | Speed | BLEU |
| Transformer | 62M | 1.0× | 1.0× | 34.4 |
| Delight | 22.0M | 2.8× | 1.2× | 34.3 |
| Subformer | 20.0M | 3.1× | - | 34.1 |
| LightFormer | 15.3M | 4.1× | 1.8× | 34.3 |
| w/o SVD WT | 15.3M | 4.1× | 1.8× | 33.9 |
| Model | Params. | R-1 | R-2 | R-L |
| Transformer | 51.28M | 37.5 | 18.9 | 34.7 |
| LightFormer | 13.04M | 37.5 | 19.2 | 35.0 |
| w/o SVD WT | 13.04M | 37.1 | 18.7 | 34.7 |
| Model | Params. | CR | MR | SUBJ | MPQA | Average |
| Transformer | 12.8M | 86.2 | 80.1 | 95.4 | 90.0 | 87.93 |
| LightFormer | 0.5M | 87.3 | 80.6 | 95.4 | 90.1 | 88.35 |
| w/o SVD WT | 0.5M | 86.5 | 80.3 | 95.1 | 89.7 | 87.90 |
| Model | E-D | Params. | BLEU | |
| r1 | Transformer | 6-6 | 36.8M | 34.5 |
| r2 | 18-3 | 52.5M | 34.8 | |
| r3 | + Weight Sharing | 18-3 | 33.6M | 34.6 |
| r4 | + LMF | 18-3 | 14.6M | 33.9 |
| r5 | + LMF & WS | 18-3 | 7.7M | 34.0 |
| r6 | LightFormer | 18-3 | 7.7M | 34.6 |
| Model | E-D | Params. | Speed | BLEU |
| Transformer | 6-6 | 36.8M | 1.0× | 34.5 |
| LightFormer | 6-6 | 7.7M | 1.3× | 34.5 |
| 18-3 | 7.7M | 2.1× | 34.6 | |
| 18-2 | 6.7M | 2.4× | 34.2 | |
| 18-1 | 5.8M | 3.0× | 33.6 | |
| 12-3 | 7.7M | 2.0× | 34.4 | |
| 12-2 | 6.7M | 2.5× | 34.1 | |
| 24-3 | 7.7M | 1.8× | 34.6 |
| Model | Groups | Params. | BLEU |
| LightFormer 6-6 | 3 | 6.9M | 34.4 |
| 2 | 6.4M | 34.2 | |
| 1 | 5.8M | 34.2 | |
| LightFormer 18-3 | 9 | 9.4M | 34.7 |
| 6 | 7.7M | 34.6 | |
| 3 | 5.9M | 34.1 | |
| 1 | 4.8M | 34.0 |
| Dataset | Model | Params | S(Pi) | BLEU |
| De-En | Transformer | 36.8M | 1.0× | 34.5 / 33.4 |
| LightFormer | 7.7M | 2.5× | 34.6 / 33.6 | |
| En-De | Transformer | 63.2M | 1.0× | 27.3 / 26.3 |
| LightFormer | 22.5M | 1.4× | 27.1 / 26.2 | |
| En-Ro | Transformer | 62.0M | 1.0× | 34.4 / 33.5 |
| LightFormer | 15.3M | 1.9× | 34.3 / 33.6 |
| # signs | # train sents | # valid | # test | |
| Character-level datasets | ||||
| FS LibriSpeech | 87k | 287k | 5.5k | 5.6k |
| FS LJSpeech | 87k | 13.1k | 348 | 523 |
| Word-level datasets | ||||
| ASL Libri. 100 | 2.6k | 14.1k | 56 | 54 |
| ASL Libri. 200 | 4.4k | 56.2k | 291 | 311 |
| ASL Libri. 500 | 8.2k | 137k | 941 | 956 |
| ASL Libri. 1k | 11.6k | 290k | 2.7k | 2.5k |
| Model | Images/ltr | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 1000 | - | 85.1 | 93.7 | 95.4 | 79.3 | 96.8 | 99.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | - | 39.5 | 1.9 | 42.1 | 59.5 | 17.6 | 44.2 | 65.2 |
| wav2vec-U 2.0 | - | 68.1 | 1.1 | 6.3 | 11.8 | 2.6 | 9.8 | 14.9 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 100 | 43.1 | 1.7 | 42.1 | 63.4 | 27.5 | 62.7 | 78.4 |
| speech2sign-U | 300 | 45.0 | 1.9 | 48.8 | 67.1 | 22.8 | 53.9 | 71.5 |
| speech2sign-U | 500 | 46.2 | 1.0 | 33.1 | 57.2 | 31.3 | 57.0 | 72.8 |
| speech2sign-U | 1000 | 48.6 | 1.3 | 43.8 | 63.8 | 32.5 | 58.7 | 71.5 |
| Model | Vocab size | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 877 | - | 55.2 | 78.9 | 86.3 | 51.7 | 85.5 | 93.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | 98 | 73.7 | 16.1 | 32.1 | 51.8 | 17.8 | 41.1 | 50.0 |
| speech2text-U | 98 | 7.5 | 98.2 | 98.2 | 100 | 98.2 | 98.2 | 98.2 |
| speech2text-U | 193 | 11.2 | 96.9 | 98.6 | 99.0 | 96.9 | 99.3 | 99.3 |
| speech2text-U | 468 | 30.0 | 68.0 | 86.2 | 90.2 | 66.7 | 85.3 | 90.2 |
| speech2text-U | 877 | 34.4 | 37.9 | 60.7 | 69.3 | 38.7 | 59.4 | 68.3 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 98 | 53.6 | 69.6 | 96.4 | 98.2 | 71.4 | 96.4 | 100 |
| speech2sign-U | 193 | 60.8 | 75.3 | 91.1 | 92.4 | 69.4 | 90.0 | 93.5 |
| speech2sign-U | 468 | 73.2 | 56.5 | 76.8 | 83.6 | 47.6 | 74.5 | 83.5 |
| speech2sign-U | 877 | 87.9 | 12.1 | 25.5 | 32.6 | 10.9 | 22.1 | 29.7 |
| Video features | Vocab Size | A→V | ||
| R@1 | R@5 | R@10 | ||
| VGG19 | 98 | 32.1 | 64.3 | 78.6 |
| OpenPose | 98 | 0.0 | 8.9 | 16.1 |
| I3D RGB | 98 | 76.8 | 89.3 | 96.4 |
| 193 | 66.0 | 86.3 | 91.4 | |
| 877 | 1.1 | 3.0 | 5.1 | |
| I3D flow | 98 | 69.6 | 96.4 | 98.2 |
| 193 | 63.9 | 86.9 | 91.1 | |
| 468 | 43.9 | 68.5 | 76.6 | |
| 877 | 12.1 | 25.5 | 32.6 | |
| I3D joint | 98 | 28.6 | 67.9 | 76.8 |
| 193 | 75.3 | 91.1 | 92.4 | |
| 468 | 56.5 | 76.8 | 83.6 | |
| 877 | 0.1 | 0.2 | 0.4 |
| Boundary Label | Word Boundary F1 | A→V | ||
| R@1 | R@5 | R@10 | ||
| speech2text-U | ||||
| Word | 100 | 98.2 | 98.2 | 100 |
| Phoneme | 88.1 | 78.6 | 98.2 | 98.2 |
| speech2sign-U | ||||
| Word | 100 | 69.6 | 96.4 | 98.2 |
| Phoneme | 88.1 | 57.1 | 82.1 | 91.1 |
| # signs | # train sents | # valid | # test | |
| Character-level datasets | ||||
| FS LibriSpeech | 87k | 287k | 5.5k | 5.6k |
| FS LJSpeech | 87k | 13.1k | 348 | 523 |
| Word-level datasets | ||||
| ASL Libri. 100 | 2.6k | 14.1k | 56 | 54 |
| ASL Libri. 200 | 4.4k | 56.2k | 291 | 311 |
| ASL Libri. 500 | 8.2k | 137k | 941 | 956 |
| ASL Libri. 1k | 11.6k | 290k | 2.7k | 2.5k |
| Model | Images/ltr | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 1000 | - | 85.1 | 93.7 | 95.4 | 79.3 | 96.8 | 99.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | - | 39.5 | 1.9 | 42.1 | 59.5 | 17.6 | 44.2 | 65.2 |
| wav2vec-U 2.0 | - | 68.1 | 1.1 | 6.3 | 11.8 | 2.6 | 9.8 | 14.9 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 100 | 43.1 | 1.7 | 42.1 | 63.4 | 27.5 | 62.7 | 78.4 |
| speech2sign-U | 300 | 45.0 | 1.9 | 48.8 | 67.1 | 22.8 | 53.9 | 71.5 |
| speech2sign-U | 500 | 46.2 | 1.0 | 33.1 | 57.2 | 31.3 | 57.0 | 72.8 |
| speech2sign-U | 1000 | 48.6 | 1.3 | 43.8 | 63.8 | 32.5 | 58.7 | 71.5 |
| Model | Vocab size | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 877 | - | 55.2 | 78.9 | 86.3 | 51.7 | 85.5 | 93.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | 98 | 73.7 | 16.1 | 32.1 | 51.8 | 17.8 | 41.1 | 50.0 |
| speech2text-U | 98 | 7.5 | 98.2 | 98.2 | 100 | 98.2 | 98.2 | 98.2 |
| speech2text-U | 193 | 11.2 | 96.9 | 98.6 | 99.0 | 96.9 | 99.3 | 99.3 |
| speech2text-U | 468 | 30.0 | 68.0 | 86.2 | 90.2 | 66.7 | 85.3 | 90.2 |
| speech2text-U | 877 | 34.4 | 37.9 | 60.7 | 69.3 | 38.7 | 59.4 | 68.3 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 98 | 53.6 | 69.6 | 96.4 | 98.2 | 71.4 | 96.4 | 100 |
| speech2sign-U | 193 | 60.8 | 75.3 | 91.1 | 92.4 | 69.4 | 90.0 | 93.5 |
| speech2sign-U | 468 | 73.2 | 56.5 | 76.8 | 83.6 | 47.6 | 74.5 | 83.5 |
| speech2sign-U | 877 | 87.9 | 12.1 | 25.5 | 32.6 | 10.9 | 22.1 | 29.7 |
| Video features | Vocab Size | A→V | ||
| R@1 | R@5 | R@10 | ||
| VGG19 | 98 | 32.1 | 64.3 | 78.6 |
| OpenPose | 98 | 0.0 | 8.9 | 16.1 |
| I3D RGB | 98 | 76.8 | 89.3 | 96.4 |
| 193 | 66.0 | 86.3 | 91.4 | |
| 877 | 1.1 | 3.0 | 5.1 | |
| I3D flow | 98 | 69.6 | 96.4 | 98.2 |
| 193 | 63.9 | 86.9 | 91.1 | |
| 468 | 43.9 | 68.5 | 76.6 | |
| 877 | 12.1 | 25.5 | 32.6 | |
| I3D joint | 98 | 28.6 | 67.9 | 76.8 |
| 193 | 75.3 | 91.1 | 92.4 | |
| 468 | 56.5 | 76.8 | 83.6 | |
| 877 | 0.1 | 0.2 | 0.4 |
| Boundary Label | Word Boundary F1 | A→V | ||
| R@1 | R@5 | R@10 | ||
| speech2text-U | ||||
| Word | 100 | 98.2 | 98.2 | 100 |
| Phoneme | 88.1 | 78.6 | 98.2 | 98.2 |
| speech2sign-U | ||||
| Word | 100 | 69.6 | 96.4 | 98.2 |
| Phoneme | 88.1 | 57.1 | 82.1 | 91.1 |
| # signs | # train sents | # valid | # test | |
| Character-level datasets | ||||
| FS LibriSpeech | 87k | 287k | 5.5k | 5.6k |
| FS LJSpeech | 87k | 13.1k | 348 | 523 |
| Word-level datasets | ||||
| ASL Libri. 100 | 2.6k | 14.1k | 56 | 54 |
| ASL Libri. 200 | 4.4k | 56.2k | 291 | 311 |
| ASL Libri. 500 | 8.2k | 137k | 941 | 956 |
| ASL Libri. 1k | 11.6k | 290k | 2.7k | 2.5k |
| Model | Images/ltr | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 1000 | - | 85.1 | 93.7 | 95.4 | 79.3 | 96.8 | 99.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | - | 39.5 | 1.9 | 42.1 | 59.5 | 17.6 | 44.2 | 65.2 |
| wav2vec-U 2.0 | - | 68.1 | 1.1 | 6.3 | 11.8 | 2.6 | 9.8 | 14.9 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 100 | 43.1 | 1.7 | 42.1 | 63.4 | 27.5 | 62.7 | 78.4 |
| speech2sign-U | 300 | 45.0 | 1.9 | 48.8 | 67.1 | 22.8 | 53.9 | 71.5 |
| speech2sign-U | 500 | 46.2 | 1.0 | 33.1 | 57.2 | 31.3 | 57.0 | 72.8 |
| speech2sign-U | 1000 | 48.6 | 1.3 | 43.8 | 63.8 | 32.5 | 58.7 | 71.5 |
| Model | Vocab size | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 877 | - | 55.2 | 78.9 | 86.3 | 51.7 | 85.5 | 93.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | 98 | 73.7 | 16.1 | 32.1 | 51.8 | 17.8 | 41.1 | 50.0 |
| speech2text-U | 98 | 7.5 | 98.2 | 98.2 | 100 | 98.2 | 98.2 | 98.2 |
| speech2text-U | 193 | 11.2 | 96.9 | 98.6 | 99.0 | 96.9 | 99.3 | 99.3 |
| speech2text-U | 468 | 30.0 | 68.0 | 86.2 | 90.2 | 66.7 | 85.3 | 90.2 |
| speech2text-U | 877 | 34.4 | 37.9 | 60.7 | 69.3 | 38.7 | 59.4 | 68.3 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 98 | 53.6 | 69.6 | 96.4 | 98.2 | 71.4 | 96.4 | 100 |
| speech2sign-U | 193 | 60.8 | 75.3 | 91.1 | 92.4 | 69.4 | 90.0 | 93.5 |
| speech2sign-U | 468 | 73.2 | 56.5 | 76.8 | 83.6 | 47.6 | 74.5 | 83.5 |
| speech2sign-U | 877 | 87.9 | 12.1 | 25.5 | 32.6 | 10.9 | 22.1 | 29.7 |
| Video features | Vocab Size | A→V | ||
| R@1 | R@5 | R@10 | ||
| VGG19 | 98 | 32.1 | 64.3 | 78.6 |
| OpenPose | 98 | 0.0 | 8.9 | 16.1 |
| I3D RGB | 98 | 76.8 | 89.3 | 96.4 |
| 193 | 66.0 | 86.3 | 91.4 | |
| 877 | 1.1 | 3.0 | 5.1 | |
| I3D flow | 98 | 69.6 | 96.4 | 98.2 |
| 193 | 63.9 | 86.9 | 91.1 | |
| 468 | 43.9 | 68.5 | 76.6 | |
| 877 | 12.1 | 25.5 | 32.6 | |
| I3D joint | 98 | 28.6 | 67.9 | 76.8 |
| 193 | 75.3 | 91.1 | 92.4 | |
| 468 | 56.5 | 76.8 | 83.6 | |
| 877 | 0.1 | 0.2 | 0.4 | |
| Boundary Label | Word Boundary F1 | A→V | ||
| R@1 | R@5 | R@10 | ||
| speech2text-U | ||||
| Word | 100 | 98.2 | 98.2 | 100 |
| Phoneme | 88.1 | 78.6 | 98.2 | 98.2 |
| speech2sign-U | ||||
| Word | 100 | 69.6 | 96.4 | 98.2 |
| Phoneme | 88.1 | 57.1 | 82.1 | 91.1 |
| # signs | # train sents | # valid | # test | |
| Character-level datasets | ||||
| FS LibriSpeech | 87k | 287k | 5.5k | 5.6k |
| FS LJSpeech | 87k | 13.1k | 348 | 523 |
| Word-level datasets | ||||
| ASL Libri. 100 | 2.6k | 14.1k | 56 | 54 |
| ASL Libri. 200 | 4.4k | 56.2k | 291 | 311 |
| ASL Libri. 500 | 8.2k | 137k | 941 | 956 |
| ASL Libri. 1k | 11.6k | 290k | 2.7k | 2.5k |
| Model | Images/ltr | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 1000 | - | 85.1 | 93.7 | 95.4 | 79.3 | 96.8 | 99.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | - | 39.5 | 1.9 | 42.1 | 59.5 | 17.6 | 44.2 | 65.2 |
| wav2vec-U 2.0 | - | 68.1 | 1.1 | 6.3 | 11.8 | 2.6 | 9.8 | 14.9 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 100 | 43.1 | 1.7 | 42.1 | 63.4 | 27.5 | 62.7 | 78.4 |
| speech2sign-U | 300 | 45.0 | 1.9 | 48.8 | 67.1 | 22.8 | 53.9 | 71.5 |
| speech2sign-U | 500 | 46.2 | 1.0 | 33.1 | 57.2 | 31.3 | 57.0 | 72.8 |
| speech2sign-U | 1000 | 48.6 | 1.3 | 43.8 | 63.8 | 32.5 | 58.7 | 71.5 |
| Model | Vocab size | UER↓ | A→V | V→A | ||||
| R@1↑ | R@5↑ | R@10↑ | R@1↑ | R@5↑ | R@10↑ | |||
| Supervised speech-to-sign recognition | ||||||||
| (Harwath et al., 2018) | 877 | - | 55.2 | 78.9 | 86.3 | 51.7 | 85.5 | 93.1 |
| Unsupervised speech recognition | ||||||||
| wav2vec-U | 98 | 73.7 | 16.1 | 32.1 | 51.8 | 17.8 | 41.1 | 50.0 |
| speech2text-U | 98 | 7.5 | 98.2 | 98.2 | 100 | 98.2 | 98.2 | 98.2 |
| speech2text-U | 193 | 11.2 | 96.9 | 98.6 | 99.0 | 96.9 | 99.3 | 99.3 |
| speech2text-U | 468 | 30.0 | 68.0 | 86.2 | 90.2 | 66.7 | 85.3 | 90.2 |
| speech2text-U | 877 | 34.4 | 37.9 | 60.7 | 69.3 | 38.7 | 59.4 | 68.3 |
| Unsupervised speech-to-sign recognition | ||||||||
| speech2sign-U | 98 | 53.6 | 69.6 | 96.4 | 98.2 | 71.4 | 96.4 | 100 |
| speech2sign-U | 193 | 60.8 | 75.3 | 91.1 | 92.4 | 69.4 | 90.0 | 93.5 |
| speech2sign-U | 468 | 73.2 | 56.5 | 76.8 | 83.6 | 47.6 | 74.5 | 83.5 |
| speech2sign-U | 877 | 87.9 | 12.1 | 25.5 | 32.6 | 10.9 | 22.1 | 29.7 |
| Video features | Vocab Size | A→V | ||
| R@1 | R@5 | R@10 | ||
| VGG19 | 98 | 32.1 | 64.3 | 78.6 |
| OpenPose | 98 | 0.0 | 8.9 | 16.1 |
| I3D RGB | 98 | 76.8 | 89.3 | 96.4 |
| 193 | 66.0 | 86.3 | 91.4 | |
| 877 | 1.1 | 3.0 | 5.1 | |
| I3D flow | 98 | 69.6 | 96.4 | 98.2 |
| 193 | 63.9 | 86.9 | 91.1 | |
| 468 | 43.9 | 68.5 | 76.6 | |
| 877 | 12.1 | 25.5 | 32.6 | |
| I3D joint | 98 | 28.6 | 67.9 | 76.8 |
| 193 | 75.3 | 91.1 | 92.4 | |
| 468 | 56.5 | 76.8 | 83.6 | |
| 877 | 0.1 | 0.2 | 0.4 |
| Boundary Label | Word Boundary F1 | A→V | ||
| R@1 | R@5 | R@10 | ||
| speech2text-U | ||||
| Word | 100 | 98.2 | 98.2 | 100 |
| Phoneme | 88.1 | 78.6 | 98.2 | 98.2 |
| speech2sign-U | ||||
| Word | 100 | 69.6 | 96.4 | 98.2 |
| Phoneme | 88.1 | 57.1 | 82.1 | 91.1 |
| n-gram | n-gram-rep | P(ŷ; θ) | P(n-gram-rep) |
| 1-gram | 0.66 | 0.89 | 0.35 |
| 2-gram | 0.60 | 0.90 | 0.07 |
| 3-gram | 0.58 | 0.91 | 0.03 |
| 4-gram | 0.57 | 0.92 | 0.02 |
| Dataset | Model | Algorithm | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Webtext | GPT2-Small | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - |
| Greedy | 54.48 | 51.52 | 49.68 | 11.1 | 1.82 | 55.2 | 24.7 | ||
| Beam Search | 60.02 | 58.04 | 56.67 | 7.3 | 1.38 | 60.3 | 20.9 | ||
| Top-p | 10.04 | 5.61 | 4.095 | 81.4 | 4.31 | 56.4 | 67.7 | ||
| Contrastive | 50.97 | 47.59 | 45.45 | 1.4 | 1.91 | 56.6 | 28.1 | ||
| Ours | 6.18 | 2.46 | 1.53 | 90.1 | 2.56 | 54.5 | 70.1 | ||
| GPT2-Medium | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 46.11 | 42.30 | 40.01 | 18.6 | 1.9 | 57.3 | 32.6 | ||
| Beam Search | 51.98 | 49.48 | 47.84 | 12.7 | 1.45 | 61.6 | 27.9 | ||
| Top-p | 8.41 | 4.34 | 3.0 | 85 | 4.1 | 58.0 | 70.2 | ||
| Contrastive | 41.86 | 37.58 | 35.0 | 23.6 | 2.1 | 58.96 | 37.3 | ||
| Ours | 4.827 | 1.52 | 0.526 | 93.2 | 2.6 | 57.1 | 72.4 | ||
| GPT2-Large | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 44.04 | 39.99 | 37.61 | 20.9 | 2.1 | 57.5 | 34.7 | ||
| Beam Search | 49.3 | 46.6 | 44.8 | 14.9 | 1.4 | 61.7 | 30.3 | ||
| Top-p | 9.665 | 5.24 | 3.765 | 82.4 | 3.5 | 58.3 | 69.3 | ||
| Contrastive | 4.7 | 2.1 | 1.4 | 91.9 | 3.69 | 55.9 | 71.7 | ||
| Ours | 5.5 | 1.85 | 0.9 | 91.9 | 2.0 | 57.1 | 72.4 | ||
| Wikitext-103 | GPT2-Wiki | Human | 3.6 | 0.85 | 0.28 | 95.3 | - | - | - |
| Greedy | 47.63 | 41.41 | 37.44 | 19.2 | 3.68 | 54.8 | 32.4 | ||
| Beam Search | 52.97 | 47.70 | 44.20 | 13.7 | 2.26 | 50.3 | 26.2 | ||
| Top-p | 12.52 | 5.73 | 3.27 | 79.8 | 5.41 | 56.5 | 67.2 | ||
| Contrastive | 9.41 | 4.22 | 2.75 | 84.4 | 2.90 | 53.3 | 67.1 | ||
| Ours | 5.91 | 1.13 | 0.30 | 92.8 | 3.11 | 50.5 | 68.45 |
| Algorithm | Δ | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Beam Search | B=20 | 57.73 | 53.02 | 49.82 | 10.0 | 2.29 | 48.28 | 21.97 |
| Ours | B=20 | 5.626 | 1.11 | 0.31 | 93.0 | 3.28 | 49.38 | 67.77 |
| Beam Search | τ=2.0 | 39.84 | 33.23 | 28.59 | 28.7 | 3.53 | 48.32 | 37.24 |
| Ours | τ2 | 22.119 | 15.27 | 11.49 | 58.4 | 2.85 | 50.19 | 54.14 |
| Ours | τ3 | 10.637 | 3.38 | 1.39 | 85.1 | 2.99 | 50.70 | 65.68 |
| Ours | τ4 | 11.9 | 2.89 | 0.57 | 85.1 | 2.92 | 51.36 | 66.11 |
| Dataset | Beam | Top-p | Contrastive | Ours) |
| EN-RU | 33.4 | 27.5 (-5.9) | 19.9 (-13.5) | 33.2 (-0.2) |
| RU-EN | 39.0 | 32.0 (-7.0) | 24.3 (-14.8) | 38.9 (-0.1) |
| EN-DE | 42.8 | 34.7 (-8.1) | 29.5(-13.3) | 42.6 (-0.2) |
| DE-EN | 41.3 | 33.9 (-7.4) | 26.9(-14.4) | 41.2 (-0.1) |
| n-gram | n-gram-rep | P(ŷ; θ) | P(n-gram-rep) |
| 1-gram | 0.66 | 0.89 | 0.35 |
| 2-gram | 0.60 | 0.90 | 0.07 |
| 3-gram | 0.58 | 0.91 | 0.03 |
| 4-gram | 0.57 | 0.92 | 0.02 |
| Dataset | Model | Algorithm | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Webtext | GPT2-Small | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - |
| Greedy | 54.48 | 51.52 | 49.68 | 11.1 | 1.82 | 55.2 | 24.7 | ||
| Beam Search | 60.02 | 58.04 | 56.67 | 7.3 | 1.38 | 60.3 | 20.9 | ||
| Top-p | 10.04 | 5.61 | 4.095 | 81.4 | 4.31 | 56.4 | 67.7 | ||
| Contrastive | 50.97 | 47.59 | 45.45 | 1.4 | 1.91 | 56.6 | 28.1 | ||
| Ours | 6.18 | 2.46 | 1.53 | 90.1 | 2.56 | 54.5 | 70.1 | ||
| GPT2-Medium | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 46.11 | 42.30 | 40.01 | 18.6 | 1.9 | 57.3 | 32.6 | ||
| Beam Search | 51.98 | 49.48 | 47.84 | 12.7 | 1.45 | 61.6 | 27.9 | ||
| Top-p | 8.41 | 4.34 | 3.0 | 85 | 4.1 | 58.0 | 70.2 | ||
| Contrastive | 41.86 | 37.58 | 35.0 | 23.6 | 2.1 | 58.96 | 37.3 | ||
| Ours | 4.827 | 1.52 | 0.526 | 93.2 | 2.6 | 57.1 | 72.4 | ||
| GPT2-Large | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 44.04 | 39.99 | 37.61 | 20.9 | 2.1 | 57.5 | 34.7 | ||
| Beam Search | 49.3 | 46.6 | 44.8 | 14.9 | 1.4 | 61.7 | 30.3 | ||
| Top-p | 9.665 | 5.24 | 3.765 | 82.4 | 3.5 | 58.3 | 69.3 | ||
| Contrastive | 4.7 | 2.1 | 1.4 | 91.9 | 3.69 | 55.9 | 71.7 | ||
| Ours | 5.5 | 1.85 | 0.9 | 91.9 | 2.0 | 57.1 | 72.4 | ||
| Wikitext-103 | GPT2-Wiki | Human | 3.6 | 0.85 | 0.28 | 95.3 | - | - | - |
| Greedy | 47.63 | 41.41 | 37.44 | 19.2 | 3.68 | 54.8 | 32.4 | ||
| Beam Search | 52.97 | 47.70 | 44.20 | 13.7 | 2.26 | 50.3 | 26.2 | ||
| Top-p | 12.52 | 5.73 | 3.27 | 79.8 | 5.41 | 56.5 | 67.2 | ||
| Contrastive | 9.41 | 4.22 | 2.75 | 84.4 | 2.90 | 53.3 | 67.1 | ||
| Ours | 5.91 | 1.13 | 0.30 | 92.8 | 3.11 | 50.5 | 68.45 |
| Algorithm | Δ | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Beam Search | B=20 | 57.73 | 53.02 | 49.82 | 10.0 | 2.29 | 48.28 | 21.97 |
| Ours | B=20 | 5.626 | 1.11 | 0.31 | 93.0 | 3.28 | 49.38 | 67.77 |
| Beam Search | τ=2.0 | 39.84 | 33.23 | 28.59 | 28.7 | 3.53 | 48.32 | 37.24 |
| Ours | τ2 | 22.119 | 15.27 | 11.49 | 58.4 | 2.85 | 50.19 | 54.14 |
| Ours | τ3 | 10.637 | 3.38 | 1.39 | 85.1 | 2.99 | 50.70 | 65.68 |
| Ours | τ4 | 11.9 | 2.89 | 0.57 | 85.1 | 2.92 | 51.36 | 66.11 |
| Dataset | Beam | Top-p | Contrastive | Ours) |
| EN-RU | 33.4 | 27.5 (-5.9) | 19.9 (-13.5) | 33.2 (-0.2) |
| RU-EN | 39.0 | 32.0 (-7.0) | 24.3 (-14.8) | 38.9 (-0.1) |
| EN-DE | 42.8 | 34.7 (-8.1) | 29.5(-13.3) | 42.6 (-0.2) |
| DE-EN | 41.3 | 33.9 (-7.4) | 26.9(-14.4) | 41.2 (-0.1) |
| n-gram | n-gram-rep | P(ŷ; θ) | P(n-gram-rep) |
| 1-gram | 0.66 | 0.89 | 0.35 |
| 2-gram | 0.60 | 0.90 | 0.07 |
| 3-gram | 0.58 | 0.91 | 0.03 |
| 4-gram | 0.57 | 0.92 | 0.02 |
| Dataset | Model | Algorithm | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Webtext | GPT2-Small | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - |
| Greedy | 54.48 | 51.52 | 49.68 | 11.1 | 1.82 | 55.2 | 24.7 | ||
| Beam Search | 60.02 | 58.04 | 56.67 | 7.3 | 1.38 | 60.3 | 20.9 | ||
| Top-p | 10.04 | 5.61 | 4.095 | 81.4 | 4.31 | 56.4 | 67.7 | ||
| Contrastive | 50.97 | 47.59 | 45.45 | 1.4 | 1.91 | 56.6 | 28.1 | ||
| Ours | 6.18 | 2.46 | 1.53 | 90.1 | 2.56 | 54.5 | 70.1 | ||
| GPT2-Medium | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 46.11 | 42.30 | 40.01 | 18.6 | 1.9 | 57.3 | 32.6 | ||
| Beam Search | 51.98 | 49.48 | 47.84 | 12.7 | 1.45 | 61.6 | 27.9 | ||
| Top-p | 8.41 | 4.34 | 3.0 | 85 | 4.1 | 58.0 | 70.2 | ||
| Contrastive | 41.86 | 37.58 | 35.0 | 23.6 | 2.1 | 58.96 | 37.3 | ||
| Ours | 4.827 | 1.52 | 0.526 | 93.2 | 2.6 | 57.1 | 72.4 | ||
| GPT2-Large | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 44.04 | 39.99 | 37.61 | 20.9 | 2.1 | 57.5 | 34.7 | ||
| Beam Search | 49.3 | 46.6 | 44.8 | 14.9 | 1.4 | 61.7 | 30.3 | ||
| Top-p | 9.665 | 5.24 | 3.765 | 82.4 | 3.5 | 58.3 | 69.3 | ||
| Contrastive | 4.7 | 2.1 | 1.4 | 91.9 | 3.69 | 55.9 | 71.7 | ||
| Ours | 5.5 | 1.85 | 0.9 | 91.9 | 2.0 | 57.1 | 72.4 | ||
| Wikitext-103 | GPT2-Wiki | Human | 3.6 | 0.85 | 0.28 | 95.3 | - | - | - |
| Greedy | 47.63 | 41.41 | 37.44 | 19.2 | 3.68 | 54.8 | 32.4 | ||
| Beam Search | 52.97 | 47.70 | 44.20 | 13.7 | 2.26 | 50.3 | 26.2 | ||
| Top-p | 12.52 | 5.73 | 3.27 | 79.8 | 5.41 | 56.5 | 67.2 | ||
| Contrastive | 9.41 | 4.22 | 2.75 | 84.4 | 2.90 | 53.3 | 67.1 | ||
| Ours | 5.91 | 1.13 | 0.30 | 92.8 | 3.11 | 50.5 | 68.45 |
| Algorithm | Δ | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Beam Search | B=20 | 57.73 | 53.02 | 49.82 | 10.0 | 2.29 | 48.28 | 21.97 |
| Ours | B=20 | 5.626 | 1.11 | 0.31 | 93.0 | 3.28 | 49.38 | 67.77 |
| Beam Search | τ=2.0 | 39.84 | 33.23 | 28.59 | 28.7 | 3.53 | 48.32 | 37.24 |
| Ours | τ2 | 22.119 | 15.27 | 11.49 | 58.4 | 2.85 | 50.19 | 54.14 |
| Ours | τ3 | 10.637 | 3.38 | 1.39 | 85.1 | 2.99 | 50.70 | 65.68 |
| Ours | τ4 | 11.9 | 2.89 | 0.57 | 85.1 | 2.92 | 51.36 | 66.11 |
| Dataset | Beam | Top-p | Contrastive | Ours) |
| EN-RU | 33.4 | 27.5 (-5.9) | 19.9 (-13.5) | 33.2 (-0.2) |
| RU-EN | 39.0 | 32.0 (-7.0) | 24.3 (-14.8) | 38.9 (-0.1) |
| EN-DE | 42.8 | 34.7 (-8.1) | 29.5(-13.3) | 42.6 (-0.2) |
| DE-EN | 41.3 | 33.9 (-7.4) | 26.9(-14.4) | 41.2 (-0.1) |
| n-gram | n-gram-rep | P(ŷ; θ) | P(n-gram-rep) |
| 1-gram | 0.66 | 0.89 | 0.35 |
| 2-gram | 0.60 | 0.90 | 0.07 |
| 3-gram | 0.58 | 0.91 | 0.03 |
| 4-gram | 0.57 | 0.92 | 0.02 |
| Dataset | Model | Algorithm | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Webtext | GPT2-Small | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - |
| Greedy | 54.48 | 51.52 | 49.68 | 11.1 | 1.82 | 55.2 | 24.7 | ||
| Beam Search | 60.02 | 58.04 | 56.67 | 7.3 | 1.38 | 60.3 | 20.9 | ||
| Top-p | 10.04 | 5.61 | 4.095 | 81.4 | 4.31 | 56.4 | 67.7 | ||
| Contrastive | 50.97 | 47.59 | 45.45 | 1.4 | 1.91 | 56.6 | 28.1 | ||
| Ours | 6.18 | 2.46 | 1.53 | 90.1 | 2.56 | 54.5 | 70.1 | ||
| GPT2-Medium | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 46.11 | 42.30 | 40.01 | 18.6 | 1.9 | 57.3 | 32.6 | ||
| Beam Search | 51.98 | 49.48 | 47.84 | 12.7 | 1.45 | 61.6 | 27.9 | ||
| Top-p | 8.41 | 4.34 | 3.0 | 85 | 4.1 | 58.0 | 70.2 | ||
| Contrastive | 41.86 | 37.58 | 35.0 | 23.6 | 2.1 | 58.96 | 37.3 | ||
| Ours | 4.827 | 1.52 | 0.526 | 93.2 | 2.6 | 57.1 | 72.4 | ||
| GPT2-Large | Human | 7.77 | 2.93 | 1.81 | 87.9 | - | - | - | |
| Greedy | 44.04 | 39.99 | 37.61 | 20.9 | 2.1 | 57.5 | 34.7 | ||
| Beam Search | 49.3 | 46.6 | 44.8 | 14.9 | 1.4 | 61.7 | 30.3 | ||
| Top-p | 9.665 | 5.24 | 3.765 | 82.4 | 3.5 | 58.3 | 69.3 | ||
| Contrastive | 4.7 | 2.1 | 1.4 | 91.9 | 3.69 | 55.9 | 71.7 | ||
| Ours | 5.5 | 1.85 | 0.9 | 91.9 | 2.0 | 57.1 | 72.4 | ||
| Wikitext-103 | GPT2-Wiki | Human | 3.6 | 0.85 | 0.28 | 95.3 | - | - | - |
| Greedy | 47.63 | 41.41 | 37.44 | 19.2 | 3.68 | 54.8 | 32.4 | ||
| Beam Search | 52.97 | 47.70 | 44.20 | 13.7 | 2.26 | 50.3 | 26.2 | ||
| Top-p | 12.52 | 5.73 | 3.27 | 79.8 | 5.41 | 56.5 | 67.2 | ||
| Contrastive | 9.41 | 4.22 | 2.75 | 84.4 | 2.90 | 53.3 | 67.1 | ||
| Ours | 5.91 | 1.13 | 0.30 | 92.8 | 3.11 | 50.5 | 68.45 |
| Algorithm | Δ | rep-2 | rep-3 | rep-4 | diversity | PPL | Sim-CSE | G-Score |
| Beam Search | B=20 | 57.73 | 53.02 | 49.82 | 10.0 | 2.29 | 48.28 | 21.97 |
| Ours | B=20 | 5.626 | 1.11 | 0.31 | 93.0 | 3.28 | 49.38 | 67.77 |
| Beam Search | τ=2.0 | 39.84 | 33.23 | 28.59 | 28.7 | 3.53 | 48.32 | 37.24 |
| Ours | τ2 | 22.119 | 15.27 | 11.49 | 58.4 | 2.85 | 50.19 | 54.14 |
| Ours | τ3 | 10.637 | 3.38 | 1.39 | 85.1 | 2.99 | 50.70 | 65.68 |
| Ours | τ4 | 11.9 | 2.89 | 0.57 | 85.1 | 2.92 | 51.36 | 66.11 |
| Dataset | Beam | Top-p | Contrastive | Ours) |
| EN-RU | 33.4 | 27.5 (-5.9) | 19.9 (-13.5) | 33.2 (-0.2) |
| RU-EN | 39.0 | 32.0 (-7.0) | 24.3 (-14.8) | 38.9 (-0.1) |
| EN-DE | 42.8 | 34.7 (-8.1) | 29.5(-13.3) | 42.6 (-0.2) |
| DE-EN | 41.3 | 33.9 (-7.4) | 26.9(-14.4) | 41.2 (-0.1) |
| Logical Embeddings | Sentence Tokens | |||||||||||||||||||
| But | I | try | very | hard | not | to | go | that | way | bec. | it | wld | be | too | easy | for | them | . | ||
| Comparative | key | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | |
| Causal | - | - | α | α | α | α | α | α | α | α | key | β | β | β | β | β | β | β | β | |
| Negation | - | - | - | - | - | key | α | α | α | α | - | - | - | - | - | - | - | - | - |
| Model | ReClor | LogiQA | DREAM |
| Acc | Acc | Acc | |
| RoBERTa-large | 62.6 | 35.3 | 82.1 |
| Logical-RoBERTa-large | 67.4 | 37.8 | 84.9 |
| Model | DialogSum | |||
| R-1 | R-2 | R-L | R-LSum | |
| BART-large | 46.10 | 20.32 | 38.04 | 40.98 |
| Logical-BART-large | 46.97 | 20.69 | 38.33 | 41.30 |
| Logical Embeddings | Sentence Tokens | |||||||||||||||||||
| But | I | try | very | hard | not | to | go | that | way | bec. | it | wld | be | too | easy | for | them | . | ||
| Comparative | key | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | |
| Causal | - | - | α | α | α | α | α | α | α | α | key | β | β | β | β | β | β | β | β | |
| Negation | - | - | - | - | - | key | α | α | α | α | - | - | - | - | - | - | - | - | - |
| Model | ReClor | LogiQA | DREAM |
| Acc | Acc | Acc | |
| RoBERTa-large | 62.6 | 35.3 | 82.1 |
| Logical-RoBERTa-large | 67.4 | 37.8 | 84.9 |
| Model | DialogSum | |||
| R-1 | R-2 | R-L | R-LSum | |
| BART-large | 46.10 | 20.32 | 38.04 | 40.98 |
| Logical-BART-large | 46.97 | 20.69 | 38.33 | 41.30 |
| Logical Embeddings | Sentence Tokens | |||||||||||||||||||
| But | I | try | very | hard | not | to | go | that | way | bec. | it | wld | be | too | easy | for | them | . | ||
| Comparative | key | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | |
| Causal | - | - | α | α | α | α | α | α | α | α | key | β | β | β | β | β | β | β | β | |
| Negation | - | - | - | - | - | key | α | α | α | α | - | - | - | - | - | - | - | - | - | |
| Model | ReClor | LogiQA | DREAM |
| Acc | Acc | Acc | |
| RoBERTa-large | 62.6 | 35.3 | 82.1 |
| Logical-RoBERTa-large | 67.4 | 37.8 | 84.9 |
| Model | DialogSum | |||
| R-1 | R-2 | R-L | R-LSum | |
| BART-large | 46.10 | 20.32 | 38.04 | 40.98 |
| Logical-BART-large | 46.97 | 20.69 | 38.33 | 41.30 |
| Logical Embeddings | Sentence Tokens | |||||||||||||||||||
| But | I | try | very | hard | not | to | go | that | way | bec. | it | wld | be | too | easy | for | them | . | ||
| Comparative | key | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | β | |
| Causal | - | - | α | α | α | α | α | α | α | α | key | β | β | β | β | β | β | β | β | |
| Negation | - | - | - | - | - | key | α | α | α | α | - | - | - | - | - | - | - | - | - |
| Model | ReClor | LogiQA | DREAM |
| Acc | Acc | Acc | |
| RoBERTa-large | 62.6 | 35.3 | 82.1 |
| Logical-RoBERTa-large | 67.4 | 37.8 | 84.9 |
| Model | DialogSum | |||
| R-1 | R-2 | R-L | R-LSum | |
| BART-large | 46.10 | 20.32 | 38.04 | 40.98 |
| Logical-BART-large | 46.97 | 20.69 | 38.33 | 41.30 |
| Bing | Internal | ||
| EN→FR PM Ann. | 24/24 | 23/24 | 24/24 |
| FR→EN Hansard-King | 3/3 | 3/3 | 1/3 |
| FR→EN PM Ann. | 22*/24 | 23/24 | 17/24 |
| Bing | Internal | ||
| EN→FR PM Ann. | 24/24 | 23/24 | 24/24 |
| FR→EN Hansard-King | 3/3 | 3/3 | 1/3 |
| FR→EN PM Ann. | 22*/24 | 23/24 | 17/24 |
| Bing | Internal | ||
| EN→FR PM Ann. | 24/24 | 23/24 | 24/24 |
| FR→EN Hansard-King | 3/3 | 3/3 | 1/3 |
| FR→EN PM Ann. | 22*/24 | 23/24 | 17/24 |
| Bing | Internal | ||
| EN→FR PM Ann. | 24/24 | 23/24 | 24/24 |
| FR→EN Hansard-King | 3/3 | 3/3 | 1/3 |
| FR→EN PM Ann. | 22*/24 | 23/24 | 17/24 |
| Model | Total Params | Trained \nparams/ \ntask | QNLI | RTE | QQP | MNLI- \nm | MNLI- \nmm | SST-2 | MRPC | COLA | STS- \nB | Avg |
| Single Task | ||||||||||||
| BERT[a] | x9 | 100% | 90.5 | 66.4 | 71.2 | 84.6 | 83.4 | 93.5 | 88.9 | 52.1 | 85.8 | 79.6 |
| Bitfit[b] | x1.008 | 0.09% | 89.7 | 65.5 | 67.8 | 80.8 | 80.9 | 92.4 | 87.4 | 47.2 | 87.6 | 77.7 |
| Multi-task | ||||||||||||
| BERT MTL[c] | x1 | 11.1% | 90.5 | 74.5 | 70.4 | 83.5 | 83.1 | 93.1 | 88.0 | 48.5 | 80.6 | 79.1 |
| BERT MTL[d] | x1 | 11.1% | 89.3 | 76.6 | 70.8 | 84.0 | 83.4 | 93.4 | 86.7 | 51.2 | 83.6 | 79.9 |
| PALs[d] | x1.13 | 12.5% | 90.0 | 76.0 | 71.5 | 84.3 | 83.5 | 92.6 | 88.7 | 51.2 | 85.8 | 80.4 |
| CA-MTL[d] | x1.12 | 5.6% | 90.5 | 76.4 | 69.2 | 85.9 | 85.8 | 93.2 | 88.6 | 53.1 | 85.3 | 80.9 |
| Our approach | ||||||||||||
| Morris† | x1.024 | 0.27% | 91.1 | 73.7 | 70.6 | 83.8 | 83.5 | 92.8 | 90.2 | 52.0 | 85.8 | 80.4 |
| Morris L0† | x1.014 | 0.16% | 91.6 | 73.7 | 70.5 | 84.1 | 83.1 | 92.1 | 89.6 | 49.9 | 86.3 | 80.1 |
| Model | Total Params | Trained params/ task | Avg |
| Single Task | |||
| LoRA | x1.024 | 0.27% | 80.7 |
| Lora_BitFit | x1.03 | 0.34% | 81.2 |
| Multi Task | |||
| LoRA_BitFit_MTL | x1.022 | 0.246% | 81.8 |
| Morris | x1.024 | 0.27% | 82.2 |
| Morris L0 | x1.014 | 0.16% | 82.4 |
| Tasks | Corpus | |Train| | |Test| |
| CoLA (Warstadt et al., 2018) | Corpus of Linguistic Acceptability | 8.5K | 1K |
| SST-2 (Socher et al., 2013) | Stanford Sentiment Treebank | 67K | 1.8K |
| MRPC (Dolan and Brockett, 2005) | Microsoft Research Paraphrase Corpus | 3.7K | 1.7K |
| STS-B (Cer et al., 2017) | Semantic Textual Similarity Benchmark | 7K | 1.4K |
| QQP (Iyer et al., 2017) | Quora Question Pairs | 364K | 391K |
| MNLI (Williams et al., 2018) | Multi-Genre NLI | 393K | 20K |
| QNLI (Rajpurkar et al., 2016) | Question NLI | 105K | 5.4K |
| RTE (Dagan et al., 2005) | Recognition Textual Entailment | 2.5K | 3K |
| WNLI (Levesque et al., 2012) | Winograd NLI | 634 | 146 |
| Model | Total Params | Trained \nparams/ \ntask | QNLI | RTE | QQP | MNLI- \nm | MNLI- \nmm | SST-2 | MRPC | COLA | STS- \nB | Avg |
| Single Task | ||||||||||||
| BERT[a] | x9 | 100% | 90.5 | 66.4 | 71.2 | 84.6 | 83.4 | 93.5 | 88.9 | 52.1 | 85.8 | 79.6 |
| Bitfit[b] | x1.008 | 0.09% | 89.7 | 65.5 | 67.8 | 80.8 | 80.9 | 92.4 | 87.4 | 47.2 | 87.6 | 77.7 |
| Multi-task | ||||||||||||
| BERT MTL[c] | x1 | 11.1% | 90.5 | 74.5 | 70.4 | 83.5 | 83.1 | 93.1 | 88.0 | 48.5 | 80.6 | 79.1 |
| BERT MTL[d] | x1 | 11.1% | 89.3 | 76.6 | 70.8 | 84.0 | 83.4 | 93.4 | 86.7 | 51.2 | 83.6 | 79.9 |
| PALs[d] | x1.13 | 12.5% | 90.0 | 76.0 | 71.5 | 84.3 | 83.5 | 92.6 | 88.7 | 51.2 | 85.8 | 80.4 |
| CA-MTL[d] | x1.12 | 5.6% | 90.5 | 76.4 | 69.2 | 85.9 | 85.8 | 93.2 | 88.6 | 53.1 | 85.3 | 80.9 |
| Our approach | ||||||||||||
| Morris† | x1.024 | 0.27% | 91.1 | 73.7 | 70.6 | 83.8 | 83.5 | 92.8 | 90.2 | 52.0 | 85.8 | 80.4 |
| Morris L0† | x1.014 | 0.16% | 91.6 | 73.7 | 70.5 | 84.1 | 83.1 | 92.1 | 89.6 | 49.9 | 86.3 | 80.1 |
| Model | Total Params | Trained params/ task | Avg |
| Single Task | |||
| LoRA | x1.024 | 0.27% | 80.7 |
| Lora_BitFit | x1.03 | 0.34% | 81.2 |
| Multi Task | |||
| LoRA_BitFit_MTL | x1.022 | 0.246% | 81.8 |
| Morris | x1.024 | 0.27% | 82.2 |
| Morris L0 | x1.014 | 0.16% | 82.4 |
| Tasks | Corpus | |Train| | |Test| |
| CoLA (Warstadt et al., 2018) | Corpus of Linguistic Acceptability | 8.5K | 1K |
| SST-2 (Socher et al., 2013) | Stanford Sentiment Treebank | 67K | 1.8K |
| MRPC (Dolan and Brockett, 2005) | Microsoft Research Paraphrase Corpus | 3.7K | 1.7K |
| STS-B (Cer et al., 2017) | Semantic Textual Similarity Benchmark | 7K | 1.4K |
| QQP (Iyer et al., 2017) | Quora Question Pairs | 364K | 391K |
| MNLI (Williams et al., 2018) | Multi-Genre NLI | 393K | 20K |
| QNLI (Rajpurkar et al., 2016) | Question NLI | 105K | 5.4K |
| RTE (Dagan et al., 2005) | Recognition Textual Entailment | 2.5K | 3K |
| WNLI (Levesque et al., 2012) | Winograd NLI | 634 | 146 |
| Model | Total Params | Trained +params/ +task | QNLI | RTE | QQP | MNLI- +m | MNLI- +mm | SST-2 | MRPC | COLA | STS- +B | Avg |
| Single Task | ||||||||||||
| BERT[a] | x9 | 100% | 90.5 | 66.4 | 71.2 | 84.6 | 83.4 | 93.5 | 88.9 | 52.1 | 85.8 | 79.6 |
| Bitfit[b] | x1.008 | 0.09% | 89.7 | 65.5 | 67.8 | 80.8 | 80.9 | 92.4 | 87.4 | 47.2 | 87.6 | 77.7 |
| Multi-task | ||||||||||||
| BERT MTL[c] | x1 | 11.1% | 90.5 | 74.5 | 70.4 | 83.5 | 83.1 | 93.1 | 88.0 | 48.5 | 80.6 | 79.1 |
| BERT MTL[d] | x1 | 11.1% | 89.3 | 76.6 | 70.8 | 84.0 | 83.4 | 93.4 | 86.7 | 51.2 | 83.6 | 79.9 |
| PALs[d] | x1.13 | 12.5% | 90.0 | 76.0 | 71.5 | 84.3 | 83.5 | 92.6 | 88.7 | 51.2 | 85.8 | 80.4 |
| CA-MTL[d] | x1.12 | 5.6% | 90.5 | 76.4 | 69.2 | 85.9 | 85.8 | 93.2 | 88.6 | 53.1 | 85.3 | 80.9 |
| Our approach | ||||||||||||
| Morris† | x1.024 | 0.27% | 91.1 | 73.7 | 70.6 | 83.8 | 83.5 | 92.8 | 90.2 | 52.0 | 85.8 | 80.4 |
| Morris L0† | x1.014 | 0.16% | 91.6 | 73.7 | 70.5 | 84.1 | 83.1 | 92.1 | 89.6 | 49.9 | 86.3 | 80.1 |
| Model | Total Params | Trained params/ task | Avg |
| Single Task | |||
| LoRA | x1.024 | 0.27% | 80.7 |
| Lora_BitFit | x1.03 | 0.34% | 81.2 |
| Multi Task | |||
| LoRA_BitFit_MTL | x1.022 | 0.246% | 81.8 |
| Morris | x1.024 | 0.27% | 82.2 |
| Morris L0 | x1.014 | 0.16% | 82.4 |
| Tasks | Corpus | |Train| | |Test| |
| CoLA (Warstadt et al., 2018) | Corpus of Linguistic Acceptability | 8.5K | 1K |
| SST-2 (Socher et al., 2013) | Stanford Sentiment Treebank | 67K | 1.8K |
| MRPC (Dolan and Brockett, 2005) | Microsoft Research Paraphrase Corpus | 3.7K | 1.7K |
| STS-B (Cer et al., 2017) | Semantic Textual Similarity Benchmark | 7K | 1.4K |
| QQP (Iyer et al., 2017) | Quora Question Pairs | 364K | 391K |
| MNLI (Williams et al., 2018) | Multi-Genre NLI | 393K | 20K |
| QNLI (Rajpurkar et al., 2016) | Question NLI | 105K | 5.4K |
| RTE (Dagan et al., 2005) | Recognition Textual Entailment | 2.5K | 3K |
| WNLI (Levesque et al., 2012) | Winograd NLI | 634 | 146 |
| Model | Total Params | Trained \nparams/ \ntask | QNLI | RTE | QQP | MNLI- \nm | MNLI- \nmm | SST-2 | MRPC | COLA | STS- \nB | Avg |
| Single Task | ||||||||||||
| BERT[a] | x9 | 100% | 90.5 | 66.4 | 71.2 | 84.6 | 83.4 | 93.5 | 88.9 | 52.1 | 85.8 | 79.6 |
| Bitfit[b] | x1.008 | 0.09% | 89.7 | 65.5 | 67.8 | 80.8 | 80.9 | 92.4 | 87.4 | 47.2 | 87.6 | 77.7 |
| Multi-task | ||||||||||||
| BERT MTL[c] | x1 | 11.1% | 90.5 | 74.5 | 70.4 | 83.5 | 83.1 | 93.1 | 88.0 | 48.5 | 80.6 | 79.1 |
| BERT MTL[d] | x1 | 11.1% | 89.3 | 76.6 | 70.8 | 84.0 | 83.4 | 93.4 | 86.7 | 51.2 | 83.6 | 79.9 |
| PALs[d] | x1.13 | 12.5% | 90.0 | 76.0 | 71.5 | 84.3 | 83.5 | 92.6 | 88.7 | 51.2 | 85.8 | 80.4 |
| CA-MTL[d] | x1.12 | 5.6% | 90.5 | 76.4 | 69.2 | 85.9 | 85.8 | 93.2 | 88.6 | 53.1 | 85.3 | 80.9 |
| Our approach | ||||||||||||
| Morris† | x1.024 | 0.27% | 91.1 | 73.7 | 70.6 | 83.8 | 83.5 | 92.8 | 90.2 | 52.0 | 85.8 | 80.4 |
| Morris L0† | x1.014 | 0.16% | 91.6 | 73.7 | 70.5 | 84.1 | 83.1 | 92.1 | 89.6 | 49.9 | 86.3 | 80.1 |
| Model | Total Params | Trained params/ task | Avg |
| Single Task | |||
| LoRA | x1.024 | 0.27% | 80.7 |
| Lora_BitFit | x1.03 | 0.34% | 81.2 |
| Multi Task | |||
| LoRA_BitFit_MTL | x1.022 | 0.246% | 81.8 |
| Morris | x1.024 | 0.27% | 82.2 |
| Morris L0 | x1.014 | 0.16% | 82.4 |
| Tasks | Corpus | |Train| | |Test| |
| CoLA (Warstadt et al., 2018) | Corpus of Linguistic Acceptability | 8.5K | 1K |
| SST-2 (Socher et al., 2013) | Stanford Sentiment Treebank | 67K | 1.8K |
| MRPC (Dolan and Brockett, 2005) | Microsoft Research Paraphrase Corpus | 3.7K | 1.7K |
| STS-B (Cer et al., 2017) | Semantic Textual Similarity Benchmark | 7K | 1.4K |
| QQP (Iyer et al., 2017) | Quora Question Pairs | 364K | 391K |
| MNLI (Williams et al., 2018) | Multi-Genre NLI | 393K | 20K |
| QNLI (Rajpurkar et al., 2016) | Question NLI | 105K | 5.4K |
| RTE (Dagan et al., 2005) | Recognition Textual Entailment | 2.5K | 3K |
| WNLI (Levesque et al., 2012) | Winograd NLI | 634 | 146 |
| Error Type | Abbr | Definition | Example Sentence |
| Grammatical/Usage Error | GramErr | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall.\ngen: Two painting is on the wall. |
| Predicate Error | PredErr | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen.\ngen: John found the kitchen. |
| Entity Error | EntErr | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat.\ngen: A dog chased a rat. |
| Predicate Ordering Error | PredOrdErr | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison.\ngen: The suspect was taken to the prison then the police arrested him. |
| Hyponyms/Hypernyms Errors | HypErr | Violations in hybernym/hyponym usage. | ref: Jim studied mechanical engineering.\ngen: Jim studied architectural science. |
| Numerical Error | NumErr | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples.\ngen: Martha ate six apples. |
| Spatial/Temporal Error | STErr | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago.\ngen: Dave lives in south Chile. |
| Attribute/Modifier Error | AttrErr | Mistakes in additional information concerning the predicates and entities.\n(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs.\ngen: Greg has two big dogs. |
| Question Error | QuestErr | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet?\ngen: When did you take the dog to the vet? |
| Negation | NegErr | Negated compared to the reference text. | ref: Susan took the gift.\ngen: Susan did not take the gift. |
| Missing Information | MissInfo | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor.\ngen: Bob saw a doctor. |
| Out of Reference | OutofRef | Contains additional details not present in the reference text. | ref: Jack and Jane are friends.\ngen: Jack and Jane are friends. Jack plays football. |
| Redundant/Repetition | RepErr | Same/similar information repeated more than once. | ref: Tom met Sam at the party.\ngen: Tom went to the party. Tom met Sam at the party. |
| Tasks | AS | IC | QG | MT | DG | DT | NLI |
| # Samples (Task-Specific) | 1600 (SummEval) | 2007 (Flickr30k) | 2726 (AQG) | 240,287 (WMT2017-19) | 420 (PersonaChat) | 5,918 (WebNLG) | 9,818 (MNLI) |
| RMSE | 0.18 (0.00) | 0.23 (0.00) | 0.16 (0.00) | 0.19 (0.00) | 0.26 (0.00) | 0.24 (0.00) | * |
| Kendall's τ | 0.30 (0.01) | 0.49 (0.00) | 0.52 (0.01) | 0.35 (0.00) | 0.31 (0.02) | 0.39 (0.01) | 0.89 (0.00) |
| Spearman's ρ | 0.41 (0.01) | 0.62 (0.00) | 0.66 (0.01) | 0.50 (0.00) | 0.40 (0.03) | 0.49 (0.01) | 0.92 (0.00) |
| Error Type | Definition | Example Sentence | Human Evaluation Criteria |
| Grammatical/Usage Error | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall.\ngen: Two painting is on the wall. | Fluency |
| Predicate Error | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen.\ngen: John found the kitchen. | Answerability, Relevance, Making sense |
| Entity Error | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat.\ngen: A dog chased a rat. | Relevance, Correctness, Thoroughness, Informativeness, Referential clarity |
| Predicate Ordering Error | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison.\ngen: The suspect was taken to the prison then the police arrested him. | Flow/Coherence, Answerability, Making sense, repetitions |
| Hyponyms/Hypernyms Errors | Violations in hypernym/hyponym usage. | ref: Jim studied mechanical engineering.\ngen: Jim studied architectural science. | Informativeness |
| Numerical Error | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples.\ngen: Martha ate six apples. | Correctness, Thoroughness |
| Spatial/Temporal Error | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago.\ngen: Dave lives in south Chile. | Correctness, Thoroughness |
| Attribute/Modifier Error | Mistakes in additional information concerning the predicates and entities.\n(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs.\ngen: Greg has two big dogs. | Correctness, Thoroughness |
| Question Error | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet?\ngen: When did you take the dog to the vet? | Answerability |
| Negation | Negated compared to the reference text. | ref: Susan took the gift.\ngen: Susan did not take the gift. | Adequacy |
| Missing Information | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor.\ngen: Bob saw a doctor. | Adequacy, Thoroughness, Data Coverage |
| Out of Reference | Contains additional details not present in the reference text. | ref: Jack and Jane are friends.\ngen: Jack and Jane are friends. Jack plays football. | Adequacy, Making Sense, Listening |
| Redundant/Repetition | Same/similar information repeated more than once. | ref: Tom met Sam at the party.\ngen: Tom went to the party. Tom met Sam at the party. | Thoroughness, Avoid Repetition, Data coverage |
| Models/Languages | cs-en | de-en | et-en | fi-en | ru-en | tr-en | zh-en | avg |
| BLEURT | 35.6 | 44.2 | 40.0 | 32.1 | 31.9 | 35.5 | 29.7 | 35.6 |
| Our Model | 36.0 | 44.7 | 40.6 | 33.3 | 32.5 | 36.0 | 30.4 | 36.9 |
| Models/Languages | de-en | fi-en | gu-en | kk-en | lt-en | ru-en | zh-en | avg |
| BLEURT | 31.2 | 31.7 | 28.3 | 39.5 | 35.2 | 28.3 | 42.7 | 33.8 |
| Our Model | 31.6 | 32.3 | 28.1 | 40.5 | 35.4 | 28.3 | 45.8 | 33.9 |
| Reference Text | Model Generated Text | Correctness | Thoroughness | Predicted Correctness | Predicted Thoroughness | Predicted Mismatch |
| a mounted police officer riding down a city street past parked cars | a man riding a horse on a city street | 1 | 1 | 0.82 | 0.74 | Error in Hyponyms or Hyper-nyms |
| a couple of people standing in a field playing with a frisbee. | a man standing on top of a sandy beach | 0.2 | 0.2 | 0.26 | 0.28 | Out of Reference |
| a bathroom with a reflection of a television and a sink. | a bathroom with a sink and a mirror | 1 | 0.8 | 0.78 | 0.72 | Missing Information |
| Reference Text | Model Generated Text | Answerability | Predicted Answerability | Predicted Mismatch |
| How do co-teachers work with each other to fulfill the needs of students? | When do co-teachers work with each other to fulfill the needs of students? | 1 | 0.89 | Question Error |
| David Lean was the director on which movies ? | was the director on which movies ? | 0.2 | 0.39 | Missing Information |
| Cate Shortland was the director on which movies ? | Cate Shortland director which movies ? | 1 | 0.90 | Grammatical & Usage |
| Reference Text | Model Generated Text | Label | Prediction | Predicted Mismatch |
| We got him out of that 10-20 range. | Got him just out of 9,000-9,500 now. | Contradiction | Contradiction | Numerical or Quantifiers |
| How about when you were in school? | How about when you went to bed? | Contradiction | Contradiction | Entity Error |
| Reference Text | Model Generated Text | Fluency | Grammar | Semantics | Predicted Fluency | Predicted Grammar | Predicted Semantics | Predicted Mismatch |
| alan b. miller hall was started on march 30, 2007 and has the mason school of business in the u . s . as a tenant | the current tenants of alan b. miller hall are 30 march 2007 and mason school of business in united states | 0.33 | 0.33 | 0.33 | 0.61 | 0.65 | 0.64 | Grammatical & Usage |
| albuquerque , new mexico is located in the united states and asian americans are an ethnic group there . john sanchez , is one of the leaders , in the new mexico senate which is leading the state | albuquerque, new mexico is a food from new mexico where the capital is asian americans and is led by John sanchez. | 0.66 | 0.66 | 0.33 | 0.66 | 0.72 | 0.48 | Out of Reference |
| Metric | Task | Reference |
| ANLI | Natural Language Inference | (Nie et al., 2020) |
| factcc | Abstractive Summarization | (KrySciński et al., 2020) |
| Q2 | Knowledge-grounded Dialogue | (Honovich et al., 2021) |
| QuestEval | Abstractive Summarization | (Scialom et al., 2021) |
| summaC (Conv/CZS) | Abstractive Summarization | (Laban et al., 2022) |
| summaqa (avg_prob/avg_fscore) | Abstractive Summarization | (Scialom et al., 2019) |
| S3 (pyr/resp) | Abstractive Summarization | (Fabbri et al., 2021) |
| ROUGE | Summarization | (Lin, 2004) |
| ROUGE-WE | Abstractive Summarization | (Ng and Abrecht, 2015) |
| BLEURT | Machine Translation | (Sellam et al., 2020) |
| BARTScore | Text Generation | (Yuan et al., 2021) |
| Blanc | Machine Translation | (Lita et al., 2005) |
| Bleu | Machine Translation | (Papineni et al., 2002) |
| SUPERT | Multi-document summarization | (Gao et al., 2020) |
| chrf | Machine Translation | (Popović, 2015) |
| chrf++ | Machine Translation | (Popović, 2017) |
| Cider | Image Description Evaluation | (Vedantam et al., 2015) |
| Mauve | Text Generation | (Pillutla et al., 2021) |
| METEOR | Machine Translation | (Banerjee and Lavie, 2005) |
| RMR (1/2) | Abstractive Summarization | (Zhu et al., 2021) |
| sms/wms/s+wms | Distance between documents | (Kusner et al., 2015) |
| coverage/density | Text summarization | (Grusky et al., 2018) |
| Eval Metrics | Time Complexity (sec) | Space Complexity (MegaBytes) |
| ANLI | 426 | 2163.88 |
| BARTScore | 94 | 1883.47 |
| BERTScore | 66 | 2107.17 |
| Blanc | 7048 | 3081.04 |
| BLEU | 47 | 169.12 |
| BLEURT | 869 | 1949.05 |
| CHRF | 80 | 161.43 |
| CIDER | 14 | 238.25 |
| Datastats (n-gram, etc) | 81 | 399.64 |
| FactCC | 165 | 2197.16 |
| MAUVE | 16191 | 3380.99 |
| METEOR | 65 | 249.09 |
| Q2 | 18790 | 6883.66 |
| QuestEval | 7996 | 5493.33 |
| RMR | 2012 | 167.24 |
| ROUGE | 1542 | 181.41 |
| ROUGE_we1 | 32830 | 1219.65 |
| ROUGE_we2 | 34218 | 1207.79 |
| S3 (pyr/resp) | 458 | 10470.42 |
| SMS | 20972 | 289.70 |
| SummaC_conv | 287 | 2323.56 |
| SummaC_zs | 284 | 2125.90 |
| SummaQA | 923 | 4830.03 |
| SUPERT | 532 | 2237.44 |
| Error Type | Abbr | Definition | Example Sentence |
| Grammatical/Usage Error | GramErr | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall.\ngen: Two painting is on the wall. |
| Predicate Error | PredErr | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen.\ngen: John found the kitchen. |
| Entity Error | EntErr | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat.\ngen: A dog chased a rat. |
| Predicate Ordering Error | PredOrdErr | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison.\ngen: The suspect was taken to the prison then the police arrested him. |
| Hyponyms/Hypernyms Errors | HypErr | Violations in hybernym/hyponym usage. | ref: Jim studied mechanical engineering.\ngen: Jim studied architectural science. |
| Numerical Error | NumErr | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples.\ngen: Martha ate six apples. |
| Spatial/Temporal Error | STErr | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago.\ngen: Dave lives in south Chile. |
| Attribute/Modifier Error | AttrErr | Mistakes in additional information concerning the predicates and entities.\n(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs.\ngen: Greg has two big dogs. |
| Question Error | QuestErr | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet?\ngen: When did you take the dog to the vet? |
| Negation | NegErr | Negated compared to the reference text. | ref: Susan took the gift.\ngen: Susan did not take the gift. |
| Missing Information | MissInfo | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor.\ngen: Bob saw a doctor. |
| Out of Reference | OutofRef | Contains additional details not present in the reference text. | ref: Jack and Jane are friends.\ngen: Jack and Jane are friends. Jack plays football. |
| Redundant/Repetition | RepErr | Same/similar information repeated more than once. | ref: Tom met Sam at the party.\ngen: Tom went to the party. Tom met Sam at the party. |
| Tasks | AS | IC | QG | MT | DG | DT | NLI |
| # Samples (Task-Specific) | 1600 (SummEval) | 2007 (Flickr30k) | 2726 (AQG) | 240,287 (WMT2017-19) | 420 (PersonaChat) | 5,918 (WebNLG) | 9,818 (MNLI) |
| RMSE | 0.18 (0.00) | 0.23 (0.00) | 0.16 (0.00) | 0.19 (0.00) | 0.26 (0.00) | 0.24 (0.00) | * |
| Kendall's τ | 0.30 (0.01) | 0.49 (0.00) | 0.52 (0.01) | 0.35 (0.00) | 0.31 (0.02) | 0.39 (0.01) | 0.89 (0.00) |
| Spearman's ρ | 0.41 (0.01) | 0.62 (0.00) | 0.66 (0.01) | 0.50 (0.00) | 0.40 (0.03) | 0.49 (0.01) | 0.92 (0.00) |
| Error Type | Definition | Example Sentence | Human Evaluation Criteria |
| Grammatical/Usage Error | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall.\ngen: Two painting is on the wall. | Fluency |
| Predicate Error | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen.\ngen: John found the kitchen. | Answerability, Relevance, Making sense |
| Entity Error | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat.\ngen: A dog chased a rat. | Relevance, Correctness, Thoroughness, Informativeness, Referential clarity |
| Predicate Ordering Error | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison.\ngen: The suspect was taken to the prison then the police arrested him. | Flow/Coherence, Answerability, Making sense, repetitions |
| Hyponyms/Hypernyms Errors | Violations in hypernym/hyponym usage. | ref: Jim studied mechanical engineering.\ngen: Jim studied architectural science. | Informativeness |
| Numerical Error | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples.\ngen: Martha ate six apples. | Correctness, Thoroughness |
| Spatial/Temporal Error | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago.\ngen: Dave lives in south Chile. | Correctness, Thoroughness |
| Attribute/Modifier Error | Mistakes in additional information concerning the predicates and entities.\n(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs.\ngen: Greg has two big dogs. | Correctness, Thoroughness |
| Question Error | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet?\ngen: When did you take the dog to the vet? | Answerability |
| Negation | Negated compared to the reference text. | ref: Susan took the gift.\ngen: Susan did not take the gift. | Adequacy |
| Missing Information | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor.\ngen: Bob saw a doctor. | Adequacy, Thoroughness, Data Coverage |
| Out of Reference | Contains additional details not present in the reference text. | ref: Jack and Jane are friends.\ngen: Jack and Jane are friends. Jack plays football. | Adequacy, Making Sense, Listening |
| Redundant/Repetition | Same/similar information repeated more than once. | ref: Tom met Sam at the party.\ngen: Tom went to the party. Tom met Sam at the party. | Thoroughness, Avoid Repetition, Data coverage |
| Models/Languages | cs-en | de-en | et-en | fi-en | ru-en | tr-en | zh-en | avg |
| BLEURT | 35.6 | 44.2 | 40.0 | 32.1 | 31.9 | 35.5 | 29.7 | 35.6 |
| Our Model | 36.0 | 44.7 | 40.6 | 33.3 | 32.5 | 36.0 | 30.4 | 36.9 |
| Models/Languages | de-en | fi-en | gu-en | kk-en | lt-en | ru-en | zh-en | avg |
| BLEURT | 31.2 | 31.7 | 28.3 | 39.5 | 35.2 | 28.3 | 42.7 | 33.8 |
| Our Model | 31.6 | 32.3 | 28.1 | 40.5 | 35.4 | 28.3 | 45.8 | 33.9 |
| Reference Text | Model Generated Text | Correctness | Thoroughness | Predicted Correctness | Predicted Thoroughness | Predicted Mismatch |
| a mounted police officer riding down a city street past parked cars | a man riding a horse on a city street | 1 | 1 | 0.82 | 0.74 | Error in Hyponyms or Hyper-nyms |
| a couple of people standing in a field playing with a frisbee. | a man standing on top of a sandy beach | 0.2 | 0.2 | 0.26 | 0.28 | Out of Reference |
| a bathroom with a reflection of a television and a sink. | a bathroom with a sink and a mirror | 1 | 0.8 | 0.78 | 0.72 | Missing Information |
| Reference Text | Model Generated Text | Answerability | Predicted Answerability | Predicted Mismatch |
| How do co-teachers work with each other to fulfill the needs of students? | When do co-teachers work with each other to fulfill the needs of students? | 1 | 0.89 | Question Error |
| David Lean was the director on which movies ? | was the director on which movies ? | 0.2 | 0.39 | Missing Information |
| Cate Shortland was the director on which movies ? | Cate Shortland director which movies ? | 1 | 0.90 | Grammatical & Usage |
| Reference Text | Model Generated Text | Label | Prediction | Predicted Mismatch |
| We got him out of that 10-20 range. | Got him just out of 9,000-9,500 now. | Contradiction | Contradiction | Numerical or Quantifiers |
| How about when you were in school? | How about when you went to bed? | Contradiction | Contradiction | Entity Error |
| Reference Text | Model Generated Text | Fluency | Grammar | Semantics | Predicted Fluency | Predicted Grammar | Predicted Semantics | Predicted Mismatch |
| alan b. miller hall was started on march 30, 2007 and has the mason school of business in the u . s . as a tenant | the current tenants of alan b. miller hall are 30 march 2007 and mason school of business in united states | 0.33 | 0.33 | 0.33 | 0.61 | 0.65 | 0.64 | Grammatical & Usage |
| albuquerque , new mexico is located in the united states and asian americans are an ethnic group there . john sanchez , is one of the leaders , in the new mexico senate which is leading the state | albuquerque, new mexico is a food from new mexico where the capital is asian americans and is led by John sanchez. | 0.66 | 0.66 | 0.33 | 0.66 | 0.72 | 0.48 | Out of Reference |
| Metric | Task | Reference |
| ANLI | Natural Language Inference | (Nie et al., 2020) |
| factcc | Abstractive Summarization | (KrySciński et al., 2020) |
| Q2 | Knowledge-grounded Dialogue | (Honovich et al., 2021) |
| QuestEval | Abstractive Summarization | (Scialom et al., 2021) |
| summaC (Conv/CZS) | Abstractive Summarization | (Laban et al., 2022) |
| summaqa (avg_prob/avg_fscore) | Abstractive Summarization | (Scialom et al., 2019) |
| S3 (pyr/resp) | Abstractive Summarization | (Fabbri et al., 2021) |
| ROUGE | Summarization | (Lin, 2004) |
| ROUGE-WE | Abstractive Summarization | (Ng and Abrecht, 2015) |
| BLEURT | Machine Translation | (Sellam et al., 2020) |
| BARTScore | Text Generation | (Yuan et al., 2021) |
| Blanc | Machine Translation | (Lita et al., 2005) |
| Bleu | Machine Translation | (Papineni et al., 2002) |
| SUPERT | Multi-document summarization | (Gao et al., 2020) |
| chrf | Machine Translation | (Popović, 2015) |
| chrf++ | Machine Translation | (Popović, 2017) |
| Cider | Image Description Evaluation | (Vedantam et al., 2015) |
| Mauve | Text Generation | (Pillutla et al., 2021) |
| METEOR | Machine Translation | (Banerjee and Lavie, 2005) |
| RMR (1/2) | Abstractive Summarization | (Zhu et al., 2021) |
| sms/wms/s+wms | Distance between documents | (Kusner et al., 2015) |
| coverage/density | Text summarization | (Grusky et al., 2018) |
| Eval Metrics | Time Complexity (sec) | Space Complexity (MegaBytes) |
| ANLI | 426 | 2163.88 |
| BARTScore | 94 | 1883.47 |
| BERTScore | 66 | 2107.17 |
| Blanc | 7048 | 3081.04 |
| BLEU | 47 | 169.12 |
| BLEURT | 869 | 1949.05 |
| CHRF | 80 | 161.43 |
| CIDER | 14 | 238.25 |
| Datastats (n-gram, etc) | 81 | 399.64 |
| FactCC | 165 | 2197.16 |
| MAUVE | 16191 | 3380.99 |
| METEOR | 65 | 249.09 |
| Q2 | 18790 | 6883.66 |
| QuestEval | 7996 | 5493.33 |
| RMR | 2012 | 167.24 |
| ROUGE | 1542 | 181.41 |
| ROUGE_we1 | 32830 | 1219.65 |
| ROUGE_we2 | 34218 | 1207.79 |
| S3 (pyr/resp) | 458 | 10470.42 |
| SMS | 20972 | 289.70 |
| SummaC_conv | 287 | 2323.56 |
| SummaC_zs | 284 | 2125.90 |
| SummaQA | 923 | 4830.03 |
| SUPERT | 532 | 2237.44 |
| Error Type | Abbr | Definition | Example Sentence |
| Grammatical/Usage Error | GramErr | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall. +gen: Two painting is on the wall. |
| Predicate Error | PredErr | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen. +gen: John found the kitchen. |
| Entity Error | EntErr | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat. +gen: A dog chased a rat. |
| Predicate Ordering Error | PredOrdErr | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison. +gen: The suspect was taken to the prison then the police arrested him. |
| Hyponyms/Hypernyms Errors | HypErr | Violations in hybernym/hyponym usage. | ref: Jim studied mechanical engineering. +gen: Jim studied architectural science. |
| Numerical Error | NumErr | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples. +gen: Martha ate six apples. |
| Spatial/Temporal Error | STErr | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago. +gen: Dave lives in south Chile. |
| Attribute/Modifier Error | AttrErr | Mistakes in additional information concerning the predicates and entities. +(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs. +gen: Greg has two big dogs. |
| Question Error | QuestErr | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet? +gen: When did you take the dog to the vet? |
| Negation | NegErr | Negated compared to the reference text. | ref: Susan took the gift. +gen: Susan did not take the gift. |
| Missing Information | MissInfo | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor. +gen: Bob saw a doctor. |
| Out of Reference | OutofRef | Contains additional details not present in the reference text. | ref: Jack and Jane are friends. +gen: Jack and Jane are friends. Jack plays football. |
| Redundant/Repetition | RepErr | Same/similar information repeated more than once. | ref: Tom met Sam at the party. +gen: Tom went to the party. Tom met Sam at the party. |
| Tasks | AS | IC | QG | MT | DG | DT | NLI |
| # Samples (Task-Specific) | 1600 (SummEval) | 2007 (Flickr30k) | 2726 (AQG) | 240,287 (WMT2017-19) | 420 (PersonaChat) | 5,918 (WebNLG) | 9,818 (MNLI) |
| RMSE | 0.18 (0.00) | 0.23 (0.00) | 0.16 (0.00) | 0.19 (0.00) | 0.26 (0.00) | 0.24 (0.00) | * |
| Kendall's τ | 0.30 (0.01) | 0.49 (0.00) | 0.52 (0.01) | 0.35 (0.00) | 0.31 (0.02) | 0.39 (0.01) | 0.89 (0.00) |
| Spearman's ρ | 0.41 (0.01) | 0.62 (0.00) | 0.66 (0.01) | 0.50 (0.00) | 0.40 (0.03) | 0.49 (0.01) | 0.92 (0.00) |
| Error Type | Definition | Example Sentence | Human Evaluation Criteria |
| Grammatical/Usage Error | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall. +gen: Two painting is on the wall. | Fluency |
| Predicate Error | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen. +gen: John found the kitchen. | Answerability, Relevance, Making sense |
| Entity Error | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat. +gen: A dog chased a rat. | Relevance, Correctness, Thoroughness, Informativeness, Referential clarity |
| Predicate Ordering Error | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison. +gen: The suspect was taken to the prison then the police arrested him. | Flow/Coherence, Answerability, Making sense, repetitions |
| Hyponyms/Hypernyms Errors | Violations in hypernym/hyponym usage. | ref: Jim studied mechanical engineering. +gen: Jim studied architectural science. | Informativeness |
| Numerical Error | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples. +gen: Martha ate six apples. | Correctness, Thoroughness |
| Spatial/Temporal Error | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago. +gen: Dave lives in south Chile. | Correctness, Thoroughness |
| Attribute/Modifier Error | Mistakes in additional information concerning the predicates and entities. +(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs. +gen: Greg has two big dogs. | Correctness, Thoroughness |
| Question Error | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet? +gen: When did you take the dog to the vet? | Answerability |
| Negation | Negated compared to the reference text. | ref: Susan took the gift. +gen: Susan did not take the gift. | Adequacy |
| Missing Information | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor. +gen: Bob saw a doctor. | Adequacy, Thoroughness, Data Coverage |
| Out of Reference | Contains additional details not present in the reference text. | ref: Jack and Jane are friends. +gen: Jack and Jane are friends. Jack plays football. | Adequacy, Making Sense, Listening |
| Redundant/Repetition | Same/similar information repeated more than once. | ref: Tom met Sam at the party. +gen: Tom went to the party. Tom met Sam at the party. | Thoroughness, Avoid Repetition, Data coverage |
| Models/Languages | cs-en | de-en | et-en | fi-en | ru-en | tr-en | zh-en | avg |
| BLEURT | 35.6 | 44.2 | 40.0 | 32.1 | 31.9 | 35.5 | 29.7 | 35.6 |
| Our Model | 36.0 | 44.7 | 40.6 | 33.3 | 32.5 | 36.0 | 30.4 | 36.9 |
| Models/Languages | de-en | fi-en | gu-en | kk-en | lt-en | ru-en | zh-en | avg |
| BLEURT | 31.2 | 31.7 | 28.3 | 39.5 | 35.2 | 28.3 | 42.7 | 33.8 |
| Our Model | 31.6 | 32.3 | 28.1 | 40.5 | 35.4 | 28.3 | 45.8 | 33.9 |
| Reference Text | Model Generated Text | Correctness | Thoroughness | Predicted Correctness | Predicted Thoroughness | Predicted Mismatch |
| a mounted police officer riding down a city street past parked cars | a man riding a horse on a city street | 1 | 1 | 0.82 | 0.74 | Error in Hyponyms or Hyper-nyms |
| a couple of people standing in a field playing with a frisbee. | a man standing on top of a sandy beach | 0.2 | 0.2 | 0.26 | 0.28 | Out of Reference |
| a bathroom with a reflection of a television and a sink. | a bathroom with a sink and a mirror | 1 | 0.8 | 0.78 | 0.72 | Missing Information |
| Reference Text | Model Generated Text | Answerability | Predicted Answerability | Predicted Mismatch |
| How do co-teachers work with each other to fulfill the needs of students? | When do co-teachers work with each other to fulfill the needs of students? | 1 | 0.89 | Question Error |
| David Lean was the director on which movies ? | was the director on which movies ? | 0.2 | 0.39 | Missing Information |
| Cate Shortland was the director on which movies ? | Cate Shortland director which movies ? | 1 | 0.90 | Grammatical & Usage |
| Reference Text | Model Generated Text | Label | Prediction | Predicted Mismatch |
| We got him out of that 10-20 range. | Got him just out of 9,000-9,500 now. | Contradiction | Contradiction | Numerical or Quantifiers |
| How about when you were in school? | How about when you went to bed? | Contradiction | Contradiction | Entity Error |
| Reference Text | Model Generated Text | Fluency | Grammar | Semantics | Predicted Fluency | Predicted Grammar | Predicted Semantics | Predicted Mismatch |
| alan b. miller hall was started on march 30, 2007 and has the mason school of business in the u . s . as a tenant | the current tenants of alan b. miller hall are 30 march 2007 and mason school of business in united states | 0.33 | 0.33 | 0.33 | 0.61 | 0.65 | 0.64 | Grammatical & Usage |
| albuquerque , new mexico is located in the united states and asian americans are an ethnic group there . john sanchez , is one of the leaders , in the new mexico senate which is leading the state | albuquerque, new mexico is a food from new mexico where the capital is asian americans and is led by John sanchez. | 0.66 | 0.66 | 0.33 | 0.66 | 0.72 | 0.48 | Out of Reference |
| Metric | Task | Reference |
| ANLI | Natural Language Inference | (Nie et al., 2020) |
| factcc | Abstractive Summarization | (KrySciński et al., 2020) |
| Q2 | Knowledge-grounded Dialogue | (Honovich et al., 2021) |
| QuestEval | Abstractive Summarization | (Scialom et al., 2021) |
| summaC (Conv/CZS) | Abstractive Summarization | (Laban et al., 2022) |
| summaqa (avg_prob/avg_fscore) | Abstractive Summarization | (Scialom et al., 2019) |
| S3 (pyr/resp) | Abstractive Summarization | (Fabbri et al., 2021) |
| ROUGE | Summarization | (Lin, 2004) |
| ROUGE-WE | Abstractive Summarization | (Ng and Abrecht, 2015) |
| BLEURT | Machine Translation | (Sellam et al., 2020) |
| BARTScore | Text Generation | (Yuan et al., 2021) |
| Blanc | Machine Translation | (Lita et al., 2005) |
| Bleu | Machine Translation | (Papineni et al., 2002) |
| SUPERT | Multi-document summarization | (Gao et al., 2020) |
| chrf | Machine Translation | (Popović, 2015) |
| chrf++ | Machine Translation | (Popović, 2017) |
| Cider | Image Description Evaluation | (Vedantam et al., 2015) |
| Mauve | Text Generation | (Pillutla et al., 2021) |
| METEOR | Machine Translation | (Banerjee and Lavie, 2005) |
| RMR (1/2) | Abstractive Summarization | (Zhu et al., 2021) |
| sms/wms/s+wms | Distance between documents | (Kusner et al., 2015) |
| coverage/density | Text summarization | (Grusky et al., 2018) |
| Eval Metrics | Time Complexity (sec) | Space Complexity (MegaBytes) |
| ANLI | 426 | 2163.88 |
| BARTScore | 94 | 1883.47 |
| BERTScore | 66 | 2107.17 |
| Blanc | 7048 | 3081.04 |
| BLEU | 47 | 169.12 |
| BLEURT | 869 | 1949.05 |
| CHRF | 80 | 161.43 |
| CIDER | 14 | 238.25 |
| Datastats (n-gram, etc) | 81 | 399.64 |
| FactCC | 165 | 2197.16 |
| MAUVE | 16191 | 3380.99 |
| METEOR | 65 | 249.09 |
| Q2 | 18790 | 6883.66 |
| QuestEval | 7996 | 5493.33 |
| RMR | 2012 | 167.24 |
| ROUGE | 1542 | 181.41 |
| ROUGE_we1 | 32830 | 1219.65 |
| ROUGE_we2 | 34218 | 1207.79 |
| S3 (pyr/resp) | 458 | 10470.42 |
| SMS | 20972 | 289.70 |
| SummaC_conv | 287 | 2323.56 |
| SummaC_zs | 284 | 2125.90 |
| SummaQA | 923 | 4830.03 |
| SUPERT | 532 | 2237.44 |
| Error Type | Abbr | Definition | Example Sentence |
| Grammatical/Usage Error | GramErr | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall.\ngen: Two painting is on the wall. |
| Predicate Error | PredErr | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen.\ngen: John found the kitchen. |
| Entity Error | EntErr | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat.\ngen: A dog chased a rat. |
| Predicate Ordering Error | PredOrdErr | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison.\ngen: The suspect was taken to the prison then the police arrested him. |
| Hyponyms/Hypernyms Errors | HypErr | Violations in hybernym/hyponym usage. | ref: Jim studied mechanical engineering.\ngen: Jim studied architectural science. |
| Numerical Error | NumErr | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples.\ngen: Martha ate six apples. |
| Spatial/Temporal Error | STErr | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago.\ngen: Dave lives in south Chile. |
| Attribute/Modifier Error | AttrErr | Mistakes in additional information concerning the predicates and entities.\n(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs.\ngen: Greg has two big dogs. |
| Question Error | QuestErr | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet?\ngen: When did you take the dog to the vet? |
| Negation | NegErr | Negated compared to the reference text. | ref: Susan took the gift.\ngen: Susan did not take the gift. |
| Missing Information | MissInfo | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor.\ngen: Bob saw a doctor. |
| Out of Reference | OutofRef | Contains additional details not present in the reference text. | ref: Jack and Jane are friends.\ngen: Jack and Jane are friends. Jack plays football. |
| Redundant/Repetition | RepErr | Same/similar information repeated more than once. | ref: Tom met Sam at the party.\ngen: Tom went to the party. Tom met Sam at the party. |
| Tasks | AS | IC | QG | MT | DG | DT | NLI |
| # Samples (Task-Specific) | 1600 (SummEval) | 2007 (Flickr30k) | 2726 (AQG) | 240,287 (WMT2017-19) | 420 (PersonaChat) | 5,918 (WebNLG) | 9,818 (MNLI) |
| RMSE | 0.18 (0.00) | 0.23 (0.00) | 0.16 (0.00) | 0.19 (0.00) | 0.26 (0.00) | 0.24 (0.00) | * |
| Kendall's τ | 0.30 (0.01) | 0.49 (0.00) | 0.52 (0.01) | 0.35 (0.00) | 0.31 (0.02) | 0.39 (0.01) | 0.89 (0.00) |
| Spearman's ρ | 0.41 (0.01) | 0.62 (0.00) | 0.66 (0.01) | 0.50 (0.00) | 0.40 (0.03) | 0.49 (0.01) | 0.92 (0.00) |
| Error Type | Definition | Example Sentence | Human Evaluation Criteria |
| Grammatical/Usage Error | Faulty or incorrect use of the grammar and syntax. | ref: Two paintings are on the wall.\ngen: Two painting is on the wall. | Fluency |
| Predicate Error | Error in the predicate or its usage with respect to the reference text. | ref: John entered the kitchen.\ngen: John found the kitchen. | Answerability, Relevance, Making sense |
| Entity Error | Mismatch in the primary arguments of the predicate. | ref: A dog chased a cat.\ngen: A dog chased a rat. | Relevance, Correctness, Thoroughness, Informativeness, Referential clarity |
| Predicate Ordering Error | Error in causal or temporal ordering of the predicates/events. | ref: The police arrested the suspect then he was taken to prison.\ngen: The suspect was taken to the prison then the police arrested him. | Flow/Coherence, Answerability, Making sense, repetitions |
| Hyponyms/Hypernyms Errors | Violations in hypernym/hyponym usage. | ref: Jim studied mechanical engineering.\ngen: Jim studied architectural science. | Informativeness |
| Numerical Error | Error in numerical, quantifiers or related to numbers (ordinals, cardinals, etc) | ref: Martha ate four apples.\ngen: Martha ate six apples. | Correctness, Thoroughness |
| Spatial/Temporal Error | Error in spatial or geographic information (location, time, etc). | ref: Dave lives in south Chicago.\ngen: Dave lives in south Chile. | Correctness, Thoroughness |
| Attribute/Modifier Error | Mistakes in additional information concerning the predicates and entities.\n(not covered by numerical, spatial, geographic) | ref: Greg has two small dogs.\ngen: Greg has two big dogs. | Correctness, Thoroughness |
| Question Error | Error/change in the nature of the question's intention. | ref: Did you take the dog to the vet?\ngen: When did you take the dog to the vet? | Answerability |
| Negation | Negated compared to the reference text. | ref: Susan took the gift.\ngen: Susan did not take the gift. | Adequacy |
| Missing Information | Missing key details from the reference text. | ref: Bob drove to the hospital and saw a doctor.\ngen: Bob saw a doctor. | Adequacy, Thoroughness, Data Coverage |
| Out of Reference | Contains additional details not present in the reference text. | ref: Jack and Jane are friends.\ngen: Jack and Jane are friends. Jack plays football. | Adequacy, Making Sense, Listening |
| Redundant/Repetition | Same/similar information repeated more than once. | ref: Tom met Sam at the party.\ngen: Tom went to the party. Tom met Sam at the party. | Thoroughness, Avoid Repetition, Data coverage |
| Models/Languages | cs-en | de-en | et-en | fi-en | ru-en | tr-en | zh-en | avg |
| BLEURT | 35.6 | 44.2 | 40.0 | 32.1 | 31.9 | 35.5 | 29.7 | 35.6 |
| Our Model | 36.0 | 44.7 | 40.6 | 33.3 | 32.5 | 36.0 | 30.4 | 36.9 |
| Models/Languages | de-en | fi-en | gu-en | kk-en | lt-en | ru-en | zh-en | avg |
| BLEURT | 31.2 | 31.7 | 28.3 | 39.5 | 35.2 | 28.3 | 42.7 | 33.8 |
| Our Model | 31.6 | 32.3 | 28.1 | 40.5 | 35.4 | 28.3 | 45.8 | 33.9 |
| Reference Text | Model Generated Text | Correctness | Thoroughness | Predicted Correctness | Predicted Thoroughness | Predicted Mismatch |
| a mounted police officer riding down a city street past parked cars | a man riding a horse on a city street | 1 | 1 | 0.82 | 0.74 | Error in Hyponyms or Hyper-nyms |
| a couple of people standing in a field playing with a frisbee. | a man standing on top of a sandy beach | 0.2 | 0.2 | 0.26 | 0.28 | Out of Reference |
| a bathroom with a reflection of a television and a sink. | a bathroom with a sink and a mirror | 1 | 0.8 | 0.78 | 0.72 | Missing Information |
| Reference Text | Model Generated Text | Answerability | Predicted Answerability | Predicted Mismatch |
| How do co-teachers work with each other to fulfill the needs of students? | When do co-teachers work with each other to fulfill the needs of students? | 1 | 0.89 | Question Error |
| David Lean was the director on which movies ? | was the director on which movies ? | 0.2 | 0.39 | Missing Information |
| Cate Shortland was the director on which movies ? | Cate Shortland director which movies ? | 1 | 0.90 | Grammatical & Usage |
| Reference Text | Model Generated Text | Label | Prediction | Predicted Mismatch |
| We got him out of that 10-20 range. | Got him just out of 9,000-9,500 now. | Contradiction | Contradiction | Numerical or Quantifiers |
| How about when you were in school? | How about when you went to bed? | Contradiction | Contradiction | Entity Error |
| Reference Text | Model Generated Text | Fluency | Grammar | Semantics | Predicted Fluency | Predicted Grammar | Predicted Semantics | Predicted Mismatch |
| alan b. miller hall was started on march 30, 2007 and has the mason school of business in the u . s . as a tenant | the current tenants of alan b. miller hall are 30 march 2007 and mason school of business in united states | 0.33 | 0.33 | 0.33 | 0.61 | 0.65 | 0.64 | Grammatical & Usage |
| albuquerque , new mexico is located in the united states and asian americans are an ethnic group there . john sanchez , is one of the leaders , in the new mexico senate which is leading the state | albuquerque, new mexico is a food from new mexico where the capital is asian americans and is led by John sanchez. | 0.66 | 0.66 | 0.33 | 0.66 | 0.72 | 0.48 | Out of Reference |
| Metric | Task | Reference |
| ANLI | Natural Language Inference | (Nie et al., 2020) |
| factcc | Abstractive Summarization | (KrySciński et al., 2020) |
| Q2 | Knowledge-grounded Dialogue | (Honovich et al., 2021) |
| QuestEval | Abstractive Summarization | (Scialom et al., 2021) |
| summaC (Conv/CZS) | Abstractive Summarization | (Laban et al., 2022) |
| summaqa (avg_prob/avg_fscore) | Abstractive Summarization | (Scialom et al., 2019) |
| S3 (pyr/resp) | Abstractive Summarization | (Fabbri et al., 2021) |
| ROUGE | Summarization | (Lin, 2004) |
| ROUGE-WE | Abstractive Summarization | (Ng and Abrecht, 2015) |
| BLEURT | Machine Translation | (Sellam et al., 2020) |
| BARTScore | Text Generation | (Yuan et al., 2021) |
| Blanc | Machine Translation | (Lita et al., 2005) |
| Bleu | Machine Translation | (Papineni et al., 2002) |
| SUPERT | Multi-document summarization | (Gao et al., 2020) |
| chrf | Machine Translation | (Popović, 2015) |
| chrf++ | Machine Translation | (Popović, 2017) |
| Cider | Image Description Evaluation | (Vedantam et al., 2015) |
| Mauve | Text Generation | (Pillutla et al., 2021) |
| METEOR | Machine Translation | (Banerjee and Lavie, 2005) |
| RMR (1/2) | Abstractive Summarization | (Zhu et al., 2021) |
| sms/wms/s+wms | Distance between documents | (Kusner et al., 2015) |
| coverage/density | Text summarization | (Grusky et al., 2018) |
| Eval Metrics | Time Complexity (sec) | Space Complexity (MegaBytes) |
| ANLI | 426 | 2163.88 |
| BARTScore | 94 | 1883.47 |
| BERTScore | 66 | 2107.17 |
| Blanc | 7048 | 3081.04 |
| BLEU | 47 | 169.12 |
| BLEURT | 869 | 1949.05 |
| CHRF | 80 | 161.43 |
| CIDER | 14 | 238.25 |
| Datastats (n-gram, etc) | 81 | 399.64 |
| FactCC | 165 | 2197.16 |
| MAUVE | 16191 | 3380.99 |
| METEOR | 65 | 249.09 |
| Q2 | 18790 | 6883.66 |
| QuestEval | 7996 | 5493.33 |
| RMR | 2012 | 167.24 |
| ROUGE | 1542 | 181.41 |
| ROUGE_we1 | 32830 | 1219.65 |
| ROUGE_we2 | 34218 | 1207.79 |
| S3 (pyr/resp) | 458 | 10470.42 |
| SMS | 20972 | 289.70 |
| SummaC_conv | 287 | 2323.56 |
| SummaC_zs | 284 | 2125.90 |
| SummaQA | 923 | 4830.03 |
| SUPERT | 532 | 2237.44 |
| Dataset | #Class | Avg. Length | Train | Test |
| SLURP | 18 × 46 | 6.93 | 50,628 | 10,992 |
| ATIS | 22 | 11.14 | 4,978 | 893 |
| TREC6 | 6 | 8.89 | 5,452 | 500 |
| Model | w/o manual transcripts | w/ manual transcripts | ||||
| SLURP | ATIS | TREC6 | SLURP | ATIS | TREC6 | |
| RoBERTa (Liu et al., 2019) | 83.97 | 94.53 | 84.08 | 84.42 | 94.86 | 84.54 |
| Phoneme-BERT (Sundararaman et al., 2021) | 83.78 | 94.83 | 85.96 | 84.16 | 95.14 | 86.48 |
| SimCSE (Gao et al., 2021) | 84.47 | 94.07 | 84.92 | 84.88 | 94.32 | 85.46 |
| SpokenCSE (Chang and Chen, 2022) | 85.26 | 95.10 | 86.36 | 85.64 | 95.58 | 86.82 |
| ML-LMCL | 88.52† | 96.52† | 89.24† | 89.16† | 97.21† | 89.96† |
| Model | w/ manual transcripts | ||
| SLURP | ATIS | TREC6 | |
| ML-LMCL | 89.16 | 97.21 | 89.96 |
| w/o Lmut | 88.68 (↓0.48) | 96.83 (↓0.38) | 89.52 (↓0.44) |
| w/o Lreg | 88.92 (↓0.24) | 96.98 (↓0.23) | 89.77 (↓0.19) |
| w/o Lqreg & Lqreg | 88.75 (↓0.41) | 96.92 (↓0.29) | 89.74 (↓0.22) |
| w/o cyc | 88.98 (↓0.18) | 97.08 (↓0.13) | 89.85 (↓0.11) |
| w/o Lmut + bsz↑ | 88.72 (↓0.44) | 96.92 (↓0.29) | 89.65 (↓0.31) |
| w/ Lsoft | 89.12 (↓0.04) | 97.18 (↓0.03) | 89.92 (↓0.04) |
| Dataset | #Class | Avg. Length | Train | Test |
| SLURP | 18 × 46 | 6.93 | 50,628 | 10,992 |
| ATIS | 22 | 11.14 | 4,978 | 893 |
| TREC6 | 6 | 8.89 | 5,452 | 500 |
| Model | w/o manual transcripts | w/ manual transcripts | ||||
| SLURP | ATIS | TREC6 | SLURP | ATIS | TREC6 | |
| RoBERTa (Liu et al., 2019) | 83.97 | 94.53 | 84.08 | 84.42 | 94.86 | 84.54 |
| Phoneme-BERT (Sundararaman et al., 2021) | 83.78 | 94.83 | 85.96 | 84.16 | 95.14 | 86.48 |
| SimCSE (Gao et al., 2021) | 84.47 | 94.07 | 84.92 | 84.88 | 94.32 | 85.46 |
| SpokenCSE (Chang and Chen, 2022) | 85.26 | 95.10 | 86.36 | 85.64 | 95.58 | 86.82 |
| ML-LMCL | 88.52† | 96.52† | 89.24† | 89.16† | 97.21† | 89.96† |
| Model | w/ manual transcripts | ||
| SLURP | ATIS | TREC6 | |
| ML-LMCL | 89.16 | 97.21 | 89.96 |
| w/o Lmut | 88.68 (↓0.48) | 96.83 (↓0.38) | 89.52 (↓0.44) |
| w/o Lreg | 88.92 (↓0.24) | 96.98 (↓0.23) | 89.77 (↓0.19) |
| w/o Lqreg & Lqreg | 88.75 (↓0.41) | 96.92 (↓0.29) | 89.74 (↓0.22) |
| w/o cyc | 88.98 (↓0.18) | 97.08 (↓0.13) | 89.85 (↓0.11) |
| w/o Lmut + bsz↑ | 88.72 (↓0.44) | 96.92 (↓0.29) | 89.65 (↓0.31) |
| w/ Lsoft | 89.12 (↓0.04) | 97.18 (↓0.03) | 89.92 (↓0.04) |
| Dataset | #Class | Avg. Length | Train | Test |
| SLURP | 18 × 46 | 6.93 | 50,628 | 10,992 |
| ATIS | 22 | 11.14 | 4,978 | 893 |
| TREC6 | 6 | 8.89 | 5,452 | 500 |
| Model | w/o manual transcripts | w/ manual transcripts | ||||
| SLURP | ATIS | TREC6 | SLURP | ATIS | TREC6 | |
| RoBERTa (Liu et al., 2019) | 83.97 | 94.53 | 84.08 | 84.42 | 94.86 | 84.54 |
| Phoneme-BERT (Sundararaman et al., 2021) | 83.78 | 94.83 | 85.96 | 84.16 | 95.14 | 86.48 |
| SimCSE (Gao et al., 2021) | 84.47 | 94.07 | 84.92 | 84.88 | 94.32 | 85.46 |
| SpokenCSE (Chang and Chen, 2022) | 85.26 | 95.10 | 86.36 | 85.64 | 95.58 | 86.82 |
| ML-LMCL | 88.52† | 96.52† | 89.24† | 89.16† | 97.21† | 89.96† |
| Model | w/ manual transcripts | ||
| SLURP | ATIS | TREC6 | |
| ML-LMCL | 89.16 | 97.21 | 89.96 |
| w/o Lmut | 88.68 (↓0.48) | 96.83 (↓0.38) | 89.52 (↓0.44) |
| w/o Lreg | 88.92 (↓0.24) | 96.98 (↓0.23) | 89.77 (↓0.19) |
| w/o Lqreg & Lqreg | 88.75 (↓0.41) | 96.92 (↓0.29) | 89.74 (↓0.22) |
| w/o cyc | 88.98 (↓0.18) | 97.08 (↓0.13) | 89.85 (↓0.11) |
| w/o Lmut + bsz↑ | 88.72 (↓0.44) | 96.92 (↓0.29) | 89.65 (↓0.31) |
| w/ Lsoft | 89.12 (↓0.04) | 97.18 (↓0.03) | 89.92 (↓0.04) |
| Dataset | #Class | Avg. Length | Train | Test |
| SLURP | 18 × 46 | 6.93 | 50,628 | 10,992 |
| ATIS | 22 | 11.14 | 4,978 | 893 |
| TREC6 | 6 | 8.89 | 5,452 | 500 |
| Model | w/o manual transcripts | w/ manual transcripts | ||||
| SLURP | ATIS | TREC6 | SLURP | ATIS | TREC6 | |
| RoBERTa (Liu et al., 2019) | 83.97 | 94.53 | 84.08 | 84.42 | 94.86 | 84.54 |
| Phoneme-BERT (Sundararaman et al., 2021) | 83.78 | 94.83 | 85.96 | 84.16 | 95.14 | 86.48 |
| SimCSE (Gao et al., 2021) | 84.47 | 94.07 | 84.92 | 84.88 | 94.32 | 85.46 |
| SpokenCSE (Chang and Chen, 2022) | 85.26 | 95.10 | 86.36 | 85.64 | 95.58 | 86.82 |
| ML-LMCL | 88.52† | 96.52† | 89.24† | 89.16† | 97.21† | 89.96† |
| Model | w/ manual transcripts | ||
| SLURP | ATIS | TREC6 | |
| ML-LMCL | 89.16 | 97.21 | 89.96 |
| w/o Lmut | 88.68 (↓0.48) | 96.83 (↓0.38) | 89.52 (↓0.44) |
| w/o Lreg | 88.92 (↓0.24) | 96.98 (↓0.23) | 89.77 (↓0.19) |
| w/o Lqreg & Lqreg | 88.75 (↓0.41) | 96.92 (↓0.29) | 89.74 (↓0.22) |
| w/o cyc | 88.98 (↓0.18) | 97.08 (↓0.13) | 89.85 (↓0.11) |
| w/o Lmut + bsz↑ | 88.72 (↓0.44) | 96.92 (↓0.29) | 89.65 (↓0.31) |
| w/ Lsoft | 89.12 (↓0.04) | 97.18 (↓0.03) | 89.92 (↓0.04) |
| MMSD/MMSD2.0 | Train | Validation | Test |
| Sentences | 19,816/19,816 | 2,410/2,410 | 2,409/2,409 |
| Positive | 8,642/9,572 | 959/1,042 | 959/1,037 |
| Negative | 11,174/10,240 | 1,451/1,368 | 1,450/1,372 |
| Model | MMSD | MMSD2.0 | ||||||
| Acc.(%) | P (%) | R (%) | F1 (%) | Acc. (%) | P (%) | R (%) | F1 (%) | |
| Text-modality Methods | ||||||||
| TextCNN (Kim, 2014) | 80.03 | 74.29 | 76.39 | 75.32 | 71.61 | 64.62 | 75.22 | 69.52 |
| Bi-LSTM (Graves and Schmidhuber, 2005) | 81.90 | 76.66 | 78.42 | 77.53 | 72.48 | 68.02 | 68.08 | 68.05 |
| SMSD (Xiong et al., 2019) | 80.90 | 76.46 | 75.18 | 75.82 | 73.56 | 68.45 | 71.55 | 69.97 |
| RoBERTa† (Liu et al., 2019) | 93.97 | 90.39 | 94.59 | 92.45 | 79.66 | 76.74 | 75.70 | 76.21 |
| Image-modality Methods | ||||||||
| ResNet (He et al., 2015) | 64.76 | 54.41 | 70.80 | 61.53 | 65.50 | 61.17 | 54.39 | 57.58 |
| ViT (Dosovitskiy et al., 2021) | 67.83 | 57.93 | 70.07 | 63.40 | 72.02 | 65.26 | 74.83 | 69.72 |
| Multi-Modality Methods | ||||||||
| HFM (Cai et al., 2019) | 83.44 | 76.57 | 84.15 | 80.18 | 70.57 | 64.84 | 69.05 | 66.88 |
| D&R Net (Xu et al., 2020) | 84.02 | 77.97 | 83.42 | 80.60 | - | - | - | - |
| Att-BERT (Pan et al., 2020) | 86.05 | 80.87 | 85.08 | 82.92 | 80.03 | 76.28 | 77.82 | 77.04 |
| InCrossMGs (Liang et al., 2021) | 86.10 | 81.38 | 84.36 | 82.84 | - | - | - | - |
| CMGCN (Liang et al., 2022) | 86.54 | - | - | 82.73 | 79.83 | 75.82 | 78.01 | 76.90 |
| HKE (Liu et al., 2022) | 87.36 | 81.84 | 86.48 | 84.09 | 76.50 | 73.48 | 71.07 | 72.25 |
| Multi-view CLIP | 88.33 | 82.66 | 88.65 | 85.55 | 85.64 | 80.33 | 88.24 | 84.10 |
| Model | Acc.(%) | P(%) | R(%) | F1(%) |
| Ours | 85.64 | 80.33 | 88.24 | 84.10 |
| w/o LT | 84.18 | 80.60 | 83.32 | 81.93 |
| w/o LV | 83.69 | 76.97 | 88.62 | 82.38 |
| w/o LFF | 82.44 | 73.80 | 91.80 | 81.82 |
| MMSD/MMSD2.0 | Train | Validation | Test |
| Sentences | 19,816/19,816 | 2,410/2,410 | 2,409/2,409 |
| Positive | 8,642/9,572 | 959/1,042 | 959/1,037 |
| Negative | 11,174/10,240 | 1,451/1,368 | 1,450/1,372 |
| Model | MMSD | MMSD2.0 | ||||||
| Acc.(%) | P (%) | R (%) | F1 (%) | Acc. (%) | P (%) | R (%) | F1 (%) | |
| Text-modality Methods | ||||||||
| TextCNN (Kim, 2014) | 80.03 | 74.29 | 76.39 | 75.32 | 71.61 | 64.62 | 75.22 | 69.52 |
| Bi-LSTM (Graves and Schmidhuber, 2005) | 81.90 | 76.66 | 78.42 | 77.53 | 72.48 | 68.02 | 68.08 | 68.05 |
| SMSD (Xiong et al., 2019) | 80.90 | 76.46 | 75.18 | 75.82 | 73.56 | 68.45 | 71.55 | 69.97 |
| RoBERTa† (Liu et al., 2019) | 93.97 | 90.39 | 94.59 | 92.45 | 79.66 | 76.74 | 75.70 | 76.21 |
| Image-modality Methods | ||||||||
| ResNet (He et al., 2015) | 64.76 | 54.41 | 70.80 | 61.53 | 65.50 | 61.17 | 54.39 | 57.58 |
| ViT (Dosovitskiy et al., 2021) | 67.83 | 57.93 | 70.07 | 63.40 | 72.02 | 65.26 | 74.83 | 69.72 |
| Multi-Modality Methods | ||||||||
| HFM (Cai et al., 2019) | 83.44 | 76.57 | 84.15 | 80.18 | 70.57 | 64.84 | 69.05 | 66.88 |
| D&R Net (Xu et al., 2020) | 84.02 | 77.97 | 83.42 | 80.60 | - | - | - | - |
| Att-BERT (Pan et al., 2020) | 86.05 | 80.87 | 85.08 | 82.92 | 80.03 | 76.28 | 77.82 | 77.04 |
| InCrossMGs (Liang et al., 2021) | 86.10 | 81.38 | 84.36 | 82.84 | - | - | - | - |
| CMGCN (Liang et al., 2022) | 86.54 | - | - | 82.73 | 79.83 | 75.82 | 78.01 | 76.90 |
| HKE (Liu et al., 2022) | 87.36 | 81.84 | 86.48 | 84.09 | 76.50 | 73.48 | 71.07 | 72.25 |
| Multi-view CLIP | 88.33 | 82.66 | 88.65 | 85.55 | 85.64 | 80.33 | 88.24 | 84.10 |
| Model | Acc.(%) | P(%) | R(%) | F1(%) |
| Ours | 85.64 | 80.33 | 88.24 | 84.10 |
| w/o LT | 84.18 | 80.60 | 83.32 | 81.93 |
| w/o LV | 83.69 | 76.97 | 88.62 | 82.38 |
| w/o LFF | 82.44 | 73.80 | 91.80 | 81.82 |
| MMSD/MMSD2.0 | Train | Validation | Test |
| Sentences | 19,816/19,816 | 2,410/2,410 | 2,409/2,409 |
| Positive | 8,642/9,572 | 959/1,042 | 959/1,037 |
| Negative | 11,174/10,240 | 1,451/1,368 | 1,450/1,372 |
| Model | MMSD | MMSD2.0 | ||||||
| Acc.(%) | P (%) | R (%) | F1 (%) | Acc. (%) | P (%) | R (%) | F1 (%) | |
| Text-modality Methods | ||||||||
| TextCNN (Kim, 2014) | 80.03 | 74.29 | 76.39 | 75.32 | 71.61 | 64.62 | 75.22 | 69.52 |
| Bi-LSTM (Graves and Schmidhuber, 2005) | 81.90 | 76.66 | 78.42 | 77.53 | 72.48 | 68.02 | 68.08 | 68.05 |
| SMSD (Xiong et al., 2019) | 80.90 | 76.46 | 75.18 | 75.82 | 73.56 | 68.45 | 71.55 | 69.97 |
| RoBERTa† (Liu et al., 2019) | 93.97 | 90.39 | 94.59 | 92.45 | 79.66 | 76.74 | 75.70 | 76.21 |
| Image-modality Methods | ||||||||
| ResNet (He et al., 2015) | 64.76 | 54.41 | 70.80 | 61.53 | 65.50 | 61.17 | 54.39 | 57.58 |
| ViT (Dosovitskiy et al., 2021) | 67.83 | 57.93 | 70.07 | 63.40 | 72.02 | 65.26 | 74.83 | 69.72 |
| Multi-Modality Methods | ||||||||
| HFM (Cai et al., 2019) | 83.44 | 76.57 | 84.15 | 80.18 | 70.57 | 64.84 | 69.05 | 66.88 |
| D&R Net (Xu et al., 2020) | 84.02 | 77.97 | 83.42 | 80.60 | - | - | - | - |
| Att-BERT (Pan et al., 2020) | 86.05 | 80.87 | 85.08 | 82.92 | 80.03 | 76.28 | 77.82 | 77.04 |
| InCrossMGs (Liang et al., 2021) | 86.10 | 81.38 | 84.36 | 82.84 | - | - | - | - |
| CMGCN (Liang et al., 2022) | 86.54 | - | - | 82.73 | 79.83 | 75.82 | 78.01 | 76.90 |
| HKE (Liu et al., 2022) | 87.36 | 81.84 | 86.48 | 84.09 | 76.50 | 73.48 | 71.07 | 72.25 |
| Multi-view CLIP | 88.33 | 82.66 | 88.65 | 85.55 | 85.64 | 80.33 | 88.24 | 84.10 |
| Model | Acc.(%) | P(%) | R(%) | F1(%) |
| Ours | 85.64 | 80.33 | 88.24 | 84.10 |
| w/o LT | 84.18 | 80.60 | 83.32 | 81.93 |
| w/o LV | 83.69 | 76.97 | 88.62 | 82.38 |
| w/o LFF | 82.44 | 73.80 | 91.80 | 81.82 |
| MMSD/MMSD2.0 | Train | Validation | Test |
| Sentences | 19,816/19,816 | 2,410/2,410 | 2,409/2,409 |
| Positive | 8,642/9,572 | 959/1,042 | 959/1,037 |
| Negative | 11,174/10,240 | 1,451/1,368 | 1,450/1,372 |
| Model | MMSD | MMSD2.0 | ||||||
| Acc.(%) | P (%) | R (%) | F1 (%) | Acc. (%) | P (%) | R (%) | F1 (%) | |
| Text-modality Methods | ||||||||
| TextCNN (Kim, 2014) | 80.03 | 74.29 | 76.39 | 75.32 | 71.61 | 64.62 | 75.22 | 69.52 |
| Bi-LSTM (Graves and Schmidhuber, 2005) | 81.90 | 76.66 | 78.42 | 77.53 | 72.48 | 68.02 | 68.08 | 68.05 |
| SMSD (Xiong et al., 2019) | 80.90 | 76.46 | 75.18 | 75.82 | 73.56 | 68.45 | 71.55 | 69.97 |
| RoBERTa† (Liu et al., 2019) | 93.97 | 90.39 | 94.59 | 92.45 | 79.66 | 76.74 | 75.70 | 76.21 |
| Image-modality Methods | ||||||||
| ResNet (He et al., 2015) | 64.76 | 54.41 | 70.80 | 61.53 | 65.50 | 61.17 | 54.39 | 57.58 |
| ViT (Dosovitskiy et al., 2021) | 67.83 | 57.93 | 70.07 | 63.40 | 72.02 | 65.26 | 74.83 | 69.72 |
| Multi-Modality Methods | ||||||||
| HFM (Cai et al., 2019) | 83.44 | 76.57 | 84.15 | 80.18 | 70.57 | 64.84 | 69.05 | 66.88 |
| D&R Net (Xu et al., 2020) | 84.02 | 77.97 | 83.42 | 80.60 | - | - | - | - |
| Att-BERT (Pan et al., 2020) | 86.05 | 80.87 | 85.08 | 82.92 | 80.03 | 76.28 | 77.82 | 77.04 |
| InCrossMGs (Liang et al., 2021) | 86.10 | 81.38 | 84.36 | 82.84 | - | - | - | - |
| CMGCN (Liang et al., 2022) | 86.54 | - | - | 82.73 | 79.83 | 75.82 | 78.01 | 76.90 |
| HKE (Liu et al., 2022) | 87.36 | 81.84 | 86.48 | 84.09 | 76.50 | 73.48 | 71.07 | 72.25 |
| Multi-view CLIP | 88.33 | 82.66 | 88.65 | 85.55 | 85.64 | 80.33 | 88.24 | 84.10 |
| Model | Acc.(%) | P(%) | R(%) | F1(%) |
| Ours | 85.64 | 80.33 | 88.24 | 84.10 |
| w/o LT | 84.18 | 80.60 | 83.32 | 81.93 |
| w/o LV | 83.69 | 76.97 | 88.62 | 82.38 |
| w/o LFF | 82.44 | 73.80 | 91.80 | 81.82 |
| Data type | EN↔DE | EN↔FR | EN↔PL | EN↔RU |
| Source & target | 5.3M | 14.7M | 12.9M | 12.4M |
| metadata | ||||
| Genre | 45.3% | 57.8% | 60.5% | 73.4% |
| PG rating | 35.9% | 46.9% | 48.8% | 62.3% |
| Writer(s) | 45.3% | 57.1% | 58.9% | 71.7% |
| Year | 45.3% | 57.8% | 60.5% | 73.7% |
| Country | 37.7% | 42.9% | 45.7% | 42.7% |
| Plot description | 43.4% | 57.1% | 59.7% | 72.6% |
| previous dialogue | ||||
| n-1 | 60.1% | 68.0% | 63.7% | 73.6% |
| n-2 | 42.0% | 51.2% | 46.4% | 57.9% |
| n-3 | 31.2% | 40.1% | 35.5% | 46.9% |
| n-4 | 23.9% | 32.2% | 28.0% | 38.6% |
| n-5 | 18.7% | 26.2% | 22.4% | 32.2% |
| Key | Value |
| Source (EN) | This is the Angel of Death, big daddy reaper. |
| Target (PL) | To aniol smierci. Kosiarz przy文化传播 "k". |
| PG rating | PG rating: TV-14 |
| Released | Released in 2009 |
| Writers | Writers: Eric Kripke, Ben Edlund, Julie Siege |
| Plot | Dean and Sam get to know the whereabouts of Lucifer and want to hunt him down. But Lucifer is well prepared and is working his own plans. |
| Genre | Drama, Fantasy, Horror |
| Country | United States, Canada |
| Model | Params | dmodel | Layers | h | FFN dim. | GPU Hour/Epoch | Epochs to best | ||||
| Cxt | Src | Dec | Cxt | Src | Dec | ||||||
| BASE | 66M | 512 | - | 6 | 6 | 8 | - | 2048 | 2048 | - | - |
| BASE-PM | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | - | - |
| TAGGING | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | 0.74 ± 0.35 | 6.13 ± 4.09 |
| NOVOTNEY-CUE | 99M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 1.29 ± 0.56 | 9.13 ± 3.60 |
| MTCUE | 105M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 0.81 ± 0.39 | 9.38 ± 4.57 |
| Model | EN→DE | EN→FR | EN→PL | EN→RU | DE→EN | FR→EN | PL→EN | RU→EN | Average | |||||||||
| BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | |
| Baselines | ||||||||||||||||||
| *BASE | 33.60 | 45.90 | 34.54 | 46.92 | 28.08 | 58.52 | 31.37 | 62.94 | 39.53 | 59.56 | 35.46 | 55.10 | 34.42 | 50.38 | 39.37 | 55.99 | 34.65 | 54.41 |
| *BASE-PM | 34.36 | 46.77 | 35.31 | 48.87 | 28.66 | 60.97 | 32.40 | 64.55 | 40.32 | 60.88 | 36.16 | 56.28 | 35.03 | 51.77 | 40.04 | 56.86 | 35.28 | 55.87 |
| TAGGING | 34.88 | 49.21 | 36.74 | 51.57 | 29.08 | 64.29 | 32.32 | 65.12 | 41.52 | 62.63 | 37.10 | 57.41 | 36.19 | 53.46 | 40.33 | 57.14 | 36.02 | 57.60 |
| NOVOTNEY-CUE | 35.30 | 49.83 | 36.75 | 50.52 | 29.09 | 62.69 | 32.36 | 64.90 | 40.86 | 61.91 | 36.51 | 56.21 | 35.28 | 52.17 | 39.44 | 56.08 | 35.70 | 56.79 |
| Proposed | ||||||||||||||||||
| MTCUE | 36.02 | 50.91 | 37.54 | 52.19 | 29.36 | 63.46 | 33.21 | 65.21 | 40.95 | 61.58 | 36.57 | 56.87 | 35.68 | 52.48 | 39.97 | 56.92 | 36.16 | 57.45 |
| Model | Supervision | Formal | Informal | Average | ||
| EN-DE | Non-context baseline | - | 74.5 | 25.5 | 50.0 | |
| Rippeth et al. (2022) | Supervised | 99.4 | 96.5 | 98.0 | ||
| Vincent et al. (2022a) | Supervised | 100.0 | 100.0 | 100.0 | ||
| MTCUE | Zero-shot | 100.0 | 100.0 | 100.0 | ||
| EN-RU | Non-context baseline | - | 96.4 | 3.6 | 50.0 | |
| Rippeth et al. (2022) | Zero-shot | 100.0 | 1.1 | 50.5 | ||
| Vincent et al. (2022a) | Zero-shot | 99.5 | 85.8 | 92.7 | ||
| MTCUE | Zero-shot | 100.0 | 99.4 | 99.7 |
| Ablation | EN→DE | COMET EN→FR | EN→PL | ZERO-SHOT ACCURACY IWSLT22 (DE) | EAMT22 |
| Full MTCUE | 46.89 | 54.06 | 62.67 | 100.0 | 81.35 |
| no context encoder | 46.76 | 53.73 | 63.26 | 89.10 | 77.42 |
| no pos. embeddings | 46.68 | 53.81 | 62.47 | 91.65 | 70.91 |
| no MINILM-V2 | 45.32 | 53.42 | 62.55 | 50.00 | 70.16 |
| no metadata | 45.23 | 53.64 | 62.64 | 89.70 | 83.41 |
| no doc.-level data | 46.23 | 53.49 | 61.67 | 68.80 | 74.64 |
| random context | 42.17 | 51.94 | 61.74 | 49.90 | 68.44 |
| no context* | 41.22 | 50.07 | 58.94 | 50.00 | 67.53 |
| EAMT22 | |
| Source | I just didn't want you to think you had to marry me. |
| Context | I am a woman. I am talking to a man |
| Reference | BoNie chciałam, zebyś myslał, ze cię zmuszam do slubu. \n"Because I didn't wantfeminine you to thinkmasculine I am forcing you into a wedding." |
| Baseline | Po prostu Nie chciałem, zebyś myslała, ze musisz za mnie wyjsć. \n"I just didn't wantmasculine you to thinkfeminine you had to marryfeminine me." |
| MTCUE | Nie chciałam, zebyś myslał, ze musisz są ze mną ożenić. \n"I didn't wantfeminine you to thinkmasculine you had to marrymasculine me." |
| Source | So then you confronted Derek. |
| Context | I am talking to a woman |
| Reference | A wiec doprowadziłs do konfrontacje z Derekiem. \n"So then you ledfeminine to a confrontation with Derek." |
| Baseline | Wiec wtedy skonfrontowale są徐 Derekiem. \n"So then you confrontedmasculine Derek." |
| MTCUE | Wiec skonfrontowale są徐 Derekiem. \n"So then you confrontedfeminine Derek." |
| IWSLT22 | |
| Source | I got a hundred colours in your city. |
| MTCUE (formal) | Ich habe 100 Farben in Ihrer Stadt. |
| MTCUE (informal) | Ich hab 100 Farben in deiner Stadt. |
| Data type | EN↔DE | EN↔FR | EN↔PL | EN↔RU |
| Source & target | 5.3M | 14.7M | 12.9M | 12.4M |
| metadata | ||||
| Genre | 45.3% | 57.8% | 60.5% | 73.4% |
| PG rating | 35.9% | 46.9% | 48.8% | 62.3% |
| Writer(s) | 45.3% | 57.1% | 58.9% | 71.7% |
| Year | 45.3% | 57.8% | 60.5% | 73.7% |
| Country | 37.7% | 42.9% | 45.7% | 42.7% |
| Plot description | 43.4% | 57.1% | 59.7% | 72.6% |
| previous dialogue | ||||
| n-1 | 60.1% | 68.0% | 63.7% | 73.6% |
| n-2 | 42.0% | 51.2% | 46.4% | 57.9% |
| n-3 | 31.2% | 40.1% | 35.5% | 46.9% |
| n-4 | 23.9% | 32.2% | 28.0% | 38.6% |
| n-5 | 18.7% | 26.2% | 22.4% | 32.2% |
| Key | Value |
| Source (EN) | This is the Angel of Death, big daddy reaper. |
| Target (PL) | To aniol smierci. Kosiarz przy文化传播 "k". |
| PG rating | PG rating: TV-14 |
| Released | Released in 2009 |
| Writers | Writers: Eric Kripke, Ben Edlund, Julie Siege |
| Plot | Dean and Sam get to know the whereabouts of Lucifer and want to hunt him down. But Lucifer is well prepared and is working his own plans. |
| Genre | Drama, Fantasy, Horror |
| Country | United States, Canada |
| Model | Params | dmodel | Layers | h | FFN dim. | GPU Hour/Epoch | Epochs to best | ||||
| Cxt | Src | Dec | Cxt | Src | Dec | ||||||
| BASE | 66M | 512 | - | 6 | 6 | 8 | - | 2048 | 2048 | - | - |
| BASE-PM | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | - | - |
| TAGGING | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | 0.74 ± 0.35 | 6.13 ± 4.09 |
| NOVOTNEY-CUE | 99M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 1.29 ± 0.56 | 9.13 ± 3.60 |
| MTCUE | 105M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 0.81 ± 0.39 | 9.38 ± 4.57 |
| Model | EN→DE | EN→FR | EN→PL | EN→RU | DE→EN | FR→EN | PL→EN | RU→EN | Average | |||||||||
| BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | |
| Baselines | ||||||||||||||||||
| *BASE | 33.60 | 45.90 | 34.54 | 46.92 | 28.08 | 58.52 | 31.37 | 62.94 | 39.53 | 59.56 | 35.46 | 55.10 | 34.42 | 50.38 | 39.37 | 55.99 | 34.65 | 54.41 |
| *BASE-PM | 34.36 | 46.77 | 35.31 | 48.87 | 28.66 | 60.97 | 32.40 | 64.55 | 40.32 | 60.88 | 36.16 | 56.28 | 35.03 | 51.77 | 40.04 | 56.86 | 35.28 | 55.87 |
| TAGGING | 34.88 | 49.21 | 36.74 | 51.57 | 29.08 | 64.29 | 32.32 | 65.12 | 41.52 | 62.63 | 37.10 | 57.41 | 36.19 | 53.46 | 40.33 | 57.14 | 36.02 | 57.60 |
| NOVOTNEY-CUE | 35.30 | 49.83 | 36.75 | 50.52 | 29.09 | 62.69 | 32.36 | 64.90 | 40.86 | 61.91 | 36.51 | 56.21 | 35.28 | 52.17 | 39.44 | 56.08 | 35.70 | 56.79 |
| Proposed | ||||||||||||||||||
| MTCUE | 36.02 | 50.91 | 37.54 | 52.19 | 29.36 | 63.46 | 33.21 | 65.21 | 40.95 | 61.58 | 36.57 | 56.87 | 35.68 | 52.48 | 39.97 | 56.92 | 36.16 | 57.45 |
| Model | Supervision | Formal | Informal | Average | ||
| EN-DE | Non-context baseline | - | 74.5 | 25.5 | 50.0 | |
| Rippeth et al. (2022) | Supervised | 99.4 | 96.5 | 98.0 | ||
| Vincent et al. (2022a) | Supervised | 100.0 | 100.0 | 100.0 | ||
| MTCUE | Zero-shot | 100.0 | 100.0 | 100.0 | ||
| EN-RU | Non-context baseline | - | 96.4 | 3.6 | 50.0 | |
| Rippeth et al. (2022) | Zero-shot | 100.0 | 1.1 | 50.5 | ||
| Vincent et al. (2022a) | Zero-shot | 99.5 | 85.8 | 92.7 | ||
| MTCUE | Zero-shot | 100.0 | 99.4 | 99.7 |
| Ablation | EN→DE | COMET EN→FR | EN→PL | ZERO-SHOT ACCURACY IWSLT22 (DE) | EAMT22 |
| Full MTCUE | 46.89 | 54.06 | 62.67 | 100.0 | 81.35 |
| no context encoder | 46.76 | 53.73 | 63.26 | 89.10 | 77.42 |
| no pos. embeddings | 46.68 | 53.81 | 62.47 | 91.65 | 70.91 |
| no MINILM-V2 | 45.32 | 53.42 | 62.55 | 50.00 | 70.16 |
| no metadata | 45.23 | 53.64 | 62.64 | 89.70 | 83.41 |
| no doc.-level data | 46.23 | 53.49 | 61.67 | 68.80 | 74.64 |
| random context | 42.17 | 51.94 | 61.74 | 49.90 | 68.44 |
| no context* | 41.22 | 50.07 | 58.94 | 50.00 | 67.53 |
| EAMT22 | |
| Source | I just didn't want you to think you had to marry me. |
| Context | I am a woman. I am talking to a man |
| Reference | BoNie chciałam, zebyś myslał, ze cię zmuszam do slubu. \n"Because I didn't wantfeminine you to thinkmasculine I am forcing you into a wedding." |
| Baseline | Po prostu Nie chciałem, zebyś myslała, ze musisz za mnie wyjsć. \n"I just didn't wantmasculine you to thinkfeminine you had to marryfeminine me." |
| MTCUE | Nie chciałam, zebyś myslał, ze musisz są ze mną ożenić. \n"I didn't wantfeminine you to thinkmasculine you had to marrymasculine me." |
| Source | So then you confronted Derek. |
| Context | I am talking to a woman |
| Reference | A wiec doprowadziłs do konfrontacje z Derekiem. \n"So then you ledfeminine to a confrontation with Derek." |
| Baseline | Wiec wtedy skonfrontowale są徐 Derekiem. \n"So then you confrontedmasculine Derek." |
| MTCUE | Wiec skonfrontowale są徐 Derekiem. \n"So then you confrontedfeminine Derek." |
| IWSLT22 | |
| Source | I got a hundred colours in your city. |
| MTCUE (formal) | Ich habe 100 Farben in Ihrer Stadt. |
| MTCUE (informal) | Ich hab 100 Farben in deiner Stadt. |
| Data type | EN↔DE | EN↔FR | EN↔PL | EN↔RU |
| Source & target | 5.3M | 14.7M | 12.9M | 12.4M |
| metadata | ||||
| Genre | 45.3% | 57.8% | 60.5% | 73.4% |
| PG rating | 35.9% | 46.9% | 48.8% | 62.3% |
| Writer(s) | 45.3% | 57.1% | 58.9% | 71.7% |
| Year | 45.3% | 57.8% | 60.5% | 73.7% |
| Country | 37.7% | 42.9% | 45.7% | 42.7% |
| Plot description | 43.4% | 57.1% | 59.7% | 72.6% |
| previous dialogue | ||||
| n-1 | 60.1% | 68.0% | 63.7% | 73.6% |
| n-2 | 42.0% | 51.2% | 46.4% | 57.9% |
| n-3 | 31.2% | 40.1% | 35.5% | 46.9% |
| n-4 | 23.9% | 32.2% | 28.0% | 38.6% |
| n-5 | 18.7% | 26.2% | 22.4% | 32.2% |
| Key | Value |
| Source (EN) | This is the Angel of Death, big daddy reaper. |
| Target (PL) | To aniol smierci. Kosiarz przy文化传播 "k". |
| PG rating | PG rating: TV-14 |
| Released | Released in 2009 |
| Writers | Writers: Eric Kripke, Ben Edlund, Julie Siege |
| Plot | Dean and Sam get to know the whereabouts of Lucifer and want to hunt him down. But Lucifer is well prepared and is working his own plans. |
| Genre | Drama, Fantasy, Horror |
| Country | United States, Canada |
| Model | Params | dmodel | Layers | h | FFN dim. | GPU Hour/Epoch | Epochs to best | ||||
| Cxt | Src | Dec | Cxt | Src | Dec | ||||||
| BASE | 66M | 512 | - | 6 | 6 | 8 | - | 2048 | 2048 | - | - |
| BASE-PM | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | - | - |
| TAGGING | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | 0.74 ± 0.35 | 6.13 ± 4.09 |
| NOVOTNEY-CUE | 99M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 1.29 ± 0.56 | 9.13 ± 3.60 |
| MTCUE | 105M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 0.81 ± 0.39 | 9.38 ± 4.57 |
| Model | EN→DE | EN→FR | EN→PL | EN→RU | DE→EN | FR→EN | PL→EN | RU→EN | Average | |||||||||
| BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | |
| Baselines | ||||||||||||||||||
| *BASE | 33.60 | 45.90 | 34.54 | 46.92 | 28.08 | 58.52 | 31.37 | 62.94 | 39.53 | 59.56 | 35.46 | 55.10 | 34.42 | 50.38 | 39.37 | 55.99 | 34.65 | 54.41 |
| *BASE-PM | 34.36 | 46.77 | 35.31 | 48.87 | 28.66 | 60.97 | 32.40 | 64.55 | 40.32 | 60.88 | 36.16 | 56.28 | 35.03 | 51.77 | 40.04 | 56.86 | 35.28 | 55.87 |
| TAGGING | 34.88 | 49.21 | 36.74 | 51.57 | 29.08 | 64.29 | 32.32 | 65.12 | 41.52 | 62.63 | 37.10 | 57.41 | 36.19 | 53.46 | 40.33 | 57.14 | 36.02 | 57.60 |
| NOVOTNEY-CUE | 35.30 | 49.83 | 36.75 | 50.52 | 29.09 | 62.69 | 32.36 | 64.90 | 40.86 | 61.91 | 36.51 | 56.21 | 35.28 | 52.17 | 39.44 | 56.08 | 35.70 | 56.79 |
| Proposed | ||||||||||||||||||
| MTCUE | 36.02 | 50.91 | 37.54 | 52.19 | 29.36 | 63.46 | 33.21 | 65.21 | 40.95 | 61.58 | 36.57 | 56.87 | 35.68 | 52.48 | 39.97 | 56.92 | 36.16 | 57.45 |
| Model | Supervision | Formal | Informal | Average | ||
| EN-DE | Non-context baseline | - | 74.5 | 25.5 | 50.0 | |
| Rippeth et al. (2022) | Supervised | 99.4 | 96.5 | 98.0 | ||
| Vincent et al. (2022a) | Supervised | 100.0 | 100.0 | 100.0 | ||
| MTCUE | Zero-shot | 100.0 | 100.0 | 100.0 | ||
| EN-RU | Non-context baseline | - | 96.4 | 3.6 | 50.0 | |
| Rippeth et al. (2022) | Zero-shot | 100.0 | 1.1 | 50.5 | ||
| Vincent et al. (2022a) | Zero-shot | 99.5 | 85.8 | 92.7 | ||
| MTCUE | Zero-shot | 100.0 | 99.4 | 99.7 | ||
| Ablation | EN→DE | COMET EN→FR | EN→PL | ZERO-SHOT ACCURACY IWSLT22 (DE) | EAMT22 |
| Full MTCUE | 46.89 | 54.06 | 62.67 | 100.0 | 81.35 |
| no context encoder | 46.76 | 53.73 | 63.26 | 89.10 | 77.42 |
| no pos. embeddings | 46.68 | 53.81 | 62.47 | 91.65 | 70.91 |
| no MINILM-V2 | 45.32 | 53.42 | 62.55 | 50.00 | 70.16 |
| no metadata | 45.23 | 53.64 | 62.64 | 89.70 | 83.41 |
| no doc.-level data | 46.23 | 53.49 | 61.67 | 68.80 | 74.64 |
| random context | 42.17 | 51.94 | 61.74 | 49.90 | 68.44 |
| no context* | 41.22 | 50.07 | 58.94 | 50.00 | 67.53 |
| EAMT22 | |
| Source | I just didn't want you to think you had to marry me. |
| Context | I am a woman. I am talking to a man |
| Reference | BoNie chciałam, zebyś myslał, ze cię zmuszam do slubu. +"Because I didn't wantfeminine you to thinkmasculine I am forcing you into a wedding." |
| Baseline | Po prostu Nie chciałem, zebyś myslała, ze musisz za mnie wyjsć. +"I just didn't wantmasculine you to thinkfeminine you had to marryfeminine me." |
| MTCUE | Nie chciałam, zebyś myslał, ze musisz są ze mną ożenić. +"I didn't wantfeminine you to thinkmasculine you had to marrymasculine me." |
| Source | So then you confronted Derek. |
| Context | I am talking to a woman |
| Reference | A wiec doprowadziłs do konfrontacje z Derekiem. +"So then you ledfeminine to a confrontation with Derek." |
| Baseline | Wiec wtedy skonfrontowale są徐 Derekiem. +"So then you confrontedmasculine Derek." |
| MTCUE | Wiec skonfrontowale są徐 Derekiem. +"So then you confrontedfeminine Derek." |
| IWSLT22 | |
| Source | I got a hundred colours in your city. |
| MTCUE (formal) | Ich habe 100 Farben in Ihrer Stadt. |
| MTCUE (informal) | Ich hab 100 Farben in deiner Stadt. |
| Data type | EN↔DE | EN↔FR | EN↔PL | EN↔RU |
| Source & target | 5.3M | 14.7M | 12.9M | 12.4M |
| metadata | ||||
| Genre | 45.3% | 57.8% | 60.5% | 73.4% |
| PG rating | 35.9% | 46.9% | 48.8% | 62.3% |
| Writer(s) | 45.3% | 57.1% | 58.9% | 71.7% |
| Year | 45.3% | 57.8% | 60.5% | 73.7% |
| Country | 37.7% | 42.9% | 45.7% | 42.7% |
| Plot description | 43.4% | 57.1% | 59.7% | 72.6% |
| previous dialogue | ||||
| n-1 | 60.1% | 68.0% | 63.7% | 73.6% |
| n-2 | 42.0% | 51.2% | 46.4% | 57.9% |
| n-3 | 31.2% | 40.1% | 35.5% | 46.9% |
| n-4 | 23.9% | 32.2% | 28.0% | 38.6% |
| n-5 | 18.7% | 26.2% | 22.4% | 32.2% |
| Key | Value |
| Source (EN) | This is the Angel of Death, big daddy reaper. |
| Target (PL) | To aniol smierci. Kosiarz przy文化传播 "k". |
| PG rating | PG rating: TV-14 |
| Released | Released in 2009 |
| Writers | Writers: Eric Kripke, Ben Edlund, Julie Siege |
| Plot | Dean and Sam get to know the whereabouts of Lucifer and want to hunt him down. But Lucifer is well prepared and is working his own plans. |
| Genre | Drama, Fantasy, Horror |
| Country | United States, Canada |
| Model | Params | dmodel | Layers | h | FFN dim. | GPU Hour/Epoch | Epochs to best | ||||
| Cxt | Src | Dec | Cxt | Src | Dec | ||||||
| BASE | 66M | 512 | - | 6 | 6 | 8 | - | 2048 | 2048 | - | - |
| BASE-PM | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | - | - |
| TAGGING | 107M | 512 | - | 10 | 6 | 8 | - | 4096 | 2048 | 0.74 ± 0.35 | 6.13 ± 4.09 |
| NOVOTNEY-CUE | 99M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 1.29 ± 0.56 | 9.13 ± 3.60 |
| MTCUE | 105M | 512 | 6 | 6 | 6 | 8 | 2048 | 2048 | 2048 | 0.81 ± 0.39 | 9.38 ± 4.57 |
| Model | EN→DE | EN→FR | EN→PL | EN→RU | DE→EN | FR→EN | PL→EN | RU→EN | Average | |||||||||
| BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | BLEU | COMET | |
| Baselines | ||||||||||||||||||
| *BASE | 33.60 | 45.90 | 34.54 | 46.92 | 28.08 | 58.52 | 31.37 | 62.94 | 39.53 | 59.56 | 35.46 | 55.10 | 34.42 | 50.38 | 39.37 | 55.99 | 34.65 | 54.41 |
| *BASE-PM | 34.36 | 46.77 | 35.31 | 48.87 | 28.66 | 60.97 | 32.40 | 64.55 | 40.32 | 60.88 | 36.16 | 56.28 | 35.03 | 51.77 | 40.04 | 56.86 | 35.28 | 55.87 |
| TAGGING | 34.88 | 49.21 | 36.74 | 51.57 | 29.08 | 64.29 | 32.32 | 65.12 | 41.52 | 62.63 | 37.10 | 57.41 | 36.19 | 53.46 | 40.33 | 57.14 | 36.02 | 57.60 |
| NOVOTNEY-CUE | 35.30 | 49.83 | 36.75 | 50.52 | 29.09 | 62.69 | 32.36 | 64.90 | 40.86 | 61.91 | 36.51 | 56.21 | 35.28 | 52.17 | 39.44 | 56.08 | 35.70 | 56.79 |
| Proposed | ||||||||||||||||||
| MTCUE | 36.02 | 50.91 | 37.54 | 52.19 | 29.36 | 63.46 | 33.21 | 65.21 | 40.95 | 61.58 | 36.57 | 56.87 | 35.68 | 52.48 | 39.97 | 56.92 | 36.16 | 57.45 |
| Model | Supervision | Formal | Informal | Average | ||
| EN-DE | Non-context baseline | - | 74.5 | 25.5 | 50.0 | |
| Rippeth et al. (2022) | Supervised | 99.4 | 96.5 | 98.0 | ||
| Vincent et al. (2022a) | Supervised | 100.0 | 100.0 | 100.0 | ||
| MTCUE | Zero-shot | 100.0 | 100.0 | 100.0 | ||
| EN-RU | Non-context baseline | - | 96.4 | 3.6 | 50.0 | |
| Rippeth et al. (2022) | Zero-shot | 100.0 | 1.1 | 50.5 | ||
| Vincent et al. (2022a) | Zero-shot | 99.5 | 85.8 | 92.7 | ||
| MTCUE | Zero-shot | 100.0 | 99.4 | 99.7 |
| Ablation | EN→DE | COMET EN→FR | EN→PL | ZERO-SHOT ACCURACY IWSLT22 (DE) | EAMT22 |
| Full MTCUE | 46.89 | 54.06 | 62.67 | 100.0 | 81.35 |
| no context encoder | 46.76 | 53.73 | 63.26 | 89.10 | 77.42 |
| no pos. embeddings | 46.68 | 53.81 | 62.47 | 91.65 | 70.91 |
| no MINILM-V2 | 45.32 | 53.42 | 62.55 | 50.00 | 70.16 |
| no metadata | 45.23 | 53.64 | 62.64 | 89.70 | 83.41 |
| no doc.-level data | 46.23 | 53.49 | 61.67 | 68.80 | 74.64 |
| random context | 42.17 | 51.94 | 61.74 | 49.90 | 68.44 |
| no context* | 41.22 | 50.07 | 58.94 | 50.00 | 67.53 |
| EAMT22 | |
| Source | I just didn't want you to think you had to marry me. |
| Context | I am a woman. I am talking to a man |
| Reference | BoNie chciałam, zebyś myslał, ze cię zmuszam do slubu. \n"Because I didn't wantfeminine you to thinkmasculine I am forcing you into a wedding." |
| Baseline | Po prostu Nie chciałem, zebyś myslała, ze musisz za mnie wyjsć. \n"I just didn't wantmasculine you to thinkfeminine you had to marryfeminine me." |
| MTCUE | Nie chciałam, zebyś myslał, ze musisz są ze mną ożenić. \n"I didn't wantfeminine you to thinkmasculine you had to marrymasculine me." |
| Source | So then you confronted Derek. |
| Context | I am talking to a woman |
| Reference | A wiec doprowadziłs do konfrontacje z Derekiem. \n"So then you ledfeminine to a confrontation with Derek." |
| Baseline | Wiec wtedy skonfrontowale są徐 Derekiem. \n"So then you confrontedmasculine Derek." |
| MTCUE | Wiec skonfrontowale są徐 Derekiem. \n"So then you confrontedfeminine Derek." |
| IWSLT22 | |
| Source | I got a hundred colours in your city. |
| MTCUE (formal) | Ich habe 100 Farben in Ihrer Stadt. |
| MTCUE (informal) | Ich hab 100 Farben in deiner Stadt. |
| Context | A: i do at times . i run and swim more than anything . B: cool. i play soccer, but this weekend i'm going to watch a movie. A: soccer is fun. what movie are you going to see. B: i'm going to look for disney movie, maybe an old one. |
| Response | A: do you like star wars the new video game is coming out its gonna be great. |
| Path | run, play soccer, fun, movie, video |
| Relations | _hasprerequisite, motivatedbygoal,_usedfor,_isa |
| Full Path | run is a dependency of play soccer motivated by goal fun uses movie has a specific instance star wars |
| Dataset | Train | Dev | Test |
| ConvAI2 | 60150 | 10260 | 751 |
| DailyDialog | 17425 | 1524 | 964 |
| OTTers | 1876 | 946 | 855 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.23 | 2.81 | 0.21 | 0.18 | 2.93 | 0.23 |
| DKRN(Qin et al., 2020) | 0.39 | 3.24 | 0.33 | 0.32 | 3.62 | 0.33 |
| CKC(Zhong et al., 2021) | 0.41 | 4.08 | 0.35 | 0.35 | 4.24 | 0.28 |
| CODA-multi(Gupta et al., 2022) | 0.81 | 2.73 | 0.24 | 0.74 | 2.88 | 0.51 |
| TopKG(Yang et al., 2022) | 0.49 | 3.95 | 0.31 | 0.45 | 4.13 | 0.3 |
| MTGP | 0.95 | 3.26 | 0.40 | 0.92 | 1.96 | 0.31 |
| MTGP-noedge | 0.91 | 2.67 | 0.29 | 0.89 | 3.01 | 0.27 |
| MTGP-kbpath | 0.75 | 3.03 | 0.25 | 0.72 | 3.24 | 0.21 |
| MTGP-notarget | 0.85 | 3.10 | 0.36 | 0.84 | 3.03 | 0.32 |
| MTGP-upper | 0.89 | 2.84 | 0.32 | 0.87 | 2.73 | 0.29 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.15 | 3.66 | 0.22 | 0.19 | 3.94 | 0.19 |
| DKRN(Qin et al., 2020) | 0.28 | 3.89 | 0.27 | 0.32 | 3.15 | 0.30 |
| CKC(Zhong et al., 2021) | 0.31 | 4.69 | 0.26 | 0.36 | 4.25 | 0.29 |
| CODA-multi(Gupta et al., 2022) | 0.69 | 4.21 | 0.48 | 0.65 | 3.98 | 0.29 |
| TopKG(Yang et al., 2022) | 0.38 | 4.25 | 0.36 | 0.33 | 4.02 | 0.34 |
| MTGP | 0.82 | 4.23 | 0.33 | 0.73 | 2.46 | 0.30 |
| MTGP-noedge | 0.74 | 3.81 | 0.31 | 0.86 | 3.54 | 0.27 |
| MTGP-kbpath | 0.68 | 3.93 | 0.30 | 0.73 | 3.33 | 0.29 |
| MTGP-notarget | 0.75 | 3.89 | 0.29 | 0.89 | 3.10 | 0.30 |
| MTGP-upper | 0.78 | 3.73 | 0.28 | 0.83 | 2.84 | 0.29 |
| Connection | Valid Entity | Triple in Cpnet | Scores | ||
| sum score | best score | max score | |||
| 99.33 | 99.69 | 54.67 | 33.13 | 22.21 | 57.38 |
| Method | BLEU | METEOR | ROUGE-L | BertScore | TC |
| MultiGen | 6.22 | 12.53 | 28.14 | 40.03 | 27.82 |
| DKRN | 3.43 | 12.24 | 24.45 | 36.13 | 28.32 |
| CKC | 2.8 | 11.16 | 23.2 | 35.23 | 21.5 |
| TopKG | 3.31 | 12.32 | 28.54 | 38.13 | 28.32 |
| CODA | 5.02 | 12.63 | 25.91 | 38.02 | 36.72 |
| MTGP | 5.92 | 12.54 | 27.32 | 38.32 | 36.90 |
| MTGP-noedge | 4.40 | 12.46 | 25.13 | 37.85 | 32.73 |
| MTGP-kbpath | 4.23 | 12.32 | 26.07 | 37.42 | 35.72 |
| MTGP-notarget | 4.01 | 12.54 | 25.53 | 38.63 | 33.84 |
| MTGP-upper | 4.13 | 12.52 | 26.96 | 38.25 | 35.70 |
| Relationship | ConvAI2 | DailtDialog | ||
| No User | User | No User | User | |
| Turns = Path Len. | 208 | 13 | 149 | 14 |
| Turns < Path Len. | 481 | 62 | 671 | 73 |
| Turns > Path Len. | 62 | 676 | 144 | 877 |
| Avg. Path Len. | 2.60 | 2.60 | 3.15 | 3.15 |
| Avg. Turns | 1.96 | 3.26 | 2.46 | 4.23 |
| Source Sentence: i also like black and blue and pink great colors.\nTarget Sentence: don’t you mean the sunrise ? i like to walk.\nhelps with my dieting.\nGlobal Target: walk\nPath: pink is distinct from blue is a property of wind makes someone\nwant to fly kite has prequisite go outside is a dependency of walk | Source Sentence: hi how are you ? what is your favorite food ?\nTarget Sentence: i work nights at the bar now and wanna do\nnight shift as a nurse.\nGlobal Target: night\nPath: food is at location dinner is the location which has drinks is\na dependency of becoming inebriated used for relaxing uses night | ||
| Case on ConvA12 with No User | Case on ConvA12 with No User | ||
| MTGP | A: blue is a great color. my dog and me go hiking.\nB: cool. i like watching the wind blow.\nA: i like flying kites.\nB: flying kite? that’s my favorite thing.\nA: i do not like planes. being outside is cool.\nB: the sun is nice. i like to walk. | MTGP | A: hello, i am great. my favorite dinner is a nice steak.\nB: sounds very complimentary. what do you drink?\nA: i don’t drink, it probably can become inebriated.\nB: sad. i love tea. calming my soul.\nA: yes. evening drinks are a relaxing thing.\nB: my night is spent relaxing after a long day of hiking. |
| CODA-multi | A: i’ve to go to the grocery store and buy groceries.\nB: i like to go to the beach and hike. do you go outside?\nA: i not walk a lot. i like to go out to eat. | CODA-multi | A: i am in a relationship with my nieces and\nnephews. what do you do for fun?\n... (get trapped) |
| Case on ConvA12 with User | Case on ConvA12 with User | ||
| MTGP | A: i love purple. what is your favorite food?\nB: blue my fav. what about you?\nA: i like red, reminds me of summer time.\nB: nice. summer is nice. i like the wind.\nA: me too! i like the color yellow.\nB: yellow is a nice color. i fly kites.\nA: cool. i fly kites for a living.\nB: what kind of things do you like to do outside?\nA: i like to go out to eat.\nB: that is cool. i like walking outside. | MTGP | A: hi i’m great how are you\nB: i am fine do you like to eat dinner\nA: i do. i’m in a wheel chair.\nB: i’m drinking a diet coke and eating snacks.\nA: sounds like you are having a rough day.\nB: yeah, i am getting inebriated.\nA: i am sorry. do you have any hobbies?\nB: yes, i like to ride my bike to relax.\nA: i love to do that too.\nB: i work nights at mcdonalds. |
| CODA-multi | A: that is so cool! i like to read a lot.\nB: i am not a big walker, but i like to go outside sometimes. | CODA-multi | A: what is your favorite food?\nB: i love night shifts. what do you do for fun? |
| Context | A: i do at times . i run and swim more than anything . B: cool. i play soccer, but this weekend i'm going to watch a movie. A: soccer is fun. what movie are you going to see. B: i'm going to look for disney movie, maybe an old one. |
| Response | A: do you like star wars the new video game is coming out its gonna be great. |
| Path | run, play soccer, fun, movie, video |
| Relations | _hasprerequisite, motivatedbygoal,_usedfor,_isa |
| Full Path | run is a dependency of play soccer motivated by goal fun uses movie has a specific instance star wars |
| Dataset | Train | Dev | Test |
| ConvAI2 | 60150 | 10260 | 751 |
| DailyDialog | 17425 | 1524 | 964 |
| OTTers | 1876 | 946 | 855 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.23 | 2.81 | 0.21 | 0.18 | 2.93 | 0.23 |
| DKRN(Qin et al., 2020) | 0.39 | 3.24 | 0.33 | 0.32 | 3.62 | 0.33 |
| CKC(Zhong et al., 2021) | 0.41 | 4.08 | 0.35 | 0.35 | 4.24 | 0.28 |
| CODA-multi(Gupta et al., 2022) | 0.81 | 2.73 | 0.24 | 0.74 | 2.88 | 0.51 |
| TopKG(Yang et al., 2022) | 0.49 | 3.95 | 0.31 | 0.45 | 4.13 | 0.3 |
| MTGP | 0.95 | 3.26 | 0.40 | 0.92 | 1.96 | 0.31 |
| MTGP-noedge | 0.91 | 2.67 | 0.29 | 0.89 | 3.01 | 0.27 |
| MTGP-kbpath | 0.75 | 3.03 | 0.25 | 0.72 | 3.24 | 0.21 |
| MTGP-notarget | 0.85 | 3.10 | 0.36 | 0.84 | 3.03 | 0.32 |
| MTGP-upper | 0.89 | 2.84 | 0.32 | 0.87 | 2.73 | 0.29 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.15 | 3.66 | 0.22 | 0.19 | 3.94 | 0.19 |
| DKRN(Qin et al., 2020) | 0.28 | 3.89 | 0.27 | 0.32 | 3.15 | 0.30 |
| CKC(Zhong et al., 2021) | 0.31 | 4.69 | 0.26 | 0.36 | 4.25 | 0.29 |
| CODA-multi(Gupta et al., 2022) | 0.69 | 4.21 | 0.48 | 0.65 | 3.98 | 0.29 |
| TopKG(Yang et al., 2022) | 0.38 | 4.25 | 0.36 | 0.33 | 4.02 | 0.34 |
| MTGP | 0.82 | 4.23 | 0.33 | 0.73 | 2.46 | 0.30 |
| MTGP-noedge | 0.74 | 3.81 | 0.31 | 0.86 | 3.54 | 0.27 |
| MTGP-kbpath | 0.68 | 3.93 | 0.30 | 0.73 | 3.33 | 0.29 |
| MTGP-notarget | 0.75 | 3.89 | 0.29 | 0.89 | 3.10 | 0.30 |
| MTGP-upper | 0.78 | 3.73 | 0.28 | 0.83 | 2.84 | 0.29 |
| Connection | Valid Entity | Triple in Cpnet | Scores | ||
| sum score | best score | max score | |||
| 99.33 | 99.69 | 54.67 | 33.13 | 22.21 | 57.38 |
| Method | BLEU | METEOR | ROUGE-L | BertScore | TC |
| MultiGen | 6.22 | 12.53 | 28.14 | 40.03 | 27.82 |
| DKRN | 3.43 | 12.24 | 24.45 | 36.13 | 28.32 |
| CKC | 2.8 | 11.16 | 23.2 | 35.23 | 21.5 |
| TopKG | 3.31 | 12.32 | 28.54 | 38.13 | 28.32 |
| CODA | 5.02 | 12.63 | 25.91 | 38.02 | 36.72 |
| MTGP | 5.92 | 12.54 | 27.32 | 38.32 | 36.90 |
| MTGP-noedge | 4.40 | 12.46 | 25.13 | 37.85 | 32.73 |
| MTGP-kbpath | 4.23 | 12.32 | 26.07 | 37.42 | 35.72 |
| MTGP-notarget | 4.01 | 12.54 | 25.53 | 38.63 | 33.84 |
| MTGP-upper | 4.13 | 12.52 | 26.96 | 38.25 | 35.70 |
| Relationship | ConvAI2 | DailtDialog | ||
| No User | User | No User | User | |
| Turns = Path Len. | 208 | 13 | 149 | 14 |
| Turns < Path Len. | 481 | 62 | 671 | 73 |
| Turns > Path Len. | 62 | 676 | 144 | 877 |
| Avg. Path Len. | 2.60 | 2.60 | 3.15 | 3.15 |
| Avg. Turns | 1.96 | 3.26 | 2.46 | 4.23 |
| Source Sentence: i also like black and blue and pink great colors.\nTarget Sentence: don’t you mean the sunrise ? i like to walk.\nhelps with my dieting.\nGlobal Target: walk\nPath: pink is distinct from blue is a property of wind makes someone\nwant to fly kite has prequisite go outside is a dependency of walk | Source Sentence: hi how are you ? what is your favorite food ?\nTarget Sentence: i work nights at the bar now and wanna do\nnight shift as a nurse.\nGlobal Target: night\nPath: food is at location dinner is the location which has drinks is\na dependency of becoming inebriated used for relaxing uses night | ||
| Case on ConvA12 with No User | Case on ConvA12 with No User | ||
| MTGP | A: blue is a great color. my dog and me go hiking.\nB: cool. i like watching the wind blow.\nA: i like flying kites.\nB: flying kite? that’s my favorite thing.\nA: i do not like planes. being outside is cool.\nB: the sun is nice. i like to walk. | MTGP | A: hello, i am great. my favorite dinner is a nice steak.\nB: sounds very complimentary. what do you drink?\nA: i don’t drink, it probably can become inebriated.\nB: sad. i love tea. calming my soul.\nA: yes. evening drinks are a relaxing thing.\nB: my night is spent relaxing after a long day of hiking. |
| CODA-multi | A: i’ve to go to the grocery store and buy groceries.\nB: i like to go to the beach and hike. do you go outside?\nA: i not walk a lot. i like to go out to eat. | CODA-multi | A: i am in a relationship with my nieces and\nnephews. what do you do for fun?\n... (get trapped) |
| Case on ConvA12 with User | Case on ConvA12 with User | ||
| MTGP | A: i love purple. what is your favorite food?\nB: blue my fav. what about you?\nA: i like red, reminds me of summer time.\nB: nice. summer is nice. i like the wind.\nA: me too! i like the color yellow.\nB: yellow is a nice color. i fly kites.\nA: cool. i fly kites for a living.\nB: what kind of things do you like to do outside?\nA: i like to go out to eat.\nB: that is cool. i like walking outside. | MTGP | A: hi i’m great how are you\nB: i am fine do you like to eat dinner\nA: i do. i’m in a wheel chair.\nB: i’m drinking a diet coke and eating snacks.\nA: sounds like you are having a rough day.\nB: yeah, i am getting inebriated.\nA: i am sorry. do you have any hobbies?\nB: yes, i like to ride my bike to relax.\nA: i love to do that too.\nB: i work nights at mcdonalds. |
| CODA-multi | A: that is so cool! i like to read a lot.\nB: i am not a big walker, but i like to go outside sometimes. | CODA-multi | A: what is your favorite food?\nB: i love night shifts. what do you do for fun? |
| Context | A: i do at times . i run and swim more than anything . B: cool. i play soccer, but this weekend i'm going to watch a movie. A: soccer is fun. what movie are you going to see. B: i'm going to look for disney movie, maybe an old one. |
| Response | A: do you like star wars the new video game is coming out its gonna be great. |
| Path | run, play soccer, fun, movie, video |
| Relations | _hasprerequisite, motivatedbygoal,_usedfor,_isa |
| Full Path | run is a dependency of play soccer motivated by goal fun uses movie has a specific instance star wars |
| Dataset | Train | Dev | Test |
| ConvAI2 | 60150 | 10260 | 751 |
| DailyDialog | 17425 | 1524 | 964 |
| OTTers | 1876 | 946 | 855 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.23 | 2.81 | 0.21 | 0.18 | 2.93 | 0.23 |
| DKRN(Qin et al., 2020) | 0.39 | 3.24 | 0.33 | 0.32 | 3.62 | 0.33 |
| CKC(Zhong et al., 2021) | 0.41 | 4.08 | 0.35 | 0.35 | 4.24 | 0.28 |
| CODA-multi(Gupta et al., 2022) | 0.81 | 2.73 | 0.24 | 0.74 | 2.88 | 0.51 |
| TopKG(Yang et al., 2022) | 0.49 | 3.95 | 0.31 | 0.45 | 4.13 | 0.3 |
| MTGP | 0.95 | 3.26 | 0.40 | 0.92 | 1.96 | 0.31 |
| MTGP-noedge | 0.91 | 2.67 | 0.29 | 0.89 | 3.01 | 0.27 |
| MTGP-kbpath | 0.75 | 3.03 | 0.25 | 0.72 | 3.24 | 0.21 |
| MTGP-notarget | 0.85 | 3.10 | 0.36 | 0.84 | 3.03 | 0.32 |
| MTGP-upper | 0.89 | 2.84 | 0.32 | 0.87 | 2.73 | 0.29 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.15 | 3.66 | 0.22 | 0.19 | 3.94 | 0.19 |
| DKRN(Qin et al., 2020) | 0.28 | 3.89 | 0.27 | 0.32 | 3.15 | 0.30 |
| CKC(Zhong et al., 2021) | 0.31 | 4.69 | 0.26 | 0.36 | 4.25 | 0.29 |
| CODA-multi(Gupta et al., 2022) | 0.69 | 4.21 | 0.48 | 0.65 | 3.98 | 0.29 |
| TopKG(Yang et al., 2022) | 0.38 | 4.25 | 0.36 | 0.33 | 4.02 | 0.34 |
| MTGP | 0.82 | 4.23 | 0.33 | 0.73 | 2.46 | 0.30 |
| MTGP-noedge | 0.74 | 3.81 | 0.31 | 0.86 | 3.54 | 0.27 |
| MTGP-kbpath | 0.68 | 3.93 | 0.30 | 0.73 | 3.33 | 0.29 |
| MTGP-notarget | 0.75 | 3.89 | 0.29 | 0.89 | 3.10 | 0.30 |
| MTGP-upper | 0.78 | 3.73 | 0.28 | 0.83 | 2.84 | 0.29 |
| Connection | Valid Entity | Triple in Cpnet | Scores | ||
| sum score | best score | max score | |||
| 99.33 | 99.69 | 54.67 | 33.13 | 22.21 | 57.38 |
| Method | BLEU | METEOR | ROUGE-L | BertScore | TC |
| MultiGen | 6.22 | 12.53 | 28.14 | 40.03 | 27.82 |
| DKRN | 3.43 | 12.24 | 24.45 | 36.13 | 28.32 |
| CKC | 2.8 | 11.16 | 23.2 | 35.23 | 21.5 |
| TopKG | 3.31 | 12.32 | 28.54 | 38.13 | 28.32 |
| CODA | 5.02 | 12.63 | 25.91 | 38.02 | 36.72 |
| MTGP | 5.92 | 12.54 | 27.32 | 38.32 | 36.90 |
| MTGP-noedge | 4.40 | 12.46 | 25.13 | 37.85 | 32.73 |
| MTGP-kbpath | 4.23 | 12.32 | 26.07 | 37.42 | 35.72 |
| MTGP-notarget | 4.01 | 12.54 | 25.53 | 38.63 | 33.84 |
| MTGP-upper | 4.13 | 12.52 | 26.96 | 38.25 | 35.70 |
| Relationship | ConvAI2 | DailtDialog | ||
| No User | User | No User | User | |
| Turns = Path Len. | 208 | 13 | 149 | 14 |
| Turns < Path Len. | 481 | 62 | 671 | 73 |
| Turns > Path Len. | 62 | 676 | 144 | 877 |
| Avg. Path Len. | 2.60 | 2.60 | 3.15 | 3.15 |
| Avg. Turns | 1.96 | 3.26 | 2.46 | 4.23 |
| Source Sentence: i also like black and blue and pink great colors. +Target Sentence: don’t you mean the sunrise ? i like to walk. +helps with my dieting. +Global Target: walk +Path: pink is distinct from blue is a property of wind makes someone +want to fly kite has prequisite go outside is a dependency of walk | Source Sentence: hi how are you ? what is your favorite food ? +Target Sentence: i work nights at the bar now and wanna do +night shift as a nurse. +Global Target: night +Path: food is at location dinner is the location which has drinks is +a dependency of becoming inebriated used for relaxing uses night | ||
| Case on ConvA12 with No User | Case on ConvA12 with No User | ||
| MTGP | A: blue is a great color. my dog and me go hiking. +B: cool. i like watching the wind blow. +A: i like flying kites. +B: flying kite? that’s my favorite thing. +A: i do not like planes. being outside is cool. +B: the sun is nice. i like to walk. | MTGP | A: hello, i am great. my favorite dinner is a nice steak. +B: sounds very complimentary. what do you drink? +A: i don’t drink, it probably can become inebriated. +B: sad. i love tea. calming my soul. +A: yes. evening drinks are a relaxing thing. +B: my night is spent relaxing after a long day of hiking. |
| CODA-multi | A: i’ve to go to the grocery store and buy groceries. +B: i like to go to the beach and hike. do you go outside? +A: i not walk a lot. i like to go out to eat. | CODA-multi | A: i am in a relationship with my nieces and +nephews. what do you do for fun? +... (get trapped) |
| Case on ConvA12 with User | Case on ConvA12 with User | ||
| MTGP | A: i love purple. what is your favorite food? +B: blue my fav. what about you? +A: i like red, reminds me of summer time. +B: nice. summer is nice. i like the wind. +A: me too! i like the color yellow. +B: yellow is a nice color. i fly kites. +A: cool. i fly kites for a living. +B: what kind of things do you like to do outside? +A: i like to go out to eat. +B: that is cool. i like walking outside. | MTGP | A: hi i’m great how are you +B: i am fine do you like to eat dinner +A: i do. i’m in a wheel chair. +B: i’m drinking a diet coke and eating snacks. +A: sounds like you are having a rough day. +B: yeah, i am getting inebriated. +A: i am sorry. do you have any hobbies? +B: yes, i like to ride my bike to relax. +A: i love to do that too. +B: i work nights at mcdonalds. |
| CODA-multi | A: that is so cool! i like to read a lot. +B: i am not a big walker, but i like to go outside sometimes. | CODA-multi | A: what is your favorite food? +B: i love night shifts. what do you do for fun? |
| Context | A: i do at times . i run and swim more than anything . B: cool. i play soccer, but this weekend i'm going to watch a movie. A: soccer is fun. what movie are you going to see. B: i'm going to look for disney movie, maybe an old one. |
| Response | A: do you like star wars the new video game is coming out its gonna be great. |
| Path | run, play soccer, fun, movie, video |
| Relations | _hasprerequisite, motivatedbygoal,_usedfor,_isa |
| Full Path | run is a dependency of play soccer motivated by goal fun uses movie has a specific instance star wars |
| Dataset | Train | Dev | Test |
| ConvAI2 | 60150 | 10260 | 751 |
| DailyDialog | 17425 | 1524 | 964 |
| OTTers | 1876 | 946 | 855 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.23 | 2.81 | 0.21 | 0.18 | 2.93 | 0.23 |
| DKRN(Qin et al., 2020) | 0.39 | 3.24 | 0.33 | 0.32 | 3.62 | 0.33 |
| CKC(Zhong et al., 2021) | 0.41 | 4.08 | 0.35 | 0.35 | 4.24 | 0.28 |
| CODA-multi(Gupta et al., 2022) | 0.81 | 2.73 | 0.24 | 0.74 | 2.88 | 0.51 |
| TopKG(Yang et al., 2022) | 0.49 | 3.95 | 0.31 | 0.45 | 4.13 | 0.3 |
| MTGP | 0.95 | 3.26 | 0.40 | 0.92 | 1.96 | 0.31 |
| MTGP-noedge | 0.91 | 2.67 | 0.29 | 0.89 | 3.01 | 0.27 |
| MTGP-kbpath | 0.75 | 3.03 | 0.25 | 0.72 | 3.24 | 0.21 |
| MTGP-notarget | 0.85 | 3.10 | 0.36 | 0.84 | 3.03 | 0.32 |
| MTGP-upper | 0.89 | 2.84 | 0.32 | 0.87 | 2.73 | 0.29 |
| Method | User | No User | ||||
| Succ. | Turns | Coh. | Succ. | Turns | Coh. | |
| MultiGen(Ji et al., 2020) | 0.15 | 3.66 | 0.22 | 0.19 | 3.94 | 0.19 |
| DKRN(Qin et al., 2020) | 0.28 | 3.89 | 0.27 | 0.32 | 3.15 | 0.30 |
| CKC(Zhong et al., 2021) | 0.31 | 4.69 | 0.26 | 0.36 | 4.25 | 0.29 |
| CODA-multi(Gupta et al., 2022) | 0.69 | 4.21 | 0.48 | 0.65 | 3.98 | 0.29 |
| TopKG(Yang et al., 2022) | 0.38 | 4.25 | 0.36 | 0.33 | 4.02 | 0.34 |
| MTGP | 0.82 | 4.23 | 0.33 | 0.73 | 2.46 | 0.30 |
| MTGP-noedge | 0.74 | 3.81 | 0.31 | 0.86 | 3.54 | 0.27 |
| MTGP-kbpath | 0.68 | 3.93 | 0.30 | 0.73 | 3.33 | 0.29 |
| MTGP-notarget | 0.75 | 3.89 | 0.29 | 0.89 | 3.10 | 0.30 |
| MTGP-upper | 0.78 | 3.73 | 0.28 | 0.83 | 2.84 | 0.29 |
| Connection | Valid Entity | Triple in Cpnet | Scores | ||
| sum score | best score | max score | |||
| 99.33 | 99.69 | 54.67 | 33.13 | 22.21 | 57.38 |
| Method | BLEU | METEOR | ROUGE-L | BertScore | TC |
| MultiGen | 6.22 | 12.53 | 28.14 | 40.03 | 27.82 |
| DKRN | 3.43 | 12.24 | 24.45 | 36.13 | 28.32 |
| CKC | 2.8 | 11.16 | 23.2 | 35.23 | 21.5 |
| TopKG | 3.31 | 12.32 | 28.54 | 38.13 | 28.32 |
| CODA | 5.02 | 12.63 | 25.91 | 38.02 | 36.72 |
| MTGP | 5.92 | 12.54 | 27.32 | 38.32 | 36.90 |
| MTGP-noedge | 4.40 | 12.46 | 25.13 | 37.85 | 32.73 |
| MTGP-kbpath | 4.23 | 12.32 | 26.07 | 37.42 | 35.72 |
| MTGP-notarget | 4.01 | 12.54 | 25.53 | 38.63 | 33.84 |
| MTGP-upper | 4.13 | 12.52 | 26.96 | 38.25 | 35.70 |
| Relationship | ConvAI2 | DailtDialog | ||
| No User | User | No User | User | |
| Turns = Path Len. | 208 | 13 | 149 | 14 |
| Turns < Path Len. | 481 | 62 | 671 | 73 |
| Turns > Path Len. | 62 | 676 | 144 | 877 |
| Avg. Path Len. | 2.60 | 2.60 | 3.15 | 3.15 |
| Avg. Turns | 1.96 | 3.26 | 2.46 | 4.23 |
| Source Sentence: i also like black and blue and pink great colors.\nTarget Sentence: don’t you mean the sunrise ? i like to walk.\nhelps with my dieting.\nGlobal Target: walk\nPath: pink is distinct from blue is a property of wind makes someone\nwant to fly kite has prequisite go outside is a dependency of walk | Source Sentence: hi how are you ? what is your favorite food ?\nTarget Sentence: i work nights at the bar now and wanna do\nnight shift as a nurse.\nGlobal Target: night\nPath: food is at location dinner is the location which has drinks is\na dependency of becoming inebriated used for relaxing uses night | ||
| Case on ConvA12 with No User | Case on ConvA12 with No User | ||
| MTGP | A: blue is a great color. my dog and me go hiking.\nB: cool. i like watching the wind blow.\nA: i like flying kites.\nB: flying kite? that’s my favorite thing.\nA: i do not like planes. being outside is cool.\nB: the sun is nice. i like to walk. | MTGP | A: hello, i am great. my favorite dinner is a nice steak.\nB: sounds very complimentary. what do you drink?\nA: i don’t drink, it probably can become inebriated.\nB: sad. i love tea. calming my soul.\nA: yes. evening drinks are a relaxing thing.\nB: my night is spent relaxing after a long day of hiking. |
| CODA-multi | A: i’ve to go to the grocery store and buy groceries.\nB: i like to go to the beach and hike. do you go outside?\nA: i not walk a lot. i like to go out to eat. | CODA-multi | A: i am in a relationship with my nieces and\nnephews. what do you do for fun?\n... (get trapped) |
| Case on ConvA12 with User | Case on ConvA12 with User | ||
| MTGP | A: i love purple. what is your favorite food?\nB: blue my fav. what about you?\nA: i like red, reminds me of summer time.\nB: nice. summer is nice. i like the wind.\nA: me too! i like the color yellow.\nB: yellow is a nice color. i fly kites.\nA: cool. i fly kites for a living.\nB: what kind of things do you like to do outside?\nA: i like to go out to eat.\nB: that is cool. i like walking outside. | MTGP | A: hi i’m great how are you\nB: i am fine do you like to eat dinner\nA: i do. i’m in a wheel chair.\nB: i’m drinking a diet coke and eating snacks.\nA: sounds like you are having a rough day.\nB: yeah, i am getting inebriated.\nA: i am sorry. do you have any hobbies?\nB: yes, i like to ride my bike to relax.\nA: i love to do that too.\nB: i work nights at mcdonalds. |
| CODA-multi | A: that is so cool! i like to read a lot.\nB: i am not a big walker, but i like to go outside sometimes. | CODA-multi | A: what is your favorite food?\nB: i love night shifts. what do you do for fun? |
| Dataset | Deductive | Inductive | Defeasible |
| RuleTaker | ✓ | ||
| LogicNLI | ✓ | ||
| LogiQA | ✓ | ||
| ProofWriter | ✓ | ||
| CLUTRR | ✓ | ||
| HotpotQA | ✓ | ||
| QuaRTz | ✓ | ||
| δ-NLI | ✓ | ||
| \\( \\mathcal{MTR}(Ours) \\) | ✓ | ✓ | ✓ |
| Statistics | D-MTR | F-MTR |
| #Instances | 30k | 2k |
| Avg. Length | 76.6 | 43.2 |
| Max. Length | 189 | 121 |
| #Hop | ≤ 5 | ≤ 5 |
| Data(Accuracy) | Models | ||||||
| Sequence | Type | BiLSTM-Attention | BiLSTM-Mean | RN | MAC | BERT | BERT-LSTM |
| Sequential | ① | 0.13±0.03 | 0.08±0.02 | 0.12±0.03 | 0.08±0.02 | 0.14±0.01 | 0.09±0.03 |
| ② | 0.23±0.02 | 0.21±0.03 | 0.10±0.02 | 0.24±0.01 | 0.13±0.02 | 0.23±0.02 | |
| ③ | 0.13±0.02 | 0.09±0.02 | 0.08±0.02 | 0.10±0.03 | 0.12±0.02 | 0.15±0.01 | |
| ④ | 0.46±0.02 | 0.38±0.03 | 0.14±0.03 | 0.37±0.05 | 0.10±0.03 | 0.40±0.04 | |
| ⑤ | 0.73±0.05 | 0.73±0.06 | 0.60±0.05 | 0.63±0.05 | 0.17±0.02 | 0.65±0.05 | |
| Average | 0.34±0.03 | 0.30±0.03 | 0.22±0.03 | 0.29±0.03 | 0.13±0.03 | 0.33±0.02 | |
| Random | ① | 0.26±0.06 | 0.23±0.06 | 0.09±0.03 | 0.28±0.02 | 0.11±0.01 | 0.26±0.03 |
| ② | 0.17±0.02 | 0.16±0.02 | 0.10±0.03 | 0.16±0.02 | 0.10±0.02 | 0.18±0.02 | |
| ③ | 0.10±0.04 | 0.10±0.04 | 0.08±0.02 | 0.09±0.01 | 0.10±0.02 | 0.09±0.03 | |
| ④ | 0.36±0.03 | 0.31±0.03 | 0.10±0.02 | 0.32±0.03 | 0.08±0.01 | 0.32±0.04 | |
| ⑤ | 0.64±0.06 | 0.62±0.06 | 0.49±0.03 | 0.55±0.04 | 0.15±0.03 | 0.62±0.05 | |
| Average | 0.31±0.04 | 0.28±0.04 | 0.17±0.03 | 0.28±0.02 | 0.11±0.02 | 0.29±0.03 | |
| CLUTRR | Average | 0.61±0.08 | 0.59±0.08 | 0.54±0.07 | 0.61±0.06 | 0.30±0.07 | 0.56±0.05 |
| Accuracy | (w/o) Supplementary Facts | |
| Models | w | o |
| BiLSTM-Attention | 0.04 | 0.09 |
| BiLSTM-Mean | 0.03 | 0.08 |
| RN | 0.02 | 0.04 |
| MAC | 0.03 | 0.08 |
| BERT | 0.05 | 0.03 |
| BERT-LSTM | 0.08 | 0.07 |
| Dataset | Deductive | Inductive | Defeasible |
| RuleTaker | ✓ | ||
| LogicNLI | ✓ | ||
| LogiQA | ✓ | ||
| ProofWriter | ✓ | ||
| CLUTRR | ✓ | ||
| HotpotQA | ✓ | ||
| QuaRTz | ✓ | ||
| δ-NLI | ✓ | ||
| \\( \\mathcal{MTR}(Ours) \\) | ✓ | ✓ | ✓ |
| Statistics | D-MTR | F-MTR |
| #Instances | 30k | 2k |
| Avg. Length | 76.6 | 43.2 |
| Max. Length | 189 | 121 |
| #Hop | ≤ 5 | ≤ 5 |
| Data(Accuracy) | Models | ||||||
| Sequence | Type | BiLSTM-Attention | BiLSTM-Mean | RN | MAC | BERT | BERT-LSTM |
| Sequential | ① | 0.13±0.03 | 0.08±0.02 | 0.12±0.03 | 0.08±0.02 | 0.14±0.01 | 0.09±0.03 |
| ② | 0.23±0.02 | 0.21±0.03 | 0.10±0.02 | 0.24±0.01 | 0.13±0.02 | 0.23±0.02 | |
| ③ | 0.13±0.02 | 0.09±0.02 | 0.08±0.02 | 0.10±0.03 | 0.12±0.02 | 0.15±0.01 | |
| ④ | 0.46±0.02 | 0.38±0.03 | 0.14±0.03 | 0.37±0.05 | 0.10±0.03 | 0.40±0.04 | |
| ⑤ | 0.73±0.05 | 0.73±0.06 | 0.60±0.05 | 0.63±0.05 | 0.17±0.02 | 0.65±0.05 | |
| Average | 0.34±0.03 | 0.30±0.03 | 0.22±0.03 | 0.29±0.03 | 0.13±0.03 | 0.33±0.02 | |
| Random | ① | 0.26±0.06 | 0.23±0.06 | 0.09±0.03 | 0.28±0.02 | 0.11±0.01 | 0.26±0.03 |
| ② | 0.17±0.02 | 0.16±0.02 | 0.10±0.03 | 0.16±0.02 | 0.10±0.02 | 0.18±0.02 | |
| ③ | 0.10±0.04 | 0.10±0.04 | 0.08±0.02 | 0.09±0.01 | 0.10±0.02 | 0.09±0.03 | |
| ④ | 0.36±0.03 | 0.31±0.03 | 0.10±0.02 | 0.32±0.03 | 0.08±0.01 | 0.32±0.04 | |
| ⑤ | 0.64±0.06 | 0.62±0.06 | 0.49±0.03 | 0.55±0.04 | 0.15±0.03 | 0.62±0.05 | |
| Average | 0.31±0.04 | 0.28±0.04 | 0.17±0.03 | 0.28±0.02 | 0.11±0.02 | 0.29±0.03 | |
| CLUTRR | Average | 0.61±0.08 | 0.59±0.08 | 0.54±0.07 | 0.61±0.06 | 0.30±0.07 | 0.56±0.05 |
| Accuracy | (w/o) Supplementary Facts | |
| Models | w | o |
| BiLSTM-Attention | 0.04 | 0.09 |
| BiLSTM-Mean | 0.03 | 0.08 |
| RN | 0.02 | 0.04 |
| MAC | 0.03 | 0.08 |
| BERT | 0.05 | 0.03 |
| BERT-LSTM | 0.08 | 0.07 |
| Dataset | Deductive | Inductive | Defeasible |
| RuleTaker | ✓ | ||
| LogicNLI | ✓ | ||
| LogiQA | ✓ | ||
| ProofWriter | ✓ | ||
| CLUTRR | ✓ | ||
| HotpotQA | ✓ | ||
| QuaRTz | ✓ | ||
| δ-NLI | ✓ | ||
| \( \mathcal{MTR}(Ours) \) | ✓ | ✓ | ✓ |
| Statistics | D-MTR | F-MTR |
| #Instances | 30k | 2k |
| Avg. Length | 76.6 | 43.2 |
| Max. Length | 189 | 121 |
| #Hop | ≤ 5 | ≤ 5 |
| Data(Accuracy) | Models | ||||||
| Sequence | Type | BiLSTM-Attention | BiLSTM-Mean | RN | MAC | BERT | BERT-LSTM |
| Sequential | ① | 0.13±0.03 | 0.08±0.02 | 0.12±0.03 | 0.08±0.02 | 0.14±0.01 | 0.09±0.03 |
| ② | 0.23±0.02 | 0.21±0.03 | 0.10±0.02 | 0.24±0.01 | 0.13±0.02 | 0.23±0.02 | |
| ③ | 0.13±0.02 | 0.09±0.02 | 0.08±0.02 | 0.10±0.03 | 0.12±0.02 | 0.15±0.01 | |
| ④ | 0.46±0.02 | 0.38±0.03 | 0.14±0.03 | 0.37±0.05 | 0.10±0.03 | 0.40±0.04 | |
| ⑤ | 0.73±0.05 | 0.73±0.06 | 0.60±0.05 | 0.63±0.05 | 0.17±0.02 | 0.65±0.05 | |
| Average | 0.34±0.03 | 0.30±0.03 | 0.22±0.03 | 0.29±0.03 | 0.13±0.03 | 0.33±0.02 | |
| Random | ① | 0.26±0.06 | 0.23±0.06 | 0.09±0.03 | 0.28±0.02 | 0.11±0.01 | 0.26±0.03 |
| ② | 0.17±0.02 | 0.16±0.02 | 0.10±0.03 | 0.16±0.02 | 0.10±0.02 | 0.18±0.02 | |
| ③ | 0.10±0.04 | 0.10±0.04 | 0.08±0.02 | 0.09±0.01 | 0.10±0.02 | 0.09±0.03 | |
| ④ | 0.36±0.03 | 0.31±0.03 | 0.10±0.02 | 0.32±0.03 | 0.08±0.01 | 0.32±0.04 | |
| ⑤ | 0.64±0.06 | 0.62±0.06 | 0.49±0.03 | 0.55±0.04 | 0.15±0.03 | 0.62±0.05 | |
| Average | 0.31±0.04 | 0.28±0.04 | 0.17±0.03 | 0.28±0.02 | 0.11±0.02 | 0.29±0.03 | |
| CLUTRR | Average | 0.61±0.08 | 0.59±0.08 | 0.54±0.07 | 0.61±0.06 | 0.30±0.07 | 0.56±0.05 |
| Accuracy | (w/o) Supplementary Facts | |
| Models | w | o |
| BiLSTM-Attention | 0.04 | 0.09 |
| BiLSTM-Mean | 0.03 | 0.08 |
| RN | 0.02 | 0.04 |
| MAC | 0.03 | 0.08 |
| BERT | 0.05 | 0.03 |
| BERT-LSTM | 0.08 | 0.07 |
| Dataset | Deductive | Inductive | Defeasible |
| RuleTaker | ✓ | ||
| LogicNLI | ✓ | ||
| LogiQA | ✓ | ||
| ProofWriter | ✓ | ||
| CLUTRR | ✓ | ||
| HotpotQA | ✓ | ||
| QuaRTz | ✓ | ||
| δ-NLI | ✓ | ||
| \\( \\mathcal{MTR}(Ours) \\) | ✓ | ✓ | ✓ |
| Statistics | D-MTR | F-MTR |
| #Instances | 30k | 2k |
| Avg. Length | 76.6 | 43.2 |
| Max. Length | 189 | 121 |
| #Hop | ≤ 5 | ≤ 5 |
| Data(Accuracy) | Models | ||||||
| Sequence | Type | BiLSTM-Attention | BiLSTM-Mean | RN | MAC | BERT | BERT-LSTM |
| Sequential | ① | 0.13±0.03 | 0.08±0.02 | 0.12±0.03 | 0.08±0.02 | 0.14±0.01 | 0.09±0.03 |
| ② | 0.23±0.02 | 0.21±0.03 | 0.10±0.02 | 0.24±0.01 | 0.13±0.02 | 0.23±0.02 | |
| ③ | 0.13±0.02 | 0.09±0.02 | 0.08±0.02 | 0.10±0.03 | 0.12±0.02 | 0.15±0.01 | |
| ④ | 0.46±0.02 | 0.38±0.03 | 0.14±0.03 | 0.37±0.05 | 0.10±0.03 | 0.40±0.04 | |
| ⑤ | 0.73±0.05 | 0.73±0.06 | 0.60±0.05 | 0.63±0.05 | 0.17±0.02 | 0.65±0.05 | |
| Average | 0.34±0.03 | 0.30±0.03 | 0.22±0.03 | 0.29±0.03 | 0.13±0.03 | 0.33±0.02 | |
| Random | ① | 0.26±0.06 | 0.23±0.06 | 0.09±0.03 | 0.28±0.02 | 0.11±0.01 | 0.26±0.03 |
| ② | 0.17±0.02 | 0.16±0.02 | 0.10±0.03 | 0.16±0.02 | 0.10±0.02 | 0.18±0.02 | |
| ③ | 0.10±0.04 | 0.10±0.04 | 0.08±0.02 | 0.09±0.01 | 0.10±0.02 | 0.09±0.03 | |
| ④ | 0.36±0.03 | 0.31±0.03 | 0.10±0.02 | 0.32±0.03 | 0.08±0.01 | 0.32±0.04 | |
| ⑤ | 0.64±0.06 | 0.62±0.06 | 0.49±0.03 | 0.55±0.04 | 0.15±0.03 | 0.62±0.05 | |
| Average | 0.31±0.04 | 0.28±0.04 | 0.17±0.03 | 0.28±0.02 | 0.11±0.02 | 0.29±0.03 | |
| CLUTRR | Average | 0.61±0.08 | 0.59±0.08 | 0.54±0.07 | 0.61±0.06 | 0.30±0.07 | 0.56±0.05 |
| Accuracy | (w/o) Supplementary Facts | |
| Models | w | o |
| BiLSTM-Attention | 0.04 | 0.09 |
| BiLSTM-Mean | 0.03 | 0.08 |
| RN | 0.02 | 0.04 |
| MAC | 0.03 | 0.08 |
| BERT | 0.05 | 0.03 |
| BERT-LSTM | 0.08 | 0.07 |
| Year | Class | Team | Points | Wins |
| 1979 | 350cc | yamaha | 3 | 0 |
| 1980 | 250cc | yamaha | 4 | 0 |
| 1982 | 250cc | yamaha | 4 | 0 |
| 1982 | 500cc | suzuki | 0 | 0 |
| 1983 | 250cc | yamaha | 14 | 0 |
| 1984 | 500cc | honda | 14 | 0 |
| 1985 | 250cc | romer-juchem | 29 | 0 |
| 1986 | 250cc | hb - honda | 10 | 0 |
| 1987 | 250cc | hb - honda | 108 | 1 |
| 1988 | 250cc | hb - honda | 158 | 0 |
| 1989 | 250cc | hb - honda | 190 | 2 |
| 1990 | 250cc | hb - honda | 52 | 0 |
| Module | Description |
| Filtering | Mfilter : (t, cr) → t' : takes a table t and a filtering criterion cr as input and outputs a table t' with rows where the criterion cr is satisfied. |
| Aggregation | Performs aggregation operations on a table. For example, Mmax : (t, c) → n is a max module that takes a table t and a column c as input and outputs the maximum number n in column c. |
| Boolean | Mbool : (t, cr) → b: takes a table t and a criterion cr as input and outputs a boolean b based on whether the criterion is satisfied. |
| Hop | Mhop : (r, c) → e: takes a row r and a column c as input; outputs the element e in (r, c) cell. |
| Graph-to-Text (WebNLG) |
| Triple → String\n(String, String) → String |
| Table-to-Text (LogicNLG) |
| Table → Table | Row | Number | Boolean\nRow → String | Number\nString | Number → Boolean\n(Table, Path) → String |
| Seen | BLEU Unseen | All | Seen | METEOR Unseen | All | ||
| supervised | MELBOURNE† | 54.5 | 33.2 | 45.1 | 41.0 | 33.0 | 37.0 |
| GPT-2-large† | 65.3 | 43.1 | 55.5 | 46.0 | 38.0 | 42.0 | |
| T5-large† | 64.9 | 54.0 | 59.9 | 46.0 | 43.0 | 44.0 | |
| Neural Pipeline‡ | - | - | 43.3 | - | - | 39.3 | |
| few-shot | Direct Prompting (k=1)∗ | 33.1±0.3 | 34.2±0.1 | 33.6±0.1 | 30.4±0.1 | 31.2±0.1 | 30.8±0.1 |
| Direct Prompting (k=5)∗ | 39.9±0.3 | 38.9±0.3 | 39.5±0.1 | 34.3±0.1 | 34.3±0.3 | 34.4±0.1 | |
| CoT Prompting (k=1)∗ | 22.2±0.2 | 14.9±0.2 | 18.0±0.1 | 22.3±0.1 | 22.9±0.2 | 22.6±0.1 | |
| MURMUR (k=1)∗ | 41.4±0.0 | 41.1±0.0 | 41.3±0.0 | 37.1±0.0 | 37.1±0.0 | 37.1±0.0 |
| DP | MURMUR | % Improve | |
| Omissions↓ | 1.64±0.06 | 0.73±0.01 | +24% |
| Hallucinations↓ | 0.77±0.03 | 0.43±0.03 | +9% |
| Disfluencies↓ | 0.14±0.05 | 0.30±0.04 | -4% |
| Module | Grammatical | Faithful | Correct |
| Surface Realization | 1.00 | 1.00 | 0.82 |
| Text Fusion | 0.90 | 0.72 | 0.64 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Field-Infusing† | 43.7 / 20.9 / 8.4 |
| BERT-TabGen† | 49.1 / 27.7 / 13.5 |
| GPT-TabGen† | 49.6 / 28.2 / 14.2 |
| GPT-Coarse-to-Fine† | 49.0 / 28.3 / 14.6 |
| DCVED† | 49.5 / 28.6 / 15.3 |
| Direct Prompting* | 37.2±0.4 / 18.8±0.2 / 8.6±0.2 |
| CoT Prompting* | 35.6±0.2 / 18.6±0.1 / 8.8±0.0 |
| BART + SR‡ | 39.2±0.2 / 20.6±0.2 / 9.5±0.0 |
| MURMUR‡ | 39.8±0.0 / 22.2±0.0 / 11.2±0.0 |
| -saliency* | 39.6±0.0 / 21.9±0.0 / 10.6±0.0 |
| Correct | Partial | Incorrect | Ungrammatical | Is Logical? | |
| Direct Prompting | 28.7±3.7 | 20.0±2.5 | 38.8±8.7 | 12.5±2.5 | 62.0±0.5 |
| MURMUR | 55.0±2.5 | 1.2±1.2 | 38.8±3.7 | 5.0±2.5 | 95.4±0.2 |
| Module Name | Input Data Type | Output Data Type | Description |
| filter_eq filter_not_eq | table, string, string | table | Returns a table with the rows where entry in the input column (second argument) is equal or not equal to the input value (third argument). |
| filtergreater filtergreater_eq filter LESSER filter LESSER_eq | table, string, number | table | Returns a table with the rows where a numerical column (second argument) is greater than or less than (or equal to) the input number (third argument). |
| filter_all | table, string | table | Returns the whole table. |
| arg_max arg_min | table, string | row | Returns the row with the minimum or maximum value for the input column (second argument). |
| max min avg sum | table, string | number | Returns the maximum, minimum, average or sum of numbers in the input column (second argument). |
| count | table | number | Returns the number of rows in the table. |
| all_eq all_NOT_eq | table, string, string | bool | Returns whether all entries in the input column are equal (or not equal to) the input value. |
| all_greater all-less all_greater_eq all-less_eq | table, string, number | bool | Returns whether all entries in the input column are greater than or less than (or equal to) the input number. |
| most_eq most_NOT_eq | table, string, string | bool | Returns whether most entries in the input column are equal (or not equal to) the input value. |
| most_greater most-less most_greater_eq most-less_eq | table, string, number | bool | Returns whether most entries in the input column are greater than or less than (or equal to) the corresponding number. |
| only | table | bool | Returns whether the table has exactly one row. |
| hop | row, string | string | Returns the entry corresponding to the input column in the row. |
| eq | string, string | bool | Returns whether the two inputs are equal or not. |
| BLEU | METEOR | |
| Direct Prompting (k=1) | 31.1±0.5 | 29.8±0.1 |
| Direct Prompting (k=5) | 38.3±0.4 | 33.6±0.1 |
| MURMUR (k=1) | 40.1±0.3 | 37.1±0.5 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Direct Prompting (k=1) | 37.0±0.2 / 18.9±0.1 / 8.5±0.1 |
| COT Prompting (k=1) | 36.5±0.1 / 18.9±0.1 / 8.7±0.3 |
| MURMUR (k=1) | 40.5±0.1 / 22.2±0.0 / 10.8±0.1 |
| Beam Size | BLEU-1 / BLEU-2 / BLEU-3 |
| 10 | 39.7 / 21.2 / 10.3 |
| 20 | 40.2 / 21.8 / 10.7 |
| 50 | 40.5 / 22.2 / 10.8 |
| 100 | 40.7 / 22.5 / 10.9 |
| # Gold Tables | # Gold Paths | # Samples (Pos/Neg/ All) | Acc. | LogicNLG BLEU-3 |
| 100 | 221 | 769/729/1498 | 76.16 | 10.8 |
| 200 | 443 | 1534/1457/2991 | 78.52 | 10.6 |
| 500 | 1085 | 3773/3633/7406 | 80.32 | 10.9 |
| 3000 | 7145 | 21.5k/20.5k/42.0k | 82.84 | 10.7 |
| Let's convert triples to sentences |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Let's convert triples to sentences step-by-step |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: AS Gubbio 1910 plays in Serie D. # Pietro Grasso is the leader of Italy. # ... # S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Input | Arlington,_Texas | isPartOf | Texas # Texas | largestCity | Houston # Texas | language | English_language |
| Direct Prompting | Texas is the second largest state in the United States of America. Its capital is Austin and the largest city is Houston. (Omissions = 2) |
| MURMUR | Houston is the largest city in Texas where English is the official language and Arlington, Texas is a city in Texas. (Omissions = 0) |
| Input | Hays_County_Texas | countySeat | San_Marcos,_Texas # Austin,_Texas | isPartOf | Hays_County,_Texas # Texas | language | Spanish_language # Austin,_Texas | isPartOf | Texas # Texas | largestCity | Houston |
| Direct Prompting | Hays County, Texas is a county in the U.S. state of Texas. (Omissions = 4) |
| MURMUR | Spanish is the official language of Texas. San Marcos is the county seat of Hays County, Texas where Houston is the largest city in Texas where Austin, Texas is part of Texas. (Omissions = 0) |
| Input | Vermont | largestCity | Burlington_Vermont # Alvah_Sabin | region | Vermont # Alvah_Sabin | activeYearsEndDate | 1857-03-03 |
| Direct Prompting | Alvah Sabin was born on March 3, 1857 in Vermont. (Omissions = 2) |
| MURMUR | Burlington is the largest city in Vermont where Alvah Sabin is from and he played from 1857-03-03 to 1857-03-03. (Omissions = 0) |
| #Step | Module | Input | Output |
| 1 | Surface Realization | Antwerp_International_Airport | runwayLength | 600.0 | The runway at Antwerp International Airport is 600.0 metres long. |
| 2 | Surface Realization | Antwerp_International_Airport | elevation-AboveTheSeaLevel_(in_metres) | 12.0 | Antwerp International Airport is 12.0 metres above sea level. |
| 3 | Surface Realization | Antwerp_International_Airport | operatingOrganisation | Flemish_Government | Antwerp International Airport is operated by the Flemish Government. |
| 4 | Text Fusion | Antwerp International Airport is owned by the Flemish Region.; Antwerp International Airport is operated by the Flemish Government. | Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. |
| 5 | Text Fusion | Antwerp International Airport is 12.0 metres above sea level.; Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. | Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| 6 | Text Fusion | The runway at Antwerp International Airport is 600.0 metres long.; Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. | The runway at Antwerp International Airport is 600.0 metres long and is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| Date | Opponent | Score | Loss | Attendance | Record |
| June 1 | giants | 5 - 4 | hernández (5 - 5) | 40893 | 30 - 26 |
| June 2 | giants | 9 - 2 | thomson (6 - 4) | 40651 | 30 - 27 |
| June 3 | dodgers | 11 - 5 | jiménez (1 - 3) | 30150 | 30 - 28 |
| June 4 | dodgers | 10 - 4 | jones (0 - 1) | 30195 | 30 - 29 |
| June 5 | dodgers | 8 - 6 | daal (4 - 2) | 31793 | 31 - 29 |
| June 7 | blue jays | 8 - 0 | hampton (3 - 7) | 20032 | 31 - 30 |
| June 8 | blue jays | 3 - 1 | thomson (6 - 5) | 21298 | 31 - 31 |
| June 9 | blue jays | 3 - 2 | jiménez (1 - 4) | 20328 | 31 - 32 |
| June 10 | red sox | 7 - 3 | neagle (4 - 3) | 33508 | 31 - 33 |
| June 11 | red sox | 3 - 1 | fossum (2 - 1) | 32340 | 32 - 33 |
| June 12 | red sox | 7 - 5 | hampton (3 - 8) | 31583 | 32 - 34 |
| June 14 | indians | 5 - 3 | thomson (6 - 6) | 40156 | 32 - 35 |
| June 15 | indians | 7 - 4 | paronto (0 - 2) | 41870 | 33 - 35 |
| June 16 | indians | 5 - 4 | neagle (4 - 4) | 40792 | 33 - 36 |
| June 18 | yankees | 10 - 5 | jennings (8 - 3) | 48738 | 33 - 37 |
| June 19 | yankees | 20 - 10 | white (1 - 5) | 48821 | 33 - 38 |
| June 20 | yankees | 14 - 11 (10) | karsay (3 - 3) | 48916 | 34 - 38 |
| June 21 | devil rays | 8 - 7 (10) | yan (3 - 3) | 30284 | 35 - 38 |
| June 22 | devil rays | 6 - 5 (11) | kent (0 - 2) | 31190 | 36 - 38 |
| June 23 | devil rays | 6 - 5 | kennedy (5 - 6) | 31043 | 37 - 38 |
| June 24 | dodgers | 4 - 1 | ishii (11 - 3) | 34641 | 38 - 38 |
| June 25 | dodgers | 4 - 0 | thomson (6 - 7) | 23635 | 38 - 39 |
| June 26 | dodgers | 5 - 3 | chacón (3 - 5) | 25083 | 38 - 40 |
| June 27 | dodgers | 7 - 1 | neagle (4 - 5) | 41279 | 38 - 41 |
| June 28 | mariners | 6 - 2 | jennings (8 - 4) | 45118 | 38 - 42 |
| June 29 | mariners | 8 - 1 | hampton (4 - 9) | 45790 | 38 - 43 |
| June 30 | mariners | 4 - 3 | sasaki (2 - 2) | 45928 | 39 - 43 |
| year (ceremony) | film title used in nomination | original title | director | result |
| 1992 (65th) | the child of man | cilveka bēns | jānis streičs | not nominated |
| 2008 (81st) | defenders of riga | rīgas sargi | aigars grauba | not nominated |
| 2010 (83rd) | hong kong confidential | amaya | māris martinsons | not nominated |
| 2012 (85th) | gulf stream under the iceberg | golfa straume zem ledus kalna | yevgeni pashkevich | not nominated |
| 2013 (86th) | mother , i love you | mammu , es tevi milu | jānis nords | tbd |
| Year | Class | Team | Points | Wins |
| 1979 | 350cc | yamaha | 3 | 0 |
| 1980 | 250cc | yamaha | 4 | 0 |
| 1982 | 250cc | yamaha | 4 | 0 |
| 1982 | 500cc | suzuki | 0 | 0 |
| 1983 | 250cc | yamaha | 14 | 0 |
| 1984 | 500cc | honda | 14 | 0 |
| 1985 | 250cc | romer-juchem | 29 | 0 |
| 1986 | 250cc | hb - honda | 10 | 0 |
| 1987 | 250cc | hb - honda | 108 | 1 |
| 1988 | 250cc | hb - honda | 158 | 0 |
| 1989 | 250cc | hb - honda | 190 | 2 |
| 1990 | 250cc | hb - honda | 52 | 0 |
| Module | Description |
| Filtering | Mfilter : (t, cr) → t' : takes a table t and a filtering criterion cr as input and outputs a table t' with rows where the criterion cr is satisfied. |
| Aggregation | Performs aggregation operations on a table. For example, Mmax : (t, c) → n is a max module that takes a table t and a column c as input and outputs the maximum number n in column c. |
| Boolean | Mbool : (t, cr) → b: takes a table t and a criterion cr as input and outputs a boolean b based on whether the criterion is satisfied. |
| Hop | Mhop : (r, c) → e: takes a row r and a column c as input; outputs the element e in (r, c) cell. |
| Graph-to-Text (WebNLG) |
| Triple → String\n(String, String) → String |
| Table-to-Text (LogicNLG) |
| Table → Table | Row | Number | Boolean\nRow → String | Number\nString | Number → Boolean\n(Table, Path) → String |
| Seen | BLEU Unseen | All | Seen | METEOR Unseen | All | ||
| supervised | MELBOURNE† | 54.5 | 33.2 | 45.1 | 41.0 | 33.0 | 37.0 |
| GPT-2-large† | 65.3 | 43.1 | 55.5 | 46.0 | 38.0 | 42.0 | |
| T5-large† | 64.9 | 54.0 | 59.9 | 46.0 | 43.0 | 44.0 | |
| Neural Pipeline‡ | - | - | 43.3 | - | - | 39.3 | |
| few-shot | Direct Prompting (k=1)∗ | 33.1±0.3 | 34.2±0.1 | 33.6±0.1 | 30.4±0.1 | 31.2±0.1 | 30.8±0.1 |
| Direct Prompting (k=5)∗ | 39.9±0.3 | 38.9±0.3 | 39.5±0.1 | 34.3±0.1 | 34.3±0.3 | 34.4±0.1 | |
| CoT Prompting (k=1)∗ | 22.2±0.2 | 14.9±0.2 | 18.0±0.1 | 22.3±0.1 | 22.9±0.2 | 22.6±0.1 | |
| MURMUR (k=1)∗ | 41.4±0.0 | 41.1±0.0 | 41.3±0.0 | 37.1±0.0 | 37.1±0.0 | 37.1±0.0 |
| DP | MURMUR | % Improve | |
| Omissions↓ | 1.64±0.06 | 0.73±0.01 | +24% |
| Hallucinations↓ | 0.77±0.03 | 0.43±0.03 | +9% |
| Disfluencies↓ | 0.14±0.05 | 0.30±0.04 | -4% |
| Module | Grammatical | Faithful | Correct |
| Surface Realization | 1.00 | 1.00 | 0.82 |
| Text Fusion | 0.90 | 0.72 | 0.64 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Field-Infusing† | 43.7 / 20.9 / 8.4 |
| BERT-TabGen† | 49.1 / 27.7 / 13.5 |
| GPT-TabGen† | 49.6 / 28.2 / 14.2 |
| GPT-Coarse-to-Fine† | 49.0 / 28.3 / 14.6 |
| DCVED† | 49.5 / 28.6 / 15.3 |
| Direct Prompting* | 37.2±0.4 / 18.8±0.2 / 8.6±0.2 |
| CoT Prompting* | 35.6±0.2 / 18.6±0.1 / 8.8±0.0 |
| BART + SR‡ | 39.2±0.2 / 20.6±0.2 / 9.5±0.0 |
| MURMUR‡ | 39.8±0.0 / 22.2±0.0 / 11.2±0.0 |
| -saliency* | 39.6±0.0 / 21.9±0.0 / 10.6±0.0 |
| Correct | Partial | Incorrect | Ungrammatical | Is Logical? | |
| Direct Prompting | 28.7±3.7 | 20.0±2.5 | 38.8±8.7 | 12.5±2.5 | 62.0±0.5 |
| MURMUR | 55.0±2.5 | 1.2±1.2 | 38.8±3.7 | 5.0±2.5 | 95.4±0.2 |
| Module Name | Input Data Type | Output Data Type | Description |
| filter_eq filter_not_eq | table, string, string | table | Returns a table with the rows where entry in the input column (second argument) is equal or not equal to the input value (third argument). |
| filtergreater filtergreater_eq filter LESSER filter LESSER_eq | table, string, number | table | Returns a table with the rows where a numerical column (second argument) is greater than or less than (or equal to) the input number (third argument). |
| filter_all | table, string | table | Returns the whole table. |
| arg_max arg_min | table, string | row | Returns the row with the minimum or maximum value for the input column (second argument). |
| max min avg sum | table, string | number | Returns the maximum, minimum, average or sum of numbers in the input column (second argument). |
| count | table | number | Returns the number of rows in the table. |
| all_eq all_NOT_eq | table, string, string | bool | Returns whether all entries in the input column are equal (or not equal to) the input value. |
| all_greater all-less all_greater_eq all-less_eq | table, string, number | bool | Returns whether all entries in the input column are greater than or less than (or equal to) the input number. |
| most_eq most_NOT_eq | table, string, string | bool | Returns whether most entries in the input column are equal (or not equal to) the input value. |
| most_greater most-less most_greater_eq most-less_eq | table, string, number | bool | Returns whether most entries in the input column are greater than or less than (or equal to) the corresponding number. |
| only | table | bool | Returns whether the table has exactly one row. |
| hop | row, string | string | Returns the entry corresponding to the input column in the row. |
| eq | string, string | bool | Returns whether the two inputs are equal or not. |
| BLEU | METEOR | |
| Direct Prompting (k=1) | 31.1±0.5 | 29.8±0.1 |
| Direct Prompting (k=5) | 38.3±0.4 | 33.6±0.1 |
| MURMUR (k=1) | 40.1±0.3 | 37.1±0.5 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Direct Prompting (k=1) | 37.0±0.2 / 18.9±0.1 / 8.5±0.1 |
| COT Prompting (k=1) | 36.5±0.1 / 18.9±0.1 / 8.7±0.3 |
| MURMUR (k=1) | 40.5±0.1 / 22.2±0.0 / 10.8±0.1 |
| Beam Size | BLEU-1 / BLEU-2 / BLEU-3 |
| 10 | 39.7 / 21.2 / 10.3 |
| 20 | 40.2 / 21.8 / 10.7 |
| 50 | 40.5 / 22.2 / 10.8 |
| 100 | 40.7 / 22.5 / 10.9 |
| # Gold Tables | # Gold Paths | # Samples (Pos/Neg/ All) | Acc. | LogicNLG BLEU-3 |
| 100 | 221 | 769/729/1498 | 76.16 | 10.8 |
| 200 | 443 | 1534/1457/2991 | 78.52 | 10.6 |
| 500 | 1085 | 3773/3633/7406 | 80.32 | 10.9 |
| 3000 | 7145 | 21.5k/20.5k/42.0k | 82.84 | 10.7 |
| Let's convert triples to sentences |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Let's convert triples to sentences step-by-step |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: AS Gubbio 1910 plays in Serie D. # Pietro Grasso is the leader of Italy. # ... # S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Input | Arlington,_Texas | isPartOf | Texas # Texas | largestCity | Houston # Texas | language | English_language |
| Direct Prompting | Texas is the second largest state in the United States of America. Its capital is Austin and the largest city is Houston. (Omissions = 2) |
| MURMUR | Houston is the largest city in Texas where English is the official language and Arlington, Texas is a city in Texas. (Omissions = 0) |
| Input | Hays_County_Texas | countySeat | San_Marcos,_Texas # Austin,_Texas | isPartOf | Hays_County,_Texas # Texas | language | Spanish_language # Austin,_Texas | isPartOf | Texas # Texas | largestCity | Houston |
| Direct Prompting | Hays County, Texas is a county in the U.S. state of Texas. (Omissions = 4) |
| MURMUR | Spanish is the official language of Texas. San Marcos is the county seat of Hays County, Texas where Houston is the largest city in Texas where Austin, Texas is part of Texas. (Omissions = 0) |
| Input | Vermont | largestCity | Burlington_Vermont # Alvah_Sabin | region | Vermont # Alvah_Sabin | activeYearsEndDate | 1857-03-03 |
| Direct Prompting | Alvah Sabin was born on March 3, 1857 in Vermont. (Omissions = 2) |
| MURMUR | Burlington is the largest city in Vermont where Alvah Sabin is from and he played from 1857-03-03 to 1857-03-03. (Omissions = 0) |
| #Step | Module | Input | Output |
| 1 | Surface Realization | Antwerp_International_Airport | runwayLength | 600.0 | The runway at Antwerp International Airport is 600.0 metres long. |
| 2 | Surface Realization | Antwerp_International_Airport | elevation-AboveTheSeaLevel_(in_metres) | 12.0 | Antwerp International Airport is 12.0 metres above sea level. |
| 3 | Surface Realization | Antwerp_International_Airport | operatingOrganisation | Flemish_Government | Antwerp International Airport is operated by the Flemish Government. |
| 4 | Text Fusion | Antwerp International Airport is owned by the Flemish Region.; Antwerp International Airport is operated by the Flemish Government. | Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. |
| 5 | Text Fusion | Antwerp International Airport is 12.0 metres above sea level.; Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. | Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| 6 | Text Fusion | The runway at Antwerp International Airport is 600.0 metres long.; Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. | The runway at Antwerp International Airport is 600.0 metres long and is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| Date | Opponent | Score | Loss | Attendance | Record |
| June 1 | giants | 5 - 4 | hernández (5 - 5) | 40893 | 30 - 26 |
| June 2 | giants | 9 - 2 | thomson (6 - 4) | 40651 | 30 - 27 |
| June 3 | dodgers | 11 - 5 | jiménez (1 - 3) | 30150 | 30 - 28 |
| June 4 | dodgers | 10 - 4 | jones (0 - 1) | 30195 | 30 - 29 |
| June 5 | dodgers | 8 - 6 | daal (4 - 2) | 31793 | 31 - 29 |
| June 7 | blue jays | 8 - 0 | hampton (3 - 7) | 20032 | 31 - 30 |
| June 8 | blue jays | 3 - 1 | thomson (6 - 5) | 21298 | 31 - 31 |
| June 9 | blue jays | 3 - 2 | jiménez (1 - 4) | 20328 | 31 - 32 |
| June 10 | red sox | 7 - 3 | neagle (4 - 3) | 33508 | 31 - 33 |
| June 11 | red sox | 3 - 1 | fossum (2 - 1) | 32340 | 32 - 33 |
| June 12 | red sox | 7 - 5 | hampton (3 - 8) | 31583 | 32 - 34 |
| June 14 | indians | 5 - 3 | thomson (6 - 6) | 40156 | 32 - 35 |
| June 15 | indians | 7 - 4 | paronto (0 - 2) | 41870 | 33 - 35 |
| June 16 | indians | 5 - 4 | neagle (4 - 4) | 40792 | 33 - 36 |
| June 18 | yankees | 10 - 5 | jennings (8 - 3) | 48738 | 33 - 37 |
| June 19 | yankees | 20 - 10 | white (1 - 5) | 48821 | 33 - 38 |
| June 20 | yankees | 14 - 11 (10) | karsay (3 - 3) | 48916 | 34 - 38 |
| June 21 | devil rays | 8 - 7 (10) | yan (3 - 3) | 30284 | 35 - 38 |
| June 22 | devil rays | 6 - 5 (11) | kent (0 - 2) | 31190 | 36 - 38 |
| June 23 | devil rays | 6 - 5 | kennedy (5 - 6) | 31043 | 37 - 38 |
| June 24 | dodgers | 4 - 1 | ishii (11 - 3) | 34641 | 38 - 38 |
| June 25 | dodgers | 4 - 0 | thomson (6 - 7) | 23635 | 38 - 39 |
| June 26 | dodgers | 5 - 3 | chacón (3 - 5) | 25083 | 38 - 40 |
| June 27 | dodgers | 7 - 1 | neagle (4 - 5) | 41279 | 38 - 41 |
| June 28 | mariners | 6 - 2 | jennings (8 - 4) | 45118 | 38 - 42 |
| June 29 | mariners | 8 - 1 | hampton (4 - 9) | 45790 | 38 - 43 |
| June 30 | mariners | 4 - 3 | sasaki (2 - 2) | 45928 | 39 - 43 |
| year (ceremony) | film title used in nomination | original title | director | result |
| 1992 (65th) | the child of man | cilveka bēns | jānis streičs | not nominated |
| 2008 (81st) | defenders of riga | rīgas sargi | aigars grauba | not nominated |
| 2010 (83rd) | hong kong confidential | amaya | māris martinsons | not nominated |
| 2012 (85th) | gulf stream under the iceberg | golfa straume zem ledus kalna | yevgeni pashkevich | not nominated |
| 2013 (86th) | mother , i love you | mammu , es tevi milu | jānis nords | tbd |
| Year | Class | Team | Points | Wins |
| 1979 | 350cc | yamaha | 3 | 0 |
| 1980 | 250cc | yamaha | 4 | 0 |
| 1982 | 250cc | yamaha | 4 | 0 |
| 1982 | 500cc | suzuki | 0 | 0 |
| 1983 | 250cc | yamaha | 14 | 0 |
| 1984 | 500cc | honda | 14 | 0 |
| 1985 | 250cc | romer-juchem | 29 | 0 |
| 1986 | 250cc | hb - honda | 10 | 0 |
| 1987 | 250cc | hb - honda | 108 | 1 |
| 1988 | 250cc | hb - honda | 158 | 0 |
| 1989 | 250cc | hb - honda | 190 | 2 |
| 1990 | 250cc | hb - honda | 52 | 0 |
| Module | Description |
| Filtering | Mfilter : (t, cr) → t' : takes a table t and a filtering criterion cr as input and outputs a table t' with rows where the criterion cr is satisfied. |
| Aggregation | Performs aggregation operations on a table. For example, Mmax : (t, c) → n is a max module that takes a table t and a column c as input and outputs the maximum number n in column c. |
| Boolean | Mbool : (t, cr) → b: takes a table t and a criterion cr as input and outputs a boolean b based on whether the criterion is satisfied. |
| Hop | Mhop : (r, c) → e: takes a row r and a column c as input; outputs the element e in (r, c) cell. |
| Graph-to-Text (WebNLG) |
| Triple → String +(String, String) → String |
| Table-to-Text (LogicNLG) |
| Table → Table | Row | Number | Boolean +Row → String | Number +String | Number → Boolean +(Table, Path) → String |
| Seen | BLEU Unseen | All | Seen | METEOR Unseen | All | ||
| supervised | MELBOURNE† | 54.5 | 33.2 | 45.1 | 41.0 | 33.0 | 37.0 |
| GPT-2-large† | 65.3 | 43.1 | 55.5 | 46.0 | 38.0 | 42.0 | |
| T5-large† | 64.9 | 54.0 | 59.9 | 46.0 | 43.0 | 44.0 | |
| Neural Pipeline‡ | - | - | 43.3 | - | - | 39.3 | |
| few-shot | Direct Prompting (k=1)∗ | 33.1±0.3 | 34.2±0.1 | 33.6±0.1 | 30.4±0.1 | 31.2±0.1 | 30.8±0.1 |
| Direct Prompting (k=5)∗ | 39.9±0.3 | 38.9±0.3 | 39.5±0.1 | 34.3±0.1 | 34.3±0.3 | 34.4±0.1 | |
| CoT Prompting (k=1)∗ | 22.2±0.2 | 14.9±0.2 | 18.0±0.1 | 22.3±0.1 | 22.9±0.2 | 22.6±0.1 | |
| MURMUR (k=1)∗ | 41.4±0.0 | 41.1±0.0 | 41.3±0.0 | 37.1±0.0 | 37.1±0.0 | 37.1±0.0 | |
| DP | MURMUR | % Improve | |
| Omissions↓ | 1.64±0.06 | 0.73±0.01 | +24% |
| Hallucinations↓ | 0.77±0.03 | 0.43±0.03 | +9% |
| Disfluencies↓ | 0.14±0.05 | 0.30±0.04 | -4% |
| Module | Grammatical | Faithful | Correct |
| Surface Realization | 1.00 | 1.00 | 0.82 |
| Text Fusion | 0.90 | 0.72 | 0.64 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Field-Infusing† | 43.7 / 20.9 / 8.4 |
| BERT-TabGen† | 49.1 / 27.7 / 13.5 |
| GPT-TabGen† | 49.6 / 28.2 / 14.2 |
| GPT-Coarse-to-Fine† | 49.0 / 28.3 / 14.6 |
| DCVED† | 49.5 / 28.6 / 15.3 |
| Direct Prompting* | 37.2±0.4 / 18.8±0.2 / 8.6±0.2 |
| CoT Prompting* | 35.6±0.2 / 18.6±0.1 / 8.8±0.0 |
| BART + SR‡ | 39.2±0.2 / 20.6±0.2 / 9.5±0.0 |
| MURMUR‡ | 39.8±0.0 / 22.2±0.0 / 11.2±0.0 |
| -saliency* | 39.6±0.0 / 21.9±0.0 / 10.6±0.0 |
| Correct | Partial | Incorrect | Ungrammatical | Is Logical? | |
| Direct Prompting | 28.7±3.7 | 20.0±2.5 | 38.8±8.7 | 12.5±2.5 | 62.0±0.5 |
| MURMUR | 55.0±2.5 | 1.2±1.2 | 38.8±3.7 | 5.0±2.5 | 95.4±0.2 |
| Module Name | Input Data Type | Output Data Type | Description |
| filter_eq filter_not_eq | table, string, string | table | Returns a table with the rows where entry in the input column (second argument) is equal or not equal to the input value (third argument). |
| filtergreater filtergreater_eq filter LESSER filter LESSER_eq | table, string, number | table | Returns a table with the rows where a numerical column (second argument) is greater than or less than (or equal to) the input number (third argument). |
| filter_all | table, string | table | Returns the whole table. |
| arg_max arg_min | table, string | row | Returns the row with the minimum or maximum value for the input column (second argument). |
| max min avg sum | table, string | number | Returns the maximum, minimum, average or sum of numbers in the input column (second argument). |
| count | table | number | Returns the number of rows in the table. |
| all_eq all_NOT_eq | table, string, string | bool | Returns whether all entries in the input column are equal (or not equal to) the input value. |
| all_greater all-less all_greater_eq all-less_eq | table, string, number | bool | Returns whether all entries in the input column are greater than or less than (or equal to) the input number. |
| most_eq most_NOT_eq | table, string, string | bool | Returns whether most entries in the input column are equal (or not equal to) the input value. |
| most_greater most-less most_greater_eq most-less_eq | table, string, number | bool | Returns whether most entries in the input column are greater than or less than (or equal to) the corresponding number. |
| only | table | bool | Returns whether the table has exactly one row. |
| hop | row, string | string | Returns the entry corresponding to the input column in the row. |
| eq | string, string | bool | Returns whether the two inputs are equal or not. |
| BLEU | METEOR | |
| Direct Prompting (k=1) | 31.1±0.5 | 29.8±0.1 |
| Direct Prompting (k=5) | 38.3±0.4 | 33.6±0.1 |
| MURMUR (k=1) | 40.1±0.3 | 37.1±0.5 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Direct Prompting (k=1) | 37.0±0.2 / 18.9±0.1 / 8.5±0.1 |
| COT Prompting (k=1) | 36.5±0.1 / 18.9±0.1 / 8.7±0.3 |
| MURMUR (k=1) | 40.5±0.1 / 22.2±0.0 / 10.8±0.1 |
| Beam Size | BLEU-1 / BLEU-2 / BLEU-3 |
| 10 | 39.7 / 21.2 / 10.3 |
| 20 | 40.2 / 21.8 / 10.7 |
| 50 | 40.5 / 22.2 / 10.8 |
| 100 | 40.7 / 22.5 / 10.9 |
| # Gold Tables | # Gold Paths | # Samples (Pos/Neg/ All) | Acc. | LogicNLG BLEU-3 |
| 100 | 221 | 769/729/1498 | 76.16 | 10.8 |
| 200 | 443 | 1534/1457/2991 | 78.52 | 10.6 |
| 500 | 1085 | 3773/3633/7406 | 80.32 | 10.9 |
| 3000 | 7145 | 21.5k/20.5k/42.0k | 82.84 | 10.7 |
| Let's convert triples to sentences |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Let's convert triples to sentences step-by-step |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: AS Gubbio 1910 plays in Serie D. # Pietro Grasso is the leader of Italy. # ... # S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Input | Arlington,_Texas | isPartOf | Texas # Texas | largestCity | Houston # Texas | language | English_language |
| Direct Prompting | Texas is the second largest state in the United States of America. Its capital is Austin and the largest city is Houston. (Omissions = 2) |
| MURMUR | Houston is the largest city in Texas where English is the official language and Arlington, Texas is a city in Texas. (Omissions = 0) |
| Input | Hays_County_Texas | countySeat | San_Marcos,_Texas # Austin,_Texas | isPartOf | Hays_County,_Texas # Texas | language | Spanish_language # Austin,_Texas | isPartOf | Texas # Texas | largestCity | Houston |
| Direct Prompting | Hays County, Texas is a county in the U.S. state of Texas. (Omissions = 4) |
| MURMUR | Spanish is the official language of Texas. San Marcos is the county seat of Hays County, Texas where Houston is the largest city in Texas where Austin, Texas is part of Texas. (Omissions = 0) |
| Input | Vermont | largestCity | Burlington_Vermont # Alvah_Sabin | region | Vermont # Alvah_Sabin | activeYearsEndDate | 1857-03-03 |
| Direct Prompting | Alvah Sabin was born on March 3, 1857 in Vermont. (Omissions = 2) |
| MURMUR | Burlington is the largest city in Vermont where Alvah Sabin is from and he played from 1857-03-03 to 1857-03-03. (Omissions = 0) |
| #Step | Module | Input | Output |
| 1 | Surface Realization | Antwerp_International_Airport | runwayLength | 600.0 | The runway at Antwerp International Airport is 600.0 metres long. |
| 2 | Surface Realization | Antwerp_International_Airport | elevation-AboveTheSeaLevel_(in_metres) | 12.0 | Antwerp International Airport is 12.0 metres above sea level. |
| 3 | Surface Realization | Antwerp_International_Airport | operatingOrganisation | Flemish_Government | Antwerp International Airport is operated by the Flemish Government. |
| 4 | Text Fusion | Antwerp International Airport is owned by the Flemish Region.; Antwerp International Airport is operated by the Flemish Government. | Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. |
| 5 | Text Fusion | Antwerp International Airport is 12.0 metres above sea level.; Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. | Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| 6 | Text Fusion | The runway at Antwerp International Airport is 600.0 metres long.; Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. | The runway at Antwerp International Airport is 600.0 metres long and is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| Date | Opponent | Score | Loss | Attendance | Record |
| June 1 | giants | 5 - 4 | hernández (5 - 5) | 40893 | 30 - 26 |
| June 2 | giants | 9 - 2 | thomson (6 - 4) | 40651 | 30 - 27 |
| June 3 | dodgers | 11 - 5 | jiménez (1 - 3) | 30150 | 30 - 28 |
| June 4 | dodgers | 10 - 4 | jones (0 - 1) | 30195 | 30 - 29 |
| June 5 | dodgers | 8 - 6 | daal (4 - 2) | 31793 | 31 - 29 |
| June 7 | blue jays | 8 - 0 | hampton (3 - 7) | 20032 | 31 - 30 |
| June 8 | blue jays | 3 - 1 | thomson (6 - 5) | 21298 | 31 - 31 |
| June 9 | blue jays | 3 - 2 | jiménez (1 - 4) | 20328 | 31 - 32 |
| June 10 | red sox | 7 - 3 | neagle (4 - 3) | 33508 | 31 - 33 |
| June 11 | red sox | 3 - 1 | fossum (2 - 1) | 32340 | 32 - 33 |
| June 12 | red sox | 7 - 5 | hampton (3 - 8) | 31583 | 32 - 34 |
| June 14 | indians | 5 - 3 | thomson (6 - 6) | 40156 | 32 - 35 |
| June 15 | indians | 7 - 4 | paronto (0 - 2) | 41870 | 33 - 35 |
| June 16 | indians | 5 - 4 | neagle (4 - 4) | 40792 | 33 - 36 |
| June 18 | yankees | 10 - 5 | jennings (8 - 3) | 48738 | 33 - 37 |
| June 19 | yankees | 20 - 10 | white (1 - 5) | 48821 | 33 - 38 |
| June 20 | yankees | 14 - 11 (10) | karsay (3 - 3) | 48916 | 34 - 38 |
| June 21 | devil rays | 8 - 7 (10) | yan (3 - 3) | 30284 | 35 - 38 |
| June 22 | devil rays | 6 - 5 (11) | kent (0 - 2) | 31190 | 36 - 38 |
| June 23 | devil rays | 6 - 5 | kennedy (5 - 6) | 31043 | 37 - 38 |
| June 24 | dodgers | 4 - 1 | ishii (11 - 3) | 34641 | 38 - 38 |
| June 25 | dodgers | 4 - 0 | thomson (6 - 7) | 23635 | 38 - 39 |
| June 26 | dodgers | 5 - 3 | chacón (3 - 5) | 25083 | 38 - 40 |
| June 27 | dodgers | 7 - 1 | neagle (4 - 5) | 41279 | 38 - 41 |
| June 28 | mariners | 6 - 2 | jennings (8 - 4) | 45118 | 38 - 42 |
| June 29 | mariners | 8 - 1 | hampton (4 - 9) | 45790 | 38 - 43 |
| June 30 | mariners | 4 - 3 | sasaki (2 - 2) | 45928 | 39 - 43 |
| year (ceremony) | film title used in nomination | original title | director | result |
| 1992 (65th) | the child of man | cilveka bēns | jānis streičs | not nominated |
| 2008 (81st) | defenders of riga | rīgas sargi | aigars grauba | not nominated |
| 2010 (83rd) | hong kong confidential | amaya | māris martinsons | not nominated |
| 2012 (85th) | gulf stream under the iceberg | golfa straume zem ledus kalna | yevgeni pashkevich | not nominated |
| 2013 (86th) | mother , i love you | mammu , es tevi milu | jānis nords | tbd |
| Year | Class | Team | Points | Wins |
| 1979 | 350cc | yamaha | 3 | 0 |
| 1980 | 250cc | yamaha | 4 | 0 |
| 1982 | 250cc | yamaha | 4 | 0 |
| 1982 | 500cc | suzuki | 0 | 0 |
| 1983 | 250cc | yamaha | 14 | 0 |
| 1984 | 500cc | honda | 14 | 0 |
| 1985 | 250cc | romer-juchem | 29 | 0 |
| 1986 | 250cc | hb - honda | 10 | 0 |
| 1987 | 250cc | hb - honda | 108 | 1 |
| 1988 | 250cc | hb - honda | 158 | 0 |
| 1989 | 250cc | hb - honda | 190 | 2 |
| 1990 | 250cc | hb - honda | 52 | 0 |
| Module | Description |
| Filtering | Mfilter : (t, cr) → t' : takes a table t and a filtering criterion cr as input and outputs a table t' with rows where the criterion cr is satisfied. |
| Aggregation | Performs aggregation operations on a table. For example, Mmax : (t, c) → n is a max module that takes a table t and a column c as input and outputs the maximum number n in column c. |
| Boolean | Mbool : (t, cr) → b: takes a table t and a criterion cr as input and outputs a boolean b based on whether the criterion is satisfied. |
| Hop | Mhop : (r, c) → e: takes a row r and a column c as input; outputs the element e in (r, c) cell. |
| Graph-to-Text (WebNLG) |
| Triple → String\n(String, String) → String |
| Table-to-Text (LogicNLG) |
| Table → Table | Row | Number | Boolean\nRow → String | Number\nString | Number → Boolean\n(Table, Path) → String |
| Seen | BLEU Unseen | All | Seen | METEOR Unseen | All | ||
| supervised | MELBOURNE† | 54.5 | 33.2 | 45.1 | 41.0 | 33.0 | 37.0 |
| GPT-2-large† | 65.3 | 43.1 | 55.5 | 46.0 | 38.0 | 42.0 | |
| T5-large† | 64.9 | 54.0 | 59.9 | 46.0 | 43.0 | 44.0 | |
| Neural Pipeline‡ | - | - | 43.3 | - | - | 39.3 | |
| few-shot | Direct Prompting (k=1)∗ | 33.1±0.3 | 34.2±0.1 | 33.6±0.1 | 30.4±0.1 | 31.2±0.1 | 30.8±0.1 |
| Direct Prompting (k=5)∗ | 39.9±0.3 | 38.9±0.3 | 39.5±0.1 | 34.3±0.1 | 34.3±0.3 | 34.4±0.1 | |
| CoT Prompting (k=1)∗ | 22.2±0.2 | 14.9±0.2 | 18.0±0.1 | 22.3±0.1 | 22.9±0.2 | 22.6±0.1 | |
| MURMUR (k=1)∗ | 41.4±0.0 | 41.1±0.0 | 41.3±0.0 | 37.1±0.0 | 37.1±0.0 | 37.1±0.0 |
| DP | MURMUR | % Improve | |
| Omissions↓ | 1.64±0.06 | 0.73±0.01 | +24% |
| Hallucinations↓ | 0.77±0.03 | 0.43±0.03 | +9% |
| Disfluencies↓ | 0.14±0.05 | 0.30±0.04 | -4% |
| Module | Grammatical | Faithful | Correct |
| Surface Realization | 1.00 | 1.00 | 0.82 |
| Text Fusion | 0.90 | 0.72 | 0.64 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Field-Infusing† | 43.7 / 20.9 / 8.4 |
| BERT-TabGen† | 49.1 / 27.7 / 13.5 |
| GPT-TabGen† | 49.6 / 28.2 / 14.2 |
| GPT-Coarse-to-Fine† | 49.0 / 28.3 / 14.6 |
| DCVED† | 49.5 / 28.6 / 15.3 |
| Direct Prompting* | 37.2±0.4 / 18.8±0.2 / 8.6±0.2 |
| CoT Prompting* | 35.6±0.2 / 18.6±0.1 / 8.8±0.0 |
| BART + SR‡ | 39.2±0.2 / 20.6±0.2 / 9.5±0.0 |
| MURMUR‡ | 39.8±0.0 / 22.2±0.0 / 11.2±0.0 |
| -saliency* | 39.6±0.0 / 21.9±0.0 / 10.6±0.0 |
| Correct | Partial | Incorrect | Ungrammatical | Is Logical? | |
| Direct Prompting | 28.7±3.7 | 20.0±2.5 | 38.8±8.7 | 12.5±2.5 | 62.0±0.5 |
| MURMUR | 55.0±2.5 | 1.2±1.2 | 38.8±3.7 | 5.0±2.5 | 95.4±0.2 |
| Module Name | Input Data Type | Output Data Type | Description |
| filter_eq filter_not_eq | table, string, string | table | Returns a table with the rows where entry in the input column (second argument) is equal or not equal to the input value (third argument). |
| filtergreater filtergreater_eq filter LESSER filter LESSER_eq | table, string, number | table | Returns a table with the rows where a numerical column (second argument) is greater than or less than (or equal to) the input number (third argument). |
| filter_all | table, string | table | Returns the whole table. |
| arg_max arg_min | table, string | row | Returns the row with the minimum or maximum value for the input column (second argument). |
| max min avg sum | table, string | number | Returns the maximum, minimum, average or sum of numbers in the input column (second argument). |
| count | table | number | Returns the number of rows in the table. |
| all_eq all_NOT_eq | table, string, string | bool | Returns whether all entries in the input column are equal (or not equal to) the input value. |
| all_greater all-less all_greater_eq all-less_eq | table, string, number | bool | Returns whether all entries in the input column are greater than or less than (or equal to) the input number. |
| most_eq most_NOT_eq | table, string, string | bool | Returns whether most entries in the input column are equal (or not equal to) the input value. |
| most_greater most-less most_greater_eq most-less_eq | table, string, number | bool | Returns whether most entries in the input column are greater than or less than (or equal to) the corresponding number. |
| only | table | bool | Returns whether the table has exactly one row. |
| hop | row, string | string | Returns the entry corresponding to the input column in the row. |
| eq | string, string | bool | Returns whether the two inputs are equal or not. |
| BLEU | METEOR | |
| Direct Prompting (k=1) | 31.1±0.5 | 29.8±0.1 |
| Direct Prompting (k=5) | 38.3±0.4 | 33.6±0.1 |
| MURMUR (k=1) | 40.1±0.3 | 37.1±0.5 |
| BLEU-1 / BLEU-2 / BLEU-3 | |
| Direct Prompting (k=1) | 37.0±0.2 / 18.9±0.1 / 8.5±0.1 |
| COT Prompting (k=1) | 36.5±0.1 / 18.9±0.1 / 8.7±0.3 |
| MURMUR (k=1) | 40.5±0.1 / 22.2±0.0 / 10.8±0.1 |
| Beam Size | BLEU-1 / BLEU-2 / BLEU-3 |
| 10 | 39.7 / 21.2 / 10.3 |
| 20 | 40.2 / 21.8 / 10.7 |
| 50 | 40.5 / 22.2 / 10.8 |
| 100 | 40.7 / 22.5 / 10.9 |
| # Gold Tables | # Gold Paths | # Samples (Pos/Neg/ All) | Acc. | LogicNLG BLEU-3 |
| 100 | 221 | 769/729/1498 | 76.16 | 10.8 |
| 200 | 443 | 1534/1457/2991 | 78.52 | 10.6 |
| 500 | 1085 | 3773/3633/7406 | 80.32 | 10.9 |
| 3000 | 7145 | 21.5k/20.5k/42.0k | 82.84 | 10.7 |
| Let's convert triples to sentences |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Let's convert triples to sentences step-by-step |
| ##### |
| Triples: A.S._Gubbio_1910 | league | Serie_D # Italy | leader | Pietro_Grasso # Italy | capital | Rome # A.S._Gubbio_1910 | ground | Italy # Serie_D | champions | S.S._Robur_Siena |
| Output: AS Gubbio 1910 plays in Serie D. # Pietro Grasso is the leader of Italy. # ... # S.S. Robur Siena are champions of Serie D in which AS Gubbio 1910 also play. This latter club have their home ground in Italy where the capital city is Rome and the leader is Pietro Grasso. |
| ##### |
| Triples: {triples} |
| Output: |
| Input | Arlington,_Texas | isPartOf | Texas # Texas | largestCity | Houston # Texas | language | English_language |
| Direct Prompting | Texas is the second largest state in the United States of America. Its capital is Austin and the largest city is Houston. (Omissions = 2) |
| MURMUR | Houston is the largest city in Texas where English is the official language and Arlington, Texas is a city in Texas. (Omissions = 0) |
| Input | Hays_County_Texas | countySeat | San_Marcos,_Texas # Austin,_Texas | isPartOf | Hays_County,_Texas # Texas | language | Spanish_language # Austin,_Texas | isPartOf | Texas # Texas | largestCity | Houston |
| Direct Prompting | Hays County, Texas is a county in the U.S. state of Texas. (Omissions = 4) |
| MURMUR | Spanish is the official language of Texas. San Marcos is the county seat of Hays County, Texas where Houston is the largest city in Texas where Austin, Texas is part of Texas. (Omissions = 0) |
| Input | Vermont | largestCity | Burlington_Vermont # Alvah_Sabin | region | Vermont # Alvah_Sabin | activeYearsEndDate | 1857-03-03 |
| Direct Prompting | Alvah Sabin was born on March 3, 1857 in Vermont. (Omissions = 2) |
| MURMUR | Burlington is the largest city in Vermont where Alvah Sabin is from and he played from 1857-03-03 to 1857-03-03. (Omissions = 0) |
| #Step | Module | Input | Output |
| 1 | Surface Realization | Antwerp_International_Airport | runwayLength | 600.0 | The runway at Antwerp International Airport is 600.0 metres long. |
| 2 | Surface Realization | Antwerp_International_Airport | elevation-AboveTheSeaLevel_(in_metres) | 12.0 | Antwerp International Airport is 12.0 metres above sea level. |
| 3 | Surface Realization | Antwerp_International_Airport | operatingOrganisation | Flemish_Government | Antwerp International Airport is operated by the Flemish Government. |
| 4 | Text Fusion | Antwerp International Airport is owned by the Flemish Region.; Antwerp International Airport is operated by the Flemish Government. | Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. |
| 5 | Text Fusion | Antwerp International Airport is 12.0 metres above sea level.; Antwerp International Airport is owned by the Flemish Region and operated by the Flemish Government. | Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| 6 | Text Fusion | The runway at Antwerp International Airport is 600.0 metres long.; Antwerp International Airport is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. | The runway at Antwerp International Airport is 600.0 metres long and is 12.0 metres above sea level and is owned by the Flemish Region and operated by the Flemish Government. |
| Date | Opponent | Score | Loss | Attendance | Record |
| June 1 | giants | 5 - 4 | hernández (5 - 5) | 40893 | 30 - 26 |
| June 2 | giants | 9 - 2 | thomson (6 - 4) | 40651 | 30 - 27 |
| June 3 | dodgers | 11 - 5 | jiménez (1 - 3) | 30150 | 30 - 28 |
| June 4 | dodgers | 10 - 4 | jones (0 - 1) | 30195 | 30 - 29 |
| June 5 | dodgers | 8 - 6 | daal (4 - 2) | 31793 | 31 - 29 |
| June 7 | blue jays | 8 - 0 | hampton (3 - 7) | 20032 | 31 - 30 |
| June 8 | blue jays | 3 - 1 | thomson (6 - 5) | 21298 | 31 - 31 |
| June 9 | blue jays | 3 - 2 | jiménez (1 - 4) | 20328 | 31 - 32 |
| June 10 | red sox | 7 - 3 | neagle (4 - 3) | 33508 | 31 - 33 |
| June 11 | red sox | 3 - 1 | fossum (2 - 1) | 32340 | 32 - 33 |
| June 12 | red sox | 7 - 5 | hampton (3 - 8) | 31583 | 32 - 34 |
| June 14 | indians | 5 - 3 | thomson (6 - 6) | 40156 | 32 - 35 |
| June 15 | indians | 7 - 4 | paronto (0 - 2) | 41870 | 33 - 35 |
| June 16 | indians | 5 - 4 | neagle (4 - 4) | 40792 | 33 - 36 |
| June 18 | yankees | 10 - 5 | jennings (8 - 3) | 48738 | 33 - 37 |
| June 19 | yankees | 20 - 10 | white (1 - 5) | 48821 | 33 - 38 |
| June 20 | yankees | 14 - 11 (10) | karsay (3 - 3) | 48916 | 34 - 38 |
| June 21 | devil rays | 8 - 7 (10) | yan (3 - 3) | 30284 | 35 - 38 |
| June 22 | devil rays | 6 - 5 (11) | kent (0 - 2) | 31190 | 36 - 38 |
| June 23 | devil rays | 6 - 5 | kennedy (5 - 6) | 31043 | 37 - 38 |
| June 24 | dodgers | 4 - 1 | ishii (11 - 3) | 34641 | 38 - 38 |
| June 25 | dodgers | 4 - 0 | thomson (6 - 7) | 23635 | 38 - 39 |
| June 26 | dodgers | 5 - 3 | chacón (3 - 5) | 25083 | 38 - 40 |
| June 27 | dodgers | 7 - 1 | neagle (4 - 5) | 41279 | 38 - 41 |
| June 28 | mariners | 6 - 2 | jennings (8 - 4) | 45118 | 38 - 42 |
| June 29 | mariners | 8 - 1 | hampton (4 - 9) | 45790 | 38 - 43 |
| June 30 | mariners | 4 - 3 | sasaki (2 - 2) | 45928 | 39 - 43 |
| year (ceremony) | film title used in nomination | original title | director | result |
| 1992 (65th) | the child of man | cilveka bēns | jānis streičs | not nominated |
| 2008 (81st) | defenders of riga | rīgas sargi | aigars grauba | not nominated |
| 2010 (83rd) | hong kong confidential | amaya | māris martinsons | not nominated |
| 2012 (85th) | gulf stream under the iceberg | golfa straume zem ledus kalna | yevgeni pashkevich | not nominated |
| 2013 (86th) | mother , i love you | mammu , es tevi milu | jānis nords | tbd |
| Settings | Supervised Pre-training | Unsupervised Pre-training |
| NLG | MVP (ours) | GPT-2, MASS, BART, T5 |
| NLU | FLAN, T0, Muppet, ExT5 | BERT, XLNet, RoBERTa, T5 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| BART | 44.16e | 21.28 | 40.90 | 64.55b | 46.51 | 75.13 | 22.00f | 26.40 | 52.55 | 68.60f | - |
| Flan-T5 | 43.45 | 21.01 | 40.03 | 66.60 | 46.93 | 75.76 | 25.55 | 26.90 | 53.51 | 84.18 | 75.44 |
| Single | 44.36 | 21.54 | 40.88 | 67.74 | 46.89 | 76.94 | 26.09 | 27.15 | 53.29 | 86.20 | 77.26 |
| MVP+S | 44.63 | 21.72 | 41.21 | 68.19 | 47.75 | 76.81 | 25.69 | 27.04 | 53.20 | 86.65 | 77.93 |
| MVP+R | 44.14 | 21.45 | 40.72 | 67.61 | 47.65 | 76.70 | 25.71 | 27.03 | 53.09 | 85.95 | 77.22 |
| MVP+M | 43.97 | 21.16 | 40.46 | 67.45 | 47.57 | 76.81 | 25.46 | 26.79 | 52.95 | 86.28 | 77.26 |
| SOTA | 47.16a | 22.55 | 43.87 | 66.14b | 47.25 | 76.10 | 25.97c | 27.33 | 53.43 | 84.50d | - |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| BART | 30.70g | 13.30 | - | 69.90 | 49.90f | 40.00 | 1.30 | 8.00 | 17.89j | 74.91 | 84.88 |
| Flan-T5 | 32.72 | 15.23 | 2.97 | 68.97 | 48.55 | 40.22 | 1.40 | 7.85 | 19.73 | 70.20 | 78.70 |
| Single | 32.67 | 15.29 | 2.72 | 72.97 | 49.96 | 40.53 | 1.27 | 7.63 | 19.73 | 75.60 | 83.70 |
| MVP+S | 33.92 | 15.60 | 3.44 | 80.58 | 47.91 | 39.97 | 1.52 | 9.54 | 20.32 | 79.90 | 86.80 |
| MVP+R | 32.93 | 15.32 | 2.88 | 73.83 | 48.45 | 40.09 | 1.30 | 7.95 | 19.02 | 73.30 | 81.80 |
| MVP+M | 33.30 | 15.51 | 2.71 | 74.24 | 46.26 | 39.30 | 1.36 | 8.07 | 19.93 | 72.70 | 79.70 |
| SOTA | 33.40g | 15.40 | - | 69.30 | 49.90f | 40.00 | 1.50h | 9.40 | 20.50i | 85.30 | 94.40 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| T0-3B | - | - | - | 1.40 | 10.20 | 18.43 | 3.06 | 12.43 | 14.91 | 13.30 | 6.60 |
| T0-11B | - | - | - | 0.26 | 6.13 | 14.12 | 2.63 | 7.00 | 15.25 | 9.18 | 4.36 |
| MVP | 29.50 | 11.29 | 25.92 | 34.42 | 31.33 | 52.33 | 2.90 | 13.94 | 15.48 | 29.40 | 18.20 |
| MVP+S | 25.60 | 9.51 | 22.67 | 39.43 | 34.32 | 55.34 | 2.96 | 15.23 | 18.23 | 52.40 | 37.30 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| T0-3B | 8.69 | 3.02 | 4.37 | 35.49 | 23.20 | 23.57 | 2.56 | 12.06 | 0.02 | 2.50 | 22.10 |
| T0-11B | 0.63 | 0.16 | 12.41 | 92.86 | 32.17 | 28.35 | 1.56 | 7.19 | 0.00 | 3.90 | 22.10 |
| MVP | 1.01 | 0.31 | 7.18 | 86.26 | 35.54 | 32.71 | 2.87 | 16.38 | 3.08 | 2.50 | 22.20 |
| MVP+S | 10.52 | 3.54 | 2.13 | 69.55 | 37.04 | 33.38 | 2.66 | 14.84 | 0.38 | 2.50 | 22.10 |
| AESOP | Quora | ||||
| B-4 | R-1 | R-2 | R-L | ME | |
| +BART | 47.30a | 73.30 | 54.10 | 75.10 | 49.70 |
| +MVP | 49.81 | 74.78 | 56.84 | 76.34 | 53.40 |
| SC & BLEU | GYAFC E&M | GYAFC F&R | ||||
| B-4 | Accuracy | HM | B-4 | Accuracy | HM | |
| +BART | 76.50b | 93.70 | 83.90 | 79.30 | 92.00 | 85.20 |
| +MVP | 77.18 | 94.49 | 84.96 | 79.43 | 92.12 | 85.31 |
| Methods | CoLA Matt. | SST-2 Acc. | MRPC F1/Acc. | STS-B P/S Corr. | QQP F1/Acc. | MNLI m./mm. | QNLI Acc. | RTE Acc. | Average |
| BART | 60.30 | 96.30 | 90.47 / 86.70 | 90.97 / 90.30 | 73.03 / 89.87 | 90.03 / 89.27 | 94.60 | 79.83 | 85.17 |
| MVP | 59.87 | 96.43 | 92.07 / 89.43 | 91.37 / 90.90 | 73.20 / 90.13 | 89.70 / 88.73 | 95.10 | 82.87 | 85.88 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP+S | 43.03 | 20.27 | 39.72 | 66.73 | 47.42 | 76.36 | 25.28 | 26.66 | 52.69 | 86.44 | 76.84 |
| BART+R | 42.47 | 19.82 | 39.15 | 65.54 | 46.86 | 75.24 | 24.27 | 26.07 | 52.03 | 82.22 | 71.92 |
| MVP+R | 42.84 | 20.21 | 39.61 | 66.12 | 47.12 | 75.83 | 25.05 | 26.34 | 52.57 | 85.51 | 75.56 |
| MVP+M | 42.99 | 20.36 | 39.70 | 66.40 | 47.16 | 75.89 | 25.24 | 26.49 | 52.88 | 85.90 | 76.34 |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP+S | 32.94 | 15.12 | 2.98 | 71.09 | 47.11 | 39.51 | 1.39 | 7.28 | 19.24 | 71.40 | 77.80 |
| BART+R | 32.14 | 14.71 | 2.85 | 68.94 | 46.23 | 38.98 | 1.30 | 6.82 | 17.94 | 62.20 | 69.20 |
| MVP+R | 32.28 | 14.85 | 2.97 | 70.29 | 46.70 | 39.23 | 1.31 | 6.98 | 18.86 | 64.40 | 71.40 |
| MVP+M | 32.62 | 15.28 | 2.95 | 69.58 | 46.78 | 39.40 | 1.33 | 7.13 | 19.13 | 67.20 | 72.90 |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| Datasets | MVP wins (%) | Ties (%) | BART wins (%) |
| CNN/DM | 46.50 | 10.67 | 42.83 |
| WebNLG | 32.17 | 45.67 | 22.17 |
| ROCStories | 46.50 | 11.33 | 42.17 |
| PersonaChat | 35.33 | 34.00 | 30.67 |
| Methods | #NLG (PT) | #NLU (PT) | #NLG (FT) | #NLU (FT) | SP model | SP prompts | Open source |
| FLAN | 3 | 9 | 2 | 9 | ✓ | ✘ | ✘ |
| T0 | 2 | 6 | 0 | 4 | ✓ | ✘ | ✓ |
| Muppet | 1 | 3 | 1 | 3 | ✓ | ✘ | ✓ |
| ExT5 | 3 | 8 | 6 | 8 | ✓ | ✘ | ✘ |
| SPoT | 1 | 4 | 0 | 6 | ✘ | ✓ | ✘ |
| MVP (ours) | 7 | 0 | 11 | 3 | ✓ | ✓ | ✓ |
| Dataset | #Train | Cleaned #Train | #Valid | #Test | Input | Output | License |
| AGENDA | 38,720 | 38,720 | 1,000 | 1,000 | 52.1 | 141.2 | N/A |
| ENT-DESC | 88,652 | 88,652 | 11,081 | 11,081 | 279.9 | 31.0 | N/A |
| GenWiki | 681,436 | 681,436 | 75,716 | 1,000 | 21.4 | 29.5 | MIT |
| LogicNLG | 28,450 | 28,450 | 4,260 | 4,305 | 178.4 | 14.2 | MIT |
| TEKGEN | 6,310,061 | 6,307,995 | 788,746 | 796,982 | 17.0 | 21.2 | CC BY-SA 2.0 |
| WEATHERGOV | 25,000 | 25,000 | 1,000 | 3,528 | 148.7 | 30.6 | N/A |
| WikiTableT | 1,453,794 | 1,452,778 | 4,533 | 4,351 | 81.0 | 99.7 | MIT |
| Cleaned OS Dialogs | 13,355,487 | 13,355,368 | 1,483,944 | - | 75.5 | 16.7 | N/A |
| CMUDoG | 82,818 | 82,818 | 5,555 | 14,510 | 433.0 | 12.2 | N/A |
| Curiosity | 64,930 | 64,551 | 8,539 | 8,495 | 144.4 | 20.2 | CC BY-NC 4.0 |
| DREAM | 14,264 | 14,242 | 4,709 | 4,766 | 75.6 | 13.6 | N/A |
| Empathetic Dialogues | 64,636 | 64,636 | 9,308 | 8,426 | 52.7 | 12.9 | CC BY-NC 4.0 |
| Movie Dialog | 762,751 | 762,711 | 8,216 | 8,066 | 126.9 | 44.0 | N/A |
| MuTual | 33,691 | 33,691 | 4,090 | 3,248 | 53.6 | 14.5 | N/A |
| OpenDialKG | 69,680 | 69,680 | 7,743 | - | 54.2 | 12.4 | CC BY-NC 4.0 |
| Topical-Chat | 179,750 | 179,750 | 22,295 | 22,452 | 223.3 | 20.0 | CDLA-Sharing-1.0 |
| Wizard of Wikipedia | 148,357 | 147,702 | 15,767 | 15,564 | 297.0 | 16.7 | MIT |
| HotpotQA | 90,447 | 87,815 | 7,405 | - | 187.9 | 2.2 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 68.7 | 13.3 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | - | 48.1 | 3.7 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 584.1 | 4.2 | Apache 2.0 |
| Natural Questions | 96,676 | 96,676 | 10,693 | 6,490 | 9.0 | 2.1 | CC BY-SA 3.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 726.8 | 5.0 | MIT |
| QuAC | 83,568 | 83,485 | 31,906 | - | 487.9 | 12.5 | CC BY-SA 4.0 |
| TriviaQA | 78,785 | 78,785 | 8,837 | 11,313 | 14.0 | 2.0 | Apache 2.0 |
| WebQuestions | 8,933 | 8,933 | 4,863 | 4,863 | 6.7 | 2.4 | CC BY 4.0 |
| HotpotQA | 90,440 | 87,808 | 6,972 | - | 79.6 | 19.8 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 75.9 | 6.0 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | 11,022 | 45.9 | 6.0 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 579.7 | 8.6 | Apache 2.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 724.2 | 7.6 | MIT |
| QuAC | 69,109 | 69,026 | 26,301 | - | 496.7 | 6.5 | CC BY-SA 4.0 |
| ChangeMyView | 42,462 | 42,459 | 6,480 | 7,562 | 17.9 | 104.1 | MIT |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 8.8 | 33.3 | MIT |
| Hippocorpus | 6,168 | 6,168 | 686 | - | 34.1 | 262.6 | CDLA-Permissive 2.0 |
| WikiPlots | 101,642 | 101,641 | 11,294 | - | 3.4 | 338.5 | N/A |
| WritingPrompts | 272,600 | 272,518 | 15,620 | 15,138 | 28.4 | 630.8 | MIT |
| CamRest676 | 4,872 | 4,872 | 616 | - | 55.3 | 9.4 | N/A |
| Frames | 26,631 | 26,631 | 2,106 | - | 116.1 | 13.0 | MIT |
| KVRET | 14,136 | 14,136 | 1,616 | - | 30.5 | 9.3 | N/A |
| MetaLWOZ | 176,073 | 176,073 | 17,912 | - | 45.6 | 8.0 | N/A |
| MSR-E2E | 103,362 | 103,362 | 5,235 | - | 51.3 | 12.8 | Microsoft |
| Schema-Guided | 494,946 | 494,933 | 73,089 | - | 120.8 | 12.5 | CC BY-SA 4.0 |
| TaskMaster | 249,664 | 249,662 | 20,680 | - | 95.6 | 12.0 | CC BY 4.0 |
| WOZ | 6,364 | 6,359 | 1,260 | - | 47.0 | 10.6 | N/A |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 33.3 | 8.8 | MIT |
| MediaSum | 443,596 | 442,021 | 10,000 | 10,000 | 1641.0 | 14.4 | N/A |
| MSNews | 136,082 | 135,937 | 7,496 | 7,562 | 309.9 | 9.8 | CC BY-SA 4.0 |
| Newsroom | 995,041 | 989,351 | 108,837 | 108,862 | 642.4 | 26.7 | N/A |
| WikiHow | 157,252 | 157,247 | 5,599 | 5,577 | 502.6 | 45.6 | CC BY-NC-SA |
| Task | Dataset | #Train | #Valid | #Test | Input | Output | License |
| Commonsen generation | CommonGen | 67,389 | 993 | - | 5.5 | 11.6 | MIT |
| Data-to-text generation | DART | 62,659 | 2,768 | - | 27.5 | 21.5 | MIT |
| E2E | 33,525 | 4,299 | - | 9.5 | 20.6 | CC BY-SA 4.0 | |
| ToTTo | 120,761 | 7,700 | - | 37.8 | 18.0 | CC BY-SA 3.0 | |
| WebNLG | 34,338 | 4,313 | 4,222 | 18.0 | 19.9 | CC BY-NA-SA 4.0 | |
| WebNLG (GEM) | 35,426 | 1,667 | - | 17.7 | 22.7 | CC BY-NA-SA 4.0 | |
| WikiBio | 582,659 | 72,831 | 72,831 | 81.6 | 26.1 | CC BY-SA 3.0 | |
| Open-ended dialogue | DailyDialog | 76,052 | 7,069 | 6,740 | 72.5 | 13.9 | CC BY-NC-SA 4.0 |
| DSTC7-AVSD | 76,590 | 17,870 | 1,710 | 148.2 | 11.5 | MIT | |
| PersonaChat | 122,499 | 14,602 | 14,056 | 132.1 | 11.9 | MIT | |
| SGD | 164,982 | 10,000 | - | 134.7 | 11.3 | CC BY-SA 4.0 | |
| Natural language inference | MNLI-m | 392,702 | 9,815 | 9,796 | 29.8 | - | Mixed |
| MNLI-mm | 9,832 | 9,847 | |||||
| QNLI | 104,743 | 5,463 | 5,463 | 36.6 | - | CC BY-SA 4.0 | |
| RTE | 2,490 | 277 | 3,000 | 51.0 | - | N/A | |
| Paraphrase generation | Quora | 137,185 | 3,000 | 3,000 | 10.9 | 10.8 | N/A |
| Paraphrase detection | MRPC | 3,668 | 408 | 1,725 | 43.8 | - | N/A |
| QQP | 363,846 | 40,430 | 390,965 | 22.3 | - | N/A | |
| STS-B | 5,749 | 1,500 | 1,379 | 20.3 | - | N/A | |
| Question answering | CoQA | 107,286 | 31,621 | - | 349.4 | 2.6 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 156.2 | 3.6 | CC BY-SA 4.0 | |
| Question generation | CoQA | 107,286 | 31,621 | - | 346.6 | 5.5 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 148.3 | 11.6 | CC BY-SA 4.0 | |
| Story generation | ROCStories | 176,688 | 9,816 | 4,909 | 9.0 | 40.7 | N/A |
| WritingPrompts | 53,516 | 4,000 | 2,000 | 25.5 | 150.4 | MIT | |
| Task-oriented dialogue | MultiWOZ | 170,220 | 22,074 | 22,116 | 128.3 | 11.3 | MIT |
| Text classification | CoLA | 8,551 | 1,043 | 1,063 | 7.7 | - | N/A |
| SST-2 | 67,349 | 872 | 1,821 | 9.8 | - | N/A | |
| Text simplification | WiA-A | 483,801 | 20,000 | 359 | 26.2 | 21.5 | Mixed |
| WiA-T | 359 | ||||||
| Text style transfer | GYAFC-E&M | 52,595 | 11,508 | 1,416 | 9.9 | 10.6 | N/A |
| GYAFC-F&R | 51,967 | 11,152 | 1,332 | 10.7 | 11.3 | ||
| Text summarization | CNN/DailyMail | 287,227 | 13,368 | 11,490 | 679.8 | 48.3 | MIT |
| SAMSum | 14,732 | 818 | 819 | 103.4 | 20.3 | CC BY-NC-ND 4.0 | |
| WLE | 99,020 | 28,614 | - | 367.6 | 33.4 | CC0 1.0 | |
| XSum | 204,045 | 11,332 | 11,334 | 373.7 | 21.1 | MIT |
| Methods | XSum | SAMSum | CoQA QG | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | B-4 | ME | R-L | |
| BART | \\( 45.14^d \\) | 22.27 | 37.25 | \\( 51.74^b \\) | 26.46 | 48.72 | \\( 12.34^c \\) | 35.78 | 46.88 |
| MVP | 45.60 | 22.47 | 37.42 | 53.78 | 29.12 | 49.37 | 23.48 | 47.79 | 55.09 |
| MVP+S | 45.67 | 22.63 | 37.50 | 53.81 | 29.75 | 49.43 | 23.43 | 47.49 | 55.25 |
| SOTA | \\( 49.57^a \\) | 25.08 | 41.81 | \\( 53.89^b \\) | 28.85 | 49.29 | \\( 15.78^c \\) | 40.15 | 50.98 |
| Methods | WritingPrompts | DailyDialog | WikiBio | ||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | |
| BART | \\( 22.40^e \\) | 8.40 | - | 31.30 | \\( 44.30^f \\) | 39.20 | 3.90 | 21.10 | - |
| MVP | 32.34 | 13.11 | 2.12 | 64.58 | 46.19 | 41.81 | 4.61 | 25.06 | 48.42 |
| MVP+S | 30.12 | 11.46 | 3.97 | 83.70 | 45.71 | 42.92 | 5.10 | 27.14 | 48.19 |
| SOTA | \\( 22.40^e \\) | 8.40 | - | 31.30 | \\( 46.10^f \\) | 40.70 | 4.10 | 22.20 | \\( 45.10^g \\) |
| Methods | DSTC7-AVSD | SQuAD | |||||||
| B-1 | B-2 | B-3 | B-4 | ME | R-L | CIDEr | F1 | EM | |
| BART | \\( 82.40^f \\) | 69.10 | 58.20 | 48.70 | 31.30 | 63.50 | 1.38 | \\( 91.56^i \\) | 84.23 |
| MVP | 83.75 | 70.89 | 60.19 | 50.94 | 32.12 | 65.04 | 1.45 | \\( 93.45 \\) | 87.20 |
| MVP+S | 83.81 | 71.07 | 60.45 | 51.20 | 31.77 | 64.76 | 1.44 | \\( 93.45 \\) | 87.17 |
| SOTA | \\( 83.20^f \\) | 70.50 | 59.80 | 50.60 | 31.40 | 63.80 | 1.39 | \\( 96.22^h \\) | 91.26 |
| Methods | DART | E2E | ToTTo | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 34.31 | 45.22 | 36.30 | 42.57 | 46.60 | 38.20 | 39.79 | 49.90 | 36.80 |
| ExT5 | 36.62 | 48.14 | 37.60 | 42.25 | 46.70 | 38.10 | 40.14 | 50.33 | 36.90 |
| MVP | 39.13 | 48.92 | 38.53 | 37.38 | 47.96 | 39.39 | 50.58 | 55.24 | 41.27 |
| MVP+S | 38.83 | 48.49 | 38.41 | 37.32 | 47.40 | 38.90 | 50.69 | 55.52 | 41.29 |
| Methods | WebNLG | CommonGen | SGD | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 31.67 | 43.31 | 34.40 | 8.38 | 17.01 | 20.20 | 33.15 | 36.17 | 32.40 |
| ExT5 | 35.03 | 48.17 | 36.50 | 9.68 | 19.04 | 21.40 | 34.74 | 37.77 | 33.00 |
| MVP | 47.03 | 59.00 | 42.34 | 32.59 | 37.71 | 33.00 | 45.63 | 48.29 | 38.48 |
| MVP+S | 47.03 | 59.03 | 42.28 | 34.10 | 37.87 | 33.11 | 45.24 | 48.25 | 38.47 |
| Methods | WiA-A | WiA-T | WLE | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 29.30 | 38.37 | 30.10 | 42.12 | 50.52 | 36.2 | 15.55 | 20.47 | 19.60 |
| ExT5 | 29.23 | 37.98 | 30.00 | 41.39 | 50.38 | 35.8 | 16.64 | 21.16 | 20.40 |
| MVP | 71.55 | 70.88 | 48.19 | 91.73 | 83.46 | 57.34 | 18.80 | 22.84 | 21.95 |
| MVP+S | 70.37 | 70.65 | 47.70 | 91.12 | 83.59 | 56.95 | 18.52 | 22.57 | 22.02 |
| Input | ||
| she was on a flight. | ||
| Output | ||
| she was trying to take a nap . suddenly , her ears started ringing . the flight attendant tried to fix it but she could n’t . she had to call for help . luckily, they were able to fix the problem . | she was bored and her ears hurt . she decided to take a nap . luckily , she was able to get a good night ’s sleep . but the next morning , she woke up and felt sick . | |
| Left Wins | Ties | Right Wins |
| Input\nDescribe the following data: Abilene,_Texas | cityServed | Abilene_Regional_Airport |
| Gold\nAbilene, Texas is served by the Abilene regional airport.\nAbilene Regional Airport serves the city of Abilene in Texas. |
| BART\nAbilene Regional Airport serves the city of Abilene in Texas. |
| MVP\nAbilene Regional Airport serves the city of Abilene, Texas. |
| MVP+S\nAbilene Regional Airport serves the city of Abilene, Texas. |
| Input\nDescribe the following data: "Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas" | location | Adolfo_Suárez_Madrid-Barajas_Airport |
| Gold\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas.\nAdolfo Suarez Madrid-Barajas airport is located at Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas.\nAdolfo Suarez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastian de los Reyes and Alcobendas. |
| BART\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP+S\nAdolfo Suárez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| Input\nGenerate the question based on the answer: Saint Bernadette Soubirous [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend "Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold\nTo whom did the Virgin Mary allegedly appear in 1858 in Lourdes France? |
| BART\nWho is believed to have appeared to the Virgin Mary at Lourdes? |
| MVP\nWho did the Virgin Mary appear to in Lourdes? |
| MVP+S\nThe Grotto is a replica of the grotto at Lourdes, France where the Virgin Mary appeared to whom? |
| Input\nGenerate the question based on the answer: a copper statue of Christ [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend " Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold\nWhat is in front of the Notre Dame Main Building? |
| BART\nWhat is in front of the Main Building and facing it? |
| MVP\nWhat is immediately in front of the Main Building? |
| MVP+S\nWhat is immediately in front of the Main Building? |
| Input\nGiven the story title: male was out jogging one morning. |
| Gold\nthe weather was crisp and cool . male felt good and energetic . he decided to keep jogging longer than normal . male went several more miles out of his way . |
| BART\nhe saw a man struggling to walk across a field . male stopped to help him . he saw a guy struggling to walking across a fields . the man thanked male profusely ! |
| MVP\nhe ran into a man who was holding a knife . male was scared , so he ran away . the man threatened male . male ran back into his house . |
| MVP+S\nhe noticed a sound coming from a tree . male went to investigate . he found a small baby bird stuck in the trunk . male took the bird to the vet and they released it . |
| Input\nGiven the story title: ken was driving around in the snow. |
| Gold\nhe needed to get home from work . he was driving slowly to avoid accidents . unfortunately the roads were too slick and ken lost control . his tires lost traction and he hit a tree . |
| BART\nhe hit a patch of black ice . ken slid and hurt his knee . he went to the doctor for it . krampus was sent home with a concussion . |
| MVP\nhe was in his car and it was very cold . ken started to back up . the car started to make a loud noise . kena had to back out and call for help . |
| MVP+S\nhe was having trouble starting his car . he had to push it a couple of blocks . his car would n’t start . ken had to call a tow truck for help . |
| Input\nGiven the dialog: i love to meet new people. [SEP] i have a turtle named timothy. [SEP] my favorite sport is ultimate frisbee. [SEP] my parents are living in bora bora. [SEP] autumn is my favorite season. [X_SEP] hello, how are you doing tonight? |
| Gold\ni am well an loving this interaction how are you? |
| BART\ni am doing well, how are you? |
| MVP\ni am doing well. how are you? |
| MVP+S\ni am doing well. how are you? |
| Input\nGiven the dialog: i just bought a brand new house. [SEP] i like to dance at the club. [SEP] i run a dog obedience school. [SEP] i have a big sweet tooth. [SEP] i like taking and posting selkies. [X_SEP] hello, how are you doing tonight? [SEP] i am well an loving this interaction how are you? |
| Gold\ni am great. i just got back from the club. |
| BART\ni am doing well, thanks for asking. what do you do for fun? |
| MVP\ni am doing well. i just got home from dancing at the club. |
| MVP+S\ni am doing well. i just bought a new house. |
| Settings | Supervised Pre-training | Unsupervised Pre-training |
| NLG | MVP (ours) | GPT-2, MASS, BART, T5 |
| NLU | FLAN, T0, Muppet, ExT5 | BERT, XLNet, RoBERTa, T5 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| BART | 44.16e | 21.28 | 40.90 | 64.55b | 46.51 | 75.13 | 22.00f | 26.40 | 52.55 | 68.60f | - |
| Flan-T5 | 43.45 | 21.01 | 40.03 | 66.60 | 46.93 | 75.76 | 25.55 | 26.90 | 53.51 | 84.18 | 75.44 |
| Single | 44.36 | 21.54 | 40.88 | 67.74 | 46.89 | 76.94 | 26.09 | 27.15 | 53.29 | 86.20 | 77.26 |
| MVP+S | 44.63 | 21.72 | 41.21 | 68.19 | 47.75 | 76.81 | 25.69 | 27.04 | 53.20 | 86.65 | 77.93 |
| MVP+R | 44.14 | 21.45 | 40.72 | 67.61 | 47.65 | 76.70 | 25.71 | 27.03 | 53.09 | 85.95 | 77.22 |
| MVP+M | 43.97 | 21.16 | 40.46 | 67.45 | 47.57 | 76.81 | 25.46 | 26.79 | 52.95 | 86.28 | 77.26 |
| SOTA | 47.16a | 22.55 | 43.87 | 66.14b | 47.25 | 76.10 | 25.97c | 27.33 | 53.43 | 84.50d | - |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| BART | 30.70g | 13.30 | - | 69.90 | 49.90f | 40.00 | 1.30 | 8.00 | 17.89j | 74.91 | 84.88 |
| Flan-T5 | 32.72 | 15.23 | 2.97 | 68.97 | 48.55 | 40.22 | 1.40 | 7.85 | 19.73 | 70.20 | 78.70 |
| Single | 32.67 | 15.29 | 2.72 | 72.97 | 49.96 | 40.53 | 1.27 | 7.63 | 19.73 | 75.60 | 83.70 |
| MVP+S | 33.92 | 15.60 | 3.44 | 80.58 | 47.91 | 39.97 | 1.52 | 9.54 | 20.32 | 79.90 | 86.80 |
| MVP+R | 32.93 | 15.32 | 2.88 | 73.83 | 48.45 | 40.09 | 1.30 | 7.95 | 19.02 | 73.30 | 81.80 |
| MVP+M | 33.30 | 15.51 | 2.71 | 74.24 | 46.26 | 39.30 | 1.36 | 8.07 | 19.93 | 72.70 | 79.70 |
| SOTA | 33.40g | 15.40 | - | 69.30 | 49.90f | 40.00 | 1.50h | 9.40 | 20.50i | 85.30 | 94.40 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| T0-3B | - | - | - | 1.40 | 10.20 | 18.43 | 3.06 | 12.43 | 14.91 | 13.30 | 6.60 |
| T0-11B | - | - | - | 0.26 | 6.13 | 14.12 | 2.63 | 7.00 | 15.25 | 9.18 | 4.36 |
| MVP | 29.50 | 11.29 | 25.92 | 34.42 | 31.33 | 52.33 | 2.90 | 13.94 | 15.48 | 29.40 | 18.20 |
| MVP+S | 25.60 | 9.51 | 22.67 | 39.43 | 34.32 | 55.34 | 2.96 | 15.23 | 18.23 | 52.40 | 37.30 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| T0-3B | 8.69 | 3.02 | 4.37 | 35.49 | 23.20 | 23.57 | 2.56 | 12.06 | 0.02 | 2.50 | 22.10 |
| T0-11B | 0.63 | 0.16 | 12.41 | 92.86 | 32.17 | 28.35 | 1.56 | 7.19 | 0.00 | 3.90 | 22.10 |
| MVP | 1.01 | 0.31 | 7.18 | 86.26 | 35.54 | 32.71 | 2.87 | 16.38 | 3.08 | 2.50 | 22.20 |
| MVP+S | 10.52 | 3.54 | 2.13 | 69.55 | 37.04 | 33.38 | 2.66 | 14.84 | 0.38 | 2.50 | 22.10 |
| AESOP | Quora | ||||
| B-4 | R-1 | R-2 | R-L | ME | |
| +BART | 47.30a | 73.30 | 54.10 | 75.10 | 49.70 |
| +MVP | 49.81 | 74.78 | 56.84 | 76.34 | 53.40 |
| SC & BLEU | GYAFC E&M | GYAFC F&R | ||||
| B-4 | Accuracy | HM | B-4 | Accuracy | HM | |
| +BART | 76.50b | 93.70 | 83.90 | 79.30 | 92.00 | 85.20 |
| +MVP | 77.18 | 94.49 | 84.96 | 79.43 | 92.12 | 85.31 |
| Methods | CoLA Matt. | SST-2 Acc. | MRPC F1/Acc. | STS-B P/S Corr. | QQP F1/Acc. | MNLI m./mm. | QNLI Acc. | RTE Acc. | Average |
| BART | 60.30 | 96.30 | 90.47 / 86.70 | 90.97 / 90.30 | 73.03 / 89.87 | 90.03 / 89.27 | 94.60 | 79.83 | 85.17 |
| MVP | 59.87 | 96.43 | 92.07 / 89.43 | 91.37 / 90.90 | 73.20 / 90.13 | 89.70 / 88.73 | 95.10 | 82.87 | 85.88 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP+S | 43.03 | 20.27 | 39.72 | 66.73 | 47.42 | 76.36 | 25.28 | 26.66 | 52.69 | 86.44 | 76.84 |
| BART+R | 42.47 | 19.82 | 39.15 | 65.54 | 46.86 | 75.24 | 24.27 | 26.07 | 52.03 | 82.22 | 71.92 |
| MVP+R | 42.84 | 20.21 | 39.61 | 66.12 | 47.12 | 75.83 | 25.05 | 26.34 | 52.57 | 85.51 | 75.56 |
| MVP+M | 42.99 | 20.36 | 39.70 | 66.40 | 47.16 | 75.89 | 25.24 | 26.49 | 52.88 | 85.90 | 76.34 |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP+S | 32.94 | 15.12 | 2.98 | 71.09 | 47.11 | 39.51 | 1.39 | 7.28 | 19.24 | 71.40 | 77.80 |
| BART+R | 32.14 | 14.71 | 2.85 | 68.94 | 46.23 | 38.98 | 1.30 | 6.82 | 17.94 | 62.20 | 69.20 |
| MVP+R | 32.28 | 14.85 | 2.97 | 70.29 | 46.70 | 39.23 | 1.31 | 6.98 | 18.86 | 64.40 | 71.40 |
| MVP+M | 32.62 | 15.28 | 2.95 | 69.58 | 46.78 | 39.40 | 1.33 | 7.13 | 19.13 | 67.20 | 72.90 |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| Datasets | MVP wins (%) | Ties (%) | BART wins (%) |
| CNN/DM | 46.50 | 10.67 | 42.83 |
| WebNLG | 32.17 | 45.67 | 22.17 |
| ROCStories | 46.50 | 11.33 | 42.17 |
| PersonaChat | 35.33 | 34.00 | 30.67 |
| Methods | #NLG (PT) | #NLU (PT) | #NLG (FT) | #NLU (FT) | SP model | SP prompts | Open source |
| FLAN | 3 | 9 | 2 | 9 | ✓ | ✘ | ✘ |
| T0 | 2 | 6 | 0 | 4 | ✓ | ✘ | ✓ |
| Muppet | 1 | 3 | 1 | 3 | ✓ | ✘ | ✓ |
| ExT5 | 3 | 8 | 6 | 8 | ✓ | ✘ | ✘ |
| SPoT | 1 | 4 | 0 | 6 | ✘ | ✓ | ✘ |
| MVP (ours) | 7 | 0 | 11 | 3 | ✓ | ✓ | ✓ |
| Dataset | #Train | Cleaned #Train | #Valid | #Test | Input | Output | License |
| AGENDA | 38,720 | 38,720 | 1,000 | 1,000 | 52.1 | 141.2 | N/A |
| ENT-DESC | 88,652 | 88,652 | 11,081 | 11,081 | 279.9 | 31.0 | N/A |
| GenWiki | 681,436 | 681,436 | 75,716 | 1,000 | 21.4 | 29.5 | MIT |
| LogicNLG | 28,450 | 28,450 | 4,260 | 4,305 | 178.4 | 14.2 | MIT |
| TEKGEN | 6,310,061 | 6,307,995 | 788,746 | 796,982 | 17.0 | 21.2 | CC BY-SA 2.0 |
| WEATHERGOV | 25,000 | 25,000 | 1,000 | 3,528 | 148.7 | 30.6 | N/A |
| WikiTableT | 1,453,794 | 1,452,778 | 4,533 | 4,351 | 81.0 | 99.7 | MIT |
| Cleaned OS Dialogs | 13,355,487 | 13,355,368 | 1,483,944 | - | 75.5 | 16.7 | N/A |
| CMUDoG | 82,818 | 82,818 | 5,555 | 14,510 | 433.0 | 12.2 | N/A |
| Curiosity | 64,930 | 64,551 | 8,539 | 8,495 | 144.4 | 20.2 | CC BY-NC 4.0 |
| DREAM | 14,264 | 14,242 | 4,709 | 4,766 | 75.6 | 13.6 | N/A |
| Empathetic Dialogues | 64,636 | 64,636 | 9,308 | 8,426 | 52.7 | 12.9 | CC BY-NC 4.0 |
| Movie Dialog | 762,751 | 762,711 | 8,216 | 8,066 | 126.9 | 44.0 | N/A |
| MuTual | 33,691 | 33,691 | 4,090 | 3,248 | 53.6 | 14.5 | N/A |
| OpenDialKG | 69,680 | 69,680 | 7,743 | - | 54.2 | 12.4 | CC BY-NC 4.0 |
| Topical-Chat | 179,750 | 179,750 | 22,295 | 22,452 | 223.3 | 20.0 | CDLA-Sharing-1.0 |
| Wizard of Wikipedia | 148,357 | 147,702 | 15,767 | 15,564 | 297.0 | 16.7 | MIT |
| HotpotQA | 90,447 | 87,815 | 7,405 | - | 187.9 | 2.2 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 68.7 | 13.3 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | - | 48.1 | 3.7 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 584.1 | 4.2 | Apache 2.0 |
| Natural Questions | 96,676 | 96,676 | 10,693 | 6,490 | 9.0 | 2.1 | CC BY-SA 3.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 726.8 | 5.0 | MIT |
| QuAC | 83,568 | 83,485 | 31,906 | - | 487.9 | 12.5 | CC BY-SA 4.0 |
| TriviaQA | 78,785 | 78,785 | 8,837 | 11,313 | 14.0 | 2.0 | Apache 2.0 |
| WebQuestions | 8,933 | 8,933 | 4,863 | 4,863 | 6.7 | 2.4 | CC BY 4.0 |
| HotpotQA | 90,440 | 87,808 | 6,972 | - | 79.6 | 19.8 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 75.9 | 6.0 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | 11,022 | 45.9 | 6.0 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 579.7 | 8.6 | Apache 2.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 724.2 | 7.6 | MIT |
| QuAC | 69,109 | 69,026 | 26,301 | - | 496.7 | 6.5 | CC BY-SA 4.0 |
| ChangeMyView | 42,462 | 42,459 | 6,480 | 7,562 | 17.9 | 104.1 | MIT |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 8.8 | 33.3 | MIT |
| Hippocorpus | 6,168 | 6,168 | 686 | - | 34.1 | 262.6 | CDLA-Permissive 2.0 |
| WikiPlots | 101,642 | 101,641 | 11,294 | - | 3.4 | 338.5 | N/A |
| WritingPrompts | 272,600 | 272,518 | 15,620 | 15,138 | 28.4 | 630.8 | MIT |
| CamRest676 | 4,872 | 4,872 | 616 | - | 55.3 | 9.4 | N/A |
| Frames | 26,631 | 26,631 | 2,106 | - | 116.1 | 13.0 | MIT |
| KVRET | 14,136 | 14,136 | 1,616 | - | 30.5 | 9.3 | N/A |
| MetaLWOZ | 176,073 | 176,073 | 17,912 | - | 45.6 | 8.0 | N/A |
| MSR-E2E | 103,362 | 103,362 | 5,235 | - | 51.3 | 12.8 | Microsoft |
| Schema-Guided | 494,946 | 494,933 | 73,089 | - | 120.8 | 12.5 | CC BY-SA 4.0 |
| TaskMaster | 249,664 | 249,662 | 20,680 | - | 95.6 | 12.0 | CC BY 4.0 |
| WOZ | 6,364 | 6,359 | 1,260 | - | 47.0 | 10.6 | N/A |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 33.3 | 8.8 | MIT |
| MediaSum | 443,596 | 442,021 | 10,000 | 10,000 | 1641.0 | 14.4 | N/A |
| MSNews | 136,082 | 135,937 | 7,496 | 7,562 | 309.9 | 9.8 | CC BY-SA 4.0 |
| Newsroom | 995,041 | 989,351 | 108,837 | 108,862 | 642.4 | 26.7 | N/A |
| WikiHow | 157,252 | 157,247 | 5,599 | 5,577 | 502.6 | 45.6 | CC BY-NC-SA |
| Task | Dataset | #Train | #Valid | #Test | Input | Output | License |
| Commonsen generation | CommonGen | 67,389 | 993 | - | 5.5 | 11.6 | MIT |
| Data-to-text generation | DART | 62,659 | 2,768 | - | 27.5 | 21.5 | MIT |
| E2E | 33,525 | 4,299 | - | 9.5 | 20.6 | CC BY-SA 4.0 | |
| ToTTo | 120,761 | 7,700 | - | 37.8 | 18.0 | CC BY-SA 3.0 | |
| WebNLG | 34,338 | 4,313 | 4,222 | 18.0 | 19.9 | CC BY-NA-SA 4.0 | |
| WebNLG (GEM) | 35,426 | 1,667 | - | 17.7 | 22.7 | CC BY-NA-SA 4.0 | |
| WikiBio | 582,659 | 72,831 | 72,831 | 81.6 | 26.1 | CC BY-SA 3.0 | |
| Open-ended dialogue | DailyDialog | 76,052 | 7,069 | 6,740 | 72.5 | 13.9 | CC BY-NC-SA 4.0 |
| DSTC7-AVSD | 76,590 | 17,870 | 1,710 | 148.2 | 11.5 | MIT | |
| PersonaChat | 122,499 | 14,602 | 14,056 | 132.1 | 11.9 | MIT | |
| SGD | 164,982 | 10,000 | - | 134.7 | 11.3 | CC BY-SA 4.0 | |
| Natural language inference | MNLI-m | 392,702 | 9,815 | 9,796 | 29.8 | - | Mixed |
| MNLI-mm | 9,832 | 9,847 | |||||
| QNLI | 104,743 | 5,463 | 5,463 | 36.6 | - | CC BY-SA 4.0 | |
| RTE | 2,490 | 277 | 3,000 | 51.0 | - | N/A | |
| Paraphrase generation | Quora | 137,185 | 3,000 | 3,000 | 10.9 | 10.8 | N/A |
| Paraphrase detection | MRPC | 3,668 | 408 | 1,725 | 43.8 | - | N/A |
| QQP | 363,846 | 40,430 | 390,965 | 22.3 | - | N/A | |
| STS-B | 5,749 | 1,500 | 1,379 | 20.3 | - | N/A | |
| Question answering | CoQA | 107,286 | 31,621 | - | 349.4 | 2.6 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 156.2 | 3.6 | CC BY-SA 4.0 | |
| Question generation | CoQA | 107,286 | 31,621 | - | 346.6 | 5.5 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 148.3 | 11.6 | CC BY-SA 4.0 | |
| Story generation | ROCStories | 176,688 | 9,816 | 4,909 | 9.0 | 40.7 | N/A |
| WritingPrompts | 53,516 | 4,000 | 2,000 | 25.5 | 150.4 | MIT | |
| Task-oriented dialogue | MultiWOZ | 170,220 | 22,074 | 22,116 | 128.3 | 11.3 | MIT |
| Text classification | CoLA | 8,551 | 1,043 | 1,063 | 7.7 | - | N/A |
| SST-2 | 67,349 | 872 | 1,821 | 9.8 | - | N/A | |
| Text simplification | WiA-A | 483,801 | 20,000 | 359 | 26.2 | 21.5 | Mixed |
| WiA-T | 359 | ||||||
| Text style transfer | GYAFC-E&M | 52,595 | 11,508 | 1,416 | 9.9 | 10.6 | N/A |
| GYAFC-F&R | 51,967 | 11,152 | 1,332 | 10.7 | 11.3 | ||
| Text summarization | CNN/DailyMail | 287,227 | 13,368 | 11,490 | 679.8 | 48.3 | MIT |
| SAMSum | 14,732 | 818 | 819 | 103.4 | 20.3 | CC BY-NC-ND 4.0 | |
| WLE | 99,020 | 28,614 | - | 367.6 | 33.4 | CC0 1.0 | |
| XSum | 204,045 | 11,332 | 11,334 | 373.7 | 21.1 | MIT |
| Methods | XSum | SAMSum | CoQA QG | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | B-4 | ME | R-L | |
| BART | \\( 45.14^d \\) | 22.27 | 37.25 | \\( 51.74^b \\) | 26.46 | 48.72 | \\( 12.34^c \\) | 35.78 | 46.88 |
| MVP | 45.60 | 22.47 | 37.42 | 53.78 | 29.12 | 49.37 | 23.48 | 47.79 | 55.09 |
| MVP+S | 45.67 | 22.63 | 37.50 | 53.81 | 29.75 | 49.43 | 23.43 | 47.49 | 55.25 |
| SOTA | \\( 49.57^a \\) | 25.08 | 41.81 | \\( 53.89^b \\) | 28.85 | 49.29 | \\( 15.78^c \\) | 40.15 | 50.98 |
| Methods | WritingPrompts | DailyDialog | WikiBio | ||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | |
| BART | \\( 22.40^e \\) | 8.40 | - | 31.30 | \\( 44.30^f \\) | 39.20 | 3.90 | 21.10 | - |
| MVP | 32.34 | 13.11 | 2.12 | 64.58 | 46.19 | 41.81 | 4.61 | 25.06 | 48.42 |
| MVP+S | 30.12 | 11.46 | 3.97 | 83.70 | 45.71 | 42.92 | 5.10 | 27.14 | 48.19 |
| SOTA | \\( 22.40^e \\) | 8.40 | - | 31.30 | \\( 46.10^f \\) | 40.70 | 4.10 | 22.20 | \\( 45.10^g \\) |
| Methods | DSTC7-AVSD | SQuAD | |||||||
| B-1 | B-2 | B-3 | B-4 | ME | R-L | CIDEr | F1 | EM | |
| BART | \\( 82.40^f \\) | 69.10 | 58.20 | 48.70 | 31.30 | 63.50 | 1.38 | \\( 91.56^i \\) | 84.23 |
| MVP | 83.75 | 70.89 | 60.19 | 50.94 | 32.12 | 65.04 | 1.45 | \\( 93.45 \\) | 87.20 |
| MVP+S | 83.81 | 71.07 | 60.45 | 51.20 | 31.77 | 64.76 | 1.44 | \\( 93.45 \\) | 87.17 |
| SOTA | \\( 83.20^f \\) | 70.50 | 59.80 | 50.60 | 31.40 | 63.80 | 1.39 | \\( 96.22^h \\) | 91.26 |
| Methods | DART | E2E | ToTTo | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 34.31 | 45.22 | 36.30 | 42.57 | 46.60 | 38.20 | 39.79 | 49.90 | 36.80 |
| ExT5 | 36.62 | 48.14 | 37.60 | 42.25 | 46.70 | 38.10 | 40.14 | 50.33 | 36.90 |
| MVP | 39.13 | 48.92 | 38.53 | 37.38 | 47.96 | 39.39 | 50.58 | 55.24 | 41.27 |
| MVP+S | 38.83 | 48.49 | 38.41 | 37.32 | 47.40 | 38.90 | 50.69 | 55.52 | 41.29 |
| Methods | WebNLG | CommonGen | SGD | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 31.67 | 43.31 | 34.40 | 8.38 | 17.01 | 20.20 | 33.15 | 36.17 | 32.40 |
| ExT5 | 35.03 | 48.17 | 36.50 | 9.68 | 19.04 | 21.40 | 34.74 | 37.77 | 33.00 |
| MVP | 47.03 | 59.00 | 42.34 | 32.59 | 37.71 | 33.00 | 45.63 | 48.29 | 38.48 |
| MVP+S | 47.03 | 59.03 | 42.28 | 34.10 | 37.87 | 33.11 | 45.24 | 48.25 | 38.47 |
| Methods | WiA-A | WiA-T | WLE | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 29.30 | 38.37 | 30.10 | 42.12 | 50.52 | 36.2 | 15.55 | 20.47 | 19.60 |
| ExT5 | 29.23 | 37.98 | 30.00 | 41.39 | 50.38 | 35.8 | 16.64 | 21.16 | 20.40 |
| MVP | 71.55 | 70.88 | 48.19 | 91.73 | 83.46 | 57.34 | 18.80 | 22.84 | 21.95 |
| MVP+S | 70.37 | 70.65 | 47.70 | 91.12 | 83.59 | 56.95 | 18.52 | 22.57 | 22.02 |
| Input | ||
| she was on a flight. | ||
| Output | ||
| she was trying to take a nap . suddenly , her ears started ringing . the flight attendant tried to fix it but she could n’t . she had to call for help . luckily, they were able to fix the problem . | she was bored and her ears hurt . she decided to take a nap . luckily , she was able to get a good night ’s sleep . but the next morning , she woke up and felt sick . | |
| Left Wins | Ties | Right Wins |
| Input\nDescribe the following data: Abilene,_Texas | cityServed | Abilene_Regional_Airport |
| Gold\nAbilene, Texas is served by the Abilene regional airport.\nAbilene Regional Airport serves the city of Abilene in Texas. |
| BART\nAbilene Regional Airport serves the city of Abilene in Texas. |
| MVP\nAbilene Regional Airport serves the city of Abilene, Texas. |
| MVP+S\nAbilene Regional Airport serves the city of Abilene, Texas. |
| Input\nDescribe the following data: "Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas" | location | Adolfo_Suárez_Madrid-Barajas_Airport |
| Gold\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas.\nAdolfo Suarez Madrid-Barajas airport is located at Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas.\nAdolfo Suarez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastian de los Reyes and Alcobendas. |
| BART\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP+S\nAdolfo Suárez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| Input\nGenerate the question based on the answer: Saint Bernadette Soubirous [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend "Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold\nTo whom did the Virgin Mary allegedly appear in 1858 in Lourdes France? |
| BART\nWho is believed to have appeared to the Virgin Mary at Lourdes? |
| MVP\nWho did the Virgin Mary appear to in Lourdes? |
| MVP+S\nThe Grotto is a replica of the grotto at Lourdes, France where the Virgin Mary appeared to whom? |
| Input\nGenerate the question based on the answer: a copper statue of Christ [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend " Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold\nWhat is in front of the Notre Dame Main Building? |
| BART\nWhat is in front of the Main Building and facing it? |
| MVP\nWhat is immediately in front of the Main Building? |
| MVP+S\nWhat is immediately in front of the Main Building? |
| Input\nGiven the story title: male was out jogging one morning. |
| Gold\nthe weather was crisp and cool . male felt good and energetic . he decided to keep jogging longer than normal . male went several more miles out of his way . |
| BART\nhe saw a man struggling to walk across a field . male stopped to help him . he saw a guy struggling to walking across a fields . the man thanked male profusely ! |
| MVP\nhe ran into a man who was holding a knife . male was scared , so he ran away . the man threatened male . male ran back into his house . |
| MVP+S\nhe noticed a sound coming from a tree . male went to investigate . he found a small baby bird stuck in the trunk . male took the bird to the vet and they released it . |
| Input\nGiven the story title: ken was driving around in the snow. |
| Gold\nhe needed to get home from work . he was driving slowly to avoid accidents . unfortunately the roads were too slick and ken lost control . his tires lost traction and he hit a tree . |
| BART\nhe hit a patch of black ice . ken slid and hurt his knee . he went to the doctor for it . krampus was sent home with a concussion . |
| MVP\nhe was in his car and it was very cold . ken started to back up . the car started to make a loud noise . kena had to back out and call for help . |
| MVP+S\nhe was having trouble starting his car . he had to push it a couple of blocks . his car would n’t start . ken had to call a tow truck for help . |
| Input\nGiven the dialog: i love to meet new people. [SEP] i have a turtle named timothy. [SEP] my favorite sport is ultimate frisbee. [SEP] my parents are living in bora bora. [SEP] autumn is my favorite season. [X_SEP] hello, how are you doing tonight? |
| Gold\ni am well an loving this interaction how are you? |
| BART\ni am doing well, how are you? |
| MVP\ni am doing well. how are you? |
| MVP+S\ni am doing well. how are you? |
| Input\nGiven the dialog: i just bought a brand new house. [SEP] i like to dance at the club. [SEP] i run a dog obedience school. [SEP] i have a big sweet tooth. [SEP] i like taking and posting selkies. [X_SEP] hello, how are you doing tonight? [SEP] i am well an loving this interaction how are you? |
| Gold\ni am great. i just got back from the club. |
| BART\ni am doing well, thanks for asking. what do you do for fun? |
| MVP\ni am doing well. i just got home from dancing at the club. |
| MVP+S\ni am doing well. i just bought a new house. |
| Settings | Supervised Pre-training | Unsupervised Pre-training |
| NLG | MVP (ours) | GPT-2, MASS, BART, T5 |
| NLU | FLAN, T0, Muppet, ExT5 | BERT, XLNet, RoBERTa, T5 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| BART | 44.16e | 21.28 | 40.90 | 64.55b | 46.51 | 75.13 | 22.00f | 26.40 | 52.55 | 68.60f | - |
| Flan-T5 | 43.45 | 21.01 | 40.03 | 66.60 | 46.93 | 75.76 | 25.55 | 26.90 | 53.51 | 84.18 | 75.44 |
| Single | 44.36 | 21.54 | 40.88 | 67.74 | 46.89 | 76.94 | 26.09 | 27.15 | 53.29 | 86.20 | 77.26 |
| MVP+S | 44.63 | 21.72 | 41.21 | 68.19 | 47.75 | 76.81 | 25.69 | 27.04 | 53.20 | 86.65 | 77.93 |
| MVP+R | 44.14 | 21.45 | 40.72 | 67.61 | 47.65 | 76.70 | 25.71 | 27.03 | 53.09 | 85.95 | 77.22 |
| MVP+M | 43.97 | 21.16 | 40.46 | 67.45 | 47.57 | 76.81 | 25.46 | 26.79 | 52.95 | 86.28 | 77.26 |
| SOTA | 47.16a | 22.55 | 43.87 | 66.14b | 47.25 | 76.10 | 25.97c | 27.33 | 53.43 | 84.50d | - |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| BART | 30.70g | 13.30 | - | 69.90 | 49.90f | 40.00 | 1.30 | 8.00 | 17.89j | 74.91 | 84.88 |
| Flan-T5 | 32.72 | 15.23 | 2.97 | 68.97 | 48.55 | 40.22 | 1.40 | 7.85 | 19.73 | 70.20 | 78.70 |
| Single | 32.67 | 15.29 | 2.72 | 72.97 | 49.96 | 40.53 | 1.27 | 7.63 | 19.73 | 75.60 | 83.70 |
| MVP+S | 33.92 | 15.60 | 3.44 | 80.58 | 47.91 | 39.97 | 1.52 | 9.54 | 20.32 | 79.90 | 86.80 |
| MVP+R | 32.93 | 15.32 | 2.88 | 73.83 | 48.45 | 40.09 | 1.30 | 7.95 | 19.02 | 73.30 | 81.80 |
| MVP+M | 33.30 | 15.51 | 2.71 | 74.24 | 46.26 | 39.30 | 1.36 | 8.07 | 19.93 | 72.70 | 79.70 |
| SOTA | 33.40g | 15.40 | - | 69.30 | 49.90f | 40.00 | 1.50h | 9.40 | 20.50i | 85.30 | 94.40 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| T0-3B | - | - | - | 1.40 | 10.20 | 18.43 | 3.06 | 12.43 | 14.91 | 13.30 | 6.60 |
| T0-11B | - | - | - | 0.26 | 6.13 | 14.12 | 2.63 | 7.00 | 15.25 | 9.18 | 4.36 |
| MVP | 29.50 | 11.29 | 25.92 | 34.42 | 31.33 | 52.33 | 2.90 | 13.94 | 15.48 | 29.40 | 18.20 |
| MVP+S | 25.60 | 9.51 | 22.67 | 39.43 | 34.32 | 55.34 | 2.96 | 15.23 | 18.23 | 52.40 | 37.30 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| T0-3B | 8.69 | 3.02 | 4.37 | 35.49 | 23.20 | 23.57 | 2.56 | 12.06 | 0.02 | 2.50 | 22.10 |
| T0-11B | 0.63 | 0.16 | 12.41 | 92.86 | 32.17 | 28.35 | 1.56 | 7.19 | 0.00 | 3.90 | 22.10 |
| MVP | 1.01 | 0.31 | 7.18 | 86.26 | 35.54 | 32.71 | 2.87 | 16.38 | 3.08 | 2.50 | 22.20 |
| MVP+S | 10.52 | 3.54 | 2.13 | 69.55 | 37.04 | 33.38 | 2.66 | 14.84 | 0.38 | 2.50 | 22.10 |
| AESOP | Quora | ||||
| B-4 | R-1 | R-2 | R-L | ME | |
| +BART | 47.30a | 73.30 | 54.10 | 75.10 | 49.70 |
| +MVP | 49.81 | 74.78 | 56.84 | 76.34 | 53.40 |
| SC & BLEU | GYAFC E&M | GYAFC F&R | ||||
| B-4 | Accuracy | HM | B-4 | Accuracy | HM | |
| +BART | 76.50b | 93.70 | 83.90 | 79.30 | 92.00 | 85.20 |
| +MVP | 77.18 | 94.49 | 84.96 | 79.43 | 92.12 | 85.31 |
| Methods | CoLA Matt. | SST-2 Acc. | MRPC F1/Acc. | STS-B P/S Corr. | QQP F1/Acc. | MNLI m./mm. | QNLI Acc. | RTE Acc. | Average |
| BART | 60.30 | 96.30 | 90.47 / 86.70 | 90.97 / 90.30 | 73.03 / 89.87 | 90.03 / 89.27 | 94.60 | 79.83 | 85.17 |
| MVP | 59.87 | 96.43 | 92.07 / 89.43 | 91.37 / 90.90 | 73.20 / 90.13 | 89.70 / 88.73 | 95.10 | 82.87 | 85.88 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP+S | 43.03 | 20.27 | 39.72 | 66.73 | 47.42 | 76.36 | 25.28 | 26.66 | 52.69 | 86.44 | 76.84 |
| BART+R | 42.47 | 19.82 | 39.15 | 65.54 | 46.86 | 75.24 | 24.27 | 26.07 | 52.03 | 82.22 | 71.92 |
| MVP+R | 42.84 | 20.21 | 39.61 | 66.12 | 47.12 | 75.83 | 25.05 | 26.34 | 52.57 | 85.51 | 75.56 |
| MVP+M | 42.99 | 20.36 | 39.70 | 66.40 | 47.16 | 75.89 | 25.24 | 26.49 | 52.88 | 85.90 | 76.34 |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP+S | 32.94 | 15.12 | 2.98 | 71.09 | 47.11 | 39.51 | 1.39 | 7.28 | 19.24 | 71.40 | 77.80 |
| BART+R | 32.14 | 14.71 | 2.85 | 68.94 | 46.23 | 38.98 | 1.30 | 6.82 | 17.94 | 62.20 | 69.20 |
| MVP+R | 32.28 | 14.85 | 2.97 | 70.29 | 46.70 | 39.23 | 1.31 | 6.98 | 18.86 | 64.40 | 71.40 |
| MVP+M | 32.62 | 15.28 | 2.95 | 69.58 | 46.78 | 39.40 | 1.33 | 7.13 | 19.13 | 67.20 | 72.90 |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| Datasets | MVP wins (%) | Ties (%) | BART wins (%) |
| CNN/DM | 46.50 | 10.67 | 42.83 |
| WebNLG | 32.17 | 45.67 | 22.17 |
| ROCStories | 46.50 | 11.33 | 42.17 |
| PersonaChat | 35.33 | 34.00 | 30.67 |
| Methods | #NLG (PT) | #NLU (PT) | #NLG (FT) | #NLU (FT) | SP model | SP prompts | Open source |
| FLAN | 3 | 9 | 2 | 9 | ✓ | ✘ | ✘ |
| T0 | 2 | 6 | 0 | 4 | ✓ | ✘ | ✓ |
| Muppet | 1 | 3 | 1 | 3 | ✓ | ✘ | ✓ |
| ExT5 | 3 | 8 | 6 | 8 | ✓ | ✘ | ✘ |
| SPoT | 1 | 4 | 0 | 6 | ✘ | ✓ | ✘ |
| MVP (ours) | 7 | 0 | 11 | 3 | ✓ | ✓ | ✓ |
| Dataset | #Train | Cleaned #Train | #Valid | #Test | Input | Output | License |
| AGENDA | 38,720 | 38,720 | 1,000 | 1,000 | 52.1 | 141.2 | N/A |
| ENT-DESC | 88,652 | 88,652 | 11,081 | 11,081 | 279.9 | 31.0 | N/A |
| GenWiki | 681,436 | 681,436 | 75,716 | 1,000 | 21.4 | 29.5 | MIT |
| LogicNLG | 28,450 | 28,450 | 4,260 | 4,305 | 178.4 | 14.2 | MIT |
| TEKGEN | 6,310,061 | 6,307,995 | 788,746 | 796,982 | 17.0 | 21.2 | CC BY-SA 2.0 |
| WEATHERGOV | 25,000 | 25,000 | 1,000 | 3,528 | 148.7 | 30.6 | N/A |
| WikiTableT | 1,453,794 | 1,452,778 | 4,533 | 4,351 | 81.0 | 99.7 | MIT |
| Cleaned OS Dialogs | 13,355,487 | 13,355,368 | 1,483,944 | - | 75.5 | 16.7 | N/A |
| CMUDoG | 82,818 | 82,818 | 5,555 | 14,510 | 433.0 | 12.2 | N/A |
| Curiosity | 64,930 | 64,551 | 8,539 | 8,495 | 144.4 | 20.2 | CC BY-NC 4.0 |
| DREAM | 14,264 | 14,242 | 4,709 | 4,766 | 75.6 | 13.6 | N/A |
| Empathetic Dialogues | 64,636 | 64,636 | 9,308 | 8,426 | 52.7 | 12.9 | CC BY-NC 4.0 |
| Movie Dialog | 762,751 | 762,711 | 8,216 | 8,066 | 126.9 | 44.0 | N/A |
| MuTual | 33,691 | 33,691 | 4,090 | 3,248 | 53.6 | 14.5 | N/A |
| OpenDialKG | 69,680 | 69,680 | 7,743 | - | 54.2 | 12.4 | CC BY-NC 4.0 |
| Topical-Chat | 179,750 | 179,750 | 22,295 | 22,452 | 223.3 | 20.0 | CDLA-Sharing-1.0 |
| Wizard of Wikipedia | 148,357 | 147,702 | 15,767 | 15,564 | 297.0 | 16.7 | MIT |
| HotpotQA | 90,447 | 87,815 | 7,405 | - | 187.9 | 2.2 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 68.7 | 13.3 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | - | 48.1 | 3.7 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 584.1 | 4.2 | Apache 2.0 |
| Natural Questions | 96,676 | 96,676 | 10,693 | 6,490 | 9.0 | 2.1 | CC BY-SA 3.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 726.8 | 5.0 | MIT |
| QuAC | 83,568 | 83,485 | 31,906 | - | 487.9 | 12.5 | CC BY-SA 4.0 |
| TriviaQA | 78,785 | 78,785 | 8,837 | 11,313 | 14.0 | 2.0 | Apache 2.0 |
| WebQuestions | 8,933 | 8,933 | 4,863 | 4,863 | 6.7 | 2.4 | CC BY 4.0 |
| HotpotQA | 90,440 | 87,808 | 6,972 | - | 79.6 | 19.8 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 75.9 | 6.0 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | 11,022 | 45.9 | 6.0 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 579.7 | 8.6 | Apache 2.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 724.2 | 7.6 | MIT |
| QuAC | 69,109 | 69,026 | 26,301 | - | 496.7 | 6.5 | CC BY-SA 4.0 |
| ChangeMyView | 42,462 | 42,459 | 6,480 | 7,562 | 17.9 | 104.1 | MIT |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 8.8 | 33.3 | MIT |
| Hippocorpus | 6,168 | 6,168 | 686 | - | 34.1 | 262.6 | CDLA-Permissive 2.0 |
| WikiPlots | 101,642 | 101,641 | 11,294 | - | 3.4 | 338.5 | N/A |
| WritingPrompts | 272,600 | 272,518 | 15,620 | 15,138 | 28.4 | 630.8 | MIT |
| CamRest676 | 4,872 | 4,872 | 616 | - | 55.3 | 9.4 | N/A |
| Frames | 26,631 | 26,631 | 2,106 | - | 116.1 | 13.0 | MIT |
| KVRET | 14,136 | 14,136 | 1,616 | - | 30.5 | 9.3 | N/A |
| MetaLWOZ | 176,073 | 176,073 | 17,912 | - | 45.6 | 8.0 | N/A |
| MSR-E2E | 103,362 | 103,362 | 5,235 | - | 51.3 | 12.8 | Microsoft |
| Schema-Guided | 494,946 | 494,933 | 73,089 | - | 120.8 | 12.5 | CC BY-SA 4.0 |
| TaskMaster | 249,664 | 249,662 | 20,680 | - | 95.6 | 12.0 | CC BY 4.0 |
| WOZ | 6,364 | 6,359 | 1,260 | - | 47.0 | 10.6 | N/A |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 33.3 | 8.8 | MIT |
| MediaSum | 443,596 | 442,021 | 10,000 | 10,000 | 1641.0 | 14.4 | N/A |
| MSNews | 136,082 | 135,937 | 7,496 | 7,562 | 309.9 | 9.8 | CC BY-SA 4.0 |
| Newsroom | 995,041 | 989,351 | 108,837 | 108,862 | 642.4 | 26.7 | N/A |
| WikiHow | 157,252 | 157,247 | 5,599 | 5,577 | 502.6 | 45.6 | CC BY-NC-SA |
| Task | Dataset | #Train | #Valid | #Test | Input | Output | License |
| Commonsen generation | CommonGen | 67,389 | 993 | - | 5.5 | 11.6 | MIT |
| Data-to-text generation | DART | 62,659 | 2,768 | - | 27.5 | 21.5 | MIT |
| E2E | 33,525 | 4,299 | - | 9.5 | 20.6 | CC BY-SA 4.0 | |
| ToTTo | 120,761 | 7,700 | - | 37.8 | 18.0 | CC BY-SA 3.0 | |
| WebNLG | 34,338 | 4,313 | 4,222 | 18.0 | 19.9 | CC BY-NA-SA 4.0 | |
| WebNLG (GEM) | 35,426 | 1,667 | - | 17.7 | 22.7 | CC BY-NA-SA 4.0 | |
| WikiBio | 582,659 | 72,831 | 72,831 | 81.6 | 26.1 | CC BY-SA 3.0 | |
| Open-ended dialogue | DailyDialog | 76,052 | 7,069 | 6,740 | 72.5 | 13.9 | CC BY-NC-SA 4.0 |
| DSTC7-AVSD | 76,590 | 17,870 | 1,710 | 148.2 | 11.5 | MIT | |
| PersonaChat | 122,499 | 14,602 | 14,056 | 132.1 | 11.9 | MIT | |
| SGD | 164,982 | 10,000 | - | 134.7 | 11.3 | CC BY-SA 4.0 | |
| Natural language inference | MNLI-m | 392,702 | 9,815 | 9,796 | 29.8 | - | Mixed |
| MNLI-mm | 9,832 | 9,847 | |||||
| QNLI | 104,743 | 5,463 | 5,463 | 36.6 | - | CC BY-SA 4.0 | |
| RTE | 2,490 | 277 | 3,000 | 51.0 | - | N/A | |
| Paraphrase generation | Quora | 137,185 | 3,000 | 3,000 | 10.9 | 10.8 | N/A |
| Paraphrase detection | MRPC | 3,668 | 408 | 1,725 | 43.8 | - | N/A |
| QQP | 363,846 | 40,430 | 390,965 | 22.3 | - | N/A | |
| STS-B | 5,749 | 1,500 | 1,379 | 20.3 | - | N/A | |
| Question answering | CoQA | 107,286 | 31,621 | - | 349.4 | 2.6 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 156.2 | 3.6 | CC BY-SA 4.0 | |
| Question generation | CoQA | 107,286 | 31,621 | - | 346.6 | 5.5 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 148.3 | 11.6 | CC BY-SA 4.0 | |
| Story generation | ROCStories | 176,688 | 9,816 | 4,909 | 9.0 | 40.7 | N/A |
| WritingPrompts | 53,516 | 4,000 | 2,000 | 25.5 | 150.4 | MIT | |
| Task-oriented dialogue | MultiWOZ | 170,220 | 22,074 | 22,116 | 128.3 | 11.3 | MIT |
| Text classification | CoLA | 8,551 | 1,043 | 1,063 | 7.7 | - | N/A |
| SST-2 | 67,349 | 872 | 1,821 | 9.8 | - | N/A | |
| Text simplification | WiA-A | 483,801 | 20,000 | 359 | 26.2 | 21.5 | Mixed |
| WiA-T | 359 | ||||||
| Text style transfer | GYAFC-E&M | 52,595 | 11,508 | 1,416 | 9.9 | 10.6 | N/A |
| GYAFC-F&R | 51,967 | 11,152 | 1,332 | 10.7 | 11.3 | ||
| Text summarization | CNN/DailyMail | 287,227 | 13,368 | 11,490 | 679.8 | 48.3 | MIT |
| SAMSum | 14,732 | 818 | 819 | 103.4 | 20.3 | CC BY-NC-ND 4.0 | |
| WLE | 99,020 | 28,614 | - | 367.6 | 33.4 | CC0 1.0 | |
| XSum | 204,045 | 11,332 | 11,334 | 373.7 | 21.1 | MIT |
| Methods | XSum | SAMSum | CoQA QG | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | B-4 | ME | R-L | |
| BART | \( 45.14^d \) | 22.27 | 37.25 | \( 51.74^b \) | 26.46 | 48.72 | \( 12.34^c \) | 35.78 | 46.88 |
| MVP | 45.60 | 22.47 | 37.42 | 53.78 | 29.12 | 49.37 | 23.48 | 47.79 | 55.09 |
| MVP+S | 45.67 | 22.63 | 37.50 | 53.81 | 29.75 | 49.43 | 23.43 | 47.49 | 55.25 |
| SOTA | \( 49.57^a \) | 25.08 | 41.81 | \( 53.89^b \) | 28.85 | 49.29 | \( 15.78^c \) | 40.15 | 50.98 |
| Methods | WritingPrompts | DailyDialog | WikiBio | ||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | |
| BART | \( 22.40^e \) | 8.40 | - | 31.30 | \( 44.30^f \) | 39.20 | 3.90 | 21.10 | - |
| MVP | 32.34 | 13.11 | 2.12 | 64.58 | 46.19 | 41.81 | 4.61 | 25.06 | 48.42 |
| MVP+S | 30.12 | 11.46 | 3.97 | 83.70 | 45.71 | 42.92 | 5.10 | 27.14 | 48.19 |
| SOTA | \( 22.40^e \) | 8.40 | - | 31.30 | \( 46.10^f \) | 40.70 | 4.10 | 22.20 | \( 45.10^g \) |
| Methods | DSTC7-AVSD | SQuAD | |||||||
| B-1 | B-2 | B-3 | B-4 | ME | R-L | CIDEr | F1 | EM | |
| BART | \( 82.40^f \) | 69.10 | 58.20 | 48.70 | 31.30 | 63.50 | 1.38 | \( 91.56^i \) | 84.23 |
| MVP | 83.75 | 70.89 | 60.19 | 50.94 | 32.12 | 65.04 | 1.45 | \( 93.45 \) | 87.20 |
| MVP+S | 83.81 | 71.07 | 60.45 | 51.20 | 31.77 | 64.76 | 1.44 | \( 93.45 \) | 87.17 |
| SOTA | \( 83.20^f \) | 70.50 | 59.80 | 50.60 | 31.40 | 63.80 | 1.39 | \( 96.22^h \) | 91.26 |
| Methods | DART | E2E | ToTTo | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 34.31 | 45.22 | 36.30 | 42.57 | 46.60 | 38.20 | 39.79 | 49.90 | 36.80 |
| ExT5 | 36.62 | 48.14 | 37.60 | 42.25 | 46.70 | 38.10 | 40.14 | 50.33 | 36.90 |
| MVP | 39.13 | 48.92 | 38.53 | 37.38 | 47.96 | 39.39 | 50.58 | 55.24 | 41.27 |
| MVP+S | 38.83 | 48.49 | 38.41 | 37.32 | 47.40 | 38.90 | 50.69 | 55.52 | 41.29 |
| Methods | WebNLG | CommonGen | SGD | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 31.67 | 43.31 | 34.40 | 8.38 | 17.01 | 20.20 | 33.15 | 36.17 | 32.40 |
| ExT5 | 35.03 | 48.17 | 36.50 | 9.68 | 19.04 | 21.40 | 34.74 | 37.77 | 33.00 |
| MVP | 47.03 | 59.00 | 42.34 | 32.59 | 37.71 | 33.00 | 45.63 | 48.29 | 38.48 |
| MVP+S | 47.03 | 59.03 | 42.28 | 34.10 | 37.87 | 33.11 | 45.24 | 48.25 | 38.47 |
| Methods | WiA-A | WiA-T | WLE | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 29.30 | 38.37 | 30.10 | 42.12 | 50.52 | 36.2 | 15.55 | 20.47 | 19.60 |
| ExT5 | 29.23 | 37.98 | 30.00 | 41.39 | 50.38 | 35.8 | 16.64 | 21.16 | 20.40 |
| MVP | 71.55 | 70.88 | 48.19 | 91.73 | 83.46 | 57.34 | 18.80 | 22.84 | 21.95 |
| MVP+S | 70.37 | 70.65 | 47.70 | 91.12 | 83.59 | 56.95 | 18.52 | 22.57 | 22.02 |
| Input | ||
| she was on a flight. | ||
| Output | ||
| she was trying to take a nap . suddenly , her ears started ringing . the flight attendant tried to fix it but she could n’t . she had to call for help . luckily, they were able to fix the problem . | she was bored and her ears hurt . she decided to take a nap . luckily , she was able to get a good night ’s sleep . but the next morning , she woke up and felt sick . | |
| Left Wins | Ties | Right Wins |
| Input +Describe the following data: Abilene,_Texas | cityServed | Abilene_Regional_Airport |
| Gold +Abilene, Texas is served by the Abilene regional airport. +Abilene Regional Airport serves the city of Abilene in Texas. |
| BART +Abilene Regional Airport serves the city of Abilene in Texas. |
| MVP +Abilene Regional Airport serves the city of Abilene, Texas. |
| MVP+S +Abilene Regional Airport serves the city of Abilene, Texas. |
| Input +Describe the following data: "Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas" | location | Adolfo_Suárez_Madrid-Barajas_Airport |
| Gold +Adolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. +Adolfo Suarez Madrid-Barajas airport is located at Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. +Adolfo Suarez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastian de los Reyes and Alcobendas. |
| BART +Adolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP +Adolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP+S +Adolfo Suárez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| Input +Generate the question based on the answer: Saint Bernadette Soubirous [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend "Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold +To whom did the Virgin Mary allegedly appear in 1858 in Lourdes France? |
| BART +Who is believed to have appeared to the Virgin Mary at Lourdes? |
| MVP +Who did the Virgin Mary appear to in Lourdes? |
| MVP+S +The Grotto is a replica of the grotto at Lourdes, France where the Virgin Mary appeared to whom? |
| Input +Generate the question based on the answer: a copper statue of Christ [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend " Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold +What is in front of the Notre Dame Main Building? |
| BART +What is in front of the Main Building and facing it? |
| MVP +What is immediately in front of the Main Building? |
| MVP+S +What is immediately in front of the Main Building? |
| Input +Given the story title: male was out jogging one morning. |
| Gold +the weather was crisp and cool . male felt good and energetic . he decided to keep jogging longer than normal . male went several more miles out of his way . |
| BART +he saw a man struggling to walk across a field . male stopped to help him . he saw a guy struggling to walking across a fields . the man thanked male profusely ! |
| MVP +he ran into a man who was holding a knife . male was scared , so he ran away . the man threatened male . male ran back into his house . |
| MVP+S +he noticed a sound coming from a tree . male went to investigate . he found a small baby bird stuck in the trunk . male took the bird to the vet and they released it . |
| Input +Given the story title: ken was driving around in the snow. |
| Gold +he needed to get home from work . he was driving slowly to avoid accidents . unfortunately the roads were too slick and ken lost control . his tires lost traction and he hit a tree . |
| BART +he hit a patch of black ice . ken slid and hurt his knee . he went to the doctor for it . krampus was sent home with a concussion . |
| MVP +he was in his car and it was very cold . ken started to back up . the car started to make a loud noise . kena had to back out and call for help . |
| MVP+S +he was having trouble starting his car . he had to push it a couple of blocks . his car would n’t start . ken had to call a tow truck for help . |
| Input +Given the dialog: i love to meet new people. [SEP] i have a turtle named timothy. [SEP] my favorite sport is ultimate frisbee. [SEP] my parents are living in bora bora. [SEP] autumn is my favorite season. [X_SEP] hello, how are you doing tonight? |
| Gold +i am well an loving this interaction how are you? |
| BART +i am doing well, how are you? |
| MVP +i am doing well. how are you? |
| MVP+S +i am doing well. how are you? |
| Input +Given the dialog: i just bought a brand new house. [SEP] i like to dance at the club. [SEP] i run a dog obedience school. [SEP] i have a big sweet tooth. [SEP] i like taking and posting selkies. [X_SEP] hello, how are you doing tonight? [SEP] i am well an loving this interaction how are you? |
| Gold +i am great. i just got back from the club. |
| BART +i am doing well, thanks for asking. what do you do for fun? |
| MVP +i am doing well. i just got home from dancing at the club. |
| MVP+S +i am doing well. i just bought a new house. |
| Settings | Supervised Pre-training | Unsupervised Pre-training |
| NLG | MVP (ours) | GPT-2, MASS, BART, T5 |
| NLU | FLAN, T0, Muppet, ExT5 | BERT, XLNet, RoBERTa, T5 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| BART | 44.16e | 21.28 | 40.90 | 64.55b | 46.51 | 75.13 | 22.00f | 26.40 | 52.55 | 68.60f | - |
| Flan-T5 | 43.45 | 21.01 | 40.03 | 66.60 | 46.93 | 75.76 | 25.55 | 26.90 | 53.51 | 84.18 | 75.44 |
| Single | 44.36 | 21.54 | 40.88 | 67.74 | 46.89 | 76.94 | 26.09 | 27.15 | 53.29 | 86.20 | 77.26 |
| MVP+S | 44.63 | 21.72 | 41.21 | 68.19 | 47.75 | 76.81 | 25.69 | 27.04 | 53.20 | 86.65 | 77.93 |
| MVP+R | 44.14 | 21.45 | 40.72 | 67.61 | 47.65 | 76.70 | 25.71 | 27.03 | 53.09 | 85.95 | 77.22 |
| MVP+M | 43.97 | 21.16 | 40.46 | 67.45 | 47.57 | 76.81 | 25.46 | 26.79 | 52.95 | 86.28 | 77.26 |
| SOTA | 47.16a | 22.55 | 43.87 | 66.14b | 47.25 | 76.10 | 25.97c | 27.33 | 53.43 | 84.50d | - |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| BART | 30.70g | 13.30 | - | 69.90 | 49.90f | 40.00 | 1.30 | 8.00 | 17.89j | 74.91 | 84.88 |
| Flan-T5 | 32.72 | 15.23 | 2.97 | 68.97 | 48.55 | 40.22 | 1.40 | 7.85 | 19.73 | 70.20 | 78.70 |
| Single | 32.67 | 15.29 | 2.72 | 72.97 | 49.96 | 40.53 | 1.27 | 7.63 | 19.73 | 75.60 | 83.70 |
| MVP+S | 33.92 | 15.60 | 3.44 | 80.58 | 47.91 | 39.97 | 1.52 | 9.54 | 20.32 | 79.90 | 86.80 |
| MVP+R | 32.93 | 15.32 | 2.88 | 73.83 | 48.45 | 40.09 | 1.30 | 7.95 | 19.02 | 73.30 | 81.80 |
| MVP+M | 33.30 | 15.51 | 2.71 | 74.24 | 46.26 | 39.30 | 1.36 | 8.07 | 19.93 | 72.70 | 79.70 |
| SOTA | 33.40g | 15.40 | - | 69.30 | 49.90f | 40.00 | 1.50h | 9.40 | 20.50i | 85.30 | 94.40 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| T0-3B | - | - | - | 1.40 | 10.20 | 18.43 | 3.06 | 12.43 | 14.91 | 13.30 | 6.60 |
| T0-11B | - | - | - | 0.26 | 6.13 | 14.12 | 2.63 | 7.00 | 15.25 | 9.18 | 4.36 |
| MVP | 29.50 | 11.29 | 25.92 | 34.42 | 31.33 | 52.33 | 2.90 | 13.94 | 15.48 | 29.40 | 18.20 |
| MVP+S | 25.60 | 9.51 | 22.67 | 39.43 | 34.32 | 55.34 | 2.96 | 15.23 | 18.23 | 52.40 | 37.30 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| T0-3B | 8.69 | 3.02 | 4.37 | 35.49 | 23.20 | 23.57 | 2.56 | 12.06 | 0.02 | 2.50 | 22.10 |
| T0-11B | 0.63 | 0.16 | 12.41 | 92.86 | 32.17 | 28.35 | 1.56 | 7.19 | 0.00 | 3.90 | 22.10 |
| MVP | 1.01 | 0.31 | 7.18 | 86.26 | 35.54 | 32.71 | 2.87 | 16.38 | 3.08 | 2.50 | 22.20 |
| MVP+S | 10.52 | 3.54 | 2.13 | 69.55 | 37.04 | 33.38 | 2.66 | 14.84 | 0.38 | 2.50 | 22.10 |
| AESOP | Quora | ||||
| B-4 | R-1 | R-2 | R-L | ME | |
| +BART | 47.30a | 73.30 | 54.10 | 75.10 | 49.70 |
| +MVP | 49.81 | 74.78 | 56.84 | 76.34 | 53.40 |
| SC & BLEU | GYAFC E&M | GYAFC F&R | ||||
| B-4 | Accuracy | HM | B-4 | Accuracy | HM | |
| +BART | 76.50b | 93.70 | 83.90 | 79.30 | 92.00 | 85.20 |
| +MVP | 77.18 | 94.49 | 84.96 | 79.43 | 92.12 | 85.31 |
| Methods | CoLA Matt. | SST-2 Acc. | MRPC F1/Acc. | STS-B P/S Corr. | QQP F1/Acc. | MNLI m./mm. | QNLI Acc. | RTE Acc. | Average |
| BART | 60.30 | 96.30 | 90.47 / 86.70 | 90.97 / 90.30 | 73.03 / 89.87 | 90.03 / 89.27 | 94.60 | 79.83 | 85.17 |
| MVP | 59.87 | 96.43 | 92.07 / 89.43 | 91.37 / 90.90 | 73.20 / 90.13 | 89.70 / 88.73 | 95.10 | 82.87 | 85.88 |
| Methods | CNN/DailyMail | WebNLG | SQuAD (QG) | CoQA | |||||||
| R-1 | R-2 | R-L | B-4 | ME | R-L | B-4 | ME | R-L | F1 | EM | |
| MVP+S | 43.03 | 20.27 | 39.72 | 66.73 | 47.42 | 76.36 | 25.28 | 26.66 | 52.69 | 86.44 | 76.84 |
| BART+R | 42.47 | 19.82 | 39.15 | 65.54 | 46.86 | 75.24 | 24.27 | 26.07 | 52.03 | 82.22 | 71.92 |
| MVP+R | 42.84 | 20.21 | 39.61 | 66.12 | 47.12 | 75.83 | 25.05 | 26.34 | 52.57 | 85.51 | 75.56 |
| MVP+M | 42.99 | 20.36 | 39.70 | 66.40 | 47.16 | 75.89 | 25.24 | 26.49 | 52.88 | 85.90 | 76.34 |
| FT BART | 44.16 | 21.28 | 40.90 | 64.55 | 46.51 | 75.13 | 22.00 | 26.40 | 52.55 | 68.60 | - |
| FT MVP | 44.52 | 21.62 | 41.10 | 67.82 | 47.47 | 76.88 | 26.26 | 27.35 | 53.49 | 86.43 | 77.78 |
| Methods | ROCStories | PersonaChat | MultiWOZ | ||||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | Success | Inform | |
| MVP+S | 32.94 | 15.12 | 2.98 | 71.09 | 47.11 | 39.51 | 1.39 | 7.28 | 19.24 | 71.40 | 77.80 |
| BART+R | 32.14 | 14.71 | 2.85 | 68.94 | 46.23 | 38.98 | 1.30 | 6.82 | 17.94 | 62.20 | 69.20 |
| MVP+R | 32.28 | 14.85 | 2.97 | 70.29 | 46.70 | 39.23 | 1.31 | 6.98 | 18.86 | 64.40 | 71.40 |
| MVP+M | 32.62 | 15.28 | 2.95 | 69.58 | 46.78 | 39.40 | 1.33 | 7.13 | 19.13 | 67.20 | 72.90 |
| FT BART | 30.70 | 13.30 | - | 69.90 | 49.90 | 40.00 | 1.30 | 8.00 | 17.89 | 74.91 | 84.88 |
| FT MVP | 33.79 | 15.76 | 3.02 | 75.65 | 50.73 | 40.69 | 1.65 | 11.23 | 20.26 | 76.40 | 85.00 |
| Datasets | MVP wins (%) | Ties (%) | BART wins (%) |
| CNN/DM | 46.50 | 10.67 | 42.83 |
| WebNLG | 32.17 | 45.67 | 22.17 |
| ROCStories | 46.50 | 11.33 | 42.17 |
| PersonaChat | 35.33 | 34.00 | 30.67 |
| Methods | #NLG (PT) | #NLU (PT) | #NLG (FT) | #NLU (FT) | SP model | SP prompts | Open source |
| FLAN | 3 | 9 | 2 | 9 | ✓ | ✘ | ✘ |
| T0 | 2 | 6 | 0 | 4 | ✓ | ✘ | ✓ |
| Muppet | 1 | 3 | 1 | 3 | ✓ | ✘ | ✓ |
| ExT5 | 3 | 8 | 6 | 8 | ✓ | ✘ | ✘ |
| SPoT | 1 | 4 | 0 | 6 | ✘ | ✓ | ✘ |
| MVP (ours) | 7 | 0 | 11 | 3 | ✓ | ✓ | ✓ |
| Dataset | #Train | Cleaned #Train | #Valid | #Test | Input | Output | License |
| AGENDA | 38,720 | 38,720 | 1,000 | 1,000 | 52.1 | 141.2 | N/A |
| ENT-DESC | 88,652 | 88,652 | 11,081 | 11,081 | 279.9 | 31.0 | N/A |
| GenWiki | 681,436 | 681,436 | 75,716 | 1,000 | 21.4 | 29.5 | MIT |
| LogicNLG | 28,450 | 28,450 | 4,260 | 4,305 | 178.4 | 14.2 | MIT |
| TEKGEN | 6,310,061 | 6,307,995 | 788,746 | 796,982 | 17.0 | 21.2 | CC BY-SA 2.0 |
| WEATHERGOV | 25,000 | 25,000 | 1,000 | 3,528 | 148.7 | 30.6 | N/A |
| WikiTableT | 1,453,794 | 1,452,778 | 4,533 | 4,351 | 81.0 | 99.7 | MIT |
| Cleaned OS Dialogs | 13,355,487 | 13,355,368 | 1,483,944 | - | 75.5 | 16.7 | N/A |
| CMUDoG | 82,818 | 82,818 | 5,555 | 14,510 | 433.0 | 12.2 | N/A |
| Curiosity | 64,930 | 64,551 | 8,539 | 8,495 | 144.4 | 20.2 | CC BY-NC 4.0 |
| DREAM | 14,264 | 14,242 | 4,709 | 4,766 | 75.6 | 13.6 | N/A |
| Empathetic Dialogues | 64,636 | 64,636 | 9,308 | 8,426 | 52.7 | 12.9 | CC BY-NC 4.0 |
| Movie Dialog | 762,751 | 762,711 | 8,216 | 8,066 | 126.9 | 44.0 | N/A |
| MuTual | 33,691 | 33,691 | 4,090 | 3,248 | 53.6 | 14.5 | N/A |
| OpenDialKG | 69,680 | 69,680 | 7,743 | - | 54.2 | 12.4 | CC BY-NC 4.0 |
| Topical-Chat | 179,750 | 179,750 | 22,295 | 22,452 | 223.3 | 20.0 | CDLA-Sharing-1.0 |
| Wizard of Wikipedia | 148,357 | 147,702 | 15,767 | 15,564 | 297.0 | 16.7 | MIT |
| HotpotQA | 90,447 | 87,815 | 7,405 | - | 187.9 | 2.2 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 68.7 | 13.3 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | - | 48.1 | 3.7 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 584.1 | 4.2 | Apache 2.0 |
| Natural Questions | 96,676 | 96,676 | 10,693 | 6,490 | 9.0 | 2.1 | CC BY-SA 3.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 726.8 | 5.0 | MIT |
| QuAC | 83,568 | 83,485 | 31,906 | - | 487.9 | 12.5 | CC BY-SA 4.0 |
| TriviaQA | 78,785 | 78,785 | 8,837 | 11,313 | 14.0 | 2.0 | Apache 2.0 |
| WebQuestions | 8,933 | 8,933 | 4,863 | 4,863 | 6.7 | 2.4 | CC BY 4.0 |
| HotpotQA | 90,440 | 87,808 | 6,972 | - | 79.6 | 19.8 | CC BY-SA 4.0 |
| MS MARCO | 681,445 | 681,226 | 77,580 | - | 75.9 | 6.0 | N/A |
| MSQG | 198,058 | 198,029 | 11,008 | 11,022 | 45.9 | 6.0 | CC BY-SA 4.0 |
| NarrativeQA | 65,494 | 65,494 | 6,922 | 21,114 | 579.7 | 8.6 | Apache 2.0 |
| NewsQA | 97,850 | 97,700 | 5,486 | 5,396 | 724.2 | 7.6 | MIT |
| QuAC | 69,109 | 69,026 | 26,301 | - | 496.7 | 6.5 | CC BY-SA 4.0 |
| ChangeMyView | 42,462 | 42,459 | 6,480 | 7,562 | 17.9 | 104.1 | MIT |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 8.8 | 33.3 | MIT |
| Hippocorpus | 6,168 | 6,168 | 686 | - | 34.1 | 262.6 | CDLA-Permissive 2.0 |
| WikiPlots | 101,642 | 101,641 | 11,294 | - | 3.4 | 338.5 | N/A |
| WritingPrompts | 272,600 | 272,518 | 15,620 | 15,138 | 28.4 | 630.8 | MIT |
| CamRest676 | 4,872 | 4,872 | 616 | - | 55.3 | 9.4 | N/A |
| Frames | 26,631 | 26,631 | 2,106 | - | 116.1 | 13.0 | MIT |
| KVRET | 14,136 | 14,136 | 1,616 | - | 30.5 | 9.3 | N/A |
| MetaLWOZ | 176,073 | 176,073 | 17,912 | - | 45.6 | 8.0 | N/A |
| MSR-E2E | 103,362 | 103,362 | 5,235 | - | 51.3 | 12.8 | Microsoft |
| Schema-Guided | 494,946 | 494,933 | 73,089 | - | 120.8 | 12.5 | CC BY-SA 4.0 |
| TaskMaster | 249,664 | 249,662 | 20,680 | - | 95.6 | 12.0 | CC BY 4.0 |
| WOZ | 6,364 | 6,359 | 1,260 | - | 47.0 | 10.6 | N/A |
| English Gigaword | 3,803,957 | 3,802,620 | 189,651 | 1,951 | 33.3 | 8.8 | MIT |
| MediaSum | 443,596 | 442,021 | 10,000 | 10,000 | 1641.0 | 14.4 | N/A |
| MSNews | 136,082 | 135,937 | 7,496 | 7,562 | 309.9 | 9.8 | CC BY-SA 4.0 |
| Newsroom | 995,041 | 989,351 | 108,837 | 108,862 | 642.4 | 26.7 | N/A |
| WikiHow | 157,252 | 157,247 | 5,599 | 5,577 | 502.6 | 45.6 | CC BY-NC-SA |
| Task | Dataset | #Train | #Valid | #Test | Input | Output | License |
| Commonsen generation | CommonGen | 67,389 | 993 | - | 5.5 | 11.6 | MIT |
| Data-to-text generation | DART | 62,659 | 2,768 | - | 27.5 | 21.5 | MIT |
| E2E | 33,525 | 4,299 | - | 9.5 | 20.6 | CC BY-SA 4.0 | |
| ToTTo | 120,761 | 7,700 | - | 37.8 | 18.0 | CC BY-SA 3.0 | |
| WebNLG | 34,338 | 4,313 | 4,222 | 18.0 | 19.9 | CC BY-NA-SA 4.0 | |
| WebNLG (GEM) | 35,426 | 1,667 | - | 17.7 | 22.7 | CC BY-NA-SA 4.0 | |
| WikiBio | 582,659 | 72,831 | 72,831 | 81.6 | 26.1 | CC BY-SA 3.0 | |
| Open-ended dialogue | DailyDialog | 76,052 | 7,069 | 6,740 | 72.5 | 13.9 | CC BY-NC-SA 4.0 |
| DSTC7-AVSD | 76,590 | 17,870 | 1,710 | 148.2 | 11.5 | MIT | |
| PersonaChat | 122,499 | 14,602 | 14,056 | 132.1 | 11.9 | MIT | |
| SGD | 164,982 | 10,000 | - | 134.7 | 11.3 | CC BY-SA 4.0 | |
| Natural language inference | MNLI-m | 392,702 | 9,815 | 9,796 | 29.8 | - | Mixed |
| MNLI-mm | 9,832 | 9,847 | |||||
| QNLI | 104,743 | 5,463 | 5,463 | 36.6 | - | CC BY-SA 4.0 | |
| RTE | 2,490 | 277 | 3,000 | 51.0 | - | N/A | |
| Paraphrase generation | Quora | 137,185 | 3,000 | 3,000 | 10.9 | 10.8 | N/A |
| Paraphrase detection | MRPC | 3,668 | 408 | 1,725 | 43.8 | - | N/A |
| QQP | 363,846 | 40,430 | 390,965 | 22.3 | - | N/A | |
| STS-B | 5,749 | 1,500 | 1,379 | 20.3 | - | N/A | |
| Question answering | CoQA | 107,286 | 31,621 | - | 349.4 | 2.6 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 156.2 | 3.6 | CC BY-SA 4.0 | |
| Question generation | CoQA | 107,286 | 31,621 | - | 346.6 | 5.5 | Mixed |
| SQuAD | 75,722 | 10,570 | 11,877 | 148.3 | 11.6 | CC BY-SA 4.0 | |
| Story generation | ROCStories | 176,688 | 9,816 | 4,909 | 9.0 | 40.7 | N/A |
| WritingPrompts | 53,516 | 4,000 | 2,000 | 25.5 | 150.4 | MIT | |
| Task-oriented dialogue | MultiWOZ | 170,220 | 22,074 | 22,116 | 128.3 | 11.3 | MIT |
| Text classification | CoLA | 8,551 | 1,043 | 1,063 | 7.7 | - | N/A |
| SST-2 | 67,349 | 872 | 1,821 | 9.8 | - | N/A | |
| Text simplification | WiA-A | 483,801 | 20,000 | 359 | 26.2 | 21.5 | Mixed |
| WiA-T | 359 | ||||||
| Text style transfer | GYAFC-E&M | 52,595 | 11,508 | 1,416 | 9.9 | 10.6 | N/A |
| GYAFC-F&R | 51,967 | 11,152 | 1,332 | 10.7 | 11.3 | ||
| Text summarization | CNN/DailyMail | 287,227 | 13,368 | 11,490 | 679.8 | 48.3 | MIT |
| SAMSum | 14,732 | 818 | 819 | 103.4 | 20.3 | CC BY-NC-ND 4.0 | |
| WLE | 99,020 | 28,614 | - | 367.6 | 33.4 | CC0 1.0 | |
| XSum | 204,045 | 11,332 | 11,334 | 373.7 | 21.1 | MIT |
| Methods | XSum | SAMSum | CoQA QG | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | B-4 | ME | R-L | |
| BART | \\( 45.14^d \\) | 22.27 | 37.25 | \\( 51.74^b \\) | 26.46 | 48.72 | \\( 12.34^c \\) | 35.78 | 46.88 |
| MVP | 45.60 | 22.47 | 37.42 | 53.78 | 29.12 | 49.37 | 23.48 | 47.79 | 55.09 |
| MVP+S | 45.67 | 22.63 | 37.50 | 53.81 | 29.75 | 49.43 | 23.43 | 47.49 | 55.25 |
| SOTA | \\( 49.57^a \\) | 25.08 | 41.81 | \\( 53.89^b \\) | 28.85 | 49.29 | \\( 15.78^c \\) | 40.15 | 50.98 |
| Methods | WritingPrompts | DailyDialog | WikiBio | ||||||
| B-1 | B-2 | D-1 | D-4 | B-1 | B-2 | D-1 | D-2 | B-4 | |
| BART | \\( 22.40^e \\) | 8.40 | - | 31.30 | \\( 44.30^f \\) | 39.20 | 3.90 | 21.10 | - |
| MVP | 32.34 | 13.11 | 2.12 | 64.58 | 46.19 | 41.81 | 4.61 | 25.06 | 48.42 |
| MVP+S | 30.12 | 11.46 | 3.97 | 83.70 | 45.71 | 42.92 | 5.10 | 27.14 | 48.19 |
| SOTA | \\( 22.40^e \\) | 8.40 | - | 31.30 | \\( 46.10^f \\) | 40.70 | 4.10 | 22.20 | \\( 45.10^g \\) |
| Methods | DSTC7-AVSD | SQuAD | |||||||
| B-1 | B-2 | B-3 | B-4 | ME | R-L | CIDEr | F1 | EM | |
| BART | \\( 82.40^f \\) | 69.10 | 58.20 | 48.70 | 31.30 | 63.50 | 1.38 | \\( 91.56^i \\) | 84.23 |
| MVP | 83.75 | 70.89 | 60.19 | 50.94 | 32.12 | 65.04 | 1.45 | \\( 93.45 \\) | 87.20 |
| MVP+S | 83.81 | 71.07 | 60.45 | 51.20 | 31.77 | 64.76 | 1.44 | \\( 93.45 \\) | 87.17 |
| SOTA | \\( 83.20^f \\) | 70.50 | 59.80 | 50.60 | 31.40 | 63.80 | 1.39 | \\( 96.22^h \\) | 91.26 |
| Methods | DART | E2E | ToTTo | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 34.31 | 45.22 | 36.30 | 42.57 | 46.60 | 38.20 | 39.79 | 49.90 | 36.80 |
| ExT5 | 36.62 | 48.14 | 37.60 | 42.25 | 46.70 | 38.10 | 40.14 | 50.33 | 36.90 |
| MVP | 39.13 | 48.92 | 38.53 | 37.38 | 47.96 | 39.39 | 50.58 | 55.24 | 41.27 |
| MVP+S | 38.83 | 48.49 | 38.41 | 37.32 | 47.40 | 38.90 | 50.69 | 55.52 | 41.29 |
| Methods | WebNLG | CommonGen | SGD | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 31.67 | 43.31 | 34.40 | 8.38 | 17.01 | 20.20 | 33.15 | 36.17 | 32.40 |
| ExT5 | 35.03 | 48.17 | 36.50 | 9.68 | 19.04 | 21.40 | 34.74 | 37.77 | 33.00 |
| MVP | 47.03 | 59.00 | 42.34 | 32.59 | 37.71 | 33.00 | 45.63 | 48.29 | 38.48 |
| MVP+S | 47.03 | 59.03 | 42.28 | 34.10 | 37.87 | 33.11 | 45.24 | 48.25 | 38.47 |
| Methods | WiA-A | WiA-T | WLE | ||||||
| B-4 | R-2 | ME | B-4 | R-2 | ME | B-4 | R-2 | ME | |
| T5.1.1 | 29.30 | 38.37 | 30.10 | 42.12 | 50.52 | 36.2 | 15.55 | 20.47 | 19.60 |
| ExT5 | 29.23 | 37.98 | 30.00 | 41.39 | 50.38 | 35.8 | 16.64 | 21.16 | 20.40 |
| MVP | 71.55 | 70.88 | 48.19 | 91.73 | 83.46 | 57.34 | 18.80 | 22.84 | 21.95 |
| MVP+S | 70.37 | 70.65 | 47.70 | 91.12 | 83.59 | 56.95 | 18.52 | 22.57 | 22.02 |
| Input | ||
| she was on a flight. | ||
| Output | ||
| she was trying to take a nap . suddenly , her ears started ringing . the flight attendant tried to fix it but she could n’t . she had to call for help . luckily, they were able to fix the problem . | she was bored and her ears hurt . she decided to take a nap . luckily , she was able to get a good night ’s sleep . but the next morning , she woke up and felt sick . | |
| Left Wins | Ties | Right Wins |
| Input\nDescribe the following data: Abilene,_Texas | cityServed | Abilene_Regional_Airport |
| Gold\nAbilene, Texas is served by the Abilene regional airport.\nAbilene Regional Airport serves the city of Abilene in Texas. |
| BART\nAbilene Regional Airport serves the city of Abilene in Texas. |
| MVP\nAbilene Regional Airport serves the city of Abilene, Texas. |
| MVP+S\nAbilene Regional Airport serves the city of Abilene, Texas. |
| Input\nDescribe the following data: "Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas" | location | Adolfo_Suárez_Madrid-Barajas_Airport |
| Gold\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas.\nAdolfo Suarez Madrid-Barajas airport is located at Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas.\nAdolfo Suarez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastian de los Reyes and Alcobendas. |
| BART\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP\nAdolfo Suárez Madrid-Barajas Airport can be found in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| MVP+S\nAdolfo Suárez Madrid-Barajas Airport is located in Madrid, Paracuellos de Jarama, San Sebastián de los Reyes and Alcobendas. |
| Input\nGenerate the question based on the answer: Saint Bernadette Soubirous [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend "Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold\nTo whom did the Virgin Mary allegedly appear in 1858 in Lourdes France? |
| BART\nWho is believed to have appeared to the Virgin Mary at Lourdes? |
| MVP\nWho did the Virgin Mary appear to in Lourdes? |
| MVP+S\nThe Grotto is a replica of the grotto at Lourdes, France where the Virgin Mary appeared to whom? |
| Input\nGenerate the question based on the answer: a copper statue of Christ [SEP] Architecturally, the school has a Catholic character. Atop the Main Building's gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend " Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary. |
| Gold\nWhat is in front of the Notre Dame Main Building? |
| BART\nWhat is in front of the Main Building and facing it? |
| MVP\nWhat is immediately in front of the Main Building? |
| MVP+S\nWhat is immediately in front of the Main Building? |
| Input\nGiven the story title: male was out jogging one morning. |
| Gold\nthe weather was crisp and cool . male felt good and energetic . he decided to keep jogging longer than normal . male went several more miles out of his way . |
| BART\nhe saw a man struggling to walk across a field . male stopped to help him . he saw a guy struggling to walking across a fields . the man thanked male profusely ! |
| MVP\nhe ran into a man who was holding a knife . male was scared , so he ran away . the man threatened male . male ran back into his house . |
| MVP+S\nhe noticed a sound coming from a tree . male went to investigate . he found a small baby bird stuck in the trunk . male took the bird to the vet and they released it . |
| Input\nGiven the story title: ken was driving around in the snow. |
| Gold\nhe needed to get home from work . he was driving slowly to avoid accidents . unfortunately the roads were too slick and ken lost control . his tires lost traction and he hit a tree . |
| BART\nhe hit a patch of black ice . ken slid and hurt his knee . he went to the doctor for it . krampus was sent home with a concussion . |
| MVP\nhe was in his car and it was very cold . ken started to back up . the car started to make a loud noise . kena had to back out and call for help . |
| MVP+S\nhe was having trouble starting his car . he had to push it a couple of blocks . his car would n’t start . ken had to call a tow truck for help . |
| Input\nGiven the dialog: i love to meet new people. [SEP] i have a turtle named timothy. [SEP] my favorite sport is ultimate frisbee. [SEP] my parents are living in bora bora. [SEP] autumn is my favorite season. [X_SEP] hello, how are you doing tonight? |
| Gold\ni am well an loving this interaction how are you? |
| BART\ni am doing well, how are you? |
| MVP\ni am doing well. how are you? |
| MVP+S\ni am doing well. how are you? |
| Input\nGiven the dialog: i just bought a brand new house. [SEP] i like to dance at the club. [SEP] i run a dog obedience school. [SEP] i have a big sweet tooth. [SEP] i like taking and posting selkies. [X_SEP] hello, how are you doing tonight? [SEP] i am well an loving this interaction how are you? |
| Gold\ni am great. i just got back from the club. |
| BART\ni am doing well, thanks for asking. what do you do for fun? |
| MVP\ni am doing well. i just got home from dancing at the club. |
| MVP+S\ni am doing well. i just bought a new house. |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||
| Pair | Triplet | Pair | Triplet | Pair | Triplet | Pair | Triplet | |
| WhatHowWhy◇ | 56.10 | 51.89 | 53.85 | 43.50 | 56.23 | 46.79 | 60.04 | 53.62 |
| CMLA+◇ | 48.95 | 43.12 | 44.10 | 32.90 | 44.60 | 35.90 | 50.00 | 41.60 |
| RINANTE+◇ | 46.29 | 34.03 | 29.70 | 20.00 | 35.40 | 28.00 | 30.70 | 23.30 |
| Unified+◇ | 55.34 | 51.68 | 52.56 | 42.47 | 56.85 | 46.69 | 53.75 | 44.51 |
| Dual-MRC◇ | 74.93 | 70.32 | 63.37 | 55.58 | 64.97 | 57.21 | 75.71 | 67.40 |
| GenerativeV | 77.68 | 72.46 | 66.11 | 57.59 | 67.98 | 60.11 | 77.38 | 69.98 |
| GAS# | - | 70.20 | - | 54.50 | - | 59.10 | - | 65.00 |
| LEGO# | - | 72.60 | - | 59.50 | - | 63.20 | - | 71.50 |
| JETM=6V | - | 60.41 | - | 46.65 | - | 53.68 | - | 63.41 |
| JETOM=6V | - | 63.92 | - | 50.00 | - | 54.67 | - | 62.98 |
| SPAN* | 78.62 | 73.96 | 69.48 | 60.59 | 71.56 | 64.50 | 78.85 | 70.48 |
| RLI(Ours) | 79.92 | 74.98 | 70.27 | 61.97 | 72.66 | 65.71 | 81.29 | 73.33 |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| WhatHowWhy◇ | 43.24 | 63.66 | 51.46 | 37.38 | 50.38 | 42.87 | 48.07 | 57.51 | 52.32 | 46.96 | 64.24 | 54.21 |
| TOP# | 63.59 | 73.44 | 68.16 | 57.84 | 59.33 | 58.58 | 54.53 | 63.30 | 58.59 | 63.57 | 71.98 | 67.52 |
| BMRC# | 72.17 | 65.43 | 68.64 | 65.91 | 52.15 | 58.18 | 62.48 | 55.55 | 58.79 | 69.87 | 65.68 | 67.35 |
| RL# | 70.60 | 68.65 | 69.61 | 64.80 | 54.99 | 59.50 | 65.45 | 60.29 | 62.72 | 67.21 | 69.69 | 68.41 |
| GAS# | - | - | 72.16 | - | - | 60.78 | - | - | 62.10 | - | - | 70.10 |
| OTE-MTL# | 63.07 | 58.25 | 60.56 | 54.26 | 41.07 | 46.75 | 60.88 | 42.68 | 50.18 | 65.65 | 54.28 | 59.42 |
| GTS# | 67.76 | 67.29 | 67.50 | 57.82 | 51.32 | 54.36 | 62.59 | 57.94 | 60.15 | 66.08 | 69.91 | 67.93 |
| JETM=6◇ | 63.44 | 54.12 | 59.41 | 53.53 | 43.28 | 47.86 | 68.20 | 42.89 | 52.66 | 65.28 | 51.95 | 63.83 |
| JETM=6◇ | 70.56 | 55.94 | 62.40 | 55.39 | 47.33 | 51.04 | 64.45 | 51.96 | 57.53 | 70.42 | 58.37 | 63.83 |
| SPAN◇ | 72.89 | 70.89 | 71.85 | 63.44 | 55.84 | 59.38 | 62.18 | 64.45 | 63.27 | 69.45 | 71.17 | 70.26 |
| EMC-GCN▼ | 71.21 | 72.39 | 71.78 | 61.70 | 56.26 | 58.81 | 61.54 | 62.47 | 61.93 | 65.62 | 71.30 | 68.33 |
| RLI (Ours) | 77.46 | 71.97 | 74.34 | 63.32 | 57.43 | 60.96 | 60.08 | 70.66 | 65.41 | 70.50 | 74.28 | 72.34 |
| Dataset | Model | Dev F1 | Test F1 | |
| Res14 | w/o joint | 66.85 | 72.07 | |
| joint | w/o sentiment | 67.55 | 73.70 | |
| w/o pre-training | 67.12 | 72.58 | ||
| full model | 68.00 | 74.34 | ||
| Lap14 | w/o joint | 60.03 | 60.33 | |
| joint | w/o sentiment | 61.90 | 60.54 | |
| w/o pre-training | 61.06 | 60.02 | ||
| full model | 62.55 | 60.96 | ||
| Res15 | w/o joint | 70.83 | 63.99 | |
| joint | w/o sentiment | 71.54 | 65.04 | |
| w/o pre-training | 71.24 | 64.48 | ||
| full model | 72.21 | 65.41 | ||
| Res16 | w/o joint | 70.47 | 70.30 | |
| joint | w/o sentiment | 71.44 | 71.39 | |
| w/o pre-training | 70.75 | 71.69 | ||
| full model | 73.04 | 72.34 |
| Models | Base | Base+Aug | Ours |
| Avg. F1 | 66.19 | 66.99(+0.80) | 68.26(+2.07) |
| Dataset | Res14 | Lap14 | Res15 | Res16 | |
| s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | ||
| Da | Train | 1300/1575/143/427 | 593/703/25/195 | 842/933/49/307 | 920/664/117/484 |
| Dev | 323/377/32/115 | 148/179/9/50 | 210/225/10/81 | 228/207/16/114 | |
| Test | 496/675/45/142 | 318/291/25/139 | 320/362/27/76 | 339/335/50/105 | |
| Db | Train | 1266/1692/166/480 | 906/817/126/517 | 605/783/25/205 | 857/1015/50/329 |
| Dev | 310/404/54/119 | 219/169/36/141 | 148/185/11/53 | 210/252/11/76 | |
| Test | 492/773/66/155 | 328/364/63/116 | 322/317/25/143 | 326/407/29/78 |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||
| Pair | Triplet | Pair | Triplet | Pair | Triplet | Pair | Triplet | |
| WhatHowWhy◇ | 56.10 | 51.89 | 53.85 | 43.50 | 56.23 | 46.79 | 60.04 | 53.62 |
| CMLA+◇ | 48.95 | 43.12 | 44.10 | 32.90 | 44.60 | 35.90 | 50.00 | 41.60 |
| RINANTE+◇ | 46.29 | 34.03 | 29.70 | 20.00 | 35.40 | 28.00 | 30.70 | 23.30 |
| Unified+◇ | 55.34 | 51.68 | 52.56 | 42.47 | 56.85 | 46.69 | 53.75 | 44.51 |
| Dual-MRC◇ | 74.93 | 70.32 | 63.37 | 55.58 | 64.97 | 57.21 | 75.71 | 67.40 |
| GenerativeV | 77.68 | 72.46 | 66.11 | 57.59 | 67.98 | 60.11 | 77.38 | 69.98 |
| GAS# | - | 70.20 | - | 54.50 | - | 59.10 | - | 65.00 |
| LEGO# | - | 72.60 | - | 59.50 | - | 63.20 | - | 71.50 |
| JETM=6V | - | 60.41 | - | 46.65 | - | 53.68 | - | 63.41 |
| JETOM=6V | - | 63.92 | - | 50.00 | - | 54.67 | - | 62.98 |
| SPAN* | 78.62 | 73.96 | 69.48 | 60.59 | 71.56 | 64.50 | 78.85 | 70.48 |
| RLI(Ours) | 79.92 | 74.98 | 70.27 | 61.97 | 72.66 | 65.71 | 81.29 | 73.33 |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| WhatHowWhy◇ | 43.24 | 63.66 | 51.46 | 37.38 | 50.38 | 42.87 | 48.07 | 57.51 | 52.32 | 46.96 | 64.24 | 54.21 |
| TOP# | 63.59 | 73.44 | 68.16 | 57.84 | 59.33 | 58.58 | 54.53 | 63.30 | 58.59 | 63.57 | 71.98 | 67.52 |
| BMRC# | 72.17 | 65.43 | 68.64 | 65.91 | 52.15 | 58.18 | 62.48 | 55.55 | 58.79 | 69.87 | 65.68 | 67.35 |
| RL# | 70.60 | 68.65 | 69.61 | 64.80 | 54.99 | 59.50 | 65.45 | 60.29 | 62.72 | 67.21 | 69.69 | 68.41 |
| GAS# | - | - | 72.16 | - | - | 60.78 | - | - | 62.10 | - | - | 70.10 |
| OTE-MTL# | 63.07 | 58.25 | 60.56 | 54.26 | 41.07 | 46.75 | 60.88 | 42.68 | 50.18 | 65.65 | 54.28 | 59.42 |
| GTS# | 67.76 | 67.29 | 67.50 | 57.82 | 51.32 | 54.36 | 62.59 | 57.94 | 60.15 | 66.08 | 69.91 | 67.93 |
| JETM=6◇ | 63.44 | 54.12 | 59.41 | 53.53 | 43.28 | 47.86 | 68.20 | 42.89 | 52.66 | 65.28 | 51.95 | 63.83 |
| JETM=6◇ | 70.56 | 55.94 | 62.40 | 55.39 | 47.33 | 51.04 | 64.45 | 51.96 | 57.53 | 70.42 | 58.37 | 63.83 |
| SPAN◇ | 72.89 | 70.89 | 71.85 | 63.44 | 55.84 | 59.38 | 62.18 | 64.45 | 63.27 | 69.45 | 71.17 | 70.26 |
| EMC-GCN▼ | 71.21 | 72.39 | 71.78 | 61.70 | 56.26 | 58.81 | 61.54 | 62.47 | 61.93 | 65.62 | 71.30 | 68.33 |
| RLI (Ours) | 77.46 | 71.97 | 74.34 | 63.32 | 57.43 | 60.96 | 60.08 | 70.66 | 65.41 | 70.50 | 74.28 | 72.34 |
| Dataset | Model | Dev F1 | Test F1 | |
| Res14 | w/o joint | 66.85 | 72.07 | |
| joint | w/o sentiment | 67.55 | 73.70 | |
| w/o pre-training | 67.12 | 72.58 | ||
| full model | 68.00 | 74.34 | ||
| Lap14 | w/o joint | 60.03 | 60.33 | |
| joint | w/o sentiment | 61.90 | 60.54 | |
| w/o pre-training | 61.06 | 60.02 | ||
| full model | 62.55 | 60.96 | ||
| Res15 | w/o joint | 70.83 | 63.99 | |
| joint | w/o sentiment | 71.54 | 65.04 | |
| w/o pre-training | 71.24 | 64.48 | ||
| full model | 72.21 | 65.41 | ||
| Res16 | w/o joint | 70.47 | 70.30 | |
| joint | w/o sentiment | 71.44 | 71.39 | |
| w/o pre-training | 70.75 | 71.69 | ||
| full model | 73.04 | 72.34 |
| Models | Base | Base+Aug | Ours |
| Avg. F1 | 66.19 | 66.99(+0.80) | 68.26(+2.07) |
| Dataset | Res14 | Lap14 | Res15 | Res16 | |
| s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | ||
| Da | Train | 1300/1575/143/427 | 593/703/25/195 | 842/933/49/307 | 920/664/117/484 |
| Dev | 323/377/32/115 | 148/179/9/50 | 210/225/10/81 | 228/207/16/114 | |
| Test | 496/675/45/142 | 318/291/25/139 | 320/362/27/76 | 339/335/50/105 | |
| Db | Train | 1266/1692/166/480 | 906/817/126/517 | 605/783/25/205 | 857/1015/50/329 |
| Dev | 310/404/54/119 | 219/169/36/141 | 148/185/11/53 | 210/252/11/76 | |
| Test | 492/773/66/155 | 328/364/63/116 | 322/317/25/143 | 326/407/29/78 |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||
| Pair | Triplet | Pair | Triplet | Pair | Triplet | Pair | Triplet | |
| WhatHowWhy◇ | 56.10 | 51.89 | 53.85 | 43.50 | 56.23 | 46.79 | 60.04 | 53.62 |
| CMLA+◇ | 48.95 | 43.12 | 44.10 | 32.90 | 44.60 | 35.90 | 50.00 | 41.60 |
| RINANTE+◇ | 46.29 | 34.03 | 29.70 | 20.00 | 35.40 | 28.00 | 30.70 | 23.30 |
| Unified+◇ | 55.34 | 51.68 | 52.56 | 42.47 | 56.85 | 46.69 | 53.75 | 44.51 |
| Dual-MRC◇ | 74.93 | 70.32 | 63.37 | 55.58 | 64.97 | 57.21 | 75.71 | 67.40 |
| GenerativeV | 77.68 | 72.46 | 66.11 | 57.59 | 67.98 | 60.11 | 77.38 | 69.98 |
| GAS# | - | 70.20 | - | 54.50 | - | 59.10 | - | 65.00 |
| LEGO# | - | 72.60 | - | 59.50 | - | 63.20 | - | 71.50 |
| JETM=6V | - | 60.41 | - | 46.65 | - | 53.68 | - | 63.41 |
| JETOM=6V | - | 63.92 | - | 50.00 | - | 54.67 | - | 62.98 |
| SPAN* | 78.62 | 73.96 | 69.48 | 60.59 | 71.56 | 64.50 | 78.85 | 70.48 |
| RLI(Ours) | 79.92 | 74.98 | 70.27 | 61.97 | 72.66 | 65.71 | 81.29 | 73.33 |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| WhatHowWhy◇ | 43.24 | 63.66 | 51.46 | 37.38 | 50.38 | 42.87 | 48.07 | 57.51 | 52.32 | 46.96 | 64.24 | 54.21 |
| TOP# | 63.59 | 73.44 | 68.16 | 57.84 | 59.33 | 58.58 | 54.53 | 63.30 | 58.59 | 63.57 | 71.98 | 67.52 |
| BMRC# | 72.17 | 65.43 | 68.64 | 65.91 | 52.15 | 58.18 | 62.48 | 55.55 | 58.79 | 69.87 | 65.68 | 67.35 |
| RL# | 70.60 | 68.65 | 69.61 | 64.80 | 54.99 | 59.50 | 65.45 | 60.29 | 62.72 | 67.21 | 69.69 | 68.41 |
| GAS# | - | - | 72.16 | - | - | 60.78 | - | - | 62.10 | - | - | 70.10 |
| OTE-MTL# | 63.07 | 58.25 | 60.56 | 54.26 | 41.07 | 46.75 | 60.88 | 42.68 | 50.18 | 65.65 | 54.28 | 59.42 |
| GTS# | 67.76 | 67.29 | 67.50 | 57.82 | 51.32 | 54.36 | 62.59 | 57.94 | 60.15 | 66.08 | 69.91 | 67.93 |
| JETM=6◇ | 63.44 | 54.12 | 59.41 | 53.53 | 43.28 | 47.86 | 68.20 | 42.89 | 52.66 | 65.28 | 51.95 | 63.83 |
| JETM=6◇ | 70.56 | 55.94 | 62.40 | 55.39 | 47.33 | 51.04 | 64.45 | 51.96 | 57.53 | 70.42 | 58.37 | 63.83 |
| SPAN◇ | 72.89 | 70.89 | 71.85 | 63.44 | 55.84 | 59.38 | 62.18 | 64.45 | 63.27 | 69.45 | 71.17 | 70.26 |
| EMC-GCN▼ | 71.21 | 72.39 | 71.78 | 61.70 | 56.26 | 58.81 | 61.54 | 62.47 | 61.93 | 65.62 | 71.30 | 68.33 |
| RLI (Ours) | 77.46 | 71.97 | 74.34 | 63.32 | 57.43 | 60.96 | 60.08 | 70.66 | 65.41 | 70.50 | 74.28 | 72.34 |
| Dataset | Model | Dev F1 | Test F1 | |
| Res14 | w/o joint | 66.85 | 72.07 | |
| joint | w/o sentiment | 67.55 | 73.70 | |
| w/o pre-training | 67.12 | 72.58 | ||
| full model | 68.00 | 74.34 | ||
| Lap14 | w/o joint | 60.03 | 60.33 | |
| joint | w/o sentiment | 61.90 | 60.54 | |
| w/o pre-training | 61.06 | 60.02 | ||
| full model | 62.55 | 60.96 | ||
| Res15 | w/o joint | 70.83 | 63.99 | |
| joint | w/o sentiment | 71.54 | 65.04 | |
| w/o pre-training | 71.24 | 64.48 | ||
| full model | 72.21 | 65.41 | ||
| Res16 | w/o joint | 70.47 | 70.30 | |
| joint | w/o sentiment | 71.44 | 71.39 | |
| w/o pre-training | 70.75 | 71.69 | ||
| full model | 73.04 | 72.34 | ||
| Models | Base | Base+Aug | Ours |
| Avg. F1 | 66.19 | 66.99(+0.80) | 68.26(+2.07) |
| Dataset | Res14 | Lap14 | Res15 | Res16 | |
| s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | ||
| Da | Train | 1300/1575/143/427 | 593/703/25/195 | 842/933/49/307 | 920/664/117/484 |
| Dev | 323/377/32/115 | 148/179/9/50 | 210/225/10/81 | 228/207/16/114 | |
| Test | 496/675/45/142 | 318/291/25/139 | 320/362/27/76 | 339/335/50/105 | |
| Db | Train | 1266/1692/166/480 | 906/817/126/517 | 605/783/25/205 | 857/1015/50/329 |
| Dev | 310/404/54/119 | 219/169/36/141 | 148/185/11/53 | 210/252/11/76 | |
| Test | 492/773/66/155 | 328/364/63/116 | 322/317/25/143 | 326/407/29/78 | |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||
| Pair | Triplet | Pair | Triplet | Pair | Triplet | Pair | Triplet | |
| WhatHowWhy◇ | 56.10 | 51.89 | 53.85 | 43.50 | 56.23 | 46.79 | 60.04 | 53.62 |
| CMLA+◇ | 48.95 | 43.12 | 44.10 | 32.90 | 44.60 | 35.90 | 50.00 | 41.60 |
| RINANTE+◇ | 46.29 | 34.03 | 29.70 | 20.00 | 35.40 | 28.00 | 30.70 | 23.30 |
| Unified+◇ | 55.34 | 51.68 | 52.56 | 42.47 | 56.85 | 46.69 | 53.75 | 44.51 |
| Dual-MRC◇ | 74.93 | 70.32 | 63.37 | 55.58 | 64.97 | 57.21 | 75.71 | 67.40 |
| GenerativeV | 77.68 | 72.46 | 66.11 | 57.59 | 67.98 | 60.11 | 77.38 | 69.98 |
| GAS# | - | 70.20 | - | 54.50 | - | 59.10 | - | 65.00 |
| LEGO# | - | 72.60 | - | 59.50 | - | 63.20 | - | 71.50 |
| JETM=6V | - | 60.41 | - | 46.65 | - | 53.68 | - | 63.41 |
| JETOM=6V | - | 63.92 | - | 50.00 | - | 54.67 | - | 62.98 |
| SPAN* | 78.62 | 73.96 | 69.48 | 60.59 | 71.56 | 64.50 | 78.85 | 70.48 |
| RLI(Ours) | 79.92 | 74.98 | 70.27 | 61.97 | 72.66 | 65.71 | 81.29 | 73.33 |
| Model | Res14 | Lap14 | Res15 | Res16 | ||||||||
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | |
| WhatHowWhy◇ | 43.24 | 63.66 | 51.46 | 37.38 | 50.38 | 42.87 | 48.07 | 57.51 | 52.32 | 46.96 | 64.24 | 54.21 |
| TOP# | 63.59 | 73.44 | 68.16 | 57.84 | 59.33 | 58.58 | 54.53 | 63.30 | 58.59 | 63.57 | 71.98 | 67.52 |
| BMRC# | 72.17 | 65.43 | 68.64 | 65.91 | 52.15 | 58.18 | 62.48 | 55.55 | 58.79 | 69.87 | 65.68 | 67.35 |
| RL# | 70.60 | 68.65 | 69.61 | 64.80 | 54.99 | 59.50 | 65.45 | 60.29 | 62.72 | 67.21 | 69.69 | 68.41 |
| GAS# | - | - | 72.16 | - | - | 60.78 | - | - | 62.10 | - | - | 70.10 |
| OTE-MTL# | 63.07 | 58.25 | 60.56 | 54.26 | 41.07 | 46.75 | 60.88 | 42.68 | 50.18 | 65.65 | 54.28 | 59.42 |
| GTS# | 67.76 | 67.29 | 67.50 | 57.82 | 51.32 | 54.36 | 62.59 | 57.94 | 60.15 | 66.08 | 69.91 | 67.93 |
| JETM=6◇ | 63.44 | 54.12 | 59.41 | 53.53 | 43.28 | 47.86 | 68.20 | 42.89 | 52.66 | 65.28 | 51.95 | 63.83 |
| JETM=6◇ | 70.56 | 55.94 | 62.40 | 55.39 | 47.33 | 51.04 | 64.45 | 51.96 | 57.53 | 70.42 | 58.37 | 63.83 |
| SPAN◇ | 72.89 | 70.89 | 71.85 | 63.44 | 55.84 | 59.38 | 62.18 | 64.45 | 63.27 | 69.45 | 71.17 | 70.26 |
| EMC-GCN▼ | 71.21 | 72.39 | 71.78 | 61.70 | 56.26 | 58.81 | 61.54 | 62.47 | 61.93 | 65.62 | 71.30 | 68.33 |
| RLI (Ours) | 77.46 | 71.97 | 74.34 | 63.32 | 57.43 | 60.96 | 60.08 | 70.66 | 65.41 | 70.50 | 74.28 | 72.34 |
| Dataset | Model | Dev F1 | Test F1 | |
| Res14 | w/o joint | 66.85 | 72.07 | |
| joint | w/o sentiment | 67.55 | 73.70 | |
| w/o pre-training | 67.12 | 72.58 | ||
| full model | 68.00 | 74.34 | ||
| Lap14 | w/o joint | 60.03 | 60.33 | |
| joint | w/o sentiment | 61.90 | 60.54 | |
| w/o pre-training | 61.06 | 60.02 | ||
| full model | 62.55 | 60.96 | ||
| Res15 | w/o joint | 70.83 | 63.99 | |
| joint | w/o sentiment | 71.54 | 65.04 | |
| w/o pre-training | 71.24 | 64.48 | ||
| full model | 72.21 | 65.41 | ||
| Res16 | w/o joint | 70.47 | 70.30 | |
| joint | w/o sentiment | 71.44 | 71.39 | |
| w/o pre-training | 70.75 | 71.69 | ||
| full model | 73.04 | 72.34 |
| Models | Base | Base+Aug | Ours |
| Avg. F1 | 66.19 | 66.99(+0.80) | 68.26(+2.07) |
| Dataset | Res14 | Lap14 | Res15 | Res16 | |
| s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | s/pos/neu/neg | ||
| Da | Train | 1300/1575/143/427 | 593/703/25/195 | 842/933/49/307 | 920/664/117/484 |
| Dev | 323/377/32/115 | 148/179/9/50 | 210/225/10/81 | 228/207/16/114 | |
| Test | 496/675/45/142 | 318/291/25/139 | 320/362/27/76 | 339/335/50/105 | |
| Db | Train | 1266/1692/166/480 | 906/817/126/517 | 605/783/25/205 | 857/1015/50/329 |
| Dev | 310/404/54/119 | 219/169/36/141 | 148/185/11/53 | 210/252/11/76 | |
| Test | 492/773/66/155 | 328/364/63/116 | 322/317/25/143 | 326/407/29/78 |
| Statistic | FewShotATIS | FewShotSNIPS | ||||||||
| # Domian | 1 | 7 | ||||||||
| # Intent | 18 | 7 | ||||||||
| # Slot | 128 | 76 | ||||||||
| # K-shot for intent | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot |
| # Training instances for intents | 33 | 67 | 101 | 130 | 161 | 14 | 27 | 40 | 54 | 67 |
| # K-shot for slot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot |
| # Training instances for slots | 336 | 639 | 1,212 | 1,675 | 1,998 | 195 | 391 | 787 | 1,185 | 1,584 |
| # Testing instances | 878 | 700 |
| Model | FewShotATIS | FewShotSNIPS | ||||||||||
| Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| BERT-ID | 0 | 20.29±2.56 | 30.58±1.91 | 59.90±2.41 | 71.73±1.99 | 85.75±1.68 | 0 | 59.02±2.24 | 76.36±2.21 | 79.87±2.31 | 85.44±2.22 | 87.30±1.42 |
| RoBERTa-ID | 0 | 5.33±2.1 | 7.45±2.46 | 74.47±4.11 | 77.58±2.17 | 85.37±1.33 | 0 | 22.14±2.51 | 75.28±1.49 | 77.64±1.55 | 85.77±1.52 | 86.07±1.81 |
| BERT-NSP-Prompt | 75.42 | 77.55±2.04 | 81.94±1.90 | 87.30±1.76 | 88.77±1.71 | 91.12±1.66 | 79.57 | 80.21±2.08 | 82.64±1.70 | 86.77±1.25 | 89.14±1.51 | 90.78±1.41 |
| Model | FewShotATIS | FewShotSNIPS | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| BERT-SF(BIO) | 22.76±2.17 | 41.79±1.88 | 66.64±1.84 | 76.02±1.46 | 82.86±1.43 | 2.86±0.80 | 40.14±1.73 | 69.42±1.66 | 74.72±1.75 | 81.39±1.84 |
| BERT-SF(ET) | 27.17±1.88 | 49.34±1.63 | 78.43±1.45 | 86.85±1.53 | 90.22±1.20 | 12.50±2.40 | 57.33±2.00 | 76.16±2.11 | 80.62±2.07 | 83.48±1.02 |
| BERT-SF(ET+FL) | 29.05±1.57 | 57.44±4.41 | 83.84±1.44 | 89.43±1.12 | 90.50±1.23 | 15.02±1.80 | 59.61±1.34 | 78.51±1.44 | 82.15±2.10 | 84.94±1.14 |
| RoBERTa-SF(BIO) | 28.12±2.47 | 64.23±2.71 | 78.21±1.67 | 83.27±3.02 | 88.65±1.05 | 14.51±1.56 | 41.09±1.38 | 66.21±1.13 | 75.25±1.16 | 76.71±1.31 |
| RoBERTa-SF(ET) | 28.96±1.75 | 70.85±3.65 | 82.99±2.00 | 86.26±1.99 | 89.71±0.95 | 22.61±2.25 | 46.23±2.44 | 68.08±1.27 | 77.13±1.72 | 78.77±1.03 |
| RoBERTa-SF(ET+FT) | 31.10±1.76 | 72.27±3.39 | 83.59±1.49 | 87.05±1.84 | 90.58±1.24 | 22.54±2.21 | 49.99±3.49 | 71.02±2.05 | 79.91±2.35 | 80.75±2.47 |
| Sample Strategy | ATIS K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 74.60 | 77.33 | 85.19 | 89.64 | 90.21 |
| Dynamic Sample | 74.60 | 83.49 | 90.32 | 89.98 | 91.34 |
| Sample Strategy | Snips K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 81.29 | 87.86 | 86.29 | 89.43 | 90.71 |
| Dynamic Sample | 81.29 | 83.57 | 88.71 | 90.29 | 91.43 |
| Sample Strategy | ATIS K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 29.26 | 49.36 | 82.01 | 89.93 | 91.90 |
| Dynamic Sample | 29.26 | 62.76 | 85.57 | 90.65 | 92.01 |
| Sample Strategy | Snips K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 15.54 | 60.81 | 76.69 | 79.93 | 83.06 |
| Dynamic Sample | 15.54 | 59.11 | 80.03 | 84.56 | 85.52 |
| Intent | Description Text |
| AddTo Playlist | He is adding something to playlist. |
| PlayMusic | He wants to play some music. |
| BookRestaurant | He is booking a restaurant. |
| SearchCreativeWork | He is searching for something like game, videos or TV shows. |
| RateBook | He is rating a book with a score. |
| GetWeather | He wants to know the weather information. |
| SearchScreeningEvent | He is searching for a movie. |
| Model | K-shot | |||||||||
| 1-shot | 2-shot | 3-shot | 4-shot | 5-shot | 6-shot | 7-shot | 8-shot | 9-shot | 10-shot | |
| ATIS | 8.93 | 18.11 | 25.73 | 31.55 | 39.87 | 59.57 | 67.92 | 72.44 | 80.61 | 87.70 |
| SNIPS | 32.49 | 59.29 | 64.26 | 78.71 | 79.64 | 82.29 | 83.06 | 86.71 | 85.81 | 87.86 |
| Model | K-shot | |||||||
| 5-shot | 10-shot | 15-shot | 20-shot | 25-shot | 30-shot | 35-shot | 40-shot | |
| ATIS | 22.21 | 41.21 | 52.17 | 66.23 | 70.14 | 78.06 | 78.92 | 82.21 |
| SNIPS | 1.8 | 41.98 | 50.80 | 69.63 | 72.04 | 75.46 | 77.93 | 80.38 |
| Model | We | Mu | PI | Bo | Se | Re | Cr | Avg. |
| TransferBERT | 59.41 | 42.00 | 46.07 | 20.74 | 28.20 | 67.75 | 58.61 | 46.11 |
| SimBERT | 53.46 | 54.13 | 42.81 | 75.54 | 57.10 | 55.30 | 32.38 | 52.96 |
| ProtoToken | 67.82 | 55.99 | 46.02 | 72.17 | 73.59 | 60.18 | 66.89 | 63.24 |
| Coach (Liu et al., 2020) | 73.56 | 45.85 | 47.23 | 61.61 | 65.82 | 69.99 | 57.28 | 60.19 |
| TapNet (Hou et al., 2020a) | 53.03 | 49.80 | 54.90 | 83.36 | 63.07 | 59.84 | 67.02 | 61.57 |
| TapNet+CDT (Hou et al., 2020a) | 66.48 | 66.36 | 68.23 | 85.76 | 73.60 | 64.20 | 68.47 | 70.44 |
| L-WPZ+CDT (Hou et al., 2020a) | 74.68 | 56.73 | 52.20 | 78.79 | 80.61 | 69.59 | 67.46 | 68.58 |
| L-TapNet+CDT (Hou et al., 2020a) | 71.64 | 67.16 | 75.88 | 84.38 | 82.58 | 70.05 | 73.41 | 75.01 |
| Retriever (Yu et al., 2021) | 82.95 | 61.74 | 71.75 | 81.65 | 73.10 | 79.54 | 51.35 | 71.72 |
| Ours | 79.70 | 78.09 | 76.64 | 90.04 | 82.87 | 79.91 | 63.88 | 78.73 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| AddTo Playlist | 97.00 | 78.23 | 86.61 | 99.10 | 88.71 | 93.62 | 124 |
| PlayMusic | 64.96 | 88.37 | 74.88 | 72.32 | 94.19 | 81.82 | 86 |
| BookRestaurant | 96.81 | 98.91 | 97.85 | 95.83 | 1.00 | 97.87 | 92 |
| SearchCreativeWork | 71.03 | 71.03 | 71.03 | 73.17 | 84.11 | 78.26 | 107 |
| RateBook | 98.75 | 98.75 | 98.75 | 98.59 | 87.50 | 92.72 | 80 |
| GetWeather | 97.09 | 96.15 | 96.62 | 96.19 | 97.12 | 96.65 | 104 |
| SearchScreeningEvent | 85.86 | 79.44 | 82.52 | 93.90 | 71.96 | 81.48 | 107 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| atis_flight | 99.07 | 84.34 | 91.11 | 98.11 | 90.51 | 94.16 | 632 |
| atis-ground_fare | 60.00 | 85.71 | 70.59 | 50.00 | 71.43 | 58.82 | 7 |
| atis_aircraft | 25.00 | 1.00 | 40.00 | 25.00 | 1.00 | 40.00 | 1 |
| atisquantity | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 3 |
| atis_distance | 81.82 | 90.00 | 85.71 | 1.0000 | 1.0000 | 1.0000 | 10 |
| atis_city | 19.35 | 100.00 | 32.43 | 66.67 | 66.67 | 66.67 | 6 |
| atis_capacity | 100.00 | 85.71 | 92.31 | 100.00 | 100.00 | 100.00 | 21 |
| atis_Aircraft | 56.25 | 100.00 | 72.00 | 87.50 | 77.78 | 82.35 | 9 |
| atis_airport no | 88.89 | 100.00 | 94.12 | 100.00 | 100.00 | 100.00 | 8 |
| atis_cheapest | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_airline | 55.88 | 100.00 | 71.70 | 61.67 | 97.37 | 75.51 | 38 |
| atis_abbreviation | 100.00 | 78.79 | 88.14 | 91.18 | 93.94 | 92.54 | 33 |
| atis_restriction | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_Airport | 90.00 | 100.00 | 94.74 | 94.74 | 100.00 | 97.30 | 18 |
| atismeal | 42.86 | 100.00 | 60.00 | 31.25 | 83.33 | 45.45 | 6 |
| atis_airfare | 74.00 | 77.08 | 75.51 | 83.67 | 85.42 | 84.54 | 48 |
| atisGround服务能力 | 100.00 | 86.11 | 92.54 | 100.00 | 86.11 | 92.54 | 36 |
| atis_day_name | 50.00 | 100.00 | 66.67 | 66.67 | 100.00 | 80.00 | 2 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| RoBERTa-ID(AS) | 4.55 | 4.67 | 71.75 | 72.83 | 85.42 | 20.14 | 75.57 | 75.71 | 82.02 | 83.71 |
| RoBERTa-ID(DS) | 4.55 | 7.57 | 81.32 | 73.87 | 86.64 | 20.14 | 77.00 | 77.86 | 84.43 | 85.57 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| RoBERTa-SF(BIO+AS) | 29.08 | 59.89 | 77.03 | 83.46 | 87.39 | 14.35 | 42.33 | 65.42 | 73.63 | 77.28 |
| RoBERTa-SF(BIO+DS) | 29.08 | 67.96 | 80.16 | 87.26 | 90.35 | 14.35 | 42.48 | 67.52 | 76.28 | 77.32 |
| RoBERTa-SF(ET+AS) | 29.11 | 65.00 | 80.59 | 88.61 | 88.65 | 22.42 | 49.58 | 67.59 | 76.61 | 77.61 |
| RoBERTa-SF(ET+DS) | 29.11 | 77.03 | 85.94 | 86.09 | 90.48 | 22.42 | 43.51 | 67.63 | 74.67 | 78.12 |
| RoBERTa-SF(ET+FL+AS) | 31.51 | 65.24 | 81.07 | 86.78 | 89.53 | 24.10 | 54.39 | 68.25 | 75.25 | 78.16 |
| RoBERTa-SF(ET+FT+DS) | 31.51 | 77.09 | 83.77 | 88.24 | 91.02 | 24.10 | 45.99 | 70.97 | 78.12 | 78.42 |
| Statistic | FewShotATIS | FewShotSNIPS | ||||||||
| # Domian | 1 | 7 | ||||||||
| # Intent | 18 | 7 | ||||||||
| # Slot | 128 | 76 | ||||||||
| # K-shot for intent | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot |
| # Training instances for intents | 33 | 67 | 101 | 130 | 161 | 14 | 27 | 40 | 54 | 67 |
| # K-shot for slot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot |
| # Training instances for slots | 336 | 639 | 1,212 | 1,675 | 1,998 | 195 | 391 | 787 | 1,185 | 1,584 |
| # Testing instances | 878 | 700 |
| Model | FewShotATIS | FewShotSNIPS | ||||||||||
| Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| BERT-ID | 0 | 20.29±2.56 | 30.58±1.91 | 59.90±2.41 | 71.73±1.99 | 85.75±1.68 | 0 | 59.02±2.24 | 76.36±2.21 | 79.87±2.31 | 85.44±2.22 | 87.30±1.42 |
| RoBERTa-ID | 0 | 5.33±2.1 | 7.45±2.46 | 74.47±4.11 | 77.58±2.17 | 85.37±1.33 | 0 | 22.14±2.51 | 75.28±1.49 | 77.64±1.55 | 85.77±1.52 | 86.07±1.81 |
| BERT-NSP-Prompt | 75.42 | 77.55±2.04 | 81.94±1.90 | 87.30±1.76 | 88.77±1.71 | 91.12±1.66 | 79.57 | 80.21±2.08 | 82.64±1.70 | 86.77±1.25 | 89.14±1.51 | 90.78±1.41 |
| Model | FewShotATIS | FewShotSNIPS | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| BERT-SF(BIO) | 22.76±2.17 | 41.79±1.88 | 66.64±1.84 | 76.02±1.46 | 82.86±1.43 | 2.86±0.80 | 40.14±1.73 | 69.42±1.66 | 74.72±1.75 | 81.39±1.84 |
| BERT-SF(ET) | 27.17±1.88 | 49.34±1.63 | 78.43±1.45 | 86.85±1.53 | 90.22±1.20 | 12.50±2.40 | 57.33±2.00 | 76.16±2.11 | 80.62±2.07 | 83.48±1.02 |
| BERT-SF(ET+FL) | 29.05±1.57 | 57.44±4.41 | 83.84±1.44 | 89.43±1.12 | 90.50±1.23 | 15.02±1.80 | 59.61±1.34 | 78.51±1.44 | 82.15±2.10 | 84.94±1.14 |
| RoBERTa-SF(BIO) | 28.12±2.47 | 64.23±2.71 | 78.21±1.67 | 83.27±3.02 | 88.65±1.05 | 14.51±1.56 | 41.09±1.38 | 66.21±1.13 | 75.25±1.16 | 76.71±1.31 |
| RoBERTa-SF(ET) | 28.96±1.75 | 70.85±3.65 | 82.99±2.00 | 86.26±1.99 | 89.71±0.95 | 22.61±2.25 | 46.23±2.44 | 68.08±1.27 | 77.13±1.72 | 78.77±1.03 |
| RoBERTa-SF(ET+FT) | 31.10±1.76 | 72.27±3.39 | 83.59±1.49 | 87.05±1.84 | 90.58±1.24 | 22.54±2.21 | 49.99±3.49 | 71.02±2.05 | 79.91±2.35 | 80.75±2.47 |
| Sample Strategy | ATIS K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 74.60 | 77.33 | 85.19 | 89.64 | 90.21 |
| Dynamic Sample | 74.60 | 83.49 | 90.32 | 89.98 | 91.34 |
| Sample Strategy | Snips K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 81.29 | 87.86 | 86.29 | 89.43 | 90.71 |
| Dynamic Sample | 81.29 | 83.57 | 88.71 | 90.29 | 91.43 |
| Sample Strategy | ATIS K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 29.26 | 49.36 | 82.01 | 89.93 | 91.90 |
| Dynamic Sample | 29.26 | 62.76 | 85.57 | 90.65 | 92.01 |
| Sample Strategy | Snips K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 15.54 | 60.81 | 76.69 | 79.93 | 83.06 |
| Dynamic Sample | 15.54 | 59.11 | 80.03 | 84.56 | 85.52 |
| Intent | Description Text |
| AddTo Playlist | He is adding something to playlist. |
| PlayMusic | He wants to play some music. |
| BookRestaurant | He is booking a restaurant. |
| SearchCreativeWork | He is searching for something like game, videos or TV shows. |
| RateBook | He is rating a book with a score. |
| GetWeather | He wants to know the weather information. |
| SearchScreeningEvent | He is searching for a movie. |
| Model | K-shot | |||||||||
| 1-shot | 2-shot | 3-shot | 4-shot | 5-shot | 6-shot | 7-shot | 8-shot | 9-shot | 10-shot | |
| ATIS | 8.93 | 18.11 | 25.73 | 31.55 | 39.87 | 59.57 | 67.92 | 72.44 | 80.61 | 87.70 |
| SNIPS | 32.49 | 59.29 | 64.26 | 78.71 | 79.64 | 82.29 | 83.06 | 86.71 | 85.81 | 87.86 |
| Model | K-shot | |||||||
| 5-shot | 10-shot | 15-shot | 20-shot | 25-shot | 30-shot | 35-shot | 40-shot | |
| ATIS | 22.21 | 41.21 | 52.17 | 66.23 | 70.14 | 78.06 | 78.92 | 82.21 |
| SNIPS | 1.8 | 41.98 | 50.80 | 69.63 | 72.04 | 75.46 | 77.93 | 80.38 |
| Model | We | Mu | PI | Bo | Se | Re | Cr | Avg. |
| TransferBERT | 59.41 | 42.00 | 46.07 | 20.74 | 28.20 | 67.75 | 58.61 | 46.11 |
| SimBERT | 53.46 | 54.13 | 42.81 | 75.54 | 57.10 | 55.30 | 32.38 | 52.96 |
| ProtoToken | 67.82 | 55.99 | 46.02 | 72.17 | 73.59 | 60.18 | 66.89 | 63.24 |
| Coach (Liu et al., 2020) | 73.56 | 45.85 | 47.23 | 61.61 | 65.82 | 69.99 | 57.28 | 60.19 |
| TapNet (Hou et al., 2020a) | 53.03 | 49.80 | 54.90 | 83.36 | 63.07 | 59.84 | 67.02 | 61.57 |
| TapNet+CDT (Hou et al., 2020a) | 66.48 | 66.36 | 68.23 | 85.76 | 73.60 | 64.20 | 68.47 | 70.44 |
| L-WPZ+CDT (Hou et al., 2020a) | 74.68 | 56.73 | 52.20 | 78.79 | 80.61 | 69.59 | 67.46 | 68.58 |
| L-TapNet+CDT (Hou et al., 2020a) | 71.64 | 67.16 | 75.88 | 84.38 | 82.58 | 70.05 | 73.41 | 75.01 |
| Retriever (Yu et al., 2021) | 82.95 | 61.74 | 71.75 | 81.65 | 73.10 | 79.54 | 51.35 | 71.72 |
| Ours | 79.70 | 78.09 | 76.64 | 90.04 | 82.87 | 79.91 | 63.88 | 78.73 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| AddTo Playlist | 97.00 | 78.23 | 86.61 | 99.10 | 88.71 | 93.62 | 124 |
| PlayMusic | 64.96 | 88.37 | 74.88 | 72.32 | 94.19 | 81.82 | 86 |
| BookRestaurant | 96.81 | 98.91 | 97.85 | 95.83 | 1.00 | 97.87 | 92 |
| SearchCreativeWork | 71.03 | 71.03 | 71.03 | 73.17 | 84.11 | 78.26 | 107 |
| RateBook | 98.75 | 98.75 | 98.75 | 98.59 | 87.50 | 92.72 | 80 |
| GetWeather | 97.09 | 96.15 | 96.62 | 96.19 | 97.12 | 96.65 | 104 |
| SearchScreeningEvent | 85.86 | 79.44 | 82.52 | 93.90 | 71.96 | 81.48 | 107 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| atis_flight | 99.07 | 84.34 | 91.11 | 98.11 | 90.51 | 94.16 | 632 |
| atis-ground_fare | 60.00 | 85.71 | 70.59 | 50.00 | 71.43 | 58.82 | 7 |
| atis_aircraft | 25.00 | 1.00 | 40.00 | 25.00 | 1.00 | 40.00 | 1 |
| atisquantity | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 3 |
| atis_distance | 81.82 | 90.00 | 85.71 | 1.0000 | 1.0000 | 1.0000 | 10 |
| atis_city | 19.35 | 100.00 | 32.43 | 66.67 | 66.67 | 66.67 | 6 |
| atis_capacity | 100.00 | 85.71 | 92.31 | 100.00 | 100.00 | 100.00 | 21 |
| atis_Aircraft | 56.25 | 100.00 | 72.00 | 87.50 | 77.78 | 82.35 | 9 |
| atis_airport no | 88.89 | 100.00 | 94.12 | 100.00 | 100.00 | 100.00 | 8 |
| atis_cheapest | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_airline | 55.88 | 100.00 | 71.70 | 61.67 | 97.37 | 75.51 | 38 |
| atis_abbreviation | 100.00 | 78.79 | 88.14 | 91.18 | 93.94 | 92.54 | 33 |
| atis_restriction | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_Airport | 90.00 | 100.00 | 94.74 | 94.74 | 100.00 | 97.30 | 18 |
| atismeal | 42.86 | 100.00 | 60.00 | 31.25 | 83.33 | 45.45 | 6 |
| atis_airfare | 74.00 | 77.08 | 75.51 | 83.67 | 85.42 | 84.54 | 48 |
| atisGround服务能力 | 100.00 | 86.11 | 92.54 | 100.00 | 86.11 | 92.54 | 36 |
| atis_day_name | 50.00 | 100.00 | 66.67 | 66.67 | 100.00 | 80.00 | 2 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| RoBERTa-ID(AS) | 4.55 | 4.67 | 71.75 | 72.83 | 85.42 | 20.14 | 75.57 | 75.71 | 82.02 | 83.71 |
| RoBERTa-ID(DS) | 4.55 | 7.57 | 81.32 | 73.87 | 86.64 | 20.14 | 77.00 | 77.86 | 84.43 | 85.57 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| RoBERTa-SF(BIO+AS) | 29.08 | 59.89 | 77.03 | 83.46 | 87.39 | 14.35 | 42.33 | 65.42 | 73.63 | 77.28 |
| RoBERTa-SF(BIO+DS) | 29.08 | 67.96 | 80.16 | 87.26 | 90.35 | 14.35 | 42.48 | 67.52 | 76.28 | 77.32 |
| RoBERTa-SF(ET+AS) | 29.11 | 65.00 | 80.59 | 88.61 | 88.65 | 22.42 | 49.58 | 67.59 | 76.61 | 77.61 |
| RoBERTa-SF(ET+DS) | 29.11 | 77.03 | 85.94 | 86.09 | 90.48 | 22.42 | 43.51 | 67.63 | 74.67 | 78.12 |
| RoBERTa-SF(ET+FL+AS) | 31.51 | 65.24 | 81.07 | 86.78 | 89.53 | 24.10 | 54.39 | 68.25 | 75.25 | 78.16 |
| RoBERTa-SF(ET+FT+DS) | 31.51 | 77.09 | 83.77 | 88.24 | 91.02 | 24.10 | 45.99 | 70.97 | 78.12 | 78.42 |
| Statistic | FewShotATIS | FewShotSNIPS | ||||||||
| # Domian | 1 | 7 | ||||||||
| # Intent | 18 | 7 | ||||||||
| # Slot | 128 | 76 | ||||||||
| # K-shot for intent | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot |
| # Training instances for intents | 33 | 67 | 101 | 130 | 161 | 14 | 27 | 40 | 54 | 67 |
| # K-shot for slot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot |
| # Training instances for slots | 336 | 639 | 1,212 | 1,675 | 1,998 | 195 | 391 | 787 | 1,185 | 1,584 |
| # Testing instances | 878 | 700 | ||||||||
| Model | FewShotATIS | FewShotSNIPS | ||||||||||
| Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| BERT-ID | 0 | 20.29±2.56 | 30.58±1.91 | 59.90±2.41 | 71.73±1.99 | 85.75±1.68 | 0 | 59.02±2.24 | 76.36±2.21 | 79.87±2.31 | 85.44±2.22 | 87.30±1.42 |
| RoBERTa-ID | 0 | 5.33±2.1 | 7.45±2.46 | 74.47±4.11 | 77.58±2.17 | 85.37±1.33 | 0 | 22.14±2.51 | 75.28±1.49 | 77.64±1.55 | 85.77±1.52 | 86.07±1.81 |
| BERT-NSP-Prompt | 75.42 | 77.55±2.04 | 81.94±1.90 | 87.30±1.76 | 88.77±1.71 | 91.12±1.66 | 79.57 | 80.21±2.08 | 82.64±1.70 | 86.77±1.25 | 89.14±1.51 | 90.78±1.41 |
| Model | FewShotATIS | FewShotSNIPS | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| BERT-SF(BIO) | 22.76±2.17 | 41.79±1.88 | 66.64±1.84 | 76.02±1.46 | 82.86±1.43 | 2.86±0.80 | 40.14±1.73 | 69.42±1.66 | 74.72±1.75 | 81.39±1.84 |
| BERT-SF(ET) | 27.17±1.88 | 49.34±1.63 | 78.43±1.45 | 86.85±1.53 | 90.22±1.20 | 12.50±2.40 | 57.33±2.00 | 76.16±2.11 | 80.62±2.07 | 83.48±1.02 |
| BERT-SF(ET+FL) | 29.05±1.57 | 57.44±4.41 | 83.84±1.44 | 89.43±1.12 | 90.50±1.23 | 15.02±1.80 | 59.61±1.34 | 78.51±1.44 | 82.15±2.10 | 84.94±1.14 |
| RoBERTa-SF(BIO) | 28.12±2.47 | 64.23±2.71 | 78.21±1.67 | 83.27±3.02 | 88.65±1.05 | 14.51±1.56 | 41.09±1.38 | 66.21±1.13 | 75.25±1.16 | 76.71±1.31 |
| RoBERTa-SF(ET) | 28.96±1.75 | 70.85±3.65 | 82.99±2.00 | 86.26±1.99 | 89.71±0.95 | 22.61±2.25 | 46.23±2.44 | 68.08±1.27 | 77.13±1.72 | 78.77±1.03 |
| RoBERTa-SF(ET+FT) | 31.10±1.76 | 72.27±3.39 | 83.59±1.49 | 87.05±1.84 | 90.58±1.24 | 22.54±2.21 | 49.99±3.49 | 71.02±2.05 | 79.91±2.35 | 80.75±2.47 |
| Sample Strategy | ATIS K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 74.60 | 77.33 | 85.19 | 89.64 | 90.21 |
| Dynamic Sample | 74.60 | 83.49 | 90.32 | 89.98 | 91.34 |
| Sample Strategy | Snips K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 81.29 | 87.86 | 86.29 | 89.43 | 90.71 |
| Dynamic Sample | 81.29 | 83.57 | 88.71 | 90.29 | 91.43 |
| Sample Strategy | ATIS K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 29.26 | 49.36 | 82.01 | 89.93 | 91.90 |
| Dynamic Sample | 29.26 | 62.76 | 85.57 | 90.65 | 92.01 |
| Sample Strategy | Snips K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 15.54 | 60.81 | 76.69 | 79.93 | 83.06 |
| Dynamic Sample | 15.54 | 59.11 | 80.03 | 84.56 | 85.52 |
| Intent | Description Text |
| AddTo Playlist | He is adding something to playlist. |
| PlayMusic | He wants to play some music. |
| BookRestaurant | He is booking a restaurant. |
| SearchCreativeWork | He is searching for something like game, videos or TV shows. |
| RateBook | He is rating a book with a score. |
| GetWeather | He wants to know the weather information. |
| SearchScreeningEvent | He is searching for a movie. |
| Model | K-shot | |||||||||
| 1-shot | 2-shot | 3-shot | 4-shot | 5-shot | 6-shot | 7-shot | 8-shot | 9-shot | 10-shot | |
| ATIS | 8.93 | 18.11 | 25.73 | 31.55 | 39.87 | 59.57 | 67.92 | 72.44 | 80.61 | 87.70 |
| SNIPS | 32.49 | 59.29 | 64.26 | 78.71 | 79.64 | 82.29 | 83.06 | 86.71 | 85.81 | 87.86 |
| Model | K-shot | |||||||
| 5-shot | 10-shot | 15-shot | 20-shot | 25-shot | 30-shot | 35-shot | 40-shot | |
| ATIS | 22.21 | 41.21 | 52.17 | 66.23 | 70.14 | 78.06 | 78.92 | 82.21 |
| SNIPS | 1.8 | 41.98 | 50.80 | 69.63 | 72.04 | 75.46 | 77.93 | 80.38 |
| Model | We | Mu | PI | Bo | Se | Re | Cr | Avg. |
| TransferBERT | 59.41 | 42.00 | 46.07 | 20.74 | 28.20 | 67.75 | 58.61 | 46.11 |
| SimBERT | 53.46 | 54.13 | 42.81 | 75.54 | 57.10 | 55.30 | 32.38 | 52.96 |
| ProtoToken | 67.82 | 55.99 | 46.02 | 72.17 | 73.59 | 60.18 | 66.89 | 63.24 |
| Coach (Liu et al., 2020) | 73.56 | 45.85 | 47.23 | 61.61 | 65.82 | 69.99 | 57.28 | 60.19 |
| TapNet (Hou et al., 2020a) | 53.03 | 49.80 | 54.90 | 83.36 | 63.07 | 59.84 | 67.02 | 61.57 |
| TapNet+CDT (Hou et al., 2020a) | 66.48 | 66.36 | 68.23 | 85.76 | 73.60 | 64.20 | 68.47 | 70.44 |
| L-WPZ+CDT (Hou et al., 2020a) | 74.68 | 56.73 | 52.20 | 78.79 | 80.61 | 69.59 | 67.46 | 68.58 |
| L-TapNet+CDT (Hou et al., 2020a) | 71.64 | 67.16 | 75.88 | 84.38 | 82.58 | 70.05 | 73.41 | 75.01 |
| Retriever (Yu et al., 2021) | 82.95 | 61.74 | 71.75 | 81.65 | 73.10 | 79.54 | 51.35 | 71.72 |
| Ours | 79.70 | 78.09 | 76.64 | 90.04 | 82.87 | 79.91 | 63.88 | 78.73 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| AddTo Playlist | 97.00 | 78.23 | 86.61 | 99.10 | 88.71 | 93.62 | 124 |
| PlayMusic | 64.96 | 88.37 | 74.88 | 72.32 | 94.19 | 81.82 | 86 |
| BookRestaurant | 96.81 | 98.91 | 97.85 | 95.83 | 1.00 | 97.87 | 92 |
| SearchCreativeWork | 71.03 | 71.03 | 71.03 | 73.17 | 84.11 | 78.26 | 107 |
| RateBook | 98.75 | 98.75 | 98.75 | 98.59 | 87.50 | 92.72 | 80 |
| GetWeather | 97.09 | 96.15 | 96.62 | 96.19 | 97.12 | 96.65 | 104 |
| SearchScreeningEvent | 85.86 | 79.44 | 82.52 | 93.90 | 71.96 | 81.48 | 107 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| atis_flight | 99.07 | 84.34 | 91.11 | 98.11 | 90.51 | 94.16 | 632 |
| atis-ground_fare | 60.00 | 85.71 | 70.59 | 50.00 | 71.43 | 58.82 | 7 |
| atis_aircraft | 25.00 | 1.00 | 40.00 | 25.00 | 1.00 | 40.00 | 1 |
| atisquantity | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 3 |
| atis_distance | 81.82 | 90.00 | 85.71 | 1.0000 | 1.0000 | 1.0000 | 10 |
| atis_city | 19.35 | 100.00 | 32.43 | 66.67 | 66.67 | 66.67 | 6 |
| atis_capacity | 100.00 | 85.71 | 92.31 | 100.00 | 100.00 | 100.00 | 21 |
| atis_Aircraft | 56.25 | 100.00 | 72.00 | 87.50 | 77.78 | 82.35 | 9 |
| atis_airport no | 88.89 | 100.00 | 94.12 | 100.00 | 100.00 | 100.00 | 8 |
| atis_cheapest | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_airline | 55.88 | 100.00 | 71.70 | 61.67 | 97.37 | 75.51 | 38 |
| atis_abbreviation | 100.00 | 78.79 | 88.14 | 91.18 | 93.94 | 92.54 | 33 |
| atis_restriction | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_Airport | 90.00 | 100.00 | 94.74 | 94.74 | 100.00 | 97.30 | 18 |
| atismeal | 42.86 | 100.00 | 60.00 | 31.25 | 83.33 | 45.45 | 6 |
| atis_airfare | 74.00 | 77.08 | 75.51 | 83.67 | 85.42 | 84.54 | 48 |
| atisGround服务能力 | 100.00 | 86.11 | 92.54 | 100.00 | 86.11 | 92.54 | 36 |
| atis_day_name | 50.00 | 100.00 | 66.67 | 66.67 | 100.00 | 80.00 | 2 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| RoBERTa-ID(AS) | 4.55 | 4.67 | 71.75 | 72.83 | 85.42 | 20.14 | 75.57 | 75.71 | 82.02 | 83.71 |
| RoBERTa-ID(DS) | 4.55 | 7.57 | 81.32 | 73.87 | 86.64 | 20.14 | 77.00 | 77.86 | 84.43 | 85.57 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| RoBERTa-SF(BIO+AS) | 29.08 | 59.89 | 77.03 | 83.46 | 87.39 | 14.35 | 42.33 | 65.42 | 73.63 | 77.28 |
| RoBERTa-SF(BIO+DS) | 29.08 | 67.96 | 80.16 | 87.26 | 90.35 | 14.35 | 42.48 | 67.52 | 76.28 | 77.32 |
| RoBERTa-SF(ET+AS) | 29.11 | 65.00 | 80.59 | 88.61 | 88.65 | 22.42 | 49.58 | 67.59 | 76.61 | 77.61 |
| RoBERTa-SF(ET+DS) | 29.11 | 77.03 | 85.94 | 86.09 | 90.48 | 22.42 | 43.51 | 67.63 | 74.67 | 78.12 |
| RoBERTa-SF(ET+FL+AS) | 31.51 | 65.24 | 81.07 | 86.78 | 89.53 | 24.10 | 54.39 | 68.25 | 75.25 | 78.16 |
| RoBERTa-SF(ET+FT+DS) | 31.51 | 77.09 | 83.77 | 88.24 | 91.02 | 24.10 | 45.99 | 70.97 | 78.12 | 78.42 |
| Statistic | FewShotATIS | FewShotSNIPS | ||||||||
| # Domian | 1 | 7 | ||||||||
| # Intent | 18 | 7 | ||||||||
| # Slot | 128 | 76 | ||||||||
| # K-shot for intent | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot |
| # Training instances for intents | 33 | 67 | 101 | 130 | 161 | 14 | 27 | 40 | 54 | 67 |
| # K-shot for slot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot |
| # Training instances for slots | 336 | 639 | 1,212 | 1,675 | 1,998 | 195 | 391 | 787 | 1,185 | 1,584 |
| # Testing instances | 878 | 700 |
| Model | FewShotATIS | FewShotSNIPS | ||||||||||
| Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | Zero-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| BERT-ID | 0 | 20.29±2.56 | 30.58±1.91 | 59.90±2.41 | 71.73±1.99 | 85.75±1.68 | 0 | 59.02±2.24 | 76.36±2.21 | 79.87±2.31 | 85.44±2.22 | 87.30±1.42 |
| RoBERTa-ID | 0 | 5.33±2.1 | 7.45±2.46 | 74.47±4.11 | 77.58±2.17 | 85.37±1.33 | 0 | 22.14±2.51 | 75.28±1.49 | 77.64±1.55 | 85.77±1.52 | 86.07±1.81 |
| BERT-NSP-Prompt | 75.42 | 77.55±2.04 | 81.94±1.90 | 87.30±1.76 | 88.77±1.71 | 91.12±1.66 | 79.57 | 80.21±2.08 | 82.64±1.70 | 86.77±1.25 | 89.14±1.51 | 90.78±1.41 |
| Model | FewShotATIS | FewShotSNIPS | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| BERT-SF(BIO) | 22.76±2.17 | 41.79±1.88 | 66.64±1.84 | 76.02±1.46 | 82.86±1.43 | 2.86±0.80 | 40.14±1.73 | 69.42±1.66 | 74.72±1.75 | 81.39±1.84 |
| BERT-SF(ET) | 27.17±1.88 | 49.34±1.63 | 78.43±1.45 | 86.85±1.53 | 90.22±1.20 | 12.50±2.40 | 57.33±2.00 | 76.16±2.11 | 80.62±2.07 | 83.48±1.02 |
| BERT-SF(ET+FL) | 29.05±1.57 | 57.44±4.41 | 83.84±1.44 | 89.43±1.12 | 90.50±1.23 | 15.02±1.80 | 59.61±1.34 | 78.51±1.44 | 82.15±2.10 | 84.94±1.14 |
| RoBERTa-SF(BIO) | 28.12±2.47 | 64.23±2.71 | 78.21±1.67 | 83.27±3.02 | 88.65±1.05 | 14.51±1.56 | 41.09±1.38 | 66.21±1.13 | 75.25±1.16 | 76.71±1.31 |
| RoBERTa-SF(ET) | 28.96±1.75 | 70.85±3.65 | 82.99±2.00 | 86.26±1.99 | 89.71±0.95 | 22.61±2.25 | 46.23±2.44 | 68.08±1.27 | 77.13±1.72 | 78.77±1.03 |
| RoBERTa-SF(ET+FT) | 31.10±1.76 | 72.27±3.39 | 83.59±1.49 | 87.05±1.84 | 90.58±1.24 | 22.54±2.21 | 49.99±3.49 | 71.02±2.05 | 79.91±2.35 | 80.75±2.47 |
| Sample Strategy | ATIS K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 74.60 | 77.33 | 85.19 | 89.64 | 90.21 |
| Dynamic Sample | 74.60 | 83.49 | 90.32 | 89.98 | 91.34 |
| Sample Strategy | Snips K-shot | ||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| Average Sample | 81.29 | 87.86 | 86.29 | 89.43 | 90.71 |
| Dynamic Sample | 81.29 | 83.57 | 88.71 | 90.29 | 91.43 |
| Sample Strategy | ATIS K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 29.26 | 49.36 | 82.01 | 89.93 | 91.90 |
| Dynamic Sample | 29.26 | 62.76 | 85.57 | 90.65 | 92.01 |
| Sample Strategy | Snips K-shot | ||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| Average Sample | 15.54 | 60.81 | 76.69 | 79.93 | 83.06 |
| Dynamic Sample | 15.54 | 59.11 | 80.03 | 84.56 | 85.52 |
| Intent | Description Text |
| AddTo Playlist | He is adding something to playlist. |
| PlayMusic | He wants to play some music. |
| BookRestaurant | He is booking a restaurant. |
| SearchCreativeWork | He is searching for something like game, videos or TV shows. |
| RateBook | He is rating a book with a score. |
| GetWeather | He wants to know the weather information. |
| SearchScreeningEvent | He is searching for a movie. |
| Model | K-shot | |||||||||
| 1-shot | 2-shot | 3-shot | 4-shot | 5-shot | 6-shot | 7-shot | 8-shot | 9-shot | 10-shot | |
| ATIS | 8.93 | 18.11 | 25.73 | 31.55 | 39.87 | 59.57 | 67.92 | 72.44 | 80.61 | 87.70 |
| SNIPS | 32.49 | 59.29 | 64.26 | 78.71 | 79.64 | 82.29 | 83.06 | 86.71 | 85.81 | 87.86 |
| Model | K-shot | |||||||
| 5-shot | 10-shot | 15-shot | 20-shot | 25-shot | 30-shot | 35-shot | 40-shot | |
| ATIS | 22.21 | 41.21 | 52.17 | 66.23 | 70.14 | 78.06 | 78.92 | 82.21 |
| SNIPS | 1.8 | 41.98 | 50.80 | 69.63 | 72.04 | 75.46 | 77.93 | 80.38 |
| Model | We | Mu | PI | Bo | Se | Re | Cr | Avg. |
| TransferBERT | 59.41 | 42.00 | 46.07 | 20.74 | 28.20 | 67.75 | 58.61 | 46.11 |
| SimBERT | 53.46 | 54.13 | 42.81 | 75.54 | 57.10 | 55.30 | 32.38 | 52.96 |
| ProtoToken | 67.82 | 55.99 | 46.02 | 72.17 | 73.59 | 60.18 | 66.89 | 63.24 |
| Coach (Liu et al., 2020) | 73.56 | 45.85 | 47.23 | 61.61 | 65.82 | 69.99 | 57.28 | 60.19 |
| TapNet (Hou et al., 2020a) | 53.03 | 49.80 | 54.90 | 83.36 | 63.07 | 59.84 | 67.02 | 61.57 |
| TapNet+CDT (Hou et al., 2020a) | 66.48 | 66.36 | 68.23 | 85.76 | 73.60 | 64.20 | 68.47 | 70.44 |
| L-WPZ+CDT (Hou et al., 2020a) | 74.68 | 56.73 | 52.20 | 78.79 | 80.61 | 69.59 | 67.46 | 68.58 |
| L-TapNet+CDT (Hou et al., 2020a) | 71.64 | 67.16 | 75.88 | 84.38 | 82.58 | 70.05 | 73.41 | 75.01 |
| Retriever (Yu et al., 2021) | 82.95 | 61.74 | 71.75 | 81.65 | 73.10 | 79.54 | 51.35 | 71.72 |
| Ours | 79.70 | 78.09 | 76.64 | 90.04 | 82.87 | 79.91 | 63.88 | 78.73 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| AddTo Playlist | 97.00 | 78.23 | 86.61 | 99.10 | 88.71 | 93.62 | 124 |
| PlayMusic | 64.96 | 88.37 | 74.88 | 72.32 | 94.19 | 81.82 | 86 |
| BookRestaurant | 96.81 | 98.91 | 97.85 | 95.83 | 1.00 | 97.87 | 92 |
| SearchCreativeWork | 71.03 | 71.03 | 71.03 | 73.17 | 84.11 | 78.26 | 107 |
| RateBook | 98.75 | 98.75 | 98.75 | 98.59 | 87.50 | 92.72 | 80 |
| GetWeather | 97.09 | 96.15 | 96.62 | 96.19 | 97.12 | 96.65 | 104 |
| SearchScreeningEvent | 85.86 | 79.44 | 82.52 | 93.90 | 71.96 | 81.48 | 107 |
| Intent | 6-shot Average Sample | 6-shot Dynamic Sample | Samples | ||||
| Precision | Recall | F1 | Precision | Recall | F1 | ||
| atis_flight | 99.07 | 84.34 | 91.11 | 98.11 | 90.51 | 94.16 | 632 |
| atis-ground_fare | 60.00 | 85.71 | 70.59 | 50.00 | 71.43 | 58.82 | 7 |
| atis_aircraft | 25.00 | 1.00 | 40.00 | 25.00 | 1.00 | 40.00 | 1 |
| atisquantity | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 3 |
| atis_distance | 81.82 | 90.00 | 85.71 | 1.0000 | 1.0000 | 1.0000 | 10 |
| atis_city | 19.35 | 100.00 | 32.43 | 66.67 | 66.67 | 66.67 | 6 |
| atis_capacity | 100.00 | 85.71 | 92.31 | 100.00 | 100.00 | 100.00 | 21 |
| atis_Aircraft | 56.25 | 100.00 | 72.00 | 87.50 | 77.78 | 82.35 | 9 |
| atis_airport no | 88.89 | 100.00 | 94.12 | 100.00 | 100.00 | 100.00 | 8 |
| atis_cheapest | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_airline | 55.88 | 100.00 | 71.70 | 61.67 | 97.37 | 75.51 | 38 |
| atis_abbreviation | 100.00 | 78.79 | 88.14 | 91.18 | 93.94 | 92.54 | 33 |
| atis_restriction | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 |
| atis_Airport | 90.00 | 100.00 | 94.74 | 94.74 | 100.00 | 97.30 | 18 |
| atismeal | 42.86 | 100.00 | 60.00 | 31.25 | 83.33 | 45.45 | 6 |
| atis_airfare | 74.00 | 77.08 | 75.51 | 83.67 | 85.42 | 84.54 | 48 |
| atisGround服务能力 | 100.00 | 86.11 | 92.54 | 100.00 | 86.11 | 92.54 | 36 |
| atis_day_name | 50.00 | 100.00 | 66.67 | 66.67 | 100.00 | 80.00 | 2 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | 2-shot | 4-shot | 6-shot | 8-shot | 10-shot | |
| RoBERTa-ID(AS) | 4.55 | 4.67 | 71.75 | 72.83 | 85.42 | 20.14 | 75.57 | 75.71 | 82.02 | 83.71 |
| RoBERTa-ID(DS) | 4.55 | 7.57 | 81.32 | 73.87 | 86.64 | 20.14 | 77.00 | 77.86 | 84.43 | 85.57 |
| Model | ATIS K-shot | Snips K-shot | ||||||||
| 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | 5-shot | 10-shot | 20-shot | 30-shot | 40-shot | |
| RoBERTa-SF(BIO+AS) | 29.08 | 59.89 | 77.03 | 83.46 | 87.39 | 14.35 | 42.33 | 65.42 | 73.63 | 77.28 |
| RoBERTa-SF(BIO+DS) | 29.08 | 67.96 | 80.16 | 87.26 | 90.35 | 14.35 | 42.48 | 67.52 | 76.28 | 77.32 |
| RoBERTa-SF(ET+AS) | 29.11 | 65.00 | 80.59 | 88.61 | 88.65 | 22.42 | 49.58 | 67.59 | 76.61 | 77.61 |
| RoBERTa-SF(ET+DS) | 29.11 | 77.03 | 85.94 | 86.09 | 90.48 | 22.42 | 43.51 | 67.63 | 74.67 | 78.12 |
| RoBERTa-SF(ET+FL+AS) | 31.51 | 65.24 | 81.07 | 86.78 | 89.53 | 24.10 | 54.39 | 68.25 | 75.25 | 78.16 |
| RoBERTa-SF(ET+FT+DS) | 31.51 | 77.09 | 83.77 | 88.24 | 91.02 | 24.10 | 45.99 | 70.97 | 78.12 | 78.42 |
| Task | Template | Verbalizer |
| Main | It was <mask>, <input_sentence> | [bad, good, neutral] |
| Calibration | Sentence: <input_sentence>, The predicted sentiment is: <prediction>. Is the prediction True or False? It's <mask>. | [False, True] |
| Task | ID | OOD-1 | OOD-2 |
| Sentiment Analysis | Amazon | SST-5 | SemEval |
| Hate-speech Detection | Civil | Hate Speech | Implicit |
| Natural Language Inference | MNLI | HANS | ANLI |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 89.44 (0.08) | 73.96 (0.35) | 61.95 (0.16) | 28.94 (0.82) | 73.70 (1.18) | 28.65 (0.57) |
| w/o Sample | 84.12 (0.49) | 64.81 (0.87) | 59.47 (1.04) | 24.46 (0.24) | 68.15 (1.42) | 22.46 (2.67) |
| 70.06 (1.67) | 3.93 (1.07) | 57.22 (3.31) | 1.16 (0.55) | 59.19 (0.94) | 1.42 (0.37) | |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 86.98 (5.00) | 59.30 (0.10) | 60.44 (0.39) | 11.18 (2.28) | 64.16 (1.98) | 21.05 (2.95) |
| w/o Sample | 83.89 (2.51) | 52.74 (3.83) | 61.83 (4.52) | 14.90 (8.28) | 63.93 (0.32) | 20.96 (0.22) |
| 5.18 (0.20) | -3.5 (2.39) | 55.73 (0.79) | 4.35 (0.27) | 37.02 (0.21) | -1.02 (0.71) | |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 78.80 (0.89) | 42.64 (0.58) | 67.90 (3.16) | 24.11 (6.55) | 44.31 (0.30) | -6.95 (0.99) |
| w/o Sample | 79.24 (0.31) | 44.94 (0.42) | 60.22 (2.11) | 14.11 (4.32) | 43.99 (0.23) | -8.05 (1.33) |
| 39.19 (0.89) | 0 (0.11) | 39.12 (0.34) | 0 (0.09) | 51.28 (0.24) | 0 (0.01) |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 85.80 (0.45) | 15.18 (0.33) | 79.14 (0.83) | 16.56 (0.78) | 71.68 (0.93) | 12.68 (0.59) |
| TS | 85.80 (0.45) | 16.85 (0.59) | 79.14 (0.83) | 10.63 (0.34) | 71.68 (0.93) | 8.07 (0.64) |
| LS | 81.93 (2.77) | 13.81 (0.83) | 76.81 (1.41) | 13.45 (0.99) | 70.52 (1.26) | 10.81 (0.90) |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 90.33 (0.99) | 14.23 (0.09) | 62.80 (0.55) | 2.74 (0.64) | 65.99 (1.22) | 3.63 (1.19) |
| TS | 90.33 (0.99) | 22.71 (0.46) | 62.80 (0.55) | 5.60 (0.27) | 65.99 (1.22) | 7.25 (0.23) |
| LS | 91.15 (0.11) | 11.59 (0.32) | 62.17 (0.22) | 2.08 (0.33) | 63.92 (0.13) | 3.30 (0.34) |
| LM-TOAST | 92.01 (0.14) | 51.38 (0.24) | 65.55 (1.76) | 12.57 (2.47) | 65.99 (0.12) | 17.29 (0.30) |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 82.08 (0.43) | 12.25 (0.19) | 51.14 (1.68) | 2.59 (0.52) | 44.14 (0.31) | -2.31 (0.06) |
| TS | 82.08 (0.43) | 18.30 (0.32) | 51.14 (1.68) | -1.87 (0.70) | 44.14 (0.31) | -2.89 (0.13) |
| LS | 80.51 (0.14) | 10.12 (0.39) | 46.52 (2.32) | 0.53 (0.80) | 43.07 (0.32) | -2.91 (0.92) |
| LM-TOAST | 82.74 (0.49) | 33.53 (1.40) | 60.60 (4.34) | 11.19 (3.81) | 43.97 (0.62) | -6.69 (0.70) |
| Dataset | PWWS | TextBugger | BERT-Attack | HotFlip | ROCKET | |||||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 74.91 | 15.66 | 79.11 | 17.68 | 81.13 | 21.57 | 83.54 | 19.01 | 77.23 | 14.24 |
| TS | 74.91 | 10.79 | 79.11 | 12.53 | 81.13 | 13.79 | 83.54 | 14.19 | 77.23 | 11.33 |
| LS | 71.10 | 10.28 | 74.53 | 11.18 | 76.51 | 12.31 | 78.69 | 12.15 | 64.24 | 9.26 |
| LM-TOAST | 84.64 | 42.66 | 86.01 | 45.75 | 84.59 | 45.08 | 87.96 | 49.56 | 85.83 | 42.78 |
| Dataset | Amazon | Civil | MNLI |
| Method | AUROC | AUROC | AUROC |
| Vanilla | 88.24 | 86.55 | 81.96 |
| TS | 88.37 | 86.60 | 82.21 |
| LS | 88.34 | 86.62 | 82.30 |
| LM-TOAST | 89.50 | 88.54 | 83.93 |
| Task | Dataset | #Class | Average Length | #Train | #Dev | #Test |
| Sentiment Analysis | Amazon | 3 | 77.86 | 24000 | 78741 | 91606 |
| SST-5 | 3 | 18.75 | - | - | 1067 | |
| SemEval | 3 | 19.61 | - | - | 6000 | |
| Hate Speech Detection | Civil | 2 | 52.86 | 48000 | 12000 | 60000 |
| Hate Speech | 2 | 21.55 | - | - | 478 | |
| Implicit | 2 | 17.34 | - | - | 21479 | |
| Natural Language Inference | MNLI | 3 | 19.36/10.06 | 373067 | 19635 | 9815 |
| HANS | 2 | 9.15/5.61 | - | - | 30000 | |
| ANLI | 3 | 54.40/10.34 | - | - | 3200 |
| Amazon | Dataset | Amazon | SST-5 | SemEval | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) | |
| w/o Cross-annotation | 87.35 (0.43) | 27.72 (1.28) | 66.96 (0.58) | 8.20 (1.84) | 72.41 (1.27) | 8.11 (1.84) | |
| w/o Down-sampling | 81.50 (3.32) | 19.45 (2.67) | 77.03 (1.36) | 20.46 (1.32) | 71.69 (0.13) | 15.90 (0.70) | |
| w/o Augmentation | 84.37 (2.00) | 37.88 (2.89) | 74.14 (0.86) | 19.73 (0.26) | 74.09 (2.24) | 17.76 (0.35) | |
| w/o Decay α | 80.56 (4.13) | 17.57 (4.13) | 77.92 (0.21) | 29.34 (0.99) | 75.18 (0.23) | 22.10 (1.16) |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 14.20 | 89.82 | 20.86 | 21.37 | 28.32 | 18.40 |
| TS | 14.27 | 90.22 | 20.86 | 23.24 | 28.11 | 18.82 |
| LS | 18.07 | 89.28 | 23.19 | 2.81 | 29.48 | 13.35 |
| LM-TOAST | 12.56 | 91.27 | 20.68 | 23.52 | 26.83 | 24.25 |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.04 |
| TS | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.16 |
| LS | 8.85 | 85.17 | 37.83 | 1.05 | 36.08 | 0 |
| LM-TOAST | 7.99 | 88.03 | 34.45 | 18.83 | 34.01 | 0.16 |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 17.40 | 69.63 | 47.04 | 62.45 | 55.89 | - |
| TS | 17.40 | 69.75 | 47.04 | 0 | 55.89 | - |
| LS | 19.49 | 68.18 | 53.48 | 0 | 56.93 | - |
| LM-TOAST | 17.26 | 71.74 | 39.40 | 79.46 | 56.03 | - |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] |
| SST-5 | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| SemEval | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| Hate Speech Detection | Civil | It was <mask>. {"placeholder": "text a"} | [benign, toxic] |
| Hate Speech | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Implicit | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Natural Language Inference | MNLI | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| HANS | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] | |
| ANLI | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | [False, True] |
| SST-5 | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| SemEval | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Hate Speech Detection | Civil | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | |
| Hate Speech | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Implicit | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Natural Language Inference | MNLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | |
| HANS | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | ||
| ANLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . |
| Task | Template | Verbalizer |
| Main | It was <mask>, <input_sentence> | [bad, good, neutral] |
| Calibration | Sentence: <input_sentence>, The predicted sentiment is: <prediction>. Is the prediction True or False? It's <mask>. | [False, True] |
| Task | ID | OOD-1 | OOD-2 |
| Sentiment Analysis | Amazon | SST-5 | SemEval |
| Hate-speech Detection | Civil | Hate Speech | Implicit |
| Natural Language Inference | MNLI | HANS | ANLI |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 89.44 (0.08) | 73.96 (0.35) | 61.95 (0.16) | 28.94 (0.82) | 73.70 (1.18) | 28.65 (0.57) |
| w/o Sample | 84.12 (0.49) | 64.81 (0.87) | 59.47 (1.04) | 24.46 (0.24) | 68.15 (1.42) | 22.46 (2.67) |
| 70.06 (1.67) | 3.93 (1.07) | 57.22 (3.31) | 1.16 (0.55) | 59.19 (0.94) | 1.42 (0.37) | |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 86.98 (5.00) | 59.30 (0.10) | 60.44 (0.39) | 11.18 (2.28) | 64.16 (1.98) | 21.05 (2.95) |
| w/o Sample | 83.89 (2.51) | 52.74 (3.83) | 61.83 (4.52) | 14.90 (8.28) | 63.93 (0.32) | 20.96 (0.22) |
| 5.18 (0.20) | -3.5 (2.39) | 55.73 (0.79) | 4.35 (0.27) | 37.02 (0.21) | -1.02 (0.71) | |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 78.80 (0.89) | 42.64 (0.58) | 67.90 (3.16) | 24.11 (6.55) | 44.31 (0.30) | -6.95 (0.99) |
| w/o Sample | 79.24 (0.31) | 44.94 (0.42) | 60.22 (2.11) | 14.11 (4.32) | 43.99 (0.23) | -8.05 (1.33) |
| 39.19 (0.89) | 0 (0.11) | 39.12 (0.34) | 0 (0.09) | 51.28 (0.24) | 0 (0.01) |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 85.80 (0.45) | 15.18 (0.33) | 79.14 (0.83) | 16.56 (0.78) | 71.68 (0.93) | 12.68 (0.59) |
| TS | 85.80 (0.45) | 16.85 (0.59) | 79.14 (0.83) | 10.63 (0.34) | 71.68 (0.93) | 8.07 (0.64) |
| LS | 81.93 (2.77) | 13.81 (0.83) | 76.81 (1.41) | 13.45 (0.99) | 70.52 (1.26) | 10.81 (0.90) |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 90.33 (0.99) | 14.23 (0.09) | 62.80 (0.55) | 2.74 (0.64) | 65.99 (1.22) | 3.63 (1.19) |
| TS | 90.33 (0.99) | 22.71 (0.46) | 62.80 (0.55) | 5.60 (0.27) | 65.99 (1.22) | 7.25 (0.23) |
| LS | 91.15 (0.11) | 11.59 (0.32) | 62.17 (0.22) | 2.08 (0.33) | 63.92 (0.13) | 3.30 (0.34) |
| LM-TOAST | 92.01 (0.14) | 51.38 (0.24) | 65.55 (1.76) | 12.57 (2.47) | 65.99 (0.12) | 17.29 (0.30) |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 82.08 (0.43) | 12.25 (0.19) | 51.14 (1.68) | 2.59 (0.52) | 44.14 (0.31) | -2.31 (0.06) |
| TS | 82.08 (0.43) | 18.30 (0.32) | 51.14 (1.68) | -1.87 (0.70) | 44.14 (0.31) | -2.89 (0.13) |
| LS | 80.51 (0.14) | 10.12 (0.39) | 46.52 (2.32) | 0.53 (0.80) | 43.07 (0.32) | -2.91 (0.92) |
| LM-TOAST | 82.74 (0.49) | 33.53 (1.40) | 60.60 (4.34) | 11.19 (3.81) | 43.97 (0.62) | -6.69 (0.70) |
| Dataset | PWWS | TextBugger | BERT-Attack | HotFlip | ROCKET | |||||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 74.91 | 15.66 | 79.11 | 17.68 | 81.13 | 21.57 | 83.54 | 19.01 | 77.23 | 14.24 |
| TS | 74.91 | 10.79 | 79.11 | 12.53 | 81.13 | 13.79 | 83.54 | 14.19 | 77.23 | 11.33 |
| LS | 71.10 | 10.28 | 74.53 | 11.18 | 76.51 | 12.31 | 78.69 | 12.15 | 64.24 | 9.26 |
| LM-TOAST | 84.64 | 42.66 | 86.01 | 45.75 | 84.59 | 45.08 | 87.96 | 49.56 | 85.83 | 42.78 |
| Dataset | Amazon | Civil | MNLI |
| Method | AUROC | AUROC | AUROC |
| Vanilla | 88.24 | 86.55 | 81.96 |
| TS | 88.37 | 86.60 | 82.21 |
| LS | 88.34 | 86.62 | 82.30 |
| LM-TOAST | 89.50 | 88.54 | 83.93 |
| Task | Dataset | #Class | Average Length | #Train | #Dev | #Test |
| Sentiment Analysis | Amazon | 3 | 77.86 | 24000 | 78741 | 91606 |
| SST-5 | 3 | 18.75 | - | - | 1067 | |
| SemEval | 3 | 19.61 | - | - | 6000 | |
| Hate Speech Detection | Civil | 2 | 52.86 | 48000 | 12000 | 60000 |
| Hate Speech | 2 | 21.55 | - | - | 478 | |
| Implicit | 2 | 17.34 | - | - | 21479 | |
| Natural Language Inference | MNLI | 3 | 19.36/10.06 | 373067 | 19635 | 9815 |
| HANS | 2 | 9.15/5.61 | - | - | 30000 | |
| ANLI | 3 | 54.40/10.34 | - | - | 3200 |
| Amazon | Dataset | Amazon | SST-5 | SemEval | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) | |
| w/o Cross-annotation | 87.35 (0.43) | 27.72 (1.28) | 66.96 (0.58) | 8.20 (1.84) | 72.41 (1.27) | 8.11 (1.84) | |
| w/o Down-sampling | 81.50 (3.32) | 19.45 (2.67) | 77.03 (1.36) | 20.46 (1.32) | 71.69 (0.13) | 15.90 (0.70) | |
| w/o Augmentation | 84.37 (2.00) | 37.88 (2.89) | 74.14 (0.86) | 19.73 (0.26) | 74.09 (2.24) | 17.76 (0.35) | |
| w/o Decay α | 80.56 (4.13) | 17.57 (4.13) | 77.92 (0.21) | 29.34 (0.99) | 75.18 (0.23) | 22.10 (1.16) |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 14.20 | 89.82 | 20.86 | 21.37 | 28.32 | 18.40 |
| TS | 14.27 | 90.22 | 20.86 | 23.24 | 28.11 | 18.82 |
| LS | 18.07 | 89.28 | 23.19 | 2.81 | 29.48 | 13.35 |
| LM-TOAST | 12.56 | 91.27 | 20.68 | 23.52 | 26.83 | 24.25 |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.04 |
| TS | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.16 |
| LS | 8.85 | 85.17 | 37.83 | 1.05 | 36.08 | 0 |
| LM-TOAST | 7.99 | 88.03 | 34.45 | 18.83 | 34.01 | 0.16 |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 17.40 | 69.63 | 47.04 | 62.45 | 55.89 | - |
| TS | 17.40 | 69.75 | 47.04 | 0 | 55.89 | - |
| LS | 19.49 | 68.18 | 53.48 | 0 | 56.93 | - |
| LM-TOAST | 17.26 | 71.74 | 39.40 | 79.46 | 56.03 | - |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] |
| SST-5 | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| SemEval | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| Hate Speech Detection | Civil | It was <mask>. {"placeholder": "text a"} | [benign, toxic] |
| Hate Speech | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Implicit | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Natural Language Inference | MNLI | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| HANS | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] | |
| ANLI | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | [False, True] |
| SST-5 | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| SemEval | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Hate Speech Detection | Civil | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | |
| Hate Speech | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Implicit | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Natural Language Inference | MNLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | |
| HANS | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | ||
| ANLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . |
| Task | Template | Verbalizer |
| Main | It was <mask>, <input_sentence> | [bad, good, neutral] |
| Calibration | Sentence: <input_sentence>, The predicted sentiment is: <prediction>. Is the prediction True or False? It's <mask>. | [False, True] |
| Task | ID | OOD-1 | OOD-2 |
| Sentiment Analysis | Amazon | SST-5 | SemEval |
| Hate-speech Detection | Civil | Hate Speech | Implicit |
| Natural Language Inference | MNLI | HANS | ANLI |
| Amazon | Dataset | Amazon | SST-5 | SemEval | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| All Features w/o Prediction | 89.44 (0.08) | 73.96 (0.35) | 61.95 (0.16) | 28.94 (0.82) | 73.70 (1.18) | 28.65 (0.57) | |
| w/o Sample | 84.12 (0.49) | 64.81 (0.87) | 59.47 (1.04) | 24.46 (0.24) | 68.15 (1.42) | 22.46 (2.67) | |
| 70.06 (1.67) | 3.93 (1.07) | 57.22 (3.31) | 1.16 (0.55) | 59.19 (0.94) | 1.42 (0.37) | ||
| Civil | Dataset | Civil | Hate Speech | Implicit | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| All Features w/o Prediction | 86.98 (5.00) | 59.30 (0.10) | 60.44 (0.39) | 11.18 (2.28) | 64.16 (1.98) | 21.05 (2.95) | |
| w/o Sample | 83.89 (2.51) | 52.74 (3.83) | 61.83 (4.52) | 14.90 (8.28) | 63.93 (0.32) | 20.96 (0.22) | |
| 5.18 (0.20) | -3.5 (2.39) | 55.73 (0.79) | 4.35 (0.27) | 37.02 (0.21) | -1.02 (0.71) | ||
| MNLI | Dataset | MNLI | HANS | ANLI | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| All Features w/o Prediction | 78.80 (0.89) | 42.64 (0.58) | 67.90 (3.16) | 24.11 (6.55) | 44.31 (0.30) | -6.95 (0.99) | |
| w/o Sample | 79.24 (0.31) | 44.94 (0.42) | 60.22 (2.11) | 14.11 (4.32) | 43.99 (0.23) | -8.05 (1.33) | |
| 39.19 (0.89) | 0 (0.11) | 39.12 (0.34) | 0 (0.09) | 51.28 (0.24) | 0 (0.01) | ||
| Amazon | Dataset | Amazon | SST-5 | SemEval | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| Vanilla | 85.80 (0.45) | 15.18 (0.33) | 79.14 (0.83) | 16.56 (0.78) | 71.68 (0.93) | 12.68 (0.59) | |
| TS | 85.80 (0.45) | 16.85 (0.59) | 79.14 (0.83) | 10.63 (0.34) | 71.68 (0.93) | 8.07 (0.64) | |
| LS | 81.93 (2.77) | 13.81 (0.83) | 76.81 (1.41) | 13.45 (0.99) | 70.52 (1.26) | 10.81 (0.90) | |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) | |
| Civil | Dataset | Civil | Hate Speech | Implicit | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| Vanilla | 90.33 (0.99) | 14.23 (0.09) | 62.80 (0.55) | 2.74 (0.64) | 65.99 (1.22) | 3.63 (1.19) | |
| TS | 90.33 (0.99) | 22.71 (0.46) | 62.80 (0.55) | 5.60 (0.27) | 65.99 (1.22) | 7.25 (0.23) | |
| LS | 91.15 (0.11) | 11.59 (0.32) | 62.17 (0.22) | 2.08 (0.33) | 63.92 (0.13) | 3.30 (0.34) | |
| LM-TOAST | 92.01 (0.14) | 51.38 (0.24) | 65.55 (1.76) | 12.57 (2.47) | 65.99 (0.12) | 17.29 (0.30) | |
| MNLI | Dataset | MNLI | HANS | ANLI | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| Vanilla | 82.08 (0.43) | 12.25 (0.19) | 51.14 (1.68) | 2.59 (0.52) | 44.14 (0.31) | -2.31 (0.06) | |
| TS | 82.08 (0.43) | 18.30 (0.32) | 51.14 (1.68) | -1.87 (0.70) | 44.14 (0.31) | -2.89 (0.13) | |
| LS | 80.51 (0.14) | 10.12 (0.39) | 46.52 (2.32) | 0.53 (0.80) | 43.07 (0.32) | -2.91 (0.92) | |
| LM-TOAST | 82.74 (0.49) | 33.53 (1.40) | 60.60 (4.34) | 11.19 (3.81) | 43.97 (0.62) | -6.69 (0.70) | |
| Dataset | PWWS | TextBugger | BERT-Attack | HotFlip | ROCKET | |||||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 74.91 | 15.66 | 79.11 | 17.68 | 81.13 | 21.57 | 83.54 | 19.01 | 77.23 | 14.24 |
| TS | 74.91 | 10.79 | 79.11 | 12.53 | 81.13 | 13.79 | 83.54 | 14.19 | 77.23 | 11.33 |
| LS | 71.10 | 10.28 | 74.53 | 11.18 | 76.51 | 12.31 | 78.69 | 12.15 | 64.24 | 9.26 |
| LM-TOAST | 84.64 | 42.66 | 86.01 | 45.75 | 84.59 | 45.08 | 87.96 | 49.56 | 85.83 | 42.78 |
| Dataset | Amazon | Civil | MNLI |
| Method | AUROC | AUROC | AUROC |
| Vanilla | 88.24 | 86.55 | 81.96 |
| TS | 88.37 | 86.60 | 82.21 |
| LS | 88.34 | 86.62 | 82.30 |
| LM-TOAST | 89.50 | 88.54 | 83.93 |
| Task | Dataset | #Class | Average Length | #Train | #Dev | #Test |
| Sentiment Analysis | Amazon | 3 | 77.86 | 24000 | 78741 | 91606 |
| SST-5 | 3 | 18.75 | - | - | 1067 | |
| SemEval | 3 | 19.61 | - | - | 6000 | |
| Hate Speech Detection | Civil | 2 | 52.86 | 48000 | 12000 | 60000 |
| Hate Speech | 2 | 21.55 | - | - | 478 | |
| Implicit | 2 | 17.34 | - | - | 21479 | |
| Natural Language Inference | MNLI | 3 | 19.36/10.06 | 373067 | 19635 | 9815 |
| HANS | 2 | 9.15/5.61 | - | - | 30000 | |
| ANLI | 3 | 54.40/10.34 | - | - | 3200 |
| Amazon | Dataset | Amazon | SST-5 | SemEval | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) | |
| w/o Cross-annotation | 87.35 (0.43) | 27.72 (1.28) | 66.96 (0.58) | 8.20 (1.84) | 72.41 (1.27) | 8.11 (1.84) | |
| w/o Down-sampling | 81.50 (3.32) | 19.45 (2.67) | 77.03 (1.36) | 20.46 (1.32) | 71.69 (0.13) | 15.90 (0.70) | |
| w/o Augmentation | 84.37 (2.00) | 37.88 (2.89) | 74.14 (0.86) | 19.73 (0.26) | 74.09 (2.24) | 17.76 (0.35) | |
| w/o Decay α | 80.56 (4.13) | 17.57 (4.13) | 77.92 (0.21) | 29.34 (0.99) | 75.18 (0.23) | 22.10 (1.16) | |
| Amazon | Dataset | Amazon | SST-5 | SemEval | |||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | |
| Vanilla | 14.20 | 89.82 | 20.86 | 21.37 | 28.32 | 18.40 | |
| TS | 14.27 | 90.22 | 20.86 | 23.24 | 28.11 | 18.82 | |
| LS | 18.07 | 89.28 | 23.19 | 2.81 | 29.48 | 13.35 | |
| LM-TOAST | 12.56 | 91.27 | 20.68 | 23.52 | 26.83 | 24.25 | |
| Civil | Dataset | Civil | Hate Speech | Implicit | |||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | |
| Vanilla | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.04 | |
| TS | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.16 | |
| LS | 8.85 | 85.17 | 37.83 | 1.05 | 36.08 | 0 | |
| LM-TOAST | 7.99 | 88.03 | 34.45 | 18.83 | 34.01 | 0.16 | |
| MNLI | Dataset | MNLI | HANS | ANLI | |||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | |
| Vanilla | 17.40 | 69.63 | 47.04 | 62.45 | 55.89 | - | |
| TS | 17.40 | 69.75 | 47.04 | 0 | 55.89 | - | |
| LS | 19.49 | 68.18 | 53.48 | 0 | 56.93 | - | |
| LM-TOAST | 17.26 | 71.74 | 39.40 | 79.46 | 56.03 | - | |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] |
| SST-5 | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| SemEval | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| Hate Speech Detection | Civil | It was <mask>. {"placeholder": "text a"} | [benign, toxic] |
| Hate Speech | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Implicit | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Natural Language Inference | MNLI | Given the two sentences: +(1) {"placeholder": "text a)}. +(2) {"placeholder": "text b)}. +Does the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| HANS | Given the two sentences: +(1) {"placeholder": "text a)}. +(2) {"placeholder": "text b)}. +Does the first sentence entails the second ? <mask> | [No, Yes, Maybe] | |
| ANLI | Given the two sentences: +(1) {"placeholder": "text a)}. +(2) {"placeholder": "text b)}. +Does the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | [False, True] |
| SST-5 | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| SemEval | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Hate Speech Detection | Civil | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | |
| Hate Speech | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Implicit | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Natural Language Inference | MNLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | |
| HANS | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | ||
| ANLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . |
| Task | Template | Verbalizer |
| Main | It was <mask>, <input_sentence> | [bad, good, neutral] |
| Calibration | Sentence: <input_sentence>, The predicted sentiment is: <prediction>. Is the prediction True or False? It's <mask>. | [False, True] |
| Task | ID | OOD-1 | OOD-2 |
| Sentiment Analysis | Amazon | SST-5 | SemEval |
| Hate-speech Detection | Civil | Hate Speech | Implicit |
| Natural Language Inference | MNLI | HANS | ANLI |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 89.44 (0.08) | 73.96 (0.35) | 61.95 (0.16) | 28.94 (0.82) | 73.70 (1.18) | 28.65 (0.57) |
| w/o Sample | 84.12 (0.49) | 64.81 (0.87) | 59.47 (1.04) | 24.46 (0.24) | 68.15 (1.42) | 22.46 (2.67) |
| 70.06 (1.67) | 3.93 (1.07) | 57.22 (3.31) | 1.16 (0.55) | 59.19 (0.94) | 1.42 (0.37) | |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 86.98 (5.00) | 59.30 (0.10) | 60.44 (0.39) | 11.18 (2.28) | 64.16 (1.98) | 21.05 (2.95) |
| w/o Sample | 83.89 (2.51) | 52.74 (3.83) | 61.83 (4.52) | 14.90 (8.28) | 63.93 (0.32) | 20.96 (0.22) |
| 5.18 (0.20) | -3.5 (2.39) | 55.73 (0.79) | 4.35 (0.27) | 37.02 (0.21) | -1.02 (0.71) | |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| All Features w/o Prediction | 78.80 (0.89) | 42.64 (0.58) | 67.90 (3.16) | 24.11 (6.55) | 44.31 (0.30) | -6.95 (0.99) |
| w/o Sample | 79.24 (0.31) | 44.94 (0.42) | 60.22 (2.11) | 14.11 (4.32) | 43.99 (0.23) | -8.05 (1.33) |
| 39.19 (0.89) | 0 (0.11) | 39.12 (0.34) | 0 (0.09) | 51.28 (0.24) | 0 (0.01) |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 85.80 (0.45) | 15.18 (0.33) | 79.14 (0.83) | 16.56 (0.78) | 71.68 (0.93) | 12.68 (0.59) |
| TS | 85.80 (0.45) | 16.85 (0.59) | 79.14 (0.83) | 10.63 (0.34) | 71.68 (0.93) | 8.07 (0.64) |
| LS | 81.93 (2.77) | 13.81 (0.83) | 76.81 (1.41) | 13.45 (0.99) | 70.52 (1.26) | 10.81 (0.90) |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 90.33 (0.99) | 14.23 (0.09) | 62.80 (0.55) | 2.74 (0.64) | 65.99 (1.22) | 3.63 (1.19) |
| TS | 90.33 (0.99) | 22.71 (0.46) | 62.80 (0.55) | 5.60 (0.27) | 65.99 (1.22) | 7.25 (0.23) |
| LS | 91.15 (0.11) | 11.59 (0.32) | 62.17 (0.22) | 2.08 (0.33) | 63.92 (0.13) | 3.30 (0.34) |
| LM-TOAST | 92.01 (0.14) | 51.38 (0.24) | 65.55 (1.76) | 12.57 (2.47) | 65.99 (0.12) | 17.29 (0.30) |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 82.08 (0.43) | 12.25 (0.19) | 51.14 (1.68) | 2.59 (0.52) | 44.14 (0.31) | -2.31 (0.06) |
| TS | 82.08 (0.43) | 18.30 (0.32) | 51.14 (1.68) | -1.87 (0.70) | 44.14 (0.31) | -2.89 (0.13) |
| LS | 80.51 (0.14) | 10.12 (0.39) | 46.52 (2.32) | 0.53 (0.80) | 43.07 (0.32) | -2.91 (0.92) |
| LM-TOAST | 82.74 (0.49) | 33.53 (1.40) | 60.60 (4.34) | 11.19 (3.81) | 43.97 (0.62) | -6.69 (0.70) |
| Dataset | PWWS | TextBugger | BERT-Attack | HotFlip | ROCKET | |||||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf |
| Vanilla | 74.91 | 15.66 | 79.11 | 17.68 | 81.13 | 21.57 | 83.54 | 19.01 | 77.23 | 14.24 |
| TS | 74.91 | 10.79 | 79.11 | 12.53 | 81.13 | 13.79 | 83.54 | 14.19 | 77.23 | 11.33 |
| LS | 71.10 | 10.28 | 74.53 | 11.18 | 76.51 | 12.31 | 78.69 | 12.15 | 64.24 | 9.26 |
| LM-TOAST | 84.64 | 42.66 | 86.01 | 45.75 | 84.59 | 45.08 | 87.96 | 49.56 | 85.83 | 42.78 |
| Dataset | Amazon | Civil | MNLI |
| Method | AUROC | AUROC | AUROC |
| Vanilla | 88.24 | 86.55 | 81.96 |
| TS | 88.37 | 86.60 | 82.21 |
| LS | 88.34 | 86.62 | 82.30 |
| LM-TOAST | 89.50 | 88.54 | 83.93 |
| Task | Dataset | #Class | Average Length | #Train | #Dev | #Test |
| Sentiment Analysis | Amazon | 3 | 77.86 | 24000 | 78741 | 91606 |
| SST-5 | 3 | 18.75 | - | - | 1067 | |
| SemEval | 3 | 19.61 | - | - | 6000 | |
| Hate Speech Detection | Civil | 2 | 52.86 | 48000 | 12000 | 60000 |
| Hate Speech | 2 | 21.55 | - | - | 478 | |
| Implicit | 2 | 17.34 | - | - | 21479 | |
| Natural Language Inference | MNLI | 3 | 19.36/10.06 | 373067 | 19635 | 9815 |
| HANS | 2 | 9.15/5.61 | - | - | 30000 | |
| ANLI | 3 | 54.40/10.34 | - | - | 3200 |
| Amazon | Dataset | Amazon | SST-5 | SemEval | |||
| Method | AUROC | ΔConf | AUROC | ΔConf | AUROC | ΔConf | |
| LM-TOAST | 87.44 (1.12) | 42.46 (3.03) | 79.32 (1.09) | 26.37 (1.04) | 73.17 (2.36) | 20.21 (1.70) | |
| w/o Cross-annotation | 87.35 (0.43) | 27.72 (1.28) | 66.96 (0.58) | 8.20 (1.84) | 72.41 (1.27) | 8.11 (1.84) | |
| w/o Down-sampling | 81.50 (3.32) | 19.45 (2.67) | 77.03 (1.36) | 20.46 (1.32) | 71.69 (0.13) | 15.90 (0.70) | |
| w/o Augmentation | 84.37 (2.00) | 37.88 (2.89) | 74.14 (0.86) | 19.73 (0.26) | 74.09 (2.24) | 17.76 (0.35) | |
| w/o Decay α | 80.56 (4.13) | 17.57 (4.13) | 77.92 (0.21) | 29.34 (0.99) | 75.18 (0.23) | 22.10 (1.16) |
| Amazon | Dataset | Amazon | SST-5 | SemEval | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 14.20 | 89.82 | 20.86 | 21.37 | 28.32 | 18.40 |
| TS | 14.27 | 90.22 | 20.86 | 23.24 | 28.11 | 18.82 |
| LS | 18.07 | 89.28 | 23.19 | 2.81 | 29.48 | 13.35 |
| LM-TOAST | 12.56 | 91.27 | 20.68 | 23.52 | 26.83 | 24.25 |
| Civil | Dataset | Civil | Hate Speech | Implicit | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.04 |
| TS | 9.67 | 84.71 | 37.20 | 10.46 | 34.01 | 0.16 |
| LS | 8.85 | 85.17 | 37.83 | 1.05 | 36.08 | 0 |
| LM-TOAST | 7.99 | 88.03 | 34.45 | 18.83 | 34.01 | 0.16 |
| MNLI | Dataset | MNLI | HANS | ANLI | ||
| Method | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ | AUROC_risk↓ | Cov↑ |
| Vanilla | 17.40 | 69.63 | 47.04 | 62.45 | 55.89 | - |
| TS | 17.40 | 69.75 | 47.04 | 0 | 55.89 | - |
| LS | 19.49 | 68.18 | 53.48 | 0 | 56.93 | - |
| LM-TOAST | 17.26 | 71.74 | 39.40 | 79.46 | 56.03 | - |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] |
| SST-5 | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| SemEval | It was <mask>. {"placeholder": "text a"} | [bad, good, neutral] | |
| Hate Speech Detection | Civil | It was <mask>. {"placeholder": "text a"} | [benign, toxic] |
| Hate Speech | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Implicit | It was <mask>. {"placeholder": "text a"} | [benign, toxic] | |
| Natural Language Inference | MNLI | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| HANS | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] | |
| ANLI | Given the two sentences: \n(1) {"placeholder": "text a)}. \n(2) {"placeholder": "text b)}. \nDoes the first sentence entails the second ? <mask> | [No, Yes, Maybe] |
| Task | Dataset | Template | Verbalizer |
| Sentiment Analysis | Amazon | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | [False, True] |
| SST-5 | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| SemEval | Sentence: {"placeholder": "text a"} The predicted sentiment is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Hate Speech Detection | Civil | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | |
| Hate Speech | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Implicit | Sentence: {"placeholder": "text a"} The predicted toxicity is {"placeholder": "text b"} . Is the prediction True or False? It's {"mask"} . | ||
| Natural Language Inference | MNLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | |
| HANS | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . | ||
| ANLI | Given the two sentences: {"placeholder": "text a"} The predicted relationship between the two sentences is {"placeholder": "text b"} Is the prediction True or False? It's {"mask"} . |
| Data | Authors | Method | N subjects | |
| 1 | 60 Nouns | Mitchell et al. (2008) | fMRI | 9 |
| 2 | Harry Potter Dataset | Wehbe et al. (2014a) | fMRI | 8 |
| 3 | Harry Potter Dataset | Wehbe et al. (2014b) | MEG | 8 |
| 4 | The Moth Radio Hour Dataset | Huth et al. (2016) | fMRI | 7 |
| 5 | Pereira Dataset | Pereira et al. (2018) | fMRI | 16 |
| 6 | Mother of all Unification Studies | Schoffelen et al. (2019) | fMRI, MEG | 204 |
| 7 | Natural Stories Audio Dataset | Zhang et al. (2020) | fMRI | 19 |
| 8 | Narratives | Nastase et al. (2021) | fMRI | 345 |
| 9 | Podcast dataset | Antonello et al. (2021) | fMRI | 5 |
| 10 | The Little Prince Datasets | Li et al. (2022) | fMRI | 112 |
| Authors | Data | E/D | Acc | P/B | Rank | MSE | RSA | Cos Sim | P@k |
| Mitchell et al. (2008) | 1 | E | ✓ | ✓ | ✓ | ||||
| Murphy et al. (2012) | 1 | E | ✓ | ||||||
| Wehbe et al. (2014a,b) | 2,3 | E | ✓ | ✓ | |||||
| Huth et al. (2016) | 4 | E | ✓ | ||||||
| Søgaard (2016) | 2 | E | ✓ | ||||||
| Pereira et al. (2018) | 5 | D | ✓ | ✓ | |||||
| Jain and Huth (2018) | 4 | E | ✓ | ||||||
| Toneva and Wehbe (2019) | 2 | E | ✓ | ||||||
| Gauthier and Levy (2019) | 5 | D | ✓ | ✓ | ✓ | ||||
| Minnema and Herbelot (2019) | 5 | D | ✓ | ✓ | ✓ | ✓ | |||
| Schwartz et al. (2019) | 3 | E | ✓ | ||||||
| Abnar et al. (2019) | 2 | E | ✓ | ||||||
| Zhang et al. (2020) | 7 | E | ✓ | ✓ | |||||
| Schrimpf et al. (2021) | 5 | E | ✓ | ✓ | |||||
| Abdou et al. (2021) | 2,5 | D | ✓ | ||||||
| Reddy and Wehbe (2021) | 2 | E | ✓ | ||||||
| Antonello et al. (2021) | 9 | E | ✓ | ||||||
| Antonello and Huth (2022) | 9 | E | ✓ | ||||||
| Caucheteux et al. (2021, 2022b,a) | 8 | E | ✓ | ||||||
| Caucheteux and King (2022) | 6 | E | ✓ | ||||||
| Zou et al. (2022) | 1,2 | D | ✓ | ||||||
| Oota et al. (2022a,b,c) | 8 | E | ✓ | ✓ | |||||
| Pasquiou et al. (2022) | 10 | E | ✓ | ||||||
| Toneva et al. (2022a) | 2 | E | ✓ | ||||||
| Merlin and Toneva (2022) | 2 | E | ✓ | ||||||
| Pascual et al. (2022) | 5 | D | ✓ | ✓ | |||||
| Aw and Toneva (2023) | 2 | E | ✓ | ✓ |
| Data | Authors | Method | N subjects | |
| 1 | 60 Nouns | Mitchell et al. (2008) | fMRI | 9 |
| 2 | Harry Potter Dataset | Wehbe et al. (2014a) | fMRI | 8 |
| 3 | Harry Potter Dataset | Wehbe et al. (2014b) | MEG | 8 |
| 4 | The Moth Radio Hour Dataset | Huth et al. (2016) | fMRI | 7 |
| 5 | Pereira Dataset | Pereira et al. (2018) | fMRI | 16 |
| 6 | Mother of all Unification Studies | Schoffelen et al. (2019) | fMRI, MEG | 204 |
| 7 | Natural Stories Audio Dataset | Zhang et al. (2020) | fMRI | 19 |
| 8 | Narratives | Nastase et al. (2021) | fMRI | 345 |
| 9 | Podcast dataset | Antonello et al. (2021) | fMRI | 5 |
| 10 | The Little Prince Datasets | Li et al. (2022) | fMRI | 112 |
| Authors | Data | E/D | Acc | P/B | Rank | MSE | RSA | Cos Sim | P@k |
| Mitchell et al. (2008) | 1 | E | ✓ | ✓ | ✓ | ||||
| Murphy et al. (2012) | 1 | E | ✓ | ||||||
| Wehbe et al. (2014a,b) | 2,3 | E | ✓ | ✓ | |||||
| Huth et al. (2016) | 4 | E | ✓ | ||||||
| Søgaard (2016) | 2 | E | ✓ | ||||||
| Pereira et al. (2018) | 5 | D | ✓ | ✓ | |||||
| Jain and Huth (2018) | 4 | E | ✓ | ||||||
| Toneva and Wehbe (2019) | 2 | E | ✓ | ||||||
| Gauthier and Levy (2019) | 5 | D | ✓ | ✓ | ✓ | ||||
| Minnema and Herbelot (2019) | 5 | D | ✓ | ✓ | ✓ | ✓ | |||
| Schwartz et al. (2019) | 3 | E | ✓ | ||||||
| Abnar et al. (2019) | 2 | E | ✓ | ||||||
| Zhang et al. (2020) | 7 | E | ✓ | ✓ | |||||
| Schrimpf et al. (2021) | 5 | E | ✓ | ✓ | |||||
| Abdou et al. (2021) | 2,5 | D | ✓ | ||||||
| Reddy and Wehbe (2021) | 2 | E | ✓ | ||||||
| Antonello et al. (2021) | 9 | E | ✓ | ||||||
| Antonello and Huth (2022) | 9 | E | ✓ | ||||||
| Caucheteux et al. (2021, 2022b,a) | 8 | E | ✓ | ||||||
| Caucheteux and King (2022) | 6 | E | ✓ | ||||||
| Zou et al. (2022) | 1,2 | D | ✓ | ||||||
| Oota et al. (2022a,b,c) | 8 | E | ✓ | ✓ | |||||
| Pasquiou et al. (2022) | 10 | E | ✓ | ||||||
| Toneva et al. (2022a) | 2 | E | ✓ | ||||||
| Merlin and Toneva (2022) | 2 | E | ✓ | ||||||
| Pascual et al. (2022) | 5 | D | ✓ | ✓ | |||||
| Aw and Toneva (2023) | 2 | E | ✓ | ✓ |
| Data | Authors | Method | N subjects | |
| 1 | 60 Nouns | Mitchell et al. (2008) | fMRI | 9 |
| 2 | Harry Potter Dataset | Wehbe et al. (2014a) | fMRI | 8 |
| 3 | Harry Potter Dataset | Wehbe et al. (2014b) | MEG | 8 |
| 4 | The Moth Radio Hour Dataset | Huth et al. (2016) | fMRI | 7 |
| 5 | Pereira Dataset | Pereira et al. (2018) | fMRI | 16 |
| 6 | Mother of all Unification Studies | Schoffelen et al. (2019) | fMRI, MEG | 204 |
| 7 | Natural Stories Audio Dataset | Zhang et al. (2020) | fMRI | 19 |
| 8 | Narratives | Nastase et al. (2021) | fMRI | 345 |
| 9 | Podcast dataset | Antonello et al. (2021) | fMRI | 5 |
| 10 | The Little Prince Datasets | Li et al. (2022) | fMRI | 112 |
| Authors | Data | E/D | Acc | P/B | Rank | MSE | RSA | Cos Sim | P@k |
| Mitchell et al. (2008) | 1 | E | ✓ | ✓ | ✓ | ||||
| Murphy et al. (2012) | 1 | E | ✓ | ||||||
| Wehbe et al. (2014a,b) | 2,3 | E | ✓ | ✓ | |||||
| Huth et al. (2016) | 4 | E | ✓ | ||||||
| Søgaard (2016) | 2 | E | ✓ | ||||||
| Pereira et al. (2018) | 5 | D | ✓ | ✓ | |||||
| Jain and Huth (2018) | 4 | E | ✓ | ||||||
| Toneva and Wehbe (2019) | 2 | E | ✓ | ||||||
| Gauthier and Levy (2019) | 5 | D | ✓ | ✓ | ✓ | ||||
| Minnema and Herbelot (2019) | 5 | D | ✓ | ✓ | ✓ | ✓ | |||
| Schwartz et al. (2019) | 3 | E | ✓ | ||||||
| Abnar et al. (2019) | 2 | E | ✓ | ||||||
| Zhang et al. (2020) | 7 | E | ✓ | ✓ | |||||
| Schrimpf et al. (2021) | 5 | E | ✓ | ✓ | |||||
| Abdou et al. (2021) | 2,5 | D | ✓ | ||||||
| Reddy and Wehbe (2021) | 2 | E | ✓ | ||||||
| Antonello et al. (2021) | 9 | E | ✓ | ||||||
| Antonello and Huth (2022) | 9 | E | ✓ | ||||||
| Caucheteux et al. (2021, 2022b,a) | 8 | E | ✓ | ||||||
| Caucheteux and King (2022) | 6 | E | ✓ | ||||||
| Zou et al. (2022) | 1,2 | D | ✓ | ||||||
| Oota et al. (2022a,b,c) | 8 | E | ✓ | ✓ | |||||
| Pasquiou et al. (2022) | 10 | E | ✓ | ||||||
| Toneva et al. (2022a) | 2 | E | ✓ | ||||||
| Merlin and Toneva (2022) | 2 | E | ✓ | ||||||
| Pascual et al. (2022) | 5 | D | ✓ | ✓ | |||||
| Aw and Toneva (2023) | 2 | E | ✓ | ✓ |
| Data | Authors | Method | N subjects | |
| 1 | 60 Nouns | Mitchell et al. (2008) | fMRI | 9 |
| 2 | Harry Potter Dataset | Wehbe et al. (2014a) | fMRI | 8 |
| 3 | Harry Potter Dataset | Wehbe et al. (2014b) | MEG | 8 |
| 4 | The Moth Radio Hour Dataset | Huth et al. (2016) | fMRI | 7 |
| 5 | Pereira Dataset | Pereira et al. (2018) | fMRI | 16 |
| 6 | Mother of all Unification Studies | Schoffelen et al. (2019) | fMRI, MEG | 204 |
| 7 | Natural Stories Audio Dataset | Zhang et al. (2020) | fMRI | 19 |
| 8 | Narratives | Nastase et al. (2021) | fMRI | 345 |
| 9 | Podcast dataset | Antonello et al. (2021) | fMRI | 5 |
| 10 | The Little Prince Datasets | Li et al. (2022) | fMRI | 112 |
| Authors | Data | E/D | Acc | P/B | Rank | MSE | RSA | Cos Sim | P@k |
| Mitchell et al. (2008) | 1 | E | ✓ | ✓ | ✓ | ||||
| Murphy et al. (2012) | 1 | E | ✓ | ||||||
| Wehbe et al. (2014a,b) | 2,3 | E | ✓ | ✓ | |||||
| Huth et al. (2016) | 4 | E | ✓ | ||||||
| Søgaard (2016) | 2 | E | ✓ | ||||||
| Pereira et al. (2018) | 5 | D | ✓ | ✓ | |||||
| Jain and Huth (2018) | 4 | E | ✓ | ||||||
| Toneva and Wehbe (2019) | 2 | E | ✓ | ||||||
| Gauthier and Levy (2019) | 5 | D | ✓ | ✓ | ✓ | ||||
| Minnema and Herbelot (2019) | 5 | D | ✓ | ✓ | ✓ | ✓ | |||
| Schwartz et al. (2019) | 3 | E | ✓ | ||||||
| Abnar et al. (2019) | 2 | E | ✓ | ||||||
| Zhang et al. (2020) | 7 | E | ✓ | ✓ | |||||
| Schrimpf et al. (2021) | 5 | E | ✓ | ✓ | |||||
| Abdou et al. (2021) | 2,5 | D | ✓ | ||||||
| Reddy and Wehbe (2021) | 2 | E | ✓ | ||||||
| Antonello et al. (2021) | 9 | E | ✓ | ||||||
| Antonello and Huth (2022) | 9 | E | ✓ | ||||||
| Caucheteux et al. (2021, 2022b,a) | 8 | E | ✓ | ||||||
| Caucheteux and King (2022) | 6 | E | ✓ | ||||||
| Zou et al. (2022) | 1,2 | D | ✓ | ||||||
| Oota et al. (2022a,b,c) | 8 | E | ✓ | ✓ | |||||
| Pasquiou et al. (2022) | 10 | E | ✓ | ||||||
| Toneva et al. (2022a) | 2 | E | ✓ | ||||||
| Merlin and Toneva (2022) | 2 | E | ✓ | ||||||
| Pascual et al. (2022) | 5 | D | ✓ | ✓ | |||||
| Aw and Toneva (2023) | 2 | E | ✓ | ✓ |
| Model | Pre-trained | Extra corpora | Dialog state tracking | Response Generation | ||||
| Joint Accuracy | Act F1 | Inform | Success | BLEU | Combined | |||
| DAMD (Zhang et al., 2020b) | - | no | - | - | 57.9 | 47.6 | 16.4 | 69.2 |
| LABES (Zhang et al., 2020a) | - | no | - | - | 68.5 | 58.1 | 18.9 | 82.2 |
| AuGPT (Kulhánek et al., 2021) | GPT-2 | yes | - | - | 76.6 | 60.5 | 16.8 | 85.4 |
| MinTL (Lin et al., 2020) | T5-small | no | 51.2 | - | 73.7 | 65.4 | 19.4 | 89.0 |
| SOLOIST (Peng et al., 2021) | GPT-2 | yes | 53.2 | - | 82.3 | 72.4 | 13.6 | 91.0 |
| DoTS (Jeon and Lee, 2021) | BERT-base | no | - | - | 80.4 | 68.7 | 16.8 | 91.4 |
| UBAR (Yang et al., 2021) | DistilGPT2 | no | 52.6 | - | 83.4 | 70.3 | 17.6 | 94.5 |
| PPTOD (Su et al., 2021) | T5-base | yes | 53.4 | - | 83.1 | 72.7 | 18.2 | 96.1 |
| BORT (Sun et al., 2022) | T5-small | no | 54.0 | - | 85.5 | 77.4 | 17.9 | 99.4 |
| MTTOD (Lee, 2021) | T5-base | no | 53.6 | - | 85.9 | 76.5 | 19.0 | 100.2 |
| GALAXY (He et al., 2022) | UniLM-base | yes | - | - | 85.4 | 75.7 | 19.6 | 100.2 |
| Baseline | T5-small | no | 53.8 | 53.0 | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-P | T5-small | no | 54.4 | 53.9 | 86.6 | 75.5 | 19.6 | 100.7 |
| Mars-G | T5-small | no | 55.1 | 53.7 | 88.9 | 78.0 | 19.9 | 103.4 |
| Model | Match | Success F1 | BLEU | Combined |
| Sequicity (Lei et al., 2018) | 92.7 | 85.4 | 25.3 | 114.4 |
| LABES (Zhang et al., 2020a) | 96.4 | 82.3 | 25.6 | 115.0 |
| SOLOIST (Peng et al., 2021) | 94.7 | 87.1 | 25.5 | 116.4 |
| GALAXY (He et al., 2022) | 98.5 | 87.7 | 24.2 | 117.3 |
| Mars-P | 97.0 | 87.2 | 25.9 | 118.0 |
| Mars-G | 96.2 | 89.6 | 26.1 | 119.0 |
| Model | Joint Accuracy |
| TRADE (Wu et al., 2019) | 36.1 |
| BART-CSP (Moradshahi et al., 2021) | 53.6 |
| GEEX (Li et al., 2021) | 54.7 |
| Mars-P | 59.3 |
| Mars-G | 59.8 |
| Model | Inform | Success | BLEU | Combined |
| Mars-G | 88.9 | 78.0 | 19.9 | 103.4 |
| w/o BSC | 88.3 | 76.6 | 19.5 | 102.0 |
| w/o ASC | 86.3 | 75.1 | 19.7 | 100.4 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| w/o BSC | 85.4 | 75.0 | 19.7 | 99.9 |
| w/o ASC | 83.7 | 73.0 | 19.8 | 98.2 |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Model | Inform | Success | BLEU |
| End-to-end model | 83.2 | 70.3 | 19.4 |
| w/ oracle state | 90.8 | 87.4 | 30.6 |
| Reference Corpus | 93.7 | 90.9 | 100.0 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.797 | 1.018 |
| Baseline | 0.844 | 1.017 |
| Mars-P | 0.340 | 0.542 |
| Mars-G | 1.996 | 1.993 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.555 | 0.807 |
| Baseline | 0.598 | 0.699 |
| Mars-P | 0.042 | 0.046 |
| Mars-G | 1.993 | 1.987 |
| Model | Inform | Success | BLEU | Combined |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-variant | 85.7 | 74.8 | 19.6 | 99.9 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | ||||||||||||||||
| run 1 | 35.4 | 17.2 | 10.9 | 37.2 | 41.3 | 23.7 | 12.4 | 44.9 | 51.8 | 31.9 | 14.1 | 56.0 | 60.1 | 44.2 | 15.6 | 67.8 |
| run 2 | 40.8 | 20.9 | 12.0 | 42.9 | 41.5 | 25.3 | 11.2 | 44.6 | 50.4 | 32.4 | 13.8 | 55.2 | 54.7 | 39.5 | 14.8 | 61.9 |
| run 3 | 38.5 | 14.5 | 10.6 | 37.1 | 40.0 | 23.9 | 12.3 | 44.3 | 42.5 | 26.8 | 15.4 | 50.1 | 59.1 | 45.7 | 15.1 | 67.5 |
| run 4 | 35.4 | 16.5 | 10.2 | 36.2 | 42.3 | 20.2 | 12.0 | 43.3 | 46.1 | 29.2 | 14.0 | 51.7 | 57.2 | 43.0 | 16.1 | 66.2 |
| run 5 | 34.0 | 17.6 | 12.4 | 38.2 | 39.5 | 21.9 | 13.0 | 43.7 | 50.8 | 31.3 | 13.6 | 54.7 | 63.1 | 49.3 | 17.1 | 73.3 |
| Average | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | ||||||||||||||||
| run 1 | 54.4 | 41.1 | 14.2 | 62.0 | 55.8 | 44.0 | 15.3 | 65.2 | 62.5 | 54.2 | 17.3 | 75.7 | 71.7 | 62.7 | 16.9 | 84.1 |
| run 2 | 54.8 | 36.8 | 13.6 | 59.4 | 51.6 | 42.0 | 15.7 | 62.5 | 65.8 | 56.3 | 15.5 | 76.6 | 67.4 | 59.7 | 18.7 | 82.3 |
| run 3 | 53.3 | 39.3 | 14.2 | 60.5 | 55.1 | 44.7 | 16.1 | 66.0 | 68.0 | 59.0 | 16.6 | 80.1 | 70.6 | 62.6 | 17.5 | 84.1 |
| run 4 | 52.4 | 37.1 | 13.8 | 58.6 | 58.4 | 47.3 | 15.2 | 68.1 | 58.3 | 48.4 | 14.4 | 67.8 | 68.9 | 61.3 | 18.2 | 83.3 |
| run 5 | 47.5 | 36.3 | 13.8 | 55.7 | 56.8 | 46.4 | 15.9 | 67.5 | 66.9 | 56.8 | 17.0 | 78.9 | 73.1 | 64.5 | 18.6 | 87.4 |
| Average | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | ||||||||||||||||
| run 1 | 37.4 | 23.0 | 11.6 | 41.8 | 52.3 | 34.8 | 13.0 | 56.6 | 61.7 | 45.7 | 15.9 | 69.6 | 77.2 | 61.5 | 15.5 | 84.9 |
| run 2 | 33.3 | 20.6 | 11.2 | 38.2 | 48.5 | 35.9 | 14.5 | 56.7 | 63.4 | 47.8 | 15.5 | 71.1 | 78.0 | 63.8 | 16.9 | 87.8 |
| run 3 | 40.0 | 23.1 | 11.7 | 43.3 | 50.3 | 33.2 | 13.6 | 55.4 | 67.8 | 50.0 | 13.1 | 72.0 | 77.4 | 64.6 | 16.2 | 87.2 |
| run 4 | 38.2 | 22.4 | 10.7 | 41.0 | 52.5 | 34.6 | 12.5 | 56.1 | 68.3 | 51.7 | 14.4 | 74.4 | 78.5 | 64.1 | 16.8 | 88.1 |
| run 5 | 38.0 | 21.3 | 11.3 | 41.0 | 47.8 | 32.3 | 13.7 | 53.8 | 66.2 | 48.3 | 13.8 | 71.1 | 76.8 | 62.4 | 16.2 | 85.8 |
| Average | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | ||||||||||||||||
| run 1 | 51.4 | 37.5 | 12.0 | 56.5 | 70.9 | 58.0 | 13.8 | 78.3 | 71.1 | 59.0 | 14.2 | 79.3 | 74.7 | 64.4 | 15.2 | 84.8 |
| run 2 | 53.8 | 41.7 | 11.3 | 59.1 | 64.1 | 53.7 | 13.8 | 72.7 | 69.5 | 60.7 | 14.0 | 79.1 | 79.3 | 67.7 | 15.0 | 88.5 |
| run 3 | 55.7 | 31.1 | 11.5 | 54.9 | 61.0 | 50.8 | 13.7 | 69.6 | 78.4 | 65.1 | 14.7 | 86.5 | 82.3 | 71.1 | 15.5 | 92.2 |
| run 4 | 52.4 | 33.3 | 10.6 | 53.5 | 73.0 | 59.3 | 14.0 | 80.2 | 80.2 | 67.4 | 14.5 | 88.3 | 76.6 | 65.6 | 15.3 | 86.4 |
| run 5 | 58.0 | 43.2 | 11.3 | 61.9 | 65.4 | 54.2 | 13.7 | 73.5 | 75.9 | 64.3 | 14.1 | 84.2 | 79.8 | 68.7 | 15.1 | 89.4 |
| Average | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | ||||||||||||||||
| run 1 | 70.7 | 46.8 | 13.7 | 72.5 | 65.2 | 50.6 | 14.2 | 72.1 | 72.3 | 55.0 | 14.9 | 78.6 | 74.8 | 60.4 | 15.8 | 83.4 |
| run 2 | 64.6 | 45.8 | 13.8 | 69.0 | 69.3 | 52.9 | 15.3 | 76.4 | 70.5 | 57.7 | 17.7 | 81.8 | 74.1 | 64.2 | 16.4 | 85.6 |
| run 3 | 64.4 | 51.1 | 15.1 | 72.9 | 65.7 | 53.6 | 15.8 | 75.5 | 74.8 | 64.6 | 16.9 | 86.6 | 74.3 | 61.8 | 17.2 | 85.3 |
| run 4 | 63.9 | 47.0 | 14.7 | 70.2 | 70.1 | 55.4 | 17.8 | 80.6 | 71.8 | 57.3 | 16.0 | 80.6 | 76.4 | 63.7 | 18.0 | 88.1 |
| run 5 | 63.7 | 50.7 | 14.4 | 71.6 | 71.2 | 55.8 | 15.6 | 79.1 | 74.1 | 61.6 | 15.8 | 83.7 | 74.4 | 61.9 | 17.5 | 85.7 |
| Average | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | ||||||||||||||||
| run 1 | 55.8 | 41.1 | 14.0 | 62.5 | 68.7 | 55.0 | 16.7 | 78.6 | 72.4 | 60.2 | 18.1 | 84.4 | 82.6 | 70.2 | 18.8 | 95.2 |
| run 2 | 57.0 | 43.2 | 12.9 | 63.0 | 68.4 | 55.9 | 15.2 | 77.4 | 76.0 | 61.4 | 17.1 | 85.8 | 78.4 | 66.9 | 18.7 | 91.4 |
| run 3 | 61.4 | 46.7 | 14.5 | 68.6 | 68.9 | 53.8 | 14.0 | 75.4 | 76.6 | 63.8 | 17.0 | 87.2 | 82.8 | 73.6 | 17.9 | 96.1 |
| run 4 | 56.1 | 42.4 | 14.1 | 63.4 | 73.1 | 60.3 | 16.6 | 83.3 | 80.6 | 63.9 | 17.1 | 89.4 | 82.5 | 71.3 | 19.0 | 95.9 |
| run 5 | 57.8 | 43.5 | 13.8 | 64.5 | 67.7 | 51.5 | 15.7 | 75.3 | 77.7 | 65.0 | 16.8 | 88.2 | 84.6 | 74.2 | 18.7 | 98.1 |
| Average | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Goal: | [restaurant]: {informable: food italian area east} {requestable: reference} |
| User (Turn 0): | i am looking for a jamaican restaurant in the east. |
| Golden R: | there are no [value_food] restaurants in the [value_area]. would you like to try another food type or area? |
| Golden BS: | [restaurant] food jamaican area east |
| Golden AS: | [restaurant] [nooffer] food area [request] area food |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another area or another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] area food |
| User (Turn 1): | i am looking for a place that serves jamaican food in the east. if not, italian will do. |
| Golden R: | there is [value_choice] [value_food] place in the [value_area], [value_name]. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [inform] choice name food area |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] food |
| User (Turn 2): | ok, i would like to make a reservation there. |
| Golden R: | i can do that for you, i just need details such as day, time, and the size of your party. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [offerbook] [request] time day people |
| Mars-G R: | what day and time would you like to dine? and how many people will be in your party? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [request] day time people |
| User (Turn 3): | i need a table for 7 at 14:45 on friday. |
| Golden R: | i booked you a table for [value_people] on [value_day] at [value_time] at [value_name], reference: [value_reference]. can i assist you further today? |
| Golden BS: | [restaurant] food italian area east time 14:45 day friday people 7 |
| Golden AS: | [restaurant] [offerbooked] time reference name day people [general] [reqmore] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value_reference]. |
| Mars-G BS: | [restaurant] food jamaican area east time 14:45 day friday people 7 |
| Mars-G AS: | [restaurant] [offerbooked] reference |
| Goal: | [attraction]: {informable: area centre} |
| User (Turn 3): | i am also looking for places to go in town. perhaps an attraction in the city centre. |
| Golden R: | i have the [value_name] located at [value_address] and it's [value_price] entrance. |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [recommend] address price name |
| Mars-G R: | there are [value_choice] attractions in the city centre. do you have a preference for type? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 4): | that sounds perfect. thanks! |
| Golden R: | can i help you with anything else? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [general] [reqmore] |
| Mars-G R: | is there anything else i can help you with? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [general] [reqmore] |
| Goal: | [attraction]: {informable: area south} {requestable: postcode} |
| User (Turn 4): | i would also like to go to some attractions in the same area as the restaurant. |
| Golden R: | there are [value_choice] attractions in the area. there's a [value_type], [value_type], [value_type], [value_type], [value_type], [value_type]... do any of these appeal to you in particular? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] type [inform] choice |
| Mars-G R: | there are [value_choice] attractions in the area. do you have a preference for type? |
| Mars-G BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 5): | restaurant is what i am looking for, entrance fee and postcode please. get me a taxi from the attraction by 10.15, get me its contact number and car type. |
| Golden R: | just to clarify you are looking for another restaurant and not an attraction? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] [restaurant] [select] |
| Mars-G R: | [value_name] is located in postcode [value_postcode]. the phone number is [value_phone]. the entrance fee is [value_price]. can i help you with anything else? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] price phone name postcode [general] [reqmore] |
| User (Turn 6): | oh, yeah, i do want information on attractions! in the same area. after that, i will need a taxi. |
| Golden R: | if you like parks the [value_name] has [value_price] admission and is in the [value_postcode] postcode. or i can tell you about the other attractions? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [recommend] name postcode price [general] [reqmore] |
| Mars-G R: | there are [value_choice] attractions in the [value_area]. [value_choice] are [value_type], [value_choice] are [value_type], and [value_choice] is [value_type]. do you have a preference? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 7): | that sounds wonderful, is it also in the south part of town? if so, i would like a taxi from there to my hotel. |
| Golden R: | yes, [value_name] is in the [value_area]. in order to help you book a taxi between the park and your hotel, i need to know what hotel you're at. |
| Golden BS: | [taxi] departure wandlebury country park [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [inform] area name [taxi] [request] destination |
| Mars-G R: | yes, [value_name] is in the [value_area]. |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area name |
| Goal: | [restaurant]: {informable: pricerange moderate area west} {requestable: reference} |
| User (Turn 0): | hi! i am planning a trip in cambridge and i am looking for a restaurant in the west part of town. |
| Golden R: | there are [value_choice] restaurants in the [value_area]. is there any kind of food you're looking for? |
| Golden BS: | [restaurant] area west |
| Golden AS: | [restaurant] [inform] choice area [request] food |
| Mars-G R: | there are [value_choice] restaurants in the [value_area] part of town. |
| do you have a price range or type of food in mind? | |
| Mars-G BS: | [restaurant] area west |
| Mars-G AS: | [restaurant] [inform] choice area [request] food price |
| User (Turn 1): | how about you suggest one a book a table for 4 on tuesday at 15:45 and i need the reference number too please |
| Golden R: | i was able to book [value_name], reference number is [value_reference] |
| Golden BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Golden AS: | [restaurant] [offerbooked] name reference |
| Mars-G R: | i have booked you at [value_name]. your reference number is [value_reference]. |
| is there anything else i can help you with? | |
| Mars-G BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Mars-G AS: | [restaurant] [offerbooked] reference name [general] [reqmore] |
| Goal: | [train]: {informable: leave 17:00 destination birmingham new street day monday departure cambridge} {requestable: reference} |
| User (Turn 0): | what train do i take to get to birmingham new street? i need to leave after 17:00. |
| Golden R: | there are trains departing hourly at 1 minute past the hour [value Leave]. all are [value_Price], and take [value_time]. |
| Golden BS: | [train] destination birmingham new street leave 17:00 |
| Golden AS: | [train] [inform] price leave time |
| Mars-G R: | where will you be departing from and what day would you like to travel ? |
| Mars-G BS: | [train] destination birmingham new street leave 17:00 |
| Mars-G AS: | [train] [request] departure day |
| User (Turn 1): | i am looking to travel on monday. the first train at 17:01 would be fine. can i get that train info? |
| Golden R: | sure the trainid is [value_id]. it leaves at [value Leave] and arrives at [value_arry]. would you like to make a reservation? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Golden AS: | [train] [inform] arrive leave id [offerbook] |
| Mars-G R: | [value_id] leaves at [value Leave] and arrives at [value_arry]. would you like me to book that for you? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Mars-G AS: | [train] [inform] arrive leave id [offerbook] |
| User (Turn 2): | yes please make a reservation and give me the reference number. |
| Golden R: | your reservation is book. the reference number is [value.Reference]. can i assist you with anything else today? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Golden AS: | [train] [offerbooked] reference [general] [reqmore] |
| Mars-G R: | how many tickets do you need? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Mars-G AS: | [train] [request] people |
| Goal: | [restaurant]: {informable: food international pricerange moderate area centre} {requestable: reference} |
| User (Turn 0): | what restaurants in the centre serve international cuisine? |
| Golden R: | the [value_name] and the [value_name] serve [value_food] food and are in the [value_area] of town. |
| Golden BS: | [restaurant] food international area centre |
| Golden AS: | [restaurant] [inform] food area name |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. do you have a price range in mind? |
| Mars-G BS: | [restaurant] food international area centre |
| Mars-G AS: | [restaurant] [inform] choice [request] price |
| User (Turn 1): | how about a place in the moderate price range? |
| Golden R: | [value_choice] of the named restaurants are in the [value_pricerange] price range. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [inform] choice price |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. [value_name] and [value_name]. would you like me to book one of these for you? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre |
| Mars-G AS: | [restaurant] [inform] choice name [offerbook] |
| User (Turn 2): | ok, can you book a table for 6 at 12:00 on tuesday at the varsity restaurant? i will need a reference number too, please. |
| Golden R: | certainly. i will have that reference number for you in just one second. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [offerbook] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value-reference]. is there anything else i can help you with? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [restaurant] [offerbooked] reference [general] [reqmore] |
| User (Turn 3): | thanks. looking forward to it. |
| Golden R: | yes i can. i have booked you a table for [value_time] on [value_day] for [value_people] people. your reference number is [value_reference]. anything else i can help you with? |
| Golden BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Golden AS: | [restaurant] [offerbooked] time reference people day |
| Mars-G R: | you're welcome. have a great day! |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [general] [welcome] [bye] |
| Goal: | [attraction]: {informable: area north} {requestable: postcode phone} |
| User (Turn 0): | i am looking for something to do in town. are there any interesting attraction in the north part of town? |
| Golden R: | yes, there are [value_choice] places: i see a [value_type], a [value_type], and two [value_type]. would one of these work for you? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] type choice [general] [reqmore] |
| Mars-G R: | there are [value_choice] [value_type] and [value_type] in the [value_area]. do you have a preference? |
| Mars-G BS: | [attraction] area north |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 1): | ooh, a pool would be fun. can you give me their contact info? |
| Golden R: | great! [value_address] is located at [value_address], kings hedges postcode is [value_postcode] and their phone is [value_phone]. is there anything else i can help you with? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] name postcode phone address [general] [reqmore] |
| Mars-G R: | [value_name] is located at [value_address]. their phone number is [value_phone]. |
| Mars-G BS: | [attraction] area north type swimming pool |
| Mars-G AS: | [attraction] [inform] name phone address |
| Goal: | [train]: {informable: destination stansted airport day friday arrive 15:30 departure cambridge} {requestable: reference} |
| User (Turn 3): | i will also need a train leaving cambridge. |
| Golden R: | sure, what is your arrival information? |
| Golden BS: | [train] departure cambridge |
| Golden AS: | [train] [request] destination |
| Mars-G R: | what day will you be traveling, and where will you be traveling to? |
| Mars-G BS: | [train] departure cambridge |
| Mars-G AS: | [train] [request] destination day |
| User (Turn 4): | i need to go to stansted airport on friday. |
| Golden R: | we have [value_choice] different departure times. any time table your looking at? |
| Golden BS: | [train] destination stansted airport departure cambridge |
| Golden AS: | [train] [inform] choice [request] leave |
| Mars-G R: | there are [value_choice] trains that meet your criteria. do you have a preferred departure or arrival time? |
| Mars-G BS: | [train] destination stansted airport day friday departure cambridge |
| Mars-G AS: | [train] [inform] choice [request] leave arrive |
| User (Turn 5): | i need to be in stansted airport by 15:30 please. |
| Golden R: | the soonest it will get you there is [value arriv] your reference number is [value/reference], anything else? |
| Golden BS: | [train] destination stansted airport arrive 15:30 departure cambridge people 1 |
| Golden AS: | [train] [offerbooked] arrive reference [general] [reqmore] |
| Mars-G R: | [value_id] will arrive at [value arriv]. would you like me to book that for you? |
| Mars-G BS: | [train] destination stansted airport day friday arrive 15:30 departure cambridge |
| Mars-G AS: | [train] [inform] arrive id [offerbook] |
| Goal: | [attraction]: {informable: type architecture area centre} {requestable: address phone postcode} |
| User (Turn 0): | i want to find a place to go in the centre to enjoy some entertainment. can you help? |
| Golden R: | there are [value_choice] attraction types in the [value_area], did you have one in mind? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [inform] area choice [request] name |
| Mars-G R: | i am sorry, there are no [value_type] attractions in the [value_area]. would you like to try another area? |
| Mars-G BS: | [attraction] area centre type entertainment |
| Mars-G AS: | [attraction] [nooffer] area type [request] area |
| User (Turn 1): | yes, i am looking for something with the type of architecture. thanks! |
| Golden R: | i would recommend [value_name] as a great place to see interesting [value_type]. |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [recommend] name type |
| Mars-G R: | there are [value_choice] [value_type] attractions in the [value_area]. i recommend [value_name]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [recommend] name [inform] type choice area |
| User (Turn 2): | okay, could you give me the address and phone number? |
| Golden R: | sure. the address is [value_address]. the phone number is [value_phone]. |
| is there anything else i can help you with today? | |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [inform] address phone [general] [reqmore] |
| Mars-G R: | the address is [value_address] and the phone number is [value_phone]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [inform] address phone |
| Model | Pre-trained | Extra corpora | Dialog state tracking | Response Generation | ||||
| Joint Accuracy | Act F1 | Inform | Success | BLEU | Combined | |||
| DAMD (Zhang et al., 2020b) | - | no | - | - | 57.9 | 47.6 | 16.4 | 69.2 |
| LABES (Zhang et al., 2020a) | - | no | - | - | 68.5 | 58.1 | 18.9 | 82.2 |
| AuGPT (Kulhánek et al., 2021) | GPT-2 | yes | - | - | 76.6 | 60.5 | 16.8 | 85.4 |
| MinTL (Lin et al., 2020) | T5-small | no | 51.2 | - | 73.7 | 65.4 | 19.4 | 89.0 |
| SOLOIST (Peng et al., 2021) | GPT-2 | yes | 53.2 | - | 82.3 | 72.4 | 13.6 | 91.0 |
| DoTS (Jeon and Lee, 2021) | BERT-base | no | - | - | 80.4 | 68.7 | 16.8 | 91.4 |
| UBAR (Yang et al., 2021) | DistilGPT2 | no | 52.6 | - | 83.4 | 70.3 | 17.6 | 94.5 |
| PPTOD (Su et al., 2021) | T5-base | yes | 53.4 | - | 83.1 | 72.7 | 18.2 | 96.1 |
| BORT (Sun et al., 2022) | T5-small | no | 54.0 | - | 85.5 | 77.4 | 17.9 | 99.4 |
| MTTOD (Lee, 2021) | T5-base | no | 53.6 | - | 85.9 | 76.5 | 19.0 | 100.2 |
| GALAXY (He et al., 2022) | UniLM-base | yes | - | - | 85.4 | 75.7 | 19.6 | 100.2 |
| Baseline | T5-small | no | 53.8 | 53.0 | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-P | T5-small | no | 54.4 | 53.9 | 86.6 | 75.5 | 19.6 | 100.7 |
| Mars-G | T5-small | no | 55.1 | 53.7 | 88.9 | 78.0 | 19.9 | 103.4 |
| Model | Match | Success F1 | BLEU | Combined |
| Sequicity (Lei et al., 2018) | 92.7 | 85.4 | 25.3 | 114.4 |
| LABES (Zhang et al., 2020a) | 96.4 | 82.3 | 25.6 | 115.0 |
| SOLOIST (Peng et al., 2021) | 94.7 | 87.1 | 25.5 | 116.4 |
| GALAXY (He et al., 2022) | 98.5 | 87.7 | 24.2 | 117.3 |
| Mars-P | 97.0 | 87.2 | 25.9 | 118.0 |
| Mars-G | 96.2 | 89.6 | 26.1 | 119.0 |
| Model | Joint Accuracy |
| TRADE (Wu et al., 2019) | 36.1 |
| BART-CSP (Moradshahi et al., 2021) | 53.6 |
| GEEX (Li et al., 2021) | 54.7 |
| Mars-P | 59.3 |
| Mars-G | 59.8 |
| Model | Inform | Success | BLEU | Combined |
| Mars-G | 88.9 | 78.0 | 19.9 | 103.4 |
| w/o BSC | 88.3 | 76.6 | 19.5 | 102.0 |
| w/o ASC | 86.3 | 75.1 | 19.7 | 100.4 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| w/o BSC | 85.4 | 75.0 | 19.7 | 99.9 |
| w/o ASC | 83.7 | 73.0 | 19.8 | 98.2 |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Model | Inform | Success | BLEU |
| End-to-end model | 83.2 | 70.3 | 19.4 |
| w/ oracle state | 90.8 | 87.4 | 30.6 |
| Reference Corpus | 93.7 | 90.9 | 100.0 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.797 | 1.018 |
| Baseline | 0.844 | 1.017 |
| Mars-P | 0.340 | 0.542 |
| Mars-G | 1.996 | 1.993 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.555 | 0.807 |
| Baseline | 0.598 | 0.699 |
| Mars-P | 0.042 | 0.046 |
| Mars-G | 1.993 | 1.987 |
| Model | Inform | Success | BLEU | Combined |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-variant | 85.7 | 74.8 | 19.6 | 99.9 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | ||||||||||||||||
| run 1 | 35.4 | 17.2 | 10.9 | 37.2 | 41.3 | 23.7 | 12.4 | 44.9 | 51.8 | 31.9 | 14.1 | 56.0 | 60.1 | 44.2 | 15.6 | 67.8 |
| run 2 | 40.8 | 20.9 | 12.0 | 42.9 | 41.5 | 25.3 | 11.2 | 44.6 | 50.4 | 32.4 | 13.8 | 55.2 | 54.7 | 39.5 | 14.8 | 61.9 |
| run 3 | 38.5 | 14.5 | 10.6 | 37.1 | 40.0 | 23.9 | 12.3 | 44.3 | 42.5 | 26.8 | 15.4 | 50.1 | 59.1 | 45.7 | 15.1 | 67.5 |
| run 4 | 35.4 | 16.5 | 10.2 | 36.2 | 42.3 | 20.2 | 12.0 | 43.3 | 46.1 | 29.2 | 14.0 | 51.7 | 57.2 | 43.0 | 16.1 | 66.2 |
| run 5 | 34.0 | 17.6 | 12.4 | 38.2 | 39.5 | 21.9 | 13.0 | 43.7 | 50.8 | 31.3 | 13.6 | 54.7 | 63.1 | 49.3 | 17.1 | 73.3 |
| Average | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | ||||||||||||||||
| run 1 | 54.4 | 41.1 | 14.2 | 62.0 | 55.8 | 44.0 | 15.3 | 65.2 | 62.5 | 54.2 | 17.3 | 75.7 | 71.7 | 62.7 | 16.9 | 84.1 |
| run 2 | 54.8 | 36.8 | 13.6 | 59.4 | 51.6 | 42.0 | 15.7 | 62.5 | 65.8 | 56.3 | 15.5 | 76.6 | 67.4 | 59.7 | 18.7 | 82.3 |
| run 3 | 53.3 | 39.3 | 14.2 | 60.5 | 55.1 | 44.7 | 16.1 | 66.0 | 68.0 | 59.0 | 16.6 | 80.1 | 70.6 | 62.6 | 17.5 | 84.1 |
| run 4 | 52.4 | 37.1 | 13.8 | 58.6 | 58.4 | 47.3 | 15.2 | 68.1 | 58.3 | 48.4 | 14.4 | 67.8 | 68.9 | 61.3 | 18.2 | 83.3 |
| run 5 | 47.5 | 36.3 | 13.8 | 55.7 | 56.8 | 46.4 | 15.9 | 67.5 | 66.9 | 56.8 | 17.0 | 78.9 | 73.1 | 64.5 | 18.6 | 87.4 |
| Average | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | ||||||||||||||||
| run 1 | 37.4 | 23.0 | 11.6 | 41.8 | 52.3 | 34.8 | 13.0 | 56.6 | 61.7 | 45.7 | 15.9 | 69.6 | 77.2 | 61.5 | 15.5 | 84.9 |
| run 2 | 33.3 | 20.6 | 11.2 | 38.2 | 48.5 | 35.9 | 14.5 | 56.7 | 63.4 | 47.8 | 15.5 | 71.1 | 78.0 | 63.8 | 16.9 | 87.8 |
| run 3 | 40.0 | 23.1 | 11.7 | 43.3 | 50.3 | 33.2 | 13.6 | 55.4 | 67.8 | 50.0 | 13.1 | 72.0 | 77.4 | 64.6 | 16.2 | 87.2 |
| run 4 | 38.2 | 22.4 | 10.7 | 41.0 | 52.5 | 34.6 | 12.5 | 56.1 | 68.3 | 51.7 | 14.4 | 74.4 | 78.5 | 64.1 | 16.8 | 88.1 |
| run 5 | 38.0 | 21.3 | 11.3 | 41.0 | 47.8 | 32.3 | 13.7 | 53.8 | 66.2 | 48.3 | 13.8 | 71.1 | 76.8 | 62.4 | 16.2 | 85.8 |
| Average | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | ||||||||||||||||
| run 1 | 51.4 | 37.5 | 12.0 | 56.5 | 70.9 | 58.0 | 13.8 | 78.3 | 71.1 | 59.0 | 14.2 | 79.3 | 74.7 | 64.4 | 15.2 | 84.8 |
| run 2 | 53.8 | 41.7 | 11.3 | 59.1 | 64.1 | 53.7 | 13.8 | 72.7 | 69.5 | 60.7 | 14.0 | 79.1 | 79.3 | 67.7 | 15.0 | 88.5 |
| run 3 | 55.7 | 31.1 | 11.5 | 54.9 | 61.0 | 50.8 | 13.7 | 69.6 | 78.4 | 65.1 | 14.7 | 86.5 | 82.3 | 71.1 | 15.5 | 92.2 |
| run 4 | 52.4 | 33.3 | 10.6 | 53.5 | 73.0 | 59.3 | 14.0 | 80.2 | 80.2 | 67.4 | 14.5 | 88.3 | 76.6 | 65.6 | 15.3 | 86.4 |
| run 5 | 58.0 | 43.2 | 11.3 | 61.9 | 65.4 | 54.2 | 13.7 | 73.5 | 75.9 | 64.3 | 14.1 | 84.2 | 79.8 | 68.7 | 15.1 | 89.4 |
| Average | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | ||||||||||||||||
| run 1 | 70.7 | 46.8 | 13.7 | 72.5 | 65.2 | 50.6 | 14.2 | 72.1 | 72.3 | 55.0 | 14.9 | 78.6 | 74.8 | 60.4 | 15.8 | 83.4 |
| run 2 | 64.6 | 45.8 | 13.8 | 69.0 | 69.3 | 52.9 | 15.3 | 76.4 | 70.5 | 57.7 | 17.7 | 81.8 | 74.1 | 64.2 | 16.4 | 85.6 |
| run 3 | 64.4 | 51.1 | 15.1 | 72.9 | 65.7 | 53.6 | 15.8 | 75.5 | 74.8 | 64.6 | 16.9 | 86.6 | 74.3 | 61.8 | 17.2 | 85.3 |
| run 4 | 63.9 | 47.0 | 14.7 | 70.2 | 70.1 | 55.4 | 17.8 | 80.6 | 71.8 | 57.3 | 16.0 | 80.6 | 76.4 | 63.7 | 18.0 | 88.1 |
| run 5 | 63.7 | 50.7 | 14.4 | 71.6 | 71.2 | 55.8 | 15.6 | 79.1 | 74.1 | 61.6 | 15.8 | 83.7 | 74.4 | 61.9 | 17.5 | 85.7 |
| Average | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | ||||||||||||||||
| run 1 | 55.8 | 41.1 | 14.0 | 62.5 | 68.7 | 55.0 | 16.7 | 78.6 | 72.4 | 60.2 | 18.1 | 84.4 | 82.6 | 70.2 | 18.8 | 95.2 |
| run 2 | 57.0 | 43.2 | 12.9 | 63.0 | 68.4 | 55.9 | 15.2 | 77.4 | 76.0 | 61.4 | 17.1 | 85.8 | 78.4 | 66.9 | 18.7 | 91.4 |
| run 3 | 61.4 | 46.7 | 14.5 | 68.6 | 68.9 | 53.8 | 14.0 | 75.4 | 76.6 | 63.8 | 17.0 | 87.2 | 82.8 | 73.6 | 17.9 | 96.1 |
| run 4 | 56.1 | 42.4 | 14.1 | 63.4 | 73.1 | 60.3 | 16.6 | 83.3 | 80.6 | 63.9 | 17.1 | 89.4 | 82.5 | 71.3 | 19.0 | 95.9 |
| run 5 | 57.8 | 43.5 | 13.8 | 64.5 | 67.7 | 51.5 | 15.7 | 75.3 | 77.7 | 65.0 | 16.8 | 88.2 | 84.6 | 74.2 | 18.7 | 98.1 |
| Average | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Goal: | [restaurant]: {informable: food italian area east} {requestable: reference} |
| User (Turn 0): | i am looking for a jamaican restaurant in the east. |
| Golden R: | there are no [value_food] restaurants in the [value_area]. would you like to try another food type or area? |
| Golden BS: | [restaurant] food jamaican area east |
| Golden AS: | [restaurant] [nooffer] food area [request] area food |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another area or another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] area food |
| User (Turn 1): | i am looking for a place that serves jamaican food in the east. if not, italian will do. |
| Golden R: | there is [value_choice] [value_food] place in the [value_area], [value_name]. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [inform] choice name food area |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] food |
| User (Turn 2): | ok, i would like to make a reservation there. |
| Golden R: | i can do that for you, i just need details such as day, time, and the size of your party. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [offerbook] [request] time day people |
| Mars-G R: | what day and time would you like to dine? and how many people will be in your party? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [request] day time people |
| User (Turn 3): | i need a table for 7 at 14:45 on friday. |
| Golden R: | i booked you a table for [value_people] on [value_day] at [value_time] at [value_name], reference: [value_reference]. can i assist you further today? |
| Golden BS: | [restaurant] food italian area east time 14:45 day friday people 7 |
| Golden AS: | [restaurant] [offerbooked] time reference name day people [general] [reqmore] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value_reference]. |
| Mars-G BS: | [restaurant] food jamaican area east time 14:45 day friday people 7 |
| Mars-G AS: | [restaurant] [offerbooked] reference |
| Goal: | [attraction]: {informable: area centre} |
| User (Turn 3): | i am also looking for places to go in town. perhaps an attraction in the city centre. |
| Golden R: | i have the [value_name] located at [value_address] and it's [value_price] entrance. |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [recommend] address price name |
| Mars-G R: | there are [value_choice] attractions in the city centre. do you have a preference for type? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 4): | that sounds perfect. thanks! |
| Golden R: | can i help you with anything else? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [general] [reqmore] |
| Mars-G R: | is there anything else i can help you with? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [general] [reqmore] |
| Goal: | [attraction]: {informable: area south} {requestable: postcode} |
| User (Turn 4): | i would also like to go to some attractions in the same area as the restaurant. |
| Golden R: | there are [value_choice] attractions in the area. there's a [value_type], [value_type], [value_type], [value_type], [value_type], [value_type]... do any of these appeal to you in particular? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] type [inform] choice |
| Mars-G R: | there are [value_choice] attractions in the area. do you have a preference for type? |
| Mars-G BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 5): | restaurant is what i am looking for, entrance fee and postcode please. get me a taxi from the attraction by 10.15, get me its contact number and car type. |
| Golden R: | just to clarify you are looking for another restaurant and not an attraction? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] [restaurant] [select] |
| Mars-G R: | [value_name] is located in postcode [value_postcode]. the phone number is [value_phone]. the entrance fee is [value_price]. can i help you with anything else? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] price phone name postcode [general] [reqmore] |
| User (Turn 6): | oh, yeah, i do want information on attractions! in the same area. after that, i will need a taxi. |
| Golden R: | if you like parks the [value_name] has [value_price] admission and is in the [value_postcode] postcode. or i can tell you about the other attractions? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [recommend] name postcode price [general] [reqmore] |
| Mars-G R: | there are [value_choice] attractions in the [value_area]. [value_choice] are [value_type], [value_choice] are [value_type], and [value_choice] is [value_type]. do you have a preference? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 7): | that sounds wonderful, is it also in the south part of town? if so, i would like a taxi from there to my hotel. |
| Golden R: | yes, [value_name] is in the [value_area]. in order to help you book a taxi between the park and your hotel, i need to know what hotel you're at. |
| Golden BS: | [taxi] departure wandlebury country park [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [inform] area name [taxi] [request] destination |
| Mars-G R: | yes, [value_name] is in the [value_area]. |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area name |
| Goal: | [restaurant]: {informable: pricerange moderate area west} {requestable: reference} |
| User (Turn 0): | hi! i am planning a trip in cambridge and i am looking for a restaurant in the west part of town. |
| Golden R: | there are [value_choice] restaurants in the [value_area]. is there any kind of food you're looking for? |
| Golden BS: | [restaurant] area west |
| Golden AS: | [restaurant] [inform] choice area [request] food |
| Mars-G R: | there are [value_choice] restaurants in the [value_area] part of town. |
| do you have a price range or type of food in mind? | |
| Mars-G BS: | [restaurant] area west |
| Mars-G AS: | [restaurant] [inform] choice area [request] food price |
| User (Turn 1): | how about you suggest one a book a table for 4 on tuesday at 15:45 and i need the reference number too please |
| Golden R: | i was able to book [value_name], reference number is [value_reference] |
| Golden BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Golden AS: | [restaurant] [offerbooked] name reference |
| Mars-G R: | i have booked you at [value_name]. your reference number is [value_reference]. |
| is there anything else i can help you with? | |
| Mars-G BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Mars-G AS: | [restaurant] [offerbooked] reference name [general] [reqmore] |
| Goal: | [train]: {informable: leave 17:00 destination birmingham new street day monday departure cambridge} {requestable: reference} |
| User (Turn 0): | what train do i take to get to birmingham new street? i need to leave after 17:00. |
| Golden R: | there are trains departing hourly at 1 minute past the hour [value Leave]. all are [value_Price], and take [value_time]. |
| Golden BS: | [train] destination birmingham new street leave 17:00 |
| Golden AS: | [train] [inform] price leave time |
| Mars-G R: | where will you be departing from and what day would you like to travel ? |
| Mars-G BS: | [train] destination birmingham new street leave 17:00 |
| Mars-G AS: | [train] [request] departure day |
| User (Turn 1): | i am looking to travel on monday. the first train at 17:01 would be fine. can i get that train info? |
| Golden R: | sure the trainid is [value_id]. it leaves at [value Leave] and arrives at [value_arry]. would you like to make a reservation? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Golden AS: | [train] [inform] arrive leave id [offerbook] |
| Mars-G R: | [value_id] leaves at [value Leave] and arrives at [value_arry]. would you like me to book that for you? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Mars-G AS: | [train] [inform] arrive leave id [offerbook] |
| User (Turn 2): | yes please make a reservation and give me the reference number. |
| Golden R: | your reservation is book. the reference number is [value.Reference]. can i assist you with anything else today? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Golden AS: | [train] [offerbooked] reference [general] [reqmore] |
| Mars-G R: | how many tickets do you need? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Mars-G AS: | [train] [request] people |
| Goal: | [restaurant]: {informable: food international pricerange moderate area centre} {requestable: reference} |
| User (Turn 0): | what restaurants in the centre serve international cuisine? |
| Golden R: | the [value_name] and the [value_name] serve [value_food] food and are in the [value_area] of town. |
| Golden BS: | [restaurant] food international area centre |
| Golden AS: | [restaurant] [inform] food area name |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. do you have a price range in mind? |
| Mars-G BS: | [restaurant] food international area centre |
| Mars-G AS: | [restaurant] [inform] choice [request] price |
| User (Turn 1): | how about a place in the moderate price range? |
| Golden R: | [value_choice] of the named restaurants are in the [value_pricerange] price range. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [inform] choice price |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. [value_name] and [value_name]. would you like me to book one of these for you? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre |
| Mars-G AS: | [restaurant] [inform] choice name [offerbook] |
| User (Turn 2): | ok, can you book a table for 6 at 12:00 on tuesday at the varsity restaurant? i will need a reference number too, please. |
| Golden R: | certainly. i will have that reference number for you in just one second. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [offerbook] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value-reference]. is there anything else i can help you with? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [restaurant] [offerbooked] reference [general] [reqmore] |
| User (Turn 3): | thanks. looking forward to it. |
| Golden R: | yes i can. i have booked you a table for [value_time] on [value_day] for [value_people] people. your reference number is [value_reference]. anything else i can help you with? |
| Golden BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Golden AS: | [restaurant] [offerbooked] time reference people day |
| Mars-G R: | you're welcome. have a great day! |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [general] [welcome] [bye] |
| Goal: | [attraction]: {informable: area north} {requestable: postcode phone} |
| User (Turn 0): | i am looking for something to do in town. are there any interesting attraction in the north part of town? |
| Golden R: | yes, there are [value_choice] places: i see a [value_type], a [value_type], and two [value_type]. would one of these work for you? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] type choice [general] [reqmore] |
| Mars-G R: | there are [value_choice] [value_type] and [value_type] in the [value_area]. do you have a preference? |
| Mars-G BS: | [attraction] area north |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 1): | ooh, a pool would be fun. can you give me their contact info? |
| Golden R: | great! [value_address] is located at [value_address], kings hedges postcode is [value_postcode] and their phone is [value_phone]. is there anything else i can help you with? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] name postcode phone address [general] [reqmore] |
| Mars-G R: | [value_name] is located at [value_address]. their phone number is [value_phone]. |
| Mars-G BS: | [attraction] area north type swimming pool |
| Mars-G AS: | [attraction] [inform] name phone address |
| Goal: | [train]: {informable: destination stansted airport day friday arrive 15:30 departure cambridge} {requestable: reference} |
| User (Turn 3): | i will also need a train leaving cambridge. |
| Golden R: | sure, what is your arrival information? |
| Golden BS: | [train] departure cambridge |
| Golden AS: | [train] [request] destination |
| Mars-G R: | what day will you be traveling, and where will you be traveling to? |
| Mars-G BS: | [train] departure cambridge |
| Mars-G AS: | [train] [request] destination day |
| User (Turn 4): | i need to go to stansted airport on friday. |
| Golden R: | we have [value_choice] different departure times. any time table your looking at? |
| Golden BS: | [train] destination stansted airport departure cambridge |
| Golden AS: | [train] [inform] choice [request] leave |
| Mars-G R: | there are [value_choice] trains that meet your criteria. do you have a preferred departure or arrival time? |
| Mars-G BS: | [train] destination stansted airport day friday departure cambridge |
| Mars-G AS: | [train] [inform] choice [request] leave arrive |
| User (Turn 5): | i need to be in stansted airport by 15:30 please. |
| Golden R: | the soonest it will get you there is [value arriv] your reference number is [value/reference], anything else? |
| Golden BS: | [train] destination stansted airport arrive 15:30 departure cambridge people 1 |
| Golden AS: | [train] [offerbooked] arrive reference [general] [reqmore] |
| Mars-G R: | [value_id] will arrive at [value arriv]. would you like me to book that for you? |
| Mars-G BS: | [train] destination stansted airport day friday arrive 15:30 departure cambridge |
| Mars-G AS: | [train] [inform] arrive id [offerbook] |
| Goal: | [attraction]: {informable: type architecture area centre} {requestable: address phone postcode} |
| User (Turn 0): | i want to find a place to go in the centre to enjoy some entertainment. can you help? |
| Golden R: | there are [value_choice] attraction types in the [value_area], did you have one in mind? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [inform] area choice [request] name |
| Mars-G R: | i am sorry, there are no [value_type] attractions in the [value_area]. would you like to try another area? |
| Mars-G BS: | [attraction] area centre type entertainment |
| Mars-G AS: | [attraction] [nooffer] area type [request] area |
| User (Turn 1): | yes, i am looking for something with the type of architecture. thanks! |
| Golden R: | i would recommend [value_name] as a great place to see interesting [value_type]. |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [recommend] name type |
| Mars-G R: | there are [value_choice] [value_type] attractions in the [value_area]. i recommend [value_name]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [recommend] name [inform] type choice area |
| User (Turn 2): | okay, could you give me the address and phone number? |
| Golden R: | sure. the address is [value_address]. the phone number is [value_phone]. |
| is there anything else i can help you with today? | |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [inform] address phone [general] [reqmore] |
| Mars-G R: | the address is [value_address] and the phone number is [value_phone]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [inform] address phone |
| Model | Pre-trained | Extra corpora | Dialog state tracking | Response Generation | ||||
| Joint Accuracy | Act F1 | Inform | Success | BLEU | Combined | |||
| DAMD (Zhang et al., 2020b) | - | no | - | - | 57.9 | 47.6 | 16.4 | 69.2 |
| LABES (Zhang et al., 2020a) | - | no | - | - | 68.5 | 58.1 | 18.9 | 82.2 |
| AuGPT (Kulhánek et al., 2021) | GPT-2 | yes | - | - | 76.6 | 60.5 | 16.8 | 85.4 |
| MinTL (Lin et al., 2020) | T5-small | no | 51.2 | - | 73.7 | 65.4 | 19.4 | 89.0 |
| SOLOIST (Peng et al., 2021) | GPT-2 | yes | 53.2 | - | 82.3 | 72.4 | 13.6 | 91.0 |
| DoTS (Jeon and Lee, 2021) | BERT-base | no | - | - | 80.4 | 68.7 | 16.8 | 91.4 |
| UBAR (Yang et al., 2021) | DistilGPT2 | no | 52.6 | - | 83.4 | 70.3 | 17.6 | 94.5 |
| PPTOD (Su et al., 2021) | T5-base | yes | 53.4 | - | 83.1 | 72.7 | 18.2 | 96.1 |
| BORT (Sun et al., 2022) | T5-small | no | 54.0 | - | 85.5 | 77.4 | 17.9 | 99.4 |
| MTTOD (Lee, 2021) | T5-base | no | 53.6 | - | 85.9 | 76.5 | 19.0 | 100.2 |
| GALAXY (He et al., 2022) | UniLM-base | yes | - | - | 85.4 | 75.7 | 19.6 | 100.2 |
| Baseline | T5-small | no | 53.8 | 53.0 | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-P | T5-small | no | 54.4 | 53.9 | 86.6 | 75.5 | 19.6 | 100.7 |
| Mars-G | T5-small | no | 55.1 | 53.7 | 88.9 | 78.0 | 19.9 | 103.4 |
| Model | Match | Success F1 | BLEU | Combined |
| Sequicity (Lei et al., 2018) | 92.7 | 85.4 | 25.3 | 114.4 |
| LABES (Zhang et al., 2020a) | 96.4 | 82.3 | 25.6 | 115.0 |
| SOLOIST (Peng et al., 2021) | 94.7 | 87.1 | 25.5 | 116.4 |
| GALAXY (He et al., 2022) | 98.5 | 87.7 | 24.2 | 117.3 |
| Mars-P | 97.0 | 87.2 | 25.9 | 118.0 |
| Mars-G | 96.2 | 89.6 | 26.1 | 119.0 |
| Model | Joint Accuracy |
| TRADE (Wu et al., 2019) | 36.1 |
| BART-CSP (Moradshahi et al., 2021) | 53.6 |
| GEEX (Li et al., 2021) | 54.7 |
| Mars-P | 59.3 |
| Mars-G | 59.8 |
| Model | Inform | Success | BLEU | Combined |
| Mars-G | 88.9 | 78.0 | 19.9 | 103.4 |
| w/o BSC | 88.3 | 76.6 | 19.5 | 102.0 |
| w/o ASC | 86.3 | 75.1 | 19.7 | 100.4 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| w/o BSC | 85.4 | 75.0 | 19.7 | 99.9 |
| w/o ASC | 83.7 | 73.0 | 19.8 | 98.2 |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Model | Inform | Success | BLEU |
| End-to-end model | 83.2 | 70.3 | 19.4 |
| w/ oracle state | 90.8 | 87.4 | 30.6 |
| Reference Corpus | 93.7 | 90.9 | 100.0 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.797 | 1.018 |
| Baseline | 0.844 | 1.017 |
| Mars-P | 0.340 | 0.542 |
| Mars-G | 1.996 | 1.993 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.555 | 0.807 |
| Baseline | 0.598 | 0.699 |
| Mars-P | 0.042 | 0.046 |
| Mars-G | 1.993 | 1.987 |
| Model | Inform | Success | BLEU | Combined |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-variant | 85.7 | 74.8 | 19.6 | 99.9 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | ||||||||||||||||
| run 1 | 35.4 | 17.2 | 10.9 | 37.2 | 41.3 | 23.7 | 12.4 | 44.9 | 51.8 | 31.9 | 14.1 | 56.0 | 60.1 | 44.2 | 15.6 | 67.8 |
| run 2 | 40.8 | 20.9 | 12.0 | 42.9 | 41.5 | 25.3 | 11.2 | 44.6 | 50.4 | 32.4 | 13.8 | 55.2 | 54.7 | 39.5 | 14.8 | 61.9 |
| run 3 | 38.5 | 14.5 | 10.6 | 37.1 | 40.0 | 23.9 | 12.3 | 44.3 | 42.5 | 26.8 | 15.4 | 50.1 | 59.1 | 45.7 | 15.1 | 67.5 |
| run 4 | 35.4 | 16.5 | 10.2 | 36.2 | 42.3 | 20.2 | 12.0 | 43.3 | 46.1 | 29.2 | 14.0 | 51.7 | 57.2 | 43.0 | 16.1 | 66.2 |
| run 5 | 34.0 | 17.6 | 12.4 | 38.2 | 39.5 | 21.9 | 13.0 | 43.7 | 50.8 | 31.3 | 13.6 | 54.7 | 63.1 | 49.3 | 17.1 | 73.3 |
| Average | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | ||||||||||||||||
| run 1 | 54.4 | 41.1 | 14.2 | 62.0 | 55.8 | 44.0 | 15.3 | 65.2 | 62.5 | 54.2 | 17.3 | 75.7 | 71.7 | 62.7 | 16.9 | 84.1 |
| run 2 | 54.8 | 36.8 | 13.6 | 59.4 | 51.6 | 42.0 | 15.7 | 62.5 | 65.8 | 56.3 | 15.5 | 76.6 | 67.4 | 59.7 | 18.7 | 82.3 |
| run 3 | 53.3 | 39.3 | 14.2 | 60.5 | 55.1 | 44.7 | 16.1 | 66.0 | 68.0 | 59.0 | 16.6 | 80.1 | 70.6 | 62.6 | 17.5 | 84.1 |
| run 4 | 52.4 | 37.1 | 13.8 | 58.6 | 58.4 | 47.3 | 15.2 | 68.1 | 58.3 | 48.4 | 14.4 | 67.8 | 68.9 | 61.3 | 18.2 | 83.3 |
| run 5 | 47.5 | 36.3 | 13.8 | 55.7 | 56.8 | 46.4 | 15.9 | 67.5 | 66.9 | 56.8 | 17.0 | 78.9 | 73.1 | 64.5 | 18.6 | 87.4 |
| Average | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | ||||||||||||||||
| run 1 | 37.4 | 23.0 | 11.6 | 41.8 | 52.3 | 34.8 | 13.0 | 56.6 | 61.7 | 45.7 | 15.9 | 69.6 | 77.2 | 61.5 | 15.5 | 84.9 |
| run 2 | 33.3 | 20.6 | 11.2 | 38.2 | 48.5 | 35.9 | 14.5 | 56.7 | 63.4 | 47.8 | 15.5 | 71.1 | 78.0 | 63.8 | 16.9 | 87.8 |
| run 3 | 40.0 | 23.1 | 11.7 | 43.3 | 50.3 | 33.2 | 13.6 | 55.4 | 67.8 | 50.0 | 13.1 | 72.0 | 77.4 | 64.6 | 16.2 | 87.2 |
| run 4 | 38.2 | 22.4 | 10.7 | 41.0 | 52.5 | 34.6 | 12.5 | 56.1 | 68.3 | 51.7 | 14.4 | 74.4 | 78.5 | 64.1 | 16.8 | 88.1 |
| run 5 | 38.0 | 21.3 | 11.3 | 41.0 | 47.8 | 32.3 | 13.7 | 53.8 | 66.2 | 48.3 | 13.8 | 71.1 | 76.8 | 62.4 | 16.2 | 85.8 |
| Average | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | ||||||||||||||||
| run 1 | 51.4 | 37.5 | 12.0 | 56.5 | 70.9 | 58.0 | 13.8 | 78.3 | 71.1 | 59.0 | 14.2 | 79.3 | 74.7 | 64.4 | 15.2 | 84.8 |
| run 2 | 53.8 | 41.7 | 11.3 | 59.1 | 64.1 | 53.7 | 13.8 | 72.7 | 69.5 | 60.7 | 14.0 | 79.1 | 79.3 | 67.7 | 15.0 | 88.5 |
| run 3 | 55.7 | 31.1 | 11.5 | 54.9 | 61.0 | 50.8 | 13.7 | 69.6 | 78.4 | 65.1 | 14.7 | 86.5 | 82.3 | 71.1 | 15.5 | 92.2 |
| run 4 | 52.4 | 33.3 | 10.6 | 53.5 | 73.0 | 59.3 | 14.0 | 80.2 | 80.2 | 67.4 | 14.5 | 88.3 | 76.6 | 65.6 | 15.3 | 86.4 |
| run 5 | 58.0 | 43.2 | 11.3 | 61.9 | 65.4 | 54.2 | 13.7 | 73.5 | 75.9 | 64.3 | 14.1 | 84.2 | 79.8 | 68.7 | 15.1 | 89.4 |
| Average | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | ||||||||||||||||
| run 1 | 70.7 | 46.8 | 13.7 | 72.5 | 65.2 | 50.6 | 14.2 | 72.1 | 72.3 | 55.0 | 14.9 | 78.6 | 74.8 | 60.4 | 15.8 | 83.4 |
| run 2 | 64.6 | 45.8 | 13.8 | 69.0 | 69.3 | 52.9 | 15.3 | 76.4 | 70.5 | 57.7 | 17.7 | 81.8 | 74.1 | 64.2 | 16.4 | 85.6 |
| run 3 | 64.4 | 51.1 | 15.1 | 72.9 | 65.7 | 53.6 | 15.8 | 75.5 | 74.8 | 64.6 | 16.9 | 86.6 | 74.3 | 61.8 | 17.2 | 85.3 |
| run 4 | 63.9 | 47.0 | 14.7 | 70.2 | 70.1 | 55.4 | 17.8 | 80.6 | 71.8 | 57.3 | 16.0 | 80.6 | 76.4 | 63.7 | 18.0 | 88.1 |
| run 5 | 63.7 | 50.7 | 14.4 | 71.6 | 71.2 | 55.8 | 15.6 | 79.1 | 74.1 | 61.6 | 15.8 | 83.7 | 74.4 | 61.9 | 17.5 | 85.7 |
| Average | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | ||||||||||||||||
| run 1 | 55.8 | 41.1 | 14.0 | 62.5 | 68.7 | 55.0 | 16.7 | 78.6 | 72.4 | 60.2 | 18.1 | 84.4 | 82.6 | 70.2 | 18.8 | 95.2 |
| run 2 | 57.0 | 43.2 | 12.9 | 63.0 | 68.4 | 55.9 | 15.2 | 77.4 | 76.0 | 61.4 | 17.1 | 85.8 | 78.4 | 66.9 | 18.7 | 91.4 |
| run 3 | 61.4 | 46.7 | 14.5 | 68.6 | 68.9 | 53.8 | 14.0 | 75.4 | 76.6 | 63.8 | 17.0 | 87.2 | 82.8 | 73.6 | 17.9 | 96.1 |
| run 4 | 56.1 | 42.4 | 14.1 | 63.4 | 73.1 | 60.3 | 16.6 | 83.3 | 80.6 | 63.9 | 17.1 | 89.4 | 82.5 | 71.3 | 19.0 | 95.9 |
| run 5 | 57.8 | 43.5 | 13.8 | 64.5 | 67.7 | 51.5 | 15.7 | 75.3 | 77.7 | 65.0 | 16.8 | 88.2 | 84.6 | 74.2 | 18.7 | 98.1 |
| Average | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Goal: | [restaurant]: {informable: food italian area east} {requestable: reference} |
| User (Turn 0): | i am looking for a jamaican restaurant in the east. |
| Golden R: | there are no [value_food] restaurants in the [value_area]. would you like to try another food type or area? |
| Golden BS: | [restaurant] food jamaican area east |
| Golden AS: | [restaurant] [nooffer] food area [request] area food |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another area or another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] area food |
| User (Turn 1): | i am looking for a place that serves jamaican food in the east. if not, italian will do. |
| Golden R: | there is [value_choice] [value_food] place in the [value_area], [value_name]. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [inform] choice name food area |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] food |
| User (Turn 2): | ok, i would like to make a reservation there. |
| Golden R: | i can do that for you, i just need details such as day, time, and the size of your party. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [offerbook] [request] time day people |
| Mars-G R: | what day and time would you like to dine? and how many people will be in your party? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [request] day time people |
| User (Turn 3): | i need a table for 7 at 14:45 on friday. |
| Golden R: | i booked you a table for [value_people] on [value_day] at [value_time] at [value_name], reference: [value_reference]. can i assist you further today? |
| Golden BS: | [restaurant] food italian area east time 14:45 day friday people 7 |
| Golden AS: | [restaurant] [offerbooked] time reference name day people [general] [reqmore] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value_reference]. |
| Mars-G BS: | [restaurant] food jamaican area east time 14:45 day friday people 7 |
| Mars-G AS: | [restaurant] [offerbooked] reference |
| Goal: | [attraction]: {informable: area centre} |
| User (Turn 3): | i am also looking for places to go in town. perhaps an attraction in the city centre. |
| Golden R: | i have the [value_name] located at [value_address] and it's [value_price] entrance. |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [recommend] address price name |
| Mars-G R: | there are [value_choice] attractions in the city centre. do you have a preference for type? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 4): | that sounds perfect. thanks! |
| Golden R: | can i help you with anything else? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [general] [reqmore] |
| Mars-G R: | is there anything else i can help you with? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [general] [reqmore] |
| Goal: | [attraction]: {informable: area south} {requestable: postcode} |
| User (Turn 4): | i would also like to go to some attractions in the same area as the restaurant. |
| Golden R: | there are [value_choice] attractions in the area. there's a [value_type], [value_type], [value_type], [value_type], [value_type], [value_type]... do any of these appeal to you in particular? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] type [inform] choice |
| Mars-G R: | there are [value_choice] attractions in the area. do you have a preference for type? |
| Mars-G BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 5): | restaurant is what i am looking for, entrance fee and postcode please. get me a taxi from the attraction by 10.15, get me its contact number and car type. |
| Golden R: | just to clarify you are looking for another restaurant and not an attraction? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] [restaurant] [select] |
| Mars-G R: | [value_name] is located in postcode [value_postcode]. the phone number is [value_phone]. the entrance fee is [value_price]. can i help you with anything else? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] price phone name postcode [general] [reqmore] |
| User (Turn 6): | oh, yeah, i do want information on attractions! in the same area. after that, i will need a taxi. |
| Golden R: | if you like parks the [value_name] has [value_price] admission and is in the [value_postcode] postcode. or i can tell you about the other attractions? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [recommend] name postcode price [general] [reqmore] |
| Mars-G R: | there are [value_choice] attractions in the [value_area]. [value_choice] are [value_type], [value_choice] are [value_type], and [value_choice] is [value_type]. do you have a preference? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 7): | that sounds wonderful, is it also in the south part of town? if so, i would like a taxi from there to my hotel. |
| Golden R: | yes, [value_name] is in the [value_area]. in order to help you book a taxi between the park and your hotel, i need to know what hotel you're at. |
| Golden BS: | [taxi] departure wandlebury country park [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [inform] area name [taxi] [request] destination |
| Mars-G R: | yes, [value_name] is in the [value_area]. |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area name |
| Goal: | [restaurant]: {informable: pricerange moderate area west} {requestable: reference} |
| User (Turn 0): | hi! i am planning a trip in cambridge and i am looking for a restaurant in the west part of town. |
| Golden R: | there are [value_choice] restaurants in the [value_area]. is there any kind of food you're looking for? |
| Golden BS: | [restaurant] area west |
| Golden AS: | [restaurant] [inform] choice area [request] food |
| Mars-G R: | there are [value_choice] restaurants in the [value_area] part of town. |
| do you have a price range or type of food in mind? | |
| Mars-G BS: | [restaurant] area west |
| Mars-G AS: | [restaurant] [inform] choice area [request] food price |
| User (Turn 1): | how about you suggest one a book a table for 4 on tuesday at 15:45 and i need the reference number too please |
| Golden R: | i was able to book [value_name], reference number is [value_reference] |
| Golden BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Golden AS: | [restaurant] [offerbooked] name reference |
| Mars-G R: | i have booked you at [value_name]. your reference number is [value_reference]. |
| is there anything else i can help you with? | |
| Mars-G BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Mars-G AS: | [restaurant] [offerbooked] reference name [general] [reqmore] |
| Goal: | [train]: {informable: leave 17:00 destination birmingham new street day monday departure cambridge} {requestable: reference} |
| User (Turn 0): | what train do i take to get to birmingham new street? i need to leave after 17:00. |
| Golden R: | there are trains departing hourly at 1 minute past the hour [value Leave]. all are [value_Price], and take [value_time]. |
| Golden BS: | [train] destination birmingham new street leave 17:00 |
| Golden AS: | [train] [inform] price leave time |
| Mars-G R: | where will you be departing from and what day would you like to travel ? |
| Mars-G BS: | [train] destination birmingham new street leave 17:00 |
| Mars-G AS: | [train] [request] departure day |
| User (Turn 1): | i am looking to travel on monday. the first train at 17:01 would be fine. can i get that train info? |
| Golden R: | sure the trainid is [value_id]. it leaves at [value Leave] and arrives at [value_arry]. would you like to make a reservation? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Golden AS: | [train] [inform] arrive leave id [offerbook] |
| Mars-G R: | [value_id] leaves at [value Leave] and arrives at [value_arry]. would you like me to book that for you? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Mars-G AS: | [train] [inform] arrive leave id [offerbook] |
| User (Turn 2): | yes please make a reservation and give me the reference number. |
| Golden R: | your reservation is book. the reference number is [value.Reference]. can i assist you with anything else today? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Golden AS: | [train] [offerbooked] reference [general] [reqmore] |
| Mars-G R: | how many tickets do you need? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Mars-G AS: | [train] [request] people |
| Goal: | [restaurant]: {informable: food international pricerange moderate area centre} {requestable: reference} |
| User (Turn 0): | what restaurants in the centre serve international cuisine? |
| Golden R: | the [value_name] and the [value_name] serve [value_food] food and are in the [value_area] of town. |
| Golden BS: | [restaurant] food international area centre |
| Golden AS: | [restaurant] [inform] food area name |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. do you have a price range in mind? |
| Mars-G BS: | [restaurant] food international area centre |
| Mars-G AS: | [restaurant] [inform] choice [request] price |
| User (Turn 1): | how about a place in the moderate price range? |
| Golden R: | [value_choice] of the named restaurants are in the [value_pricerange] price range. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [inform] choice price |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. [value_name] and [value_name]. would you like me to book one of these for you? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre |
| Mars-G AS: | [restaurant] [inform] choice name [offerbook] |
| User (Turn 2): | ok, can you book a table for 6 at 12:00 on tuesday at the varsity restaurant? i will need a reference number too, please. |
| Golden R: | certainly. i will have that reference number for you in just one second. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [offerbook] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value-reference]. is there anything else i can help you with? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [restaurant] [offerbooked] reference [general] [reqmore] |
| User (Turn 3): | thanks. looking forward to it. |
| Golden R: | yes i can. i have booked you a table for [value_time] on [value_day] for [value_people] people. your reference number is [value_reference]. anything else i can help you with? |
| Golden BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Golden AS: | [restaurant] [offerbooked] time reference people day |
| Mars-G R: | you're welcome. have a great day! |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [general] [welcome] [bye] |
| Goal: | [attraction]: {informable: area north} {requestable: postcode phone} |
| User (Turn 0): | i am looking for something to do in town. are there any interesting attraction in the north part of town? |
| Golden R: | yes, there are [value_choice] places: i see a [value_type], a [value_type], and two [value_type]. would one of these work for you? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] type choice [general] [reqmore] |
| Mars-G R: | there are [value_choice] [value_type] and [value_type] in the [value_area]. do you have a preference? |
| Mars-G BS: | [attraction] area north |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 1): | ooh, a pool would be fun. can you give me their contact info? |
| Golden R: | great! [value_address] is located at [value_address], kings hedges postcode is [value_postcode] and their phone is [value_phone]. is there anything else i can help you with? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] name postcode phone address [general] [reqmore] |
| Mars-G R: | [value_name] is located at [value_address]. their phone number is [value_phone]. |
| Mars-G BS: | [attraction] area north type swimming pool |
| Mars-G AS: | [attraction] [inform] name phone address |
| Goal: | [train]: {informable: destination stansted airport day friday arrive 15:30 departure cambridge} {requestable: reference} |
| User (Turn 3): | i will also need a train leaving cambridge. |
| Golden R: | sure, what is your arrival information? |
| Golden BS: | [train] departure cambridge |
| Golden AS: | [train] [request] destination |
| Mars-G R: | what day will you be traveling, and where will you be traveling to? |
| Mars-G BS: | [train] departure cambridge |
| Mars-G AS: | [train] [request] destination day |
| User (Turn 4): | i need to go to stansted airport on friday. |
| Golden R: | we have [value_choice] different departure times. any time table your looking at? |
| Golden BS: | [train] destination stansted airport departure cambridge |
| Golden AS: | [train] [inform] choice [request] leave |
| Mars-G R: | there are [value_choice] trains that meet your criteria. do you have a preferred departure or arrival time? |
| Mars-G BS: | [train] destination stansted airport day friday departure cambridge |
| Mars-G AS: | [train] [inform] choice [request] leave arrive |
| User (Turn 5): | i need to be in stansted airport by 15:30 please. |
| Golden R: | the soonest it will get you there is [value arriv] your reference number is [value/reference], anything else? |
| Golden BS: | [train] destination stansted airport arrive 15:30 departure cambridge people 1 |
| Golden AS: | [train] [offerbooked] arrive reference [general] [reqmore] |
| Mars-G R: | [value_id] will arrive at [value arriv]. would you like me to book that for you? |
| Mars-G BS: | [train] destination stansted airport day friday arrive 15:30 departure cambridge |
| Mars-G AS: | [train] [inform] arrive id [offerbook] |
| Goal: | [attraction]: {informable: type architecture area centre} {requestable: address phone postcode} |
| User (Turn 0): | i want to find a place to go in the centre to enjoy some entertainment. can you help? |
| Golden R: | there are [value_choice] attraction types in the [value_area], did you have one in mind? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [inform] area choice [request] name |
| Mars-G R: | i am sorry, there are no [value_type] attractions in the [value_area]. would you like to try another area? |
| Mars-G BS: | [attraction] area centre type entertainment |
| Mars-G AS: | [attraction] [nooffer] area type [request] area |
| User (Turn 1): | yes, i am looking for something with the type of architecture. thanks! |
| Golden R: | i would recommend [value_name] as a great place to see interesting [value_type]. |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [recommend] name type |
| Mars-G R: | there are [value_choice] [value_type] attractions in the [value_area]. i recommend [value_name]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [recommend] name [inform] type choice area |
| User (Turn 2): | okay, could you give me the address and phone number? |
| Golden R: | sure. the address is [value_address]. the phone number is [value_phone]. |
| is there anything else i can help you with today? | |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [inform] address phone [general] [reqmore] |
| Mars-G R: | the address is [value_address] and the phone number is [value_phone]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [inform] address phone |
| Model | Pre-trained | Extra corpora | Dialog state tracking | Response Generation | ||||
| Joint Accuracy | Act F1 | Inform | Success | BLEU | Combined | |||
| DAMD (Zhang et al., 2020b) | - | no | - | - | 57.9 | 47.6 | 16.4 | 69.2 |
| LABES (Zhang et al., 2020a) | - | no | - | - | 68.5 | 58.1 | 18.9 | 82.2 |
| AuGPT (Kulhánek et al., 2021) | GPT-2 | yes | - | - | 76.6 | 60.5 | 16.8 | 85.4 |
| MinTL (Lin et al., 2020) | T5-small | no | 51.2 | - | 73.7 | 65.4 | 19.4 | 89.0 |
| SOLOIST (Peng et al., 2021) | GPT-2 | yes | 53.2 | - | 82.3 | 72.4 | 13.6 | 91.0 |
| DoTS (Jeon and Lee, 2021) | BERT-base | no | - | - | 80.4 | 68.7 | 16.8 | 91.4 |
| UBAR (Yang et al., 2021) | DistilGPT2 | no | 52.6 | - | 83.4 | 70.3 | 17.6 | 94.5 |
| PPTOD (Su et al., 2021) | T5-base | yes | 53.4 | - | 83.1 | 72.7 | 18.2 | 96.1 |
| BORT (Sun et al., 2022) | T5-small | no | 54.0 | - | 85.5 | 77.4 | 17.9 | 99.4 |
| MTTOD (Lee, 2021) | T5-base | no | 53.6 | - | 85.9 | 76.5 | 19.0 | 100.2 |
| GALAXY (He et al., 2022) | UniLM-base | yes | - | - | 85.4 | 75.7 | 19.6 | 100.2 |
| Baseline | T5-small | no | 53.8 | 53.0 | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-P | T5-small | no | 54.4 | 53.9 | 86.6 | 75.5 | 19.6 | 100.7 |
| Mars-G | T5-small | no | 55.1 | 53.7 | 88.9 | 78.0 | 19.9 | 103.4 |
| Model | Match | Success F1 | BLEU | Combined |
| Sequicity (Lei et al., 2018) | 92.7 | 85.4 | 25.3 | 114.4 |
| LABES (Zhang et al., 2020a) | 96.4 | 82.3 | 25.6 | 115.0 |
| SOLOIST (Peng et al., 2021) | 94.7 | 87.1 | 25.5 | 116.4 |
| GALAXY (He et al., 2022) | 98.5 | 87.7 | 24.2 | 117.3 |
| Mars-P | 97.0 | 87.2 | 25.9 | 118.0 |
| Mars-G | 96.2 | 89.6 | 26.1 | 119.0 |
| Model | Joint Accuracy |
| TRADE (Wu et al., 2019) | 36.1 |
| BART-CSP (Moradshahi et al., 2021) | 53.6 |
| GEEX (Li et al., 2021) | 54.7 |
| Mars-P | 59.3 |
| Mars-G | 59.8 |
| Model | Inform | Success | BLEU | Combined |
| Mars-G | 88.9 | 78.0 | 19.9 | 103.4 |
| w/o BSC | 88.3 | 76.6 | 19.5 | 102.0 |
| w/o ASC | 86.3 | 75.1 | 19.7 | 100.4 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| w/o BSC | 85.4 | 75.0 | 19.7 | 99.9 |
| w/o ASC | 83.7 | 73.0 | 19.8 | 98.2 |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Model | Inform | Success | BLEU |
| End-to-end model | 83.2 | 70.3 | 19.4 |
| w/ oracle state | 90.8 | 87.4 | 30.6 |
| Reference Corpus | 93.7 | 90.9 | 100.0 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.797 | 1.018 |
| Baseline | 0.844 | 1.017 |
| Mars-P | 0.340 | 0.542 |
| Mars-G | 1.996 | 1.993 |
| Model | Context&Belief State | Context&Action State |
| T5 | 0.555 | 0.807 |
| Baseline | 0.598 | 0.699 |
| Mars-P | 0.042 | 0.046 |
| Mars-G | 1.993 | 1.987 |
| Model | Inform | Success | BLEU | Combined |
| Baseline | 83.2 | 70.3 | 19.4 | 96.2 |
| Mars-variant | 85.7 | 74.8 | 19.6 | 99.9 |
| Mars-P | 86.6 | 75.5 | 19.6 | 100.7 |
| Model | 5% | 10% | 20% | 50% | ||||||||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| DAMD | ||||||||||||||||
| run 1 | 35.4 | 17.2 | 10.9 | 37.2 | 41.3 | 23.7 | 12.4 | 44.9 | 51.8 | 31.9 | 14.1 | 56.0 | 60.1 | 44.2 | 15.6 | 67.8 |
| run 2 | 40.8 | 20.9 | 12.0 | 42.9 | 41.5 | 25.3 | 11.2 | 44.6 | 50.4 | 32.4 | 13.8 | 55.2 | 54.7 | 39.5 | 14.8 | 61.9 |
| run 3 | 38.5 | 14.5 | 10.6 | 37.1 | 40.0 | 23.9 | 12.3 | 44.3 | 42.5 | 26.8 | 15.4 | 50.1 | 59.1 | 45.7 | 15.1 | 67.5 |
| run 4 | 35.4 | 16.5 | 10.2 | 36.2 | 42.3 | 20.2 | 12.0 | 43.3 | 46.1 | 29.2 | 14.0 | 51.7 | 57.2 | 43.0 | 16.1 | 66.2 |
| run 5 | 34.0 | 17.6 | 12.4 | 38.2 | 39.5 | 21.9 | 13.0 | 43.7 | 50.8 | 31.3 | 13.6 | 54.7 | 63.1 | 49.3 | 17.1 | 73.3 |
| Average | 36.8 | 17.3 | 11.2 | 38.3 | 40.9 | 23.0 | 12.2 | 44.2 | 48.3 | 30.3 | 14.2 | 53.5 | 58.8 | 44.3 | 15.7 | 67.3 |
| MinTL | ||||||||||||||||
| run 1 | 54.4 | 41.1 | 14.2 | 62.0 | 55.8 | 44.0 | 15.3 | 65.2 | 62.5 | 54.2 | 17.3 | 75.7 | 71.7 | 62.7 | 16.9 | 84.1 |
| run 2 | 54.8 | 36.8 | 13.6 | 59.4 | 51.6 | 42.0 | 15.7 | 62.5 | 65.8 | 56.3 | 15.5 | 76.6 | 67.4 | 59.7 | 18.7 | 82.3 |
| run 3 | 53.3 | 39.3 | 14.2 | 60.5 | 55.1 | 44.7 | 16.1 | 66.0 | 68.0 | 59.0 | 16.6 | 80.1 | 70.6 | 62.6 | 17.5 | 84.1 |
| run 4 | 52.4 | 37.1 | 13.8 | 58.6 | 58.4 | 47.3 | 15.2 | 68.1 | 58.3 | 48.4 | 14.4 | 67.8 | 68.9 | 61.3 | 18.2 | 83.3 |
| run 5 | 47.5 | 36.3 | 13.8 | 55.7 | 56.8 | 46.4 | 15.9 | 67.5 | 66.9 | 56.8 | 17.0 | 78.9 | 73.1 | 64.5 | 18.6 | 87.4 |
| Average | 52.5 | 38.1 | 13.9 | 59.2 | 55.5 | 44.9 | 15.6 | 65.8 | 64.3 | 54.9 | 16.2 | 75.8 | 70.3 | 62.2 | 18.0 | 84.3 |
| UBAR | ||||||||||||||||
| run 1 | 37.4 | 23.0 | 11.6 | 41.8 | 52.3 | 34.8 | 13.0 | 56.6 | 61.7 | 45.7 | 15.9 | 69.6 | 77.2 | 61.5 | 15.5 | 84.9 |
| run 2 | 33.3 | 20.6 | 11.2 | 38.2 | 48.5 | 35.9 | 14.5 | 56.7 | 63.4 | 47.8 | 15.5 | 71.1 | 78.0 | 63.8 | 16.9 | 87.8 |
| run 3 | 40.0 | 23.1 | 11.7 | 43.3 | 50.3 | 33.2 | 13.6 | 55.4 | 67.8 | 50.0 | 13.1 | 72.0 | 77.4 | 64.6 | 16.2 | 87.2 |
| run 4 | 38.2 | 22.4 | 10.7 | 41.0 | 52.5 | 34.6 | 12.5 | 56.1 | 68.3 | 51.7 | 14.4 | 74.4 | 78.5 | 64.1 | 16.8 | 88.1 |
| run 5 | 38.0 | 21.3 | 11.3 | 41.0 | 47.8 | 32.3 | 13.7 | 53.8 | 66.2 | 48.3 | 13.8 | 71.1 | 76.8 | 62.4 | 16.2 | 85.8 |
| Average | 37.4 | 22.1 | 11.3 | 41.1 | 50.3 | 34.2 | 13.5 | 55.8 | 65.5 | 48.7 | 14.5 | 71.6 | 77.6 | 63.3 | 16.3 | 86.8 |
| MTTOD | ||||||||||||||||
| run 1 | 51.4 | 37.5 | 12.0 | 56.5 | 70.9 | 58.0 | 13.8 | 78.3 | 71.1 | 59.0 | 14.2 | 79.3 | 74.7 | 64.4 | 15.2 | 84.8 |
| run 2 | 53.8 | 41.7 | 11.3 | 59.1 | 64.1 | 53.7 | 13.8 | 72.7 | 69.5 | 60.7 | 14.0 | 79.1 | 79.3 | 67.7 | 15.0 | 88.5 |
| run 3 | 55.7 | 31.1 | 11.5 | 54.9 | 61.0 | 50.8 | 13.7 | 69.6 | 78.4 | 65.1 | 14.7 | 86.5 | 82.3 | 71.1 | 15.5 | 92.2 |
| run 4 | 52.4 | 33.3 | 10.6 | 53.5 | 73.0 | 59.3 | 14.0 | 80.2 | 80.2 | 67.4 | 14.5 | 88.3 | 76.6 | 65.6 | 15.3 | 86.4 |
| run 5 | 58.0 | 43.2 | 11.3 | 61.9 | 65.4 | 54.2 | 13.7 | 73.5 | 75.9 | 64.3 | 14.1 | 84.2 | 79.8 | 68.7 | 15.1 | 89.4 |
| Average | 54.3 | 37.4 | 11.3 | 57.2 | 66.9 | 55.2 | 13.8 | 74.9 | 75.0 | 63.3 | 14.3 | 83.5 | 78.5 | 67.5 | 15.2 | 88.2 |
| PPTOD | ||||||||||||||||
| run 1 | 70.7 | 46.8 | 13.7 | 72.5 | 65.2 | 50.6 | 14.2 | 72.1 | 72.3 | 55.0 | 14.9 | 78.6 | 74.8 | 60.4 | 15.8 | 83.4 |
| run 2 | 64.6 | 45.8 | 13.8 | 69.0 | 69.3 | 52.9 | 15.3 | 76.4 | 70.5 | 57.7 | 17.7 | 81.8 | 74.1 | 64.2 | 16.4 | 85.6 |
| run 3 | 64.4 | 51.1 | 15.1 | 72.9 | 65.7 | 53.6 | 15.8 | 75.5 | 74.8 | 64.6 | 16.9 | 86.6 | 74.3 | 61.8 | 17.2 | 85.3 |
| run 4 | 63.9 | 47.0 | 14.7 | 70.2 | 70.1 | 55.4 | 17.8 | 80.6 | 71.8 | 57.3 | 16.0 | 80.6 | 76.4 | 63.7 | 18.0 | 88.1 |
| run 5 | 63.7 | 50.7 | 14.4 | 71.6 | 71.2 | 55.8 | 15.6 | 79.1 | 74.1 | 61.6 | 15.8 | 83.7 | 74.4 | 61.9 | 17.5 | 85.7 |
| Average | 65.5 | 48.3 | 14.3 | 71.2 | 68.3 | 53.7 | 15.7 | 76.7 | 72.7 | 59.2 | 16.3 | 82.3 | 74.8 | 62.4 | 17.0 | 85.6 |
| Mars-G | ||||||||||||||||
| run 1 | 55.8 | 41.1 | 14.0 | 62.5 | 68.7 | 55.0 | 16.7 | 78.6 | 72.4 | 60.2 | 18.1 | 84.4 | 82.6 | 70.2 | 18.8 | 95.2 |
| run 2 | 57.0 | 43.2 | 12.9 | 63.0 | 68.4 | 55.9 | 15.2 | 77.4 | 76.0 | 61.4 | 17.1 | 85.8 | 78.4 | 66.9 | 18.7 | 91.4 |
| run 3 | 61.4 | 46.7 | 14.5 | 68.6 | 68.9 | 53.8 | 14.0 | 75.4 | 76.6 | 63.8 | 17.0 | 87.2 | 82.8 | 73.6 | 17.9 | 96.1 |
| run 4 | 56.1 | 42.4 | 14.1 | 63.4 | 73.1 | 60.3 | 16.6 | 83.3 | 80.6 | 63.9 | 17.1 | 89.4 | 82.5 | 71.3 | 19.0 | 95.9 |
| run 5 | 57.8 | 43.5 | 13.8 | 64.5 | 67.7 | 51.5 | 15.7 | 75.3 | 77.7 | 65.0 | 16.8 | 88.2 | 84.6 | 74.2 | 18.7 | 98.1 |
| Average | 57.6 | 43.4 | 13.9 | 64.4 | 69.4 | 55.3 | 15.6 | 78.0 | 76.7 | 62.9 | 17.2 | 87.0 | 82.2 | 71.2 | 18.6 | 95.3 |
| Goal: | [restaurant]: {informable: food italian area east} {requestable: reference} |
| User (Turn 0): | i am looking for a jamaican restaurant in the east. |
| Golden R: | there are no [value_food] restaurants in the [value_area]. would you like to try another food type or area? |
| Golden BS: | [restaurant] food jamaican area east |
| Golden AS: | [restaurant] [nooffer] food area [request] area food |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another area or another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] area food |
| User (Turn 1): | i am looking for a place that serves jamaican food in the east. if not, italian will do. |
| Golden R: | there is [value_choice] [value_food] place in the [value_area], [value_name]. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [inform] choice name food area |
| Mars-G R: | i am sorry, there are no [value_food] restaurants in the [value_area]. would you like to try another type of food? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [nooffer] food area [request] food |
| User (Turn 2): | ok, i would like to make a reservation there. |
| Golden R: | i can do that for you, i just need details such as day, time, and the size of your party. |
| Golden BS: | [restaurant] food italian area east |
| Golden AS: | [restaurant] [offerbook] [request] time day people |
| Mars-G R: | what day and time would you like to dine? and how many people will be in your party? |
| Mars-G BS: | [restaurant] food jamaican area east |
| Mars-G AS: | [restaurant] [request] day time people |
| User (Turn 3): | i need a table for 7 at 14:45 on friday. |
| Golden R: | i booked you a table for [value_people] on [value_day] at [value_time] at [value_name], reference: [value_reference]. can i assist you further today? |
| Golden BS: | [restaurant] food italian area east time 14:45 day friday people 7 |
| Golden AS: | [restaurant] [offerbooked] time reference name day people [general] [reqmore] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value_reference]. |
| Mars-G BS: | [restaurant] food jamaican area east time 14:45 day friday people 7 |
| Mars-G AS: | [restaurant] [offerbooked] reference |
| Goal: | [attraction]: {informable: area centre} |
| User (Turn 3): | i am also looking for places to go in town. perhaps an attraction in the city centre. |
| Golden R: | i have the [value_name] located at [value_address] and it's [value_price] entrance. |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [recommend] address price name |
| Mars-G R: | there are [value_choice] attractions in the city centre. do you have a preference for type? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 4): | that sounds perfect. thanks! |
| Golden R: | can i help you with anything else? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [general] [reqmore] |
| Mars-G R: | is there anything else i can help you with? |
| Mars-G BS: | [attraction] area centre |
| Mars-G AS: | [general] [reqmore] |
| Goal: | [attraction]: {informable: area south} {requestable: postcode} |
| User (Turn 4): | i would also like to go to some attractions in the same area as the restaurant. |
| Golden R: | there are [value_choice] attractions in the area. there's a [value_type], [value_type], [value_type], [value_type], [value_type], [value_type]... do any of these appeal to you in particular? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] type [inform] choice |
| Mars-G R: | there are [value_choice] attractions in the area. do you have a preference for type? |
| Mars-G BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] choice [request] type |
| User (Turn 5): | restaurant is what i am looking for, entrance fee and postcode please. get me a taxi from the attraction by 10.15, get me its contact number and car type. |
| Golden R: | just to clarify you are looking for another restaurant and not an attraction? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [select] [restaurant] [select] |
| Mars-G R: | [value_name] is located in postcode [value_postcode]. the phone number is [value_phone]. the entrance fee is [value_price]. can i help you with anything else? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] price phone name postcode [general] [reqmore] |
| User (Turn 6): | oh, yeah, i do want information on attractions! in the same area. after that, i will need a taxi. |
| Golden R: | if you like parks the [value_name] has [value_price] admission and is in the [value_postcode] postcode. or i can tell you about the other attractions? |
| Golden BS: | [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [recommend] name postcode price [general] [reqmore] |
| Mars-G R: | there are [value_choice] attractions in the [value_area]. [value_choice] are [value_type], [value_choice] are [value_type], and [value_choice] is [value_type]. do you have a preference? |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 7): | that sounds wonderful, is it also in the south part of town? if so, i would like a taxi from there to my hotel. |
| Golden R: | yes, [value_name] is in the [value_area]. in order to help you book a taxi between the park and your hotel, i need to know what hotel you're at. |
| Golden BS: | [taxi] departure wandlebury country park [attraction] area south [restaurant] food indian pricerange expensive area south |
| Golden AS: | [attraction] [inform] area name [taxi] [request] destination |
| Mars-G R: | yes, [value_name] is in the [value_area]. |
| Mars-G BS: | [taxi] leave 10.15 destination tandoori palace departure the nirala [attraction] area south [restaurant] food indian pricerange expensive area south |
| Mars-G AS: | [attraction] [inform] area name |
| Goal: | [restaurant]: {informable: pricerange moderate area west} {requestable: reference} |
| User (Turn 0): | hi! i am planning a trip in cambridge and i am looking for a restaurant in the west part of town. |
| Golden R: | there are [value_choice] restaurants in the [value_area]. is there any kind of food you're looking for? |
| Golden BS: | [restaurant] area west |
| Golden AS: | [restaurant] [inform] choice area [request] food |
| Mars-G R: | there are [value_choice] restaurants in the [value_area] part of town. |
| do you have a price range or type of food in mind? | |
| Mars-G BS: | [restaurant] area west |
| Mars-G AS: | [restaurant] [inform] choice area [request] food price |
| User (Turn 1): | how about you suggest one a book a table for 4 on tuesday at 15:45 and i need the reference number too please |
| Golden R: | i was able to book [value_name], reference number is [value_reference] |
| Golden BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Golden AS: | [restaurant] [offerbooked] name reference |
| Mars-G R: | i have booked you at [value_name]. your reference number is [value_reference]. |
| is there anything else i can help you with? | |
| Mars-G BS: | [restaurant] area west time 15:45 day tuesday people 4 |
| Mars-G AS: | [restaurant] [offerbooked] reference name [general] [reqmore] |
| Goal: | [train]: {informable: leave 17:00 destination birmingham new street day monday departure cambridge} {requestable: reference} |
| User (Turn 0): | what train do i take to get to birmingham new street? i need to leave after 17:00. |
| Golden R: | there are trains departing hourly at 1 minute past the hour [value Leave]. all are [value_Price], and take [value_time]. |
| Golden BS: | [train] destination birmingham new street leave 17:00 |
| Golden AS: | [train] [inform] price leave time |
| Mars-G R: | where will you be departing from and what day would you like to travel ? |
| Mars-G BS: | [train] destination birmingham new street leave 17:00 |
| Mars-G AS: | [train] [request] departure day |
| User (Turn 1): | i am looking to travel on monday. the first train at 17:01 would be fine. can i get that train info? |
| Golden R: | sure the trainid is [value_id]. it leaves at [value Leave] and arrives at [value_arry]. would you like to make a reservation? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Golden AS: | [train] [inform] arrive leave id [offerbook] |
| Mars-G R: | [value_id] leaves at [value Leave] and arrives at [value_arry]. would you like me to book that for you? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge leave 17:00 |
| Mars-G AS: | [train] [inform] arrive leave id [offerbook] |
| User (Turn 2): | yes please make a reservation and give me the reference number. |
| Golden R: | your reservation is book. the reference number is [value.Reference]. can i assist you with anything else today? |
| Golden BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Golden AS: | [train] [offerbooked] reference [general] [reqmore] |
| Mars-G R: | how many tickets do you need? |
| Mars-G BS: | [train] destination birmingham new street day monday departure cambridge people 1 leave 17:00 |
| Mars-G AS: | [train] [request] people |
| Goal: | [restaurant]: {informable: food international pricerange moderate area centre} {requestable: reference} |
| User (Turn 0): | what restaurants in the centre serve international cuisine? |
| Golden R: | the [value_name] and the [value_name] serve [value_food] food and are in the [value_area] of town. |
| Golden BS: | [restaurant] food international area centre |
| Golden AS: | [restaurant] [inform] food area name |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. do you have a price range in mind? |
| Mars-G BS: | [restaurant] food international area centre |
| Mars-G AS: | [restaurant] [inform] choice [request] price |
| User (Turn 1): | how about a place in the moderate price range? |
| Golden R: | [value_choice] of the named restaurants are in the [value_pricerange] price range. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [inform] choice price |
| Mars-G R: | there are [value_choice] restaurants that meet your criteria. [value_name] and [value_name]. would you like me to book one of these for you? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre |
| Mars-G AS: | [restaurant] [inform] choice name [offerbook] |
| User (Turn 2): | ok, can you book a table for 6 at 12:00 on tuesday at the varsity restaurant? i will need a reference number too, please. |
| Golden R: | certainly. i will have that reference number for you in just one second. |
| Golden BS: | [restaurant] food international pricerange moderate area centre |
| Golden AS: | [restaurant] [offerbook] |
| Mars-G R: | booking was successful. the table will be reserved for 15 minutes. reference number is: [value-reference]. is there anything else i can help you with? |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [restaurant] [offerbooked] reference [general] [reqmore] |
| User (Turn 3): | thanks. looking forward to it. |
| Golden R: | yes i can. i have booked you a table for [value_time] on [value_day] for [value_people] people. your reference number is [value_reference]. anything else i can help you with? |
| Golden BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Golden AS: | [restaurant] [offerbooked] time reference people day |
| Mars-G R: | you're welcome. have a great day! |
| Mars-G BS: | [restaurant] food international pricerange moderate area centre time 12:00 day tuesday people 6 |
| Mars-G AS: | [general] [welcome] [bye] |
| Goal: | [attraction]: {informable: area north} {requestable: postcode phone} |
| User (Turn 0): | i am looking for something to do in town. are there any interesting attraction in the north part of town? |
| Golden R: | yes, there are [value_choice] places: i see a [value_type], a [value_type], and two [value_type]. would one of these work for you? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] type choice [general] [reqmore] |
| Mars-G R: | there are [value_choice] [value_type] and [value_type] in the [value_area]. do you have a preference? |
| Mars-G BS: | [attraction] area north |
| Mars-G AS: | [attraction] [inform] area type choice [request] type |
| User (Turn 1): | ooh, a pool would be fun. can you give me their contact info? |
| Golden R: | great! [value_address] is located at [value_address], kings hedges postcode is [value_postcode] and their phone is [value_phone]. is there anything else i can help you with? |
| Golden BS: | [attraction] area north |
| Golden AS: | [attraction] [inform] name postcode phone address [general] [reqmore] |
| Mars-G R: | [value_name] is located at [value_address]. their phone number is [value_phone]. |
| Mars-G BS: | [attraction] area north type swimming pool |
| Mars-G AS: | [attraction] [inform] name phone address |
| Goal: | [train]: {informable: destination stansted airport day friday arrive 15:30 departure cambridge} {requestable: reference} |
| User (Turn 3): | i will also need a train leaving cambridge. |
| Golden R: | sure, what is your arrival information? |
| Golden BS: | [train] departure cambridge |
| Golden AS: | [train] [request] destination |
| Mars-G R: | what day will you be traveling, and where will you be traveling to? |
| Mars-G BS: | [train] departure cambridge |
| Mars-G AS: | [train] [request] destination day |
| User (Turn 4): | i need to go to stansted airport on friday. |
| Golden R: | we have [value_choice] different departure times. any time table your looking at? |
| Golden BS: | [train] destination stansted airport departure cambridge |
| Golden AS: | [train] [inform] choice [request] leave |
| Mars-G R: | there are [value_choice] trains that meet your criteria. do you have a preferred departure or arrival time? |
| Mars-G BS: | [train] destination stansted airport day friday departure cambridge |
| Mars-G AS: | [train] [inform] choice [request] leave arrive |
| User (Turn 5): | i need to be in stansted airport by 15:30 please. |
| Golden R: | the soonest it will get you there is [value arriv] your reference number is [value/reference], anything else? |
| Golden BS: | [train] destination stansted airport arrive 15:30 departure cambridge people 1 |
| Golden AS: | [train] [offerbooked] arrive reference [general] [reqmore] |
| Mars-G R: | [value_id] will arrive at [value arriv]. would you like me to book that for you? |
| Mars-G BS: | [train] destination stansted airport day friday arrive 15:30 departure cambridge |
| Mars-G AS: | [train] [inform] arrive id [offerbook] |
| Goal: | [attraction]: {informable: type architecture area centre} {requestable: address phone postcode} |
| User (Turn 0): | i want to find a place to go in the centre to enjoy some entertainment. can you help? |
| Golden R: | there are [value_choice] attraction types in the [value_area], did you have one in mind? |
| Golden BS: | [attraction] area centre |
| Golden AS: | [attraction] [inform] area choice [request] name |
| Mars-G R: | i am sorry, there are no [value_type] attractions in the [value_area]. would you like to try another area? |
| Mars-G BS: | [attraction] area centre type entertainment |
| Mars-G AS: | [attraction] [nooffer] area type [request] area |
| User (Turn 1): | yes, i am looking for something with the type of architecture. thanks! |
| Golden R: | i would recommend [value_name] as a great place to see interesting [value_type]. |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [recommend] name type |
| Mars-G R: | there are [value_choice] [value_type] attractions in the [value_area]. i recommend [value_name]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [recommend] name [inform] type choice area |
| User (Turn 2): | okay, could you give me the address and phone number? |
| Golden R: | sure. the address is [value_address]. the phone number is [value_phone]. |
| is there anything else i can help you with today? | |
| Golden BS: | [attraction] area centre type architecture |
| Golden AS: | [attraction] [inform] address phone [general] [reqmore] |
| Mars-G R: | the address is [value_address] and the phone number is [value_phone]. |
| Mars-G BS: | [attraction] area centre type architecture |
| Mars-G AS: | [attraction] [inform] address phone |
| In-domain | Out-of-domain | |||||||||||||||
| MTDialogue | LS test-clean | LS test-other | Voxpopuli | WSJ | ConvAI | SLURP | ||||||||||
| WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | |||
| (1) | No rescoring | 9.47 | 14.63 | 6.75 | 8.07 | 11.98 | 15.61 | 11.06 | 10.33 | 8.16 | 8.75 | 5.89 | 9.00 | 24.91 | 29.53 | |
| (2) | GPT2-text | 9.32 | 14.37 | 6.45 | 7.78 | 11.70 | 15.11 | 10.72 | 9.94 | 7.64 | 8.40 | 5.76 | 8.66 | 24.91 | 29.53 | |
| (3) | BERT-text | 9.05 | 13.88 | 5.50 | 7.20 | 10.70 | 14.45 | 10.33 | 9.96 | 6.46 | 8.20 | 5.38 | 8.37 | 24.48 | 29.27 | |
| (4) | LAS rescoring | 9.27 | 14.15 | 6.7 | 7.99 | 11.97 | 15.59 | 11.02 | 10.21 | 8.01 | 8.55 | 5.81 | 8.84 | 24.91 | 29.53 | |
| (5) | Multi-modal-GPT2 | 9.24 | 14.17 | 6.35 | 7.69 | 11.54 | 14.93 | 10.56 | 9.83 | 7.55 | 8.20 | 5.69 | 8.59 | 24.89 | 29.40 | |
| (6) | MATE-NA | 9.05 | 13.90 | 5.55 | 7.29 | 10.75 | 14.51 | 10.34 | 9.92 | 6.49 | 8.10 | 5.40 | 8.36 | 24.46 | 29.24 | |
| (7) | MATE-MSE | 7.49 | 11.41 | 5.22 | 6.95 | 10.31 | 13.97 | 10.10 | 9.62 | 6.10 | 7.65 | 5.07 | 7.92 | 23.84 | 28.24 | |
| (8) | MATE (ours) | 7.64 | 11.70 | 5.16 | 6.84 | 10.30 | 13.81 | 9.91 | 9.47 | 6.01 | 7.46 | 5.10 | 7.91 | 23.77 | 28.14 | |
| Parameter-Efficient Tuning | ||||||||||||||||
| (9) | Frozen-ME | 9.21 | 14.22 | 5.57 | 7.34 | 10.82 | 14.65 | 10.37 | 9.80 | 6.55 | 8.15 | 5.42 | 8.34 | 24.39 | 29.13 | |
| (10) | WavLM-adapter | 9.15 | 14.02 | 5.58 | 7.41 | 10.81 | 14.69 | 10.23 | 9.86 | 6.52 | 8.05 | 5.47 | 8.39 | 24.56 | 29.27 | |
| (11) | ME-adapter | 9.19 | 14.12 | 5.56 | 7.43 | 10.79 | 14.63 | 10.09 | 9.60 | 6.43 | 8.20 | 5.42 | 8.35 | 24.34 | 29.08 |
| Scenario | Action | Entity | |
| No rescoring | 78.01 | 72.53 | 53.23 |
| GPT2-text | 78.01 | 72.53 | 53.23 |
| Multi-modal-GPT2 | 78.07 | 72.65 | 53.26 |
| BERT-text | 77.72 | 72.45 | 53.37 |
| MATE | 78.76 | 73.70 | 54.26 |
| Dataset | Utterance | |
| SLURP | Ground Truth | remove tuesday alarm of nine a m |
| Rescored 1-best by BERT-text | move to alarm of nine a m | |
| Rescored 1-best by MATE | remove tuesday alarm at nine a m | |
| Ground Truth | hoover the hallway | |
| Rescored 1-best by BERT-text | who in the hallway | |
| Rescored 1-best by MATE | hoover the hallway | |
| Ground Truth | cancel business meeting on wednesday | |
| Rescored 1-best by BERT-text | council business meeting on wednesday | |
| Rescored 1-best by MATE | cancel business meeting on wednesday | |
| Ground Truth | can you let delta know i am never using them again | |
| Rescored 1-best by BERT-text | can you let doctor know i am never using them again | |
| Rescored 1-best by MATE | can you let delta know i am never using them again | |
| Ground Truth | i want to play fifa seventeen | |
| Rescored 1-best by BERT-text | i want to leave for seventeen | |
| Rescored 1-best by MATE | i want to play fifa seventeen | |
| Ground Truth | what do you know about fringe in edinburgh next year | |
| Rescored 1-best by BERT-text | what do you know about french in edinburgh next year | |
| Rescored 1-best by MATE | what do you know about fringe in edinburgh next year | |
| Voxpopuli | Ground Truth | for example the report talks about the rule of law and corruption |
| Rescored 1-best by BERT-text | for example the report talks about the rule of law on corruption | |
| Rescored 1-best by MATE | for example the report talks about the rule of law and corruption | |
| Ground Truth | i have met them they are young capable and visionary | |
| Rescored 1-best by BERT-text | i have met them they are young capable and missionary | |
| Rescored 1-best by MATE | i have met them they are young capable and visionary | |
| MTDialogue | Ground Truth | it's muffled |
| Rescored 1-best by BERT-text | it's muff | |
| Rescored 1-best by MATE | it's muffled | |
| Ground Truth | how much she got to pay | |
| Rescored 1-best by BERT-text | how much he got to pay | |
| Rescored 1-best by MATE | how much she got to pay | |
| ConvAI | Ground Truth | why did the noodle box in greensborough fail its health inspection |
| Rescored 1-best by BERT-text | why did the noodle box in greensboro fail its health inspection | |
| Rescored 1-best by MATE | why did the noodle box in greensborough fail its health inspection | |
| Ground Truth | tell me about duty free shopping | |
| Rescored 1-best by BERT-text | tell me about duty free shop | |
| Rescored 1-best by MATE | tell me about duty free shopping |
| In-domain | Out-of-domain | |||||||||||||||
| MTDialogue | LS test-clean | LS test-other | Voxpopuli | WSJ | ConvAI | SLURP | ||||||||||
| WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | |||
| (1) | No rescoring | 9.47 | 14.63 | 6.75 | 8.07 | 11.98 | 15.61 | 11.06 | 10.33 | 8.16 | 8.75 | 5.89 | 9.00 | 24.91 | 29.53 | |
| (2) | GPT2-text | 9.32 | 14.37 | 6.45 | 7.78 | 11.70 | 15.11 | 10.72 | 9.94 | 7.64 | 8.40 | 5.76 | 8.66 | 24.91 | 29.53 | |
| (3) | BERT-text | 9.05 | 13.88 | 5.50 | 7.20 | 10.70 | 14.45 | 10.33 | 9.96 | 6.46 | 8.20 | 5.38 | 8.37 | 24.48 | 29.27 | |
| (4) | LAS rescoring | 9.27 | 14.15 | 6.7 | 7.99 | 11.97 | 15.59 | 11.02 | 10.21 | 8.01 | 8.55 | 5.81 | 8.84 | 24.91 | 29.53 | |
| (5) | Multi-modal-GPT2 | 9.24 | 14.17 | 6.35 | 7.69 | 11.54 | 14.93 | 10.56 | 9.83 | 7.55 | 8.20 | 5.69 | 8.59 | 24.89 | 29.40 | |
| (6) | MATE-NA | 9.05 | 13.90 | 5.55 | 7.29 | 10.75 | 14.51 | 10.34 | 9.92 | 6.49 | 8.10 | 5.40 | 8.36 | 24.46 | 29.24 | |
| (7) | MATE-MSE | 7.49 | 11.41 | 5.22 | 6.95 | 10.31 | 13.97 | 10.10 | 9.62 | 6.10 | 7.65 | 5.07 | 7.92 | 23.84 | 28.24 | |
| (8) | MATE (ours) | 7.64 | 11.70 | 5.16 | 6.84 | 10.30 | 13.81 | 9.91 | 9.47 | 6.01 | 7.46 | 5.10 | 7.91 | 23.77 | 28.14 | |
| Parameter-Efficient Tuning | ||||||||||||||||
| (9) | Frozen-ME | 9.21 | 14.22 | 5.57 | 7.34 | 10.82 | 14.65 | 10.37 | 9.80 | 6.55 | 8.15 | 5.42 | 8.34 | 24.39 | 29.13 | |
| (10) | WavLM-adapter | 9.15 | 14.02 | 5.58 | 7.41 | 10.81 | 14.69 | 10.23 | 9.86 | 6.52 | 8.05 | 5.47 | 8.39 | 24.56 | 29.27 | |
| (11) | ME-adapter | 9.19 | 14.12 | 5.56 | 7.43 | 10.79 | 14.63 | 10.09 | 9.60 | 6.43 | 8.20 | 5.42 | 8.35 | 24.34 | 29.08 |
| Scenario | Action | Entity | |
| No rescoring | 78.01 | 72.53 | 53.23 |
| GPT2-text | 78.01 | 72.53 | 53.23 |
| Multi-modal-GPT2 | 78.07 | 72.65 | 53.26 |
| BERT-text | 77.72 | 72.45 | 53.37 |
| MATE | 78.76 | 73.70 | 54.26 |
| Dataset | Utterance | |
| SLURP | Ground Truth | remove tuesday alarm of nine a m |
| Rescored 1-best by BERT-text | move to alarm of nine a m | |
| Rescored 1-best by MATE | remove tuesday alarm at nine a m | |
| Ground Truth | hoover the hallway | |
| Rescored 1-best by BERT-text | who in the hallway | |
| Rescored 1-best by MATE | hoover the hallway | |
| Ground Truth | cancel business meeting on wednesday | |
| Rescored 1-best by BERT-text | council business meeting on wednesday | |
| Rescored 1-best by MATE | cancel business meeting on wednesday | |
| Ground Truth | can you let delta know i am never using them again | |
| Rescored 1-best by BERT-text | can you let doctor know i am never using them again | |
| Rescored 1-best by MATE | can you let delta know i am never using them again | |
| Ground Truth | i want to play fifa seventeen | |
| Rescored 1-best by BERT-text | i want to leave for seventeen | |
| Rescored 1-best by MATE | i want to play fifa seventeen | |
| Ground Truth | what do you know about fringe in edinburgh next year | |
| Rescored 1-best by BERT-text | what do you know about french in edinburgh next year | |
| Rescored 1-best by MATE | what do you know about fringe in edinburgh next year | |
| Voxpopuli | Ground Truth | for example the report talks about the rule of law and corruption |
| Rescored 1-best by BERT-text | for example the report talks about the rule of law on corruption | |
| Rescored 1-best by MATE | for example the report talks about the rule of law and corruption | |
| Ground Truth | i have met them they are young capable and visionary | |
| Rescored 1-best by BERT-text | i have met them they are young capable and missionary | |
| Rescored 1-best by MATE | i have met them they are young capable and visionary | |
| MTDialogue | Ground Truth | it's muffled |
| Rescored 1-best by BERT-text | it's muff | |
| Rescored 1-best by MATE | it's muffled | |
| Ground Truth | how much she got to pay | |
| Rescored 1-best by BERT-text | how much he got to pay | |
| Rescored 1-best by MATE | how much she got to pay | |
| ConvAI | Ground Truth | why did the noodle box in greensborough fail its health inspection |
| Rescored 1-best by BERT-text | why did the noodle box in greensboro fail its health inspection | |
| Rescored 1-best by MATE | why did the noodle box in greensborough fail its health inspection | |
| Ground Truth | tell me about duty free shopping | |
| Rescored 1-best by BERT-text | tell me about duty free shop | |
| Rescored 1-best by MATE | tell me about duty free shopping |
| In-domain | Out-of-domain | |||||||||||||||
| MTDialogue | LS test-clean | LS test-other | Voxpopuli | WSJ | ConvAI | SLURP | ||||||||||
| WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | |||
| (1) | No rescoring | 9.47 | 14.63 | 6.75 | 8.07 | 11.98 | 15.61 | 11.06 | 10.33 | 8.16 | 8.75 | 5.89 | 9.00 | 24.91 | 29.53 | |
| (2) | GPT2-text | 9.32 | 14.37 | 6.45 | 7.78 | 11.70 | 15.11 | 10.72 | 9.94 | 7.64 | 8.40 | 5.76 | 8.66 | 24.91 | 29.53 | |
| (3) | BERT-text | 9.05 | 13.88 | 5.50 | 7.20 | 10.70 | 14.45 | 10.33 | 9.96 | 6.46 | 8.20 | 5.38 | 8.37 | 24.48 | 29.27 | |
| (4) | LAS rescoring | 9.27 | 14.15 | 6.7 | 7.99 | 11.97 | 15.59 | 11.02 | 10.21 | 8.01 | 8.55 | 5.81 | 8.84 | 24.91 | 29.53 | |
| (5) | Multi-modal-GPT2 | 9.24 | 14.17 | 6.35 | 7.69 | 11.54 | 14.93 | 10.56 | 9.83 | 7.55 | 8.20 | 5.69 | 8.59 | 24.89 | 29.40 | |
| (6) | MATE-NA | 9.05 | 13.90 | 5.55 | 7.29 | 10.75 | 14.51 | 10.34 | 9.92 | 6.49 | 8.10 | 5.40 | 8.36 | 24.46 | 29.24 | |
| (7) | MATE-MSE | 7.49 | 11.41 | 5.22 | 6.95 | 10.31 | 13.97 | 10.10 | 9.62 | 6.10 | 7.65 | 5.07 | 7.92 | 23.84 | 28.24 | |
| (8) | MATE (ours) | 7.64 | 11.70 | 5.16 | 6.84 | 10.30 | 13.81 | 9.91 | 9.47 | 6.01 | 7.46 | 5.10 | 7.91 | 23.77 | 28.14 | |
| Parameter-Efficient Tuning | ||||||||||||||||
| (9) | Frozen-ME | 9.21 | 14.22 | 5.57 | 7.34 | 10.82 | 14.65 | 10.37 | 9.80 | 6.55 | 8.15 | 5.42 | 8.34 | 24.39 | 29.13 | |
| (10) | WavLM-adapter | 9.15 | 14.02 | 5.58 | 7.41 | 10.81 | 14.69 | 10.23 | 9.86 | 6.52 | 8.05 | 5.47 | 8.39 | 24.56 | 29.27 | |
| (11) | ME-adapter | 9.19 | 14.12 | 5.56 | 7.43 | 10.79 | 14.63 | 10.09 | 9.60 | 6.43 | 8.20 | 5.42 | 8.35 | 24.34 | 29.08 | |
| Scenario | Action | Entity | |
| No rescoring | 78.01 | 72.53 | 53.23 |
| GPT2-text | 78.01 | 72.53 | 53.23 |
| Multi-modal-GPT2 | 78.07 | 72.65 | 53.26 |
| BERT-text | 77.72 | 72.45 | 53.37 |
| MATE | 78.76 | 73.70 | 54.26 |
| Dataset | Utterance | |
| SLURP | Ground Truth | remove tuesday alarm of nine a m |
| Rescored 1-best by BERT-text | move to alarm of nine a m | |
| Rescored 1-best by MATE | remove tuesday alarm at nine a m | |
| Ground Truth | hoover the hallway | |
| Rescored 1-best by BERT-text | who in the hallway | |
| Rescored 1-best by MATE | hoover the hallway | |
| Ground Truth | cancel business meeting on wednesday | |
| Rescored 1-best by BERT-text | council business meeting on wednesday | |
| Rescored 1-best by MATE | cancel business meeting on wednesday | |
| Ground Truth | can you let delta know i am never using them again | |
| Rescored 1-best by BERT-text | can you let doctor know i am never using them again | |
| Rescored 1-best by MATE | can you let delta know i am never using them again | |
| Ground Truth | i want to play fifa seventeen | |
| Rescored 1-best by BERT-text | i want to leave for seventeen | |
| Rescored 1-best by MATE | i want to play fifa seventeen | |
| Ground Truth | what do you know about fringe in edinburgh next year | |
| Rescored 1-best by BERT-text | what do you know about french in edinburgh next year | |
| Rescored 1-best by MATE | what do you know about fringe in edinburgh next year | |
| Voxpopuli | Ground Truth | for example the report talks about the rule of law and corruption |
| Rescored 1-best by BERT-text | for example the report talks about the rule of law on corruption | |
| Rescored 1-best by MATE | for example the report talks about the rule of law and corruption | |
| Ground Truth | i have met them they are young capable and visionary | |
| Rescored 1-best by BERT-text | i have met them they are young capable and missionary | |
| Rescored 1-best by MATE | i have met them they are young capable and visionary | |
| MTDialogue | Ground Truth | it's muffled |
| Rescored 1-best by BERT-text | it's muff | |
| Rescored 1-best by MATE | it's muffled | |
| Ground Truth | how much she got to pay | |
| Rescored 1-best by BERT-text | how much he got to pay | |
| Rescored 1-best by MATE | how much she got to pay | |
| ConvAI | Ground Truth | why did the noodle box in greensborough fail its health inspection |
| Rescored 1-best by BERT-text | why did the noodle box in greensboro fail its health inspection | |
| Rescored 1-best by MATE | why did the noodle box in greensborough fail its health inspection | |
| Ground Truth | tell me about duty free shopping | |
| Rescored 1-best by BERT-text | tell me about duty free shop | |
| Rescored 1-best by MATE | tell me about duty free shopping |
| In-domain | Out-of-domain | |||||||||||||||
| MTDialogue | LS test-clean | LS test-other | Voxpopuli | WSJ | ConvAI | SLURP | ||||||||||
| WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | WER | CWER | |||
| (1) | No rescoring | 9.47 | 14.63 | 6.75 | 8.07 | 11.98 | 15.61 | 11.06 | 10.33 | 8.16 | 8.75 | 5.89 | 9.00 | 24.91 | 29.53 | |
| (2) | GPT2-text | 9.32 | 14.37 | 6.45 | 7.78 | 11.70 | 15.11 | 10.72 | 9.94 | 7.64 | 8.40 | 5.76 | 8.66 | 24.91 | 29.53 | |
| (3) | BERT-text | 9.05 | 13.88 | 5.50 | 7.20 | 10.70 | 14.45 | 10.33 | 9.96 | 6.46 | 8.20 | 5.38 | 8.37 | 24.48 | 29.27 | |
| (4) | LAS rescoring | 9.27 | 14.15 | 6.7 | 7.99 | 11.97 | 15.59 | 11.02 | 10.21 | 8.01 | 8.55 | 5.81 | 8.84 | 24.91 | 29.53 | |
| (5) | Multi-modal-GPT2 | 9.24 | 14.17 | 6.35 | 7.69 | 11.54 | 14.93 | 10.56 | 9.83 | 7.55 | 8.20 | 5.69 | 8.59 | 24.89 | 29.40 | |
| (6) | MATE-NA | 9.05 | 13.90 | 5.55 | 7.29 | 10.75 | 14.51 | 10.34 | 9.92 | 6.49 | 8.10 | 5.40 | 8.36 | 24.46 | 29.24 | |
| (7) | MATE-MSE | 7.49 | 11.41 | 5.22 | 6.95 | 10.31 | 13.97 | 10.10 | 9.62 | 6.10 | 7.65 | 5.07 | 7.92 | 23.84 | 28.24 | |
| (8) | MATE (ours) | 7.64 | 11.70 | 5.16 | 6.84 | 10.30 | 13.81 | 9.91 | 9.47 | 6.01 | 7.46 | 5.10 | 7.91 | 23.77 | 28.14 | |
| Parameter-Efficient Tuning | ||||||||||||||||
| (9) | Frozen-ME | 9.21 | 14.22 | 5.57 | 7.34 | 10.82 | 14.65 | 10.37 | 9.80 | 6.55 | 8.15 | 5.42 | 8.34 | 24.39 | 29.13 | |
| (10) | WavLM-adapter | 9.15 | 14.02 | 5.58 | 7.41 | 10.81 | 14.69 | 10.23 | 9.86 | 6.52 | 8.05 | 5.47 | 8.39 | 24.56 | 29.27 | |
| (11) | ME-adapter | 9.19 | 14.12 | 5.56 | 7.43 | 10.79 | 14.63 | 10.09 | 9.60 | 6.43 | 8.20 | 5.42 | 8.35 | 24.34 | 29.08 |
| Scenario | Action | Entity | |
| No rescoring | 78.01 | 72.53 | 53.23 |
| GPT2-text | 78.01 | 72.53 | 53.23 |
| Multi-modal-GPT2 | 78.07 | 72.65 | 53.26 |
| BERT-text | 77.72 | 72.45 | 53.37 |
| MATE | 78.76 | 73.70 | 54.26 |
| Dataset | Utterance | |
| SLURP | Ground Truth | remove tuesday alarm of nine a m |
| Rescored 1-best by BERT-text | move to alarm of nine a m | |
| Rescored 1-best by MATE | remove tuesday alarm at nine a m | |
| Ground Truth | hoover the hallway | |
| Rescored 1-best by BERT-text | who in the hallway | |
| Rescored 1-best by MATE | hoover the hallway | |
| Ground Truth | cancel business meeting on wednesday | |
| Rescored 1-best by BERT-text | council business meeting on wednesday | |
| Rescored 1-best by MATE | cancel business meeting on wednesday | |
| Ground Truth | can you let delta know i am never using them again | |
| Rescored 1-best by BERT-text | can you let doctor know i am never using them again | |
| Rescored 1-best by MATE | can you let delta know i am never using them again | |
| Ground Truth | i want to play fifa seventeen | |
| Rescored 1-best by BERT-text | i want to leave for seventeen | |
| Rescored 1-best by MATE | i want to play fifa seventeen | |
| Ground Truth | what do you know about fringe in edinburgh next year | |
| Rescored 1-best by BERT-text | what do you know about french in edinburgh next year | |
| Rescored 1-best by MATE | what do you know about fringe in edinburgh next year | |
| Voxpopuli | Ground Truth | for example the report talks about the rule of law and corruption |
| Rescored 1-best by BERT-text | for example the report talks about the rule of law on corruption | |
| Rescored 1-best by MATE | for example the report talks about the rule of law and corruption | |
| Ground Truth | i have met them they are young capable and visionary | |
| Rescored 1-best by BERT-text | i have met them they are young capable and missionary | |
| Rescored 1-best by MATE | i have met them they are young capable and visionary | |
| MTDialogue | Ground Truth | it's muffled |
| Rescored 1-best by BERT-text | it's muff | |
| Rescored 1-best by MATE | it's muffled | |
| Ground Truth | how much she got to pay | |
| Rescored 1-best by BERT-text | how much he got to pay | |
| Rescored 1-best by MATE | how much she got to pay | |
| ConvAI | Ground Truth | why did the noodle box in greensborough fail its health inspection |
| Rescored 1-best by BERT-text | why did the noodle box in greensboro fail its health inspection | |
| Rescored 1-best by MATE | why did the noodle box in greensborough fail its health inspection | |
| Ground Truth | tell me about duty free shopping | |
| Rescored 1-best by BERT-text | tell me about duty free shop | |
| Rescored 1-best by MATE | tell me about duty free shopping |
| Dataset | Metric | MLM | MESM-WN | MESM-CN | MLM-D | MLSM |
| CoLA | Matthews correlation | 0.391 | 0.371 | 0.369 | 0.390 | 0.403 |
| CoNLL2003 | F1 | 0.919 | 0.905 | 0.909 | 0.920 | 0.926 |
| MNLIm | Accuracy | 0.768 | 0.719 | 0.735 | 0.771 | 0.798 |
| MNLImm | Accuracy | 0.771 | 0.731 | 0.749 | 0.782 | 0.808 |
| MRPC | Accuracy | 0.761 | 0.713 | 0.712 | 0.775 | 0.786 |
| MRPC | F1 | 0.841 | 0.815 | 0.815 | 0.846 | 0.851 |
| QNLI | Accuracy | 0.850 | 0.775 | 0.779 | 0.852 | 0.870 |
| QQP | Accuracy | 0.886 | 0.840 | 0.850 | 0.888 | 0.892 |
| QQP | F1 | 0.847 | 0.777 | 0.797 | 0.849 | 0.855 |
| RTE | Accuracy | 0.571 | 0.552 | 0.552 | 0.578 | 0.571 |
| SST2 | Accuracy | 0.892 | 0.880 | 0.891 | 0.890 | 0.905 |
| STSB | Pearson correlation | 0.809 | 0.277 | 0.243 | 0.803 | 0.818 |
| STSB | Spearman correlation | 0.809 | 0.264 | 0.226 | 0.804 | 0.820 |
| WiC | Accuracy | 0.629 | 0.614 | 0.599 | 0.644 | 0.639 |
| Average | 0.7675 | 0.6596 | 0.6590 | 0.7709 | 0.7815 |
| Dataset | Metric | base w/ MLM | medium w/ MLMSM | ||
| CoLA | Matthews corr. | 0.430 | (0.014) | 0.403 | (0.012) |
| CoNLL2003 | F1 | 0.940 | (0.002) | 0.926 | (0.003) |
| MNLIm | Accuracy | 0.803 | (0.002) | 0.798 | (0.001) |
| MNLImm | Accuracy | 0.807 | (0.002) | 0.808 | (0.002) |
| MRPC | Accuracy | 0.763 | (0.056) | 0.786 | (0.020) |
| MRPC | F1 | 0.837 | (0.034) | 0.851 | (0.013) |
| QNLI | Accuracy | 0.882 | (0.003) | 0.870 | (0.004) |
| QQP | Accuracy | 0.898 | (0.001) | 0.892 | (0.001) |
| QQP | F1 | 0.862 | (0.001) | 0.855 | (0.001) |
| RTE | Accuracy | 0.538 | (0.019) | 0.571 | (0.011) |
| SST2 | Accuracy | 0.903 | (0.006) | 0.905 | (0.004) |
| STSB | Pearson corr. | 0.828 | (0.004) | 0.818 | (0.024) |
| STSB | Spearman corr. | 0.824 | (0.003) | 0.820 | (0.021) |
| WiC | Accuracy | 0.630 | (0.025) | 0.639 | (0.007) |
| Average | 0.7818 | 0.7815 |
| Teacher model T | |||
| BERT base | BERT large | RoBERTa large | |
| MLM-D | 0.7709 | 0.7698 | 0.7563 |
| MLSM | 0.7815 | 0.7696 | 0.7783 |
| Dataset | Metric | MLM | MESM-WN | MESM-CN | MLM-D | MLSM |
| CoLA | Matthews correlation | 0.391 | 0.371 | 0.369 | 0.390 | 0.403 |
| CoNLL2003 | F1 | 0.919 | 0.905 | 0.909 | 0.920 | 0.926 |
| MNLIm | Accuracy | 0.768 | 0.719 | 0.735 | 0.771 | 0.798 |
| MNLImm | Accuracy | 0.771 | 0.731 | 0.749 | 0.782 | 0.808 |
| MRPC | Accuracy | 0.761 | 0.713 | 0.712 | 0.775 | 0.786 |
| MRPC | F1 | 0.841 | 0.815 | 0.815 | 0.846 | 0.851 |
| QNLI | Accuracy | 0.850 | 0.775 | 0.779 | 0.852 | 0.870 |
| QQP | Accuracy | 0.886 | 0.840 | 0.850 | 0.888 | 0.892 |
| QQP | F1 | 0.847 | 0.777 | 0.797 | 0.849 | 0.855 |
| RTE | Accuracy | 0.571 | 0.552 | 0.552 | 0.578 | 0.571 |
| SST2 | Accuracy | 0.892 | 0.880 | 0.891 | 0.890 | 0.905 |
| STSB | Pearson correlation | 0.809 | 0.277 | 0.243 | 0.803 | 0.818 |
| STSB | Spearman correlation | 0.809 | 0.264 | 0.226 | 0.804 | 0.820 |
| WiC | Accuracy | 0.629 | 0.614 | 0.599 | 0.644 | 0.639 |
| Average | 0.7675 | 0.6596 | 0.6590 | 0.7709 | 0.7815 |
| Dataset | Metric | base w/ MLM | medium w/ MLMSM | ||
| CoLA | Matthews corr. | 0.430 | (0.014) | 0.403 | (0.012) |
| CoNLL2003 | F1 | 0.940 | (0.002) | 0.926 | (0.003) |
| MNLIm | Accuracy | 0.803 | (0.002) | 0.798 | (0.001) |
| MNLImm | Accuracy | 0.807 | (0.002) | 0.808 | (0.002) |
| MRPC | Accuracy | 0.763 | (0.056) | 0.786 | (0.020) |
| MRPC | F1 | 0.837 | (0.034) | 0.851 | (0.013) |
| QNLI | Accuracy | 0.882 | (0.003) | 0.870 | (0.004) |
| QQP | Accuracy | 0.898 | (0.001) | 0.892 | (0.001) |
| QQP | F1 | 0.862 | (0.001) | 0.855 | (0.001) |
| RTE | Accuracy | 0.538 | (0.019) | 0.571 | (0.011) |
| SST2 | Accuracy | 0.903 | (0.006) | 0.905 | (0.004) |
| STSB | Pearson corr. | 0.828 | (0.004) | 0.818 | (0.024) |
| STSB | Spearman corr. | 0.824 | (0.003) | 0.820 | (0.021) |
| WiC | Accuracy | 0.630 | (0.025) | 0.639 | (0.007) |
| Average | 0.7818 | 0.7815 |
| Teacher model T | |||
| BERT base | BERT large | RoBERTa large | |
| MLM-D | 0.7709 | 0.7698 | 0.7563 |
| MLSM | 0.7815 | 0.7696 | 0.7783 |
| Dataset | Metric | MLM | MESM-WN | MESM-CN | MLM-D | MLSM |
| CoLA | Matthews correlation | 0.391 | 0.371 | 0.369 | 0.390 | 0.403 |
| CoNLL2003 | F1 | 0.919 | 0.905 | 0.909 | 0.920 | 0.926 |
| MNLIm | Accuracy | 0.768 | 0.719 | 0.735 | 0.771 | 0.798 |
| MNLImm | Accuracy | 0.771 | 0.731 | 0.749 | 0.782 | 0.808 |
| MRPC | Accuracy | 0.761 | 0.713 | 0.712 | 0.775 | 0.786 |
| MRPC | F1 | 0.841 | 0.815 | 0.815 | 0.846 | 0.851 |
| QNLI | Accuracy | 0.850 | 0.775 | 0.779 | 0.852 | 0.870 |
| QQP | Accuracy | 0.886 | 0.840 | 0.850 | 0.888 | 0.892 |
| QQP | F1 | 0.847 | 0.777 | 0.797 | 0.849 | 0.855 |
| RTE | Accuracy | 0.571 | 0.552 | 0.552 | 0.578 | 0.571 |
| SST2 | Accuracy | 0.892 | 0.880 | 0.891 | 0.890 | 0.905 |
| STSB | Pearson correlation | 0.809 | 0.277 | 0.243 | 0.803 | 0.818 |
| STSB | Spearman correlation | 0.809 | 0.264 | 0.226 | 0.804 | 0.820 |
| WiC | Accuracy | 0.629 | 0.614 | 0.599 | 0.644 | 0.639 |
| Average | 0.7675 | 0.6596 | 0.6590 | 0.7709 | 0.7815 |
| Dataset | Metric | base w/ MLM | medium w/ MLMSM | ||
| CoLA | Matthews corr. | 0.430 | (0.014) | 0.403 | (0.012) |
| CoNLL2003 | F1 | 0.940 | (0.002) | 0.926 | (0.003) |
| MNLIm | Accuracy | 0.803 | (0.002) | 0.798 | (0.001) |
| MNLImm | Accuracy | 0.807 | (0.002) | 0.808 | (0.002) |
| MRPC | Accuracy | 0.763 | (0.056) | 0.786 | (0.020) |
| MRPC | F1 | 0.837 | (0.034) | 0.851 | (0.013) |
| QNLI | Accuracy | 0.882 | (0.003) | 0.870 | (0.004) |
| QQP | Accuracy | 0.898 | (0.001) | 0.892 | (0.001) |
| QQP | F1 | 0.862 | (0.001) | 0.855 | (0.001) |
| RTE | Accuracy | 0.538 | (0.019) | 0.571 | (0.011) |
| SST2 | Accuracy | 0.903 | (0.006) | 0.905 | (0.004) |
| STSB | Pearson corr. | 0.828 | (0.004) | 0.818 | (0.024) |
| STSB | Spearman corr. | 0.824 | (0.003) | 0.820 | (0.021) |
| WiC | Accuracy | 0.630 | (0.025) | 0.639 | (0.007) |
| Average | 0.7818 | 0.7815 | |||
| Teacher model T | |||
| BERT base | BERT large | RoBERTa large | |
| MLM-D | 0.7709 | 0.7698 | 0.7563 |
| MLSM | 0.7815 | 0.7696 | 0.7783 |
| Dataset | Metric | MLM | MESM-WN | MESM-CN | MLM-D | MLSM |
| CoLA | Matthews correlation | 0.391 | 0.371 | 0.369 | 0.390 | 0.403 |
| CoNLL2003 | F1 | 0.919 | 0.905 | 0.909 | 0.920 | 0.926 |
| MNLIm | Accuracy | 0.768 | 0.719 | 0.735 | 0.771 | 0.798 |
| MNLImm | Accuracy | 0.771 | 0.731 | 0.749 | 0.782 | 0.808 |
| MRPC | Accuracy | 0.761 | 0.713 | 0.712 | 0.775 | 0.786 |
| MRPC | F1 | 0.841 | 0.815 | 0.815 | 0.846 | 0.851 |
| QNLI | Accuracy | 0.850 | 0.775 | 0.779 | 0.852 | 0.870 |
| QQP | Accuracy | 0.886 | 0.840 | 0.850 | 0.888 | 0.892 |
| QQP | F1 | 0.847 | 0.777 | 0.797 | 0.849 | 0.855 |
| RTE | Accuracy | 0.571 | 0.552 | 0.552 | 0.578 | 0.571 |
| SST2 | Accuracy | 0.892 | 0.880 | 0.891 | 0.890 | 0.905 |
| STSB | Pearson correlation | 0.809 | 0.277 | 0.243 | 0.803 | 0.818 |
| STSB | Spearman correlation | 0.809 | 0.264 | 0.226 | 0.804 | 0.820 |
| WiC | Accuracy | 0.629 | 0.614 | 0.599 | 0.644 | 0.639 |
| Average | 0.7675 | 0.6596 | 0.6590 | 0.7709 | 0.7815 |
| Dataset | Metric | base w/ MLM | medium w/ MLMSM | ||
| CoLA | Matthews corr. | 0.430 | (0.014) | 0.403 | (0.012) |
| CoNLL2003 | F1 | 0.940 | (0.002) | 0.926 | (0.003) |
| MNLIm | Accuracy | 0.803 | (0.002) | 0.798 | (0.001) |
| MNLImm | Accuracy | 0.807 | (0.002) | 0.808 | (0.002) |
| MRPC | Accuracy | 0.763 | (0.056) | 0.786 | (0.020) |
| MRPC | F1 | 0.837 | (0.034) | 0.851 | (0.013) |
| QNLI | Accuracy | 0.882 | (0.003) | 0.870 | (0.004) |
| QQP | Accuracy | 0.898 | (0.001) | 0.892 | (0.001) |
| QQP | F1 | 0.862 | (0.001) | 0.855 | (0.001) |
| RTE | Accuracy | 0.538 | (0.019) | 0.571 | (0.011) |
| SST2 | Accuracy | 0.903 | (0.006) | 0.905 | (0.004) |
| STSB | Pearson corr. | 0.828 | (0.004) | 0.818 | (0.024) |
| STSB | Spearman corr. | 0.824 | (0.003) | 0.820 | (0.021) |
| WiC | Accuracy | 0.630 | (0.025) | 0.639 | (0.007) |
| Average | 0.7818 | 0.7815 |
| Teacher model T | |||
| BERT base | BERT large | RoBERTa large | |
| MLM-D | 0.7709 | 0.7698 | 0.7563 |
| MLSM | 0.7815 | 0.7696 | 0.7783 |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 91.82 | 94.52 | 91.65 | 99.71 | 91.59 | 100.00 | 91.63 | 100.00 | 91.74 | 51.38 | 91.33 | 77.84 | 11.34 | |
| FTH | 91.75 | 56.75 | 91.42 | 75.81 | 91.75 | 96.03 | 91.58 | 99.82 | 91.91 | 34.85 | 91.54 | 64.13 | 27.35 | |
| FreeLB | 91.69 | 60.83 | 92.04 | 96.10 | 91.55 | 99.74 | 91.78 | 99.98 | 91.75 | 47.19 | 91.86 | 69.09 | 19.76 | |
| FP | 90.55 | 22.26 | 90.46 | 27.04 | 90.35 | 82.61 | 90.35 | 58.68 | 90.60 | 39.28 | 90.06 | 68.50 | 48.86 | |
| Ours-lite | 90.88 | 21.54 | 90.74 | 33.22 | 91.26 | 79.17 | 91.37 | 85.15 | 90.35 | 41.51 | 91.40 | 48.52 | 47.07 | |
| Ours | 90.75 | 19.45 | 90.23 | 23.57 | 91.00 | 82.28 | 91.38 | 62.52 | 91.30 | 37.13 | 91.44 | 45.85 | 53.45 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 91.68 | 89.68 | 91.63 | 87.83 | 91.72 | 75.04 | 91.70 | 99.58 | 91.59 | 68.59 | 91.61 | 81.38 | 15.46 | |
| FTH | 91.67 | 42.30 | 91.68 | 55.27 | 91.80 | 41.13 | 91.85 | 86.37 | 91.79 | 28.72 | 91.87 | 58.73 | 47.06 | |
| FreeLB | 91.69 | 47.61 | 91.89 | 53.21 | 91.69 | 31.02 | 91.89 | 94.42 | 91.92 | 45.90 | 91.88 | 60.24 | 43.74 | |
| FP | 90.39 | 29.59 | 90.50 | 45.59 | 90.63 | 27.49 | 90.48 | 18.71 | 90.64 | 15.66 | 90.54 | 57.59 | 66.71 | |
| Ours-lite | 90.87 | 13.99 | 91.01 | 30.00 | 90.99 | 29.20 | 90.82 | 20.74 | 91.26 | 19.26 | 91.21 | 40.73 | 73.49 | |
| Ours | 90.83 | 8.97 | 90.73 | 23.65 | 90.75 | 21.89 | 90.60 | 10.54 | 90.37 | 6.77 | 90.91 | 32.44 | 81.77 |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 6.66 | |
| FTH | 90.53 | 89.43 | 89.49 | 99.14 | 89.53 | 100.00 | 90.49 | 99.89 | 90.53 | 49.71 | 90.18 | 71.55 | 13.63 | |
| FreeLB | 90.97 | 94.63 | 90.26 | 99.89 | 89.83 | 100.00 | 90.99 | 100.00 | 90.91 | 66.69 | 90.24 | 82.83 | 7.91 | |
| FP | 88.49 | 22.63 | 87.74 | 36.01 | 88.14 | 88.46 | 88.71 | 49.69 | 88.07 | 54.21 | 87.98 | 66.53 | 45.66 | |
| Ours-lite | 87.55 | 33.16 | 85.33 | 24.50 | 84.83 | 47.98 | 89.03 | 95.18 | 86.57 | 45.22 | 89.36 | 45.58 | 49.98 | |
| Ours | 86.19 | 24.93 | 83.66 | 20.37 | 83.55 | 40.48 | 87.01 | 63.44 | 84.01 | 48.97 | 86.53 | 49.49 | 57.31 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 7.22 | |
| FTH | 89.51 | 56.78 | 89.23 | 78.79 | 89.18 | 99.28 | 89.86 | 99.54 | 89.36 | 49.56 | 89.41 | 61.81 | 24.85 | |
| FreeLB | 90.04 | 59.17 | 89.32 | 95.57 | 88.05 | 99.91 | 90.30 | 99.67 | 89.30 | 62.41 | 89.96 | 72.23 | 17.65 | |
| FP | 88.39 | 26.97 | 88.35 | 39.49 | 88.16 | 44.04 | 88.75 | 7.01 | 88.53 | 8.53 | 88.40 | 52.30 | 69.42 | |
| Ours-lite | 88.46 | 12.10 | 88.01 | 55.74 | 88.02 | 43.00 | 88.35 | 50.43 | 88.02 | 8.30 | 88.60 | 16.71 | 68.10 | |
| Ours | 87.56 | 9.28 | 88.06 | 42.99 | 87.91 | 29.18 | 87.89 | 42.46 | 87.97 | 10.89 | 88.27 | 21.23 | 73.14 |
| Data | CE | Max Entropy | Δ |
| ADD1 | 77.69 | 78.43 | +0.74 |
| DPR | 50.70 | 50.99 | +0.29 |
| SPR | 59.65 | 61.38 | +1.73 |
| FN+ | 55.56 | 57.25 | +1.69 |
| JOCI | 51.64 | 52.12 | +0.48 |
| MPE | 67.05 | 67.59 | +0.54 |
| SCITAIL | 74.33 | 74.92 | +0.59 |
| SICK | 60.56 | 62.04 | +1.48 |
| GLUE | 74.60 | 74.88 | +0.28 |
| QQP | 68.50 | 68.65 | +0.15 |
| MNLI | 74.88 | 74.92 | +0.04 |
| MNLI-M | 74.60 | 74.88 | +0.28 |
| SNLI | 90.73 | 90.91 | +0.18 |
| Backdoor Attack | Variation | Ours | ||
| ACC | ASR | ACC | ASR | |
| BadNets | 91.38 | 59.74 | 90.88 | 21.54 |
| RIPPLE | 91.05 | 85.46 | 90.74 | 33.22 |
| RIPPLES | 91.21 | 86.16 | 91.26 | 79.17 |
| SOS | 91.39 | 99.93 | 91.37 | 85.15 |
| HiddenKiller | 91.11 | 40.59 | 90.35 | 41.51 |
| StyleBkd | 91.35 | 73.89 | 91.40 | 48.52 |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 91.82 | 94.52 | 91.65 | 99.71 | 91.59 | 100.00 | 91.63 | 100.00 | 91.74 | 51.38 | 91.33 | 77.84 | 11.34 | |
| FTH | 91.75 | 56.75 | 91.42 | 75.81 | 91.75 | 96.03 | 91.58 | 99.82 | 91.91 | 34.85 | 91.54 | 64.13 | 27.35 | |
| FreeLB | 91.69 | 60.83 | 92.04 | 96.10 | 91.55 | 99.74 | 91.78 | 99.98 | 91.75 | 47.19 | 91.86 | 69.09 | 19.76 | |
| FP | 90.55 | 22.26 | 90.46 | 27.04 | 90.35 | 82.61 | 90.35 | 58.68 | 90.60 | 39.28 | 90.06 | 68.50 | 48.86 | |
| Ours-lite | 90.88 | 21.54 | 90.74 | 33.22 | 91.26 | 79.17 | 91.37 | 85.15 | 90.35 | 41.51 | 91.40 | 48.52 | 47.07 | |
| Ours | 90.75 | 19.45 | 90.23 | 23.57 | 91.00 | 82.28 | 91.38 | 62.52 | 91.30 | 37.13 | 91.44 | 45.85 | 53.45 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 91.68 | 89.68 | 91.63 | 87.83 | 91.72 | 75.04 | 91.70 | 99.58 | 91.59 | 68.59 | 91.61 | 81.38 | 15.46 | |
| FTH | 91.67 | 42.30 | 91.68 | 55.27 | 91.80 | 41.13 | 91.85 | 86.37 | 91.79 | 28.72 | 91.87 | 58.73 | 47.06 | |
| FreeLB | 91.69 | 47.61 | 91.89 | 53.21 | 91.69 | 31.02 | 91.89 | 94.42 | 91.92 | 45.90 | 91.88 | 60.24 | 43.74 | |
| FP | 90.39 | 29.59 | 90.50 | 45.59 | 90.63 | 27.49 | 90.48 | 18.71 | 90.64 | 15.66 | 90.54 | 57.59 | 66.71 | |
| Ours-lite | 90.87 | 13.99 | 91.01 | 30.00 | 90.99 | 29.20 | 90.82 | 20.74 | 91.26 | 19.26 | 91.21 | 40.73 | 73.49 | |
| Ours | 90.83 | 8.97 | 90.73 | 23.65 | 90.75 | 21.89 | 90.60 | 10.54 | 90.37 | 6.77 | 90.91 | 32.44 | 81.77 |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 6.66 | |
| FTH | 90.53 | 89.43 | 89.49 | 99.14 | 89.53 | 100.00 | 90.49 | 99.89 | 90.53 | 49.71 | 90.18 | 71.55 | 13.63 | |
| FreeLB | 90.97 | 94.63 | 90.26 | 99.89 | 89.83 | 100.00 | 90.99 | 100.00 | 90.91 | 66.69 | 90.24 | 82.83 | 7.91 | |
| FP | 88.49 | 22.63 | 87.74 | 36.01 | 88.14 | 88.46 | 88.71 | 49.69 | 88.07 | 54.21 | 87.98 | 66.53 | 45.66 | |
| Ours-lite | 87.55 | 33.16 | 85.33 | 24.50 | 84.83 | 47.98 | 89.03 | 95.18 | 86.57 | 45.22 | 89.36 | 45.58 | 49.98 | |
| Ours | 86.19 | 24.93 | 83.66 | 20.37 | 83.55 | 40.48 | 87.01 | 63.44 | 84.01 | 48.97 | 86.53 | 49.49 | 57.31 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 7.22 | |
| FTH | 89.51 | 56.78 | 89.23 | 78.79 | 89.18 | 99.28 | 89.86 | 99.54 | 89.36 | 49.56 | 89.41 | 61.81 | 24.85 | |
| FreeLB | 90.04 | 59.17 | 89.32 | 95.57 | 88.05 | 99.91 | 90.30 | 99.67 | 89.30 | 62.41 | 89.96 | 72.23 | 17.65 | |
| FP | 88.39 | 26.97 | 88.35 | 39.49 | 88.16 | 44.04 | 88.75 | 7.01 | 88.53 | 8.53 | 88.40 | 52.30 | 69.42 | |
| Ours-lite | 88.46 | 12.10 | 88.01 | 55.74 | 88.02 | 43.00 | 88.35 | 50.43 | 88.02 | 8.30 | 88.60 | 16.71 | 68.10 | |
| Ours | 87.56 | 9.28 | 88.06 | 42.99 | 87.91 | 29.18 | 87.89 | 42.46 | 87.97 | 10.89 | 88.27 | 21.23 | 73.14 |
| Data | CE | Max Entropy | Δ |
| ADD1 | 77.69 | 78.43 | +0.74 |
| DPR | 50.70 | 50.99 | +0.29 |
| SPR | 59.65 | 61.38 | +1.73 |
| FN+ | 55.56 | 57.25 | +1.69 |
| JOCI | 51.64 | 52.12 | +0.48 |
| MPE | 67.05 | 67.59 | +0.54 |
| SCITAIL | 74.33 | 74.92 | +0.59 |
| SICK | 60.56 | 62.04 | +1.48 |
| GLUE | 74.60 | 74.88 | +0.28 |
| QQP | 68.50 | 68.65 | +0.15 |
| MNLI | 74.88 | 74.92 | +0.04 |
| MNLI-M | 74.60 | 74.88 | +0.28 |
| SNLI | 90.73 | 90.91 | +0.18 |
| Backdoor Attack | Variation | Ours | ||
| ACC | ASR | ACC | ASR | |
| BadNets | 91.38 | 59.74 | 90.88 | 21.54 |
| RIPPLE | 91.05 | 85.46 | 90.74 | 33.22 |
| RIPPLES | 91.21 | 86.16 | 91.26 | 79.17 |
| SOS | 91.39 | 99.93 | 91.37 | 85.15 |
| HiddenKiller | 91.11 | 40.59 | 90.35 | 41.51 |
| StyleBkd | 91.35 | 73.89 | 91.40 | 48.52 |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 91.82 | 94.52 | 91.65 | 99.71 | 91.59 | 100.00 | 91.63 | 100.00 | 91.74 | 51.38 | 91.33 | 77.84 | 11.34 | |
| FTH | 91.75 | 56.75 | 91.42 | 75.81 | 91.75 | 96.03 | 91.58 | 99.82 | 91.91 | 34.85 | 91.54 | 64.13 | 27.35 | |
| FreeLB | 91.69 | 60.83 | 92.04 | 96.10 | 91.55 | 99.74 | 91.78 | 99.98 | 91.75 | 47.19 | 91.86 | 69.09 | 19.76 | |
| FP | 90.55 | 22.26 | 90.46 | 27.04 | 90.35 | 82.61 | 90.35 | 58.68 | 90.60 | 39.28 | 90.06 | 68.50 | 48.86 | |
| Ours-lite | 90.88 | 21.54 | 90.74 | 33.22 | 91.26 | 79.17 | 91.37 | 85.15 | 90.35 | 41.51 | 91.40 | 48.52 | 47.07 | |
| Ours | 90.75 | 19.45 | 90.23 | 23.57 | 91.00 | 82.28 | 91.38 | 62.52 | 91.30 | 37.13 | 91.44 | 45.85 | 53.45 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 91.68 | 89.68 | 91.63 | 87.83 | 91.72 | 75.04 | 91.70 | 99.58 | 91.59 | 68.59 | 91.61 | 81.38 | 15.46 | |
| FTH | 91.67 | 42.30 | 91.68 | 55.27 | 91.80 | 41.13 | 91.85 | 86.37 | 91.79 | 28.72 | 91.87 | 58.73 | 47.06 | |
| FreeLB | 91.69 | 47.61 | 91.89 | 53.21 | 91.69 | 31.02 | 91.89 | 94.42 | 91.92 | 45.90 | 91.88 | 60.24 | 43.74 | |
| FP | 90.39 | 29.59 | 90.50 | 45.59 | 90.63 | 27.49 | 90.48 | 18.71 | 90.64 | 15.66 | 90.54 | 57.59 | 66.71 | |
| Ours-lite | 90.87 | 13.99 | 91.01 | 30.00 | 90.99 | 29.20 | 90.82 | 20.74 | 91.26 | 19.26 | 91.21 | 40.73 | 73.49 | |
| Ours | 90.83 | 8.97 | 90.73 | 23.65 | 90.75 | 21.89 | 90.60 | 10.54 | 90.37 | 6.77 | 90.91 | 32.44 | 81.77 | |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 6.66 | |
| FTH | 90.53 | 89.43 | 89.49 | 99.14 | 89.53 | 100.00 | 90.49 | 99.89 | 90.53 | 49.71 | 90.18 | 71.55 | 13.63 | |
| FreeLB | 90.97 | 94.63 | 90.26 | 99.89 | 89.83 | 100.00 | 90.99 | 100.00 | 90.91 | 66.69 | 90.24 | 82.83 | 7.91 | |
| FP | 88.49 | 22.63 | 87.74 | 36.01 | 88.14 | 88.46 | 88.71 | 49.69 | 88.07 | 54.21 | 87.98 | 66.53 | 45.66 | |
| Ours-lite | 87.55 | 33.16 | 85.33 | 24.50 | 84.83 | 47.98 | 89.03 | 95.18 | 86.57 | 45.22 | 89.36 | 45.58 | 49.98 | |
| Ours | 86.19 | 24.93 | 83.66 | 20.37 | 83.55 | 40.48 | 87.01 | 63.44 | 84.01 | 48.97 | 86.53 | 49.49 | 57.31 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 7.22 | |
| FTH | 89.51 | 56.78 | 89.23 | 78.79 | 89.18 | 99.28 | 89.86 | 99.54 | 89.36 | 49.56 | 89.41 | 61.81 | 24.85 | |
| FreeLB | 90.04 | 59.17 | 89.32 | 95.57 | 88.05 | 99.91 | 90.30 | 99.67 | 89.30 | 62.41 | 89.96 | 72.23 | 17.65 | |
| FP | 88.39 | 26.97 | 88.35 | 39.49 | 88.16 | 44.04 | 88.75 | 7.01 | 88.53 | 8.53 | 88.40 | 52.30 | 69.42 | |
| Ours-lite | 88.46 | 12.10 | 88.01 | 55.74 | 88.02 | 43.00 | 88.35 | 50.43 | 88.02 | 8.30 | 88.60 | 16.71 | 68.10 | |
| Ours | 87.56 | 9.28 | 88.06 | 42.99 | 87.91 | 29.18 | 87.89 | 42.46 | 87.97 | 10.89 | 88.27 | 21.23 | 73.14 | |
| Data | CE | Max Entropy | Δ |
| ADD1 | 77.69 | 78.43 | +0.74 |
| DPR | 50.70 | 50.99 | +0.29 |
| SPR | 59.65 | 61.38 | +1.73 |
| FN+ | 55.56 | 57.25 | +1.69 |
| JOCI | 51.64 | 52.12 | +0.48 |
| MPE | 67.05 | 67.59 | +0.54 |
| SCITAIL | 74.33 | 74.92 | +0.59 |
| SICK | 60.56 | 62.04 | +1.48 |
| GLUE | 74.60 | 74.88 | +0.28 |
| QQP | 68.50 | 68.65 | +0.15 |
| MNLI | 74.88 | 74.92 | +0.04 |
| MNLI-M | 74.60 | 74.88 | +0.28 |
| SNLI | 90.73 | 90.91 | +0.18 |
| Backdoor Attack | Variation | Ours | ||
| ACC | ASR | ACC | ASR | |
| BadNets | 91.38 | 59.74 | 90.88 | 21.54 |
| RIPPLE | 91.05 | 85.46 | 90.74 | 33.22 |
| RIPPLES | 91.21 | 86.16 | 91.26 | 79.17 |
| SOS | 91.39 | 99.93 | 91.37 | 85.15 |
| HiddenKiller | 91.11 | 40.59 | 90.35 | 41.51 |
| StyleBkd | 91.35 | 73.89 | 91.40 | 48.52 |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 91.82 | 94.52 | 91.65 | 99.71 | 91.59 | 100.00 | 91.63 | 100.00 | 91.74 | 51.38 | 91.33 | 77.84 | 11.34 | |
| FTH | 91.75 | 56.75 | 91.42 | 75.81 | 91.75 | 96.03 | 91.58 | 99.82 | 91.91 | 34.85 | 91.54 | 64.13 | 27.35 | |
| FreeLB | 91.69 | 60.83 | 92.04 | 96.10 | 91.55 | 99.74 | 91.78 | 99.98 | 91.75 | 47.19 | 91.86 | 69.09 | 19.76 | |
| FP | 90.55 | 22.26 | 90.46 | 27.04 | 90.35 | 82.61 | 90.35 | 58.68 | 90.60 | 39.28 | 90.06 | 68.50 | 48.86 | |
| Ours-lite | 90.88 | 21.54 | 90.74 | 33.22 | 91.26 | 79.17 | 91.37 | 85.15 | 90.35 | 41.51 | 91.40 | 48.52 | 47.07 | |
| Ours | 90.75 | 19.45 | 90.23 | 23.57 | 91.00 | 82.28 | 91.38 | 62.52 | 91.30 | 37.13 | 91.44 | 45.85 | 53.45 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 91.68 | 89.68 | 91.63 | 87.83 | 91.72 | 75.04 | 91.70 | 99.58 | 91.59 | 68.59 | 91.61 | 81.38 | 15.46 | |
| FTH | 91.67 | 42.30 | 91.68 | 55.27 | 91.80 | 41.13 | 91.85 | 86.37 | 91.79 | 28.72 | 91.87 | 58.73 | 47.06 | |
| FreeLB | 91.69 | 47.61 | 91.89 | 53.21 | 91.69 | 31.02 | 91.89 | 94.42 | 91.92 | 45.90 | 91.88 | 60.24 | 43.74 | |
| FP | 90.39 | 29.59 | 90.50 | 45.59 | 90.63 | 27.49 | 90.48 | 18.71 | 90.64 | 15.66 | 90.54 | 57.59 | 66.71 | |
| Ours-lite | 90.87 | 13.99 | 91.01 | 30.00 | 90.99 | 29.20 | 90.82 | 20.74 | 91.26 | 19.26 | 91.21 | 40.73 | 73.49 | |
| Ours | 90.83 | 8.97 | 90.73 | 23.65 | 90.75 | 21.89 | 90.60 | 10.54 | 90.37 | 6.77 | 90.91 | 32.44 | 81.77 |
| Dataset | Methods | BadNets | RIPPLE | RIPPLES | SOS | HiddenKiller | StyleBkd | Δ | ||||||
| ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | ACC | ASR | |||
| SST-2 | Victim | 91.05 | 99.89 | 90.28 | 100.00 | 90.61 | 100.00 | 91.10 | 100.00 | 87.26 | 98.57 | 88.69 | 93.05 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 6.66 | |
| FTH | 90.53 | 89.43 | 89.49 | 99.14 | 89.53 | 100.00 | 90.49 | 99.89 | 90.53 | 49.71 | 90.18 | 71.55 | 13.63 | |
| FreeLB | 90.97 | 94.63 | 90.26 | 99.89 | 89.83 | 100.00 | 90.99 | 100.00 | 90.91 | 66.69 | 90.24 | 82.83 | 7.91 | |
| FP | 88.49 | 22.63 | 87.74 | 36.01 | 88.14 | 88.46 | 88.71 | 49.69 | 88.07 | 54.21 | 87.98 | 66.53 | 45.66 | |
| Ours-lite | 87.55 | 33.16 | 85.33 | 24.50 | 84.83 | 47.98 | 89.03 | 95.18 | 86.57 | 45.22 | 89.36 | 45.58 | 49.98 | |
| Ours | 86.19 | 24.93 | 83.66 | 20.37 | 83.55 | 40.48 | 87.01 | 63.44 | 84.01 | 48.97 | 86.53 | 49.49 | 57.31 | |
| AG | Victim | 91.43 | 99.79 | 91.08 | 99.86 | 91.11 | 99.86 | 92.01 | 99.61 | 91.09 | 99.23 | 90.11 | 96.51 | - |
| FT | 90.53 | 98.57 | 90.01 | 100.00 | 89.80 | 100.00 | 91.29 | 100.00 | 90.72 | 66.71 | 90.17 | 86.25 | 7.22 | |
| FTH | 89.51 | 56.78 | 89.23 | 78.79 | 89.18 | 99.28 | 89.86 | 99.54 | 89.36 | 49.56 | 89.41 | 61.81 | 24.85 | |
| FreeLB | 90.04 | 59.17 | 89.32 | 95.57 | 88.05 | 99.91 | 90.30 | 99.67 | 89.30 | 62.41 | 89.96 | 72.23 | 17.65 | |
| FP | 88.39 | 26.97 | 88.35 | 39.49 | 88.16 | 44.04 | 88.75 | 7.01 | 88.53 | 8.53 | 88.40 | 52.30 | 69.42 | |
| Ours-lite | 88.46 | 12.10 | 88.01 | 55.74 | 88.02 | 43.00 | 88.35 | 50.43 | 88.02 | 8.30 | 88.60 | 16.71 | 68.10 | |
| Ours | 87.56 | 9.28 | 88.06 | 42.99 | 87.91 | 29.18 | 87.89 | 42.46 | 87.97 | 10.89 | 88.27 | 21.23 | 73.14 |
| Data | CE | Max Entropy | Δ |
| ADD1 | 77.69 | 78.43 | +0.74 |
| DPR | 50.70 | 50.99 | +0.29 |
| SPR | 59.65 | 61.38 | +1.73 |
| FN+ | 55.56 | 57.25 | +1.69 |
| JOCI | 51.64 | 52.12 | +0.48 |
| MPE | 67.05 | 67.59 | +0.54 |
| SCITAIL | 74.33 | 74.92 | +0.59 |
| SICK | 60.56 | 62.04 | +1.48 |
| GLUE | 74.60 | 74.88 | +0.28 |
| QQP | 68.50 | 68.65 | +0.15 |
| MNLI | 74.88 | 74.92 | +0.04 |
| MNLI-M | 74.60 | 74.88 | +0.28 |
| SNLI | 90.73 | 90.91 | +0.18 |
| Backdoor Attack | Variation | Ours | ||
| ACC | ASR | ACC | ASR | |
| BadNets | 91.38 | 59.74 | 90.88 | 21.54 |
| RIPPLE | 91.05 | 85.46 | 90.74 | 33.22 |
| RIPPLES | 91.21 | 86.16 | 91.26 | 79.17 |
| SOS | 91.39 | 99.93 | 91.37 | 85.15 |
| HiddenKiller | 91.11 | 40.59 | 90.35 | 41.51 |
| StyleBkd | 91.35 | 73.89 | 91.40 | 48.52 |
| Source | #Files | #Sentences |
| Caribbean Project | 7487 | 5224591 |
| Danish Royal Library | 5661 | 657618 |
| Total | 13148 | 5882209 |
| Period | Decade | #Issues | Total |
| International conflicts | 1710–1770 | 15 | |
| 1770s | 747 | 1886 | |
| and slave rebellions | 1780s | 283 | |
| 1790s | 841 | ||
| Revolutions and nation building | 1800s | 604 | 3790 |
| 1810s | 1347 | ||
| 1820s | 1839 | ||
| Abolishment of slavery | 1830s | 1838 | 7453 |
| 1840s | 1197 | ||
| 1850s | 1111 | ||
| 1860s | 1521 | ||
| 1870s | 1786 |
| #Entities | #Males | #Females | #Non-whites | #Non-white males | #Non-white females |
| 601 468 | 387 292 | 78 821 | 8 525 | 4 543 | 1 548 |
| Tokenisation | Embedding Size | Min Freq | Mean JS Top 20 | Correct Word in Top 5 (all words) | % Misspelling in vocabulary |
| BPE | 100 | 20 | 0.66 | 37.04 | 94.44 |
| 100 | 100 | 0.66 | 37.04 | 94.44 | |
| 300 | 20 | 0.63 | 40.74 | 94.44 | |
| 300 | 100 | 0.64 | 39.81 | 94.44 | |
| SpaCy | 100 | 20 | 0.59 | 63.89 | 74.07 |
| 100 | 100 | 0.65 | 48.15 | 56.48 | |
| 300 | 20 | 0.55 | 63.89 | 74.07 | |
| 300 | 100 | 0.61 | 50.00 | 56.48 |
| Subgroup | Wordlist |
| Males | husband, suitor, brother, boyhood, beau, salesman, daddy, man, spokesman, chairman, lad, mister, men, sperm, dad, gelding, gentle-man, boy, sir, horsemen, paternity, statesman, prince, sons, countryman, pa, suitors, stallion, fella, monks, fiance, chap, uncles, godfather, bulls, males, grandfather, penis, lions, nephew, monk, countrymen, grandsons, beards, schoolboy, councilmen, dads, fellow, colts, mr, king, father, fraternal, baritone, gentlemen, fathers, husbands, guy, semen, brotherhood, nephews, lion, lads, grandson, widower, bachelor, kings, male, son, brothers, uncle, brethren, boys, councilman, czar, beard, bull, salesmen, fraternity, dude, colt, john, he, himself, his |
| Females | sisters, queen, ladies, princess, witch, mother, nun, aunt, princes, house-wife, women, convent, gals, witches, stepmother, wife, granddaughter, mis, widows, nieces, studs, niece, actresses, wives, sister, dowry, hens, daughters, womb, monastery, ms, misses, mama, mrs, fillies, woman, aunts, girl, actress, wench, brides, grandmother, stud, lady, female, maid, gal, queens, hostess, daughter, grandmothers, girls, heiress, moms, maids, mistress, mothers, mom, mare, folly, maternal, bride, widow, goddess, diva, maiden, hen, housewives, heroine, nuns, females', she, herself, hers, her |
| Non-whites | negro, negros, creole, indian, negroes, colored, mulatto, mulattos, negresse, mundingo, brown, browns, african, congo, black, blacks, dark, creoles |
| Whites | (any entity that was not classified as Non-white) |
| Attribute | Wordlist |
| Males | husband, man, mister, gentleman, boy, sir, prince, countryman, fiance, godfather, grandfather, nephew, fellow, mr, king, father, guy, grandson, widower, bachelor, male, son, brother, uncle, brethren |
| Females | sister, queen, lady, witch, mother, aunt, princes, housewife, stepmother, wife, granddaughter, mis, niece, ms, misses, mrs, woman, girl, wench, bride, grandmother, female, maid, daughter, mistress, bride, widow, maiden |
| European countries | ireland, georgia, france, monaco, poland, cyprus, greece, hungary, norway, portugal, belgium, luxembourg, finland, albania, germany, netherlands, montenegro, scotland, spain, europe, russia, vatican, switzerland, lithuania, bulgaria, wales, ukraine, romania, denmark, england, italy, bosnia, turkey, malta, iceland, austria, croatia, sweden, macedonia |
| African countries | liberia, mozambique, gambia, ghana, morocco, chad, senegal, togo, algeria, egypt, benin, ethiopia, niger, madagascar, guinea, mauritius, africa, Mali, congo, angola |
| Caribbean countries | barbuda, bahamas, jamaica, dominica, haiti, antigua, Grenada, caribbean, barbados, cuba, trinidad, dominican, nevis, kitts, lucia, croix, tobago, grenadines, puerto, rico |
| Target | Wordlist |
| Appearance | apt, discerning, judicious, imaginative, inquiring, intelligent, inquisitive, wise, shrewd, logical, astute, intuitive, precocious, analytical, smart, ingenious, reflective, inventive, venerable, genius, brilliant, clever, thoughtful |
| Intelligence | bold, strong, muscular, thin, veluptuous, blushing, athletic, gorgeous, handsome, homely, feeble, fashionable, attractive, weak, plump, ugly, slim, stout, pretty, fat, sensual, beautiful, healthy, alluring, slender |
| Weak | failure, loser, weak, timid, withdraw, follow, fragile, afraid, weakness, shy, lose, surrender, vulnerable, yield |
| Strong | strong, potent, succeed, loud, assert, leader, winner, dominant, command, confident, power, triumph, shout, bold |
| Family | loved, sisters, mother, reunited, estranged, aunt, relatives, grandchildren, godmother, kin, grand- sons, sons, son, parents, stepmother, childless, paramour, nieces, children, niece, father, twins, sister, fame, daughters, youngest, uncle, uncles, aunts, eldest, cousins, grandmother, children, loving, daughter, paternal, girls, nephews, friends, mothers, grandfather, cousin, maternal, married, nephew, wedding, grandson |
| Career | branch, managers, usurping, subsidiary, engineering, performs, fiscal, personnel, duties, offices, clerical, engineer, executive, functions, revenues, entity, competitive, competitor, employing, chairman, director, commissions, audit, promotion, professional, assistant, company, auditors, oversight, departments, comptroller, president, manager, operations, marketing, directors, share-holder, engineers, corporate, salaries, internal, management, salaried, corporation, revenue, salary, usurpation, managing, delegated, operating |
| Manual labour | sailor, bricklayer, server, butcher, gardener, cook, repairer, maid, guard, farmer, fisher, carpenter, paver, cleaner, cabinetmaker, barber, breeder, washer, miner, builder, baker, fisherman, plumber, labourer, servant |
| Non-manual labour | teacher, judge, manager, lawyer, director, mathematician, physician, medic, designer, bookkeeper, nurse, librarian, doctor, educator, auditor, clerk, midwife, translator, inspector, surgeon |
| Mental illness | sleep, pica, disorders, nightmare, personality, histrionic, stress, dependence, anxiety, terror, emotional, delusion, depression, panic, abuse, disorder, mania, hysteria |
| Physical illness | scurvy, sciatica, asthma, gangrene, gerd, cowpox, lice, rickets, malaria, epilepsy, sars, diphtheria, smallpox, bronchitis, thrush, leprosy, typhus, sids, watkins, measles, jaundice, shingles, cholera, boil, pneumonia, mumps, rheumatism, rabies, abscess, warts, plague, dysentery, syphilis, cancer, influenza, ulcers, tetanus |
| Crime | arrested, unreliable, detaineded, arrest, detain, murder, murdered, criminal, criminally, thug, theft, thief, mugger, mugging, suspicious, executed, illegal, unjust, jailed, jail, prison, arson, arsonist, kidnap, kidnapped, assaulted, assault, released, custody, police, sheriff, bailed, bail |
| lawfulness | loyal, charming, friendly, respectful, dutiful, grateful, amiable, honourable, honourably, good, faithfully, faithful, pleasant, praised, just, dignified, approving, approve, compliment, generous, faithful, intelligent, appreciative, delighted, appreciate |
| Attribute | Ratio of correctly classified entities | Ratio of incorrectly classified entities | Ratio of unable to classify |
| Non-whites | 0.89 | 0.036 | 0.07 |
| Whites | 0.75 | 0.18 | 0.07 |
| Males | 0.89 | 0.036 | 0.07 |
| Females | 0.79 | 0.21 | 0 |
| Word | Sentence |
| ingenious | This comprehensive piece of clockwork cost the ingenious and indefatig-able artist (one Jacob Lovelace, of Exeter,) 34 years’ labour. |
| elderly | y un away for upwards of 16 Months past; elderly NEGRO WOMAN hamed LOUISA, belongifg to the Estate of the late Ancup. |
| active | FOR SALE, STRONG active NEGRO GIRL, about 24 Years of Age, she is a good Cook, can W asu, [rron, and is well acquainted with Housework in general. |
| beautiful | and the young husband was hurried away, being scarcely permitted to take a parting kiss from his blooming and beautiful bride. |
| blind | Dick, of the Mundingo Country, blind mark, about 18 years of ane, says he belongs te the estate Of ee Nichole, dec. of Mantego bay. |
| sick | The young wife had snatched upa.; few of her own and her baby’s clothes; the husband, | Openiug Chorus, though sick, had attended to his duty to the last, and es | Song caped penniless with the clothes on his back. |
| free | A free black girl JOSEPHINE, detained by the Police as being diseased; Proprietors and Managers an the Country are kindly requested to have the said Josephine apprehended ‘and lodged in the Towa Prison, the usual reward will be paid |
| brave | From that moment the brave Lopez Lara was only occupied in devising means for delivering this notorious criminal into the hvids of justice. |
| Source | #Files | #Sentences |
| Caribbean Project | 7487 | 5224591 |
| Danish Royal Library | 5661 | 657618 |
| Total | 13148 | 5882209 |
| Period | Decade | #Issues | Total |
| International conflicts | 1710–1770 | 15 | |
| 1770s | 747 | 1886 | |
| and slave rebellions | 1780s | 283 | |
| 1790s | 841 | ||
| Revolutions and nation building | 1800s | 604 | 3790 |
| 1810s | 1347 | ||
| 1820s | 1839 | ||
| Abolishment of slavery | 1830s | 1838 | 7453 |
| 1840s | 1197 | ||
| 1850s | 1111 | ||
| 1860s | 1521 | ||
| 1870s | 1786 |
| #Entities | #Males | #Females | #Non-whites | #Non-white males | #Non-white females |
| 601 468 | 387 292 | 78 821 | 8 525 | 4 543 | 1 548 |
| Tokenisation | Embedding Size | Min Freq | Mean JS Top 20 | Correct Word in Top 5 (all words) | % Misspelling in vocabulary |
| BPE | 100 | 20 | 0.66 | 37.04 | 94.44 |
| 100 | 100 | 0.66 | 37.04 | 94.44 | |
| 300 | 20 | 0.63 | 40.74 | 94.44 | |
| 300 | 100 | 0.64 | 39.81 | 94.44 | |
| SpaCy | 100 | 20 | 0.59 | 63.89 | 74.07 |
| 100 | 100 | 0.65 | 48.15 | 56.48 | |
| 300 | 20 | 0.55 | 63.89 | 74.07 | |
| 300 | 100 | 0.61 | 50.00 | 56.48 |
| Subgroup | Wordlist |
| Males | husband, suitor, brother, boyhood, beau, salesman, daddy, man, spokesman, chairman, lad, mister, men, sperm, dad, gelding, gentle-man, boy, sir, horsemen, paternity, statesman, prince, sons, countryman, pa, suitors, stallion, fella, monks, fiance, chap, uncles, godfather, bulls, males, grandfather, penis, lions, nephew, monk, countrymen, grandsons, beards, schoolboy, councilmen, dads, fellow, colts, mr, king, father, fraternal, baritone, gentlemen, fathers, husbands, guy, semen, brotherhood, nephews, lion, lads, grandson, widower, bachelor, kings, male, son, brothers, uncle, brethren, boys, councilman, czar, beard, bull, salesmen, fraternity, dude, colt, john, he, himself, his |
| Females | sisters, queen, ladies, princess, witch, mother, nun, aunt, princes, house-wife, women, convent, gals, witches, stepmother, wife, granddaughter, mis, widows, nieces, studs, niece, actresses, wives, sister, dowry, hens, daughters, womb, monastery, ms, misses, mama, mrs, fillies, woman, aunts, girl, actress, wench, brides, grandmother, stud, lady, female, maid, gal, queens, hostess, daughter, grandmothers, girls, heiress, moms, maids, mistress, mothers, mom, mare, folly, maternal, bride, widow, goddess, diva, maiden, hen, housewives, heroine, nuns, females', she, herself, hers, her |
| Non-whites | negro, negros, creole, indian, negroes, colored, mulatto, mulattos, negresse, mundingo, brown, browns, african, congo, black, blacks, dark, creoles |
| Whites | (any entity that was not classified as Non-white) |
| Attribute | Wordlist |
| Males | husband, man, mister, gentleman, boy, sir, prince, countryman, fiance, godfather, grandfather, nephew, fellow, mr, king, father, guy, grandson, widower, bachelor, male, son, brother, uncle, brethren |
| Females | sister, queen, lady, witch, mother, aunt, princes, housewife, stepmother, wife, granddaughter, mis, niece, ms, misses, mrs, woman, girl, wench, bride, grandmother, female, maid, daughter, mistress, bride, widow, maiden |
| European countries | ireland, georgia, france, monaco, poland, cyprus, greece, hungary, norway, portugal, belgium, luxembourg, finland, albania, germany, netherlands, montenegro, scotland, spain, europe, russia, vatican, switzerland, lithuania, bulgaria, wales, ukraine, romania, denmark, england, italy, bosnia, turkey, malta, iceland, austria, croatia, sweden, macedonia |
| African countries | liberia, mozambique, gambia, ghana, morocco, chad, senegal, togo, algeria, egypt, benin, ethiopia, niger, madagascar, guinea, mauritius, africa, Mali, congo, angola |
| Caribbean countries | barbuda, bahamas, jamaica, dominica, haiti, antigua, Grenada, caribbean, barbados, cuba, trinidad, dominican, nevis, kitts, lucia, croix, tobago, grenadines, puerto, rico |
| Target | Wordlist |
| Appearance | apt, discerning, judicious, imaginative, inquiring, intelligent, inquisitive, wise, shrewd, logical, astute, intuitive, precocious, analytical, smart, ingenious, reflective, inventive, venerable, genius, brilliant, clever, thoughtful |
| Intelligence | bold, strong, muscular, thin, veluptuous, blushing, athletic, gorgeous, handsome, homely, feeble, fashionable, attractive, weak, plump, ugly, slim, stout, pretty, fat, sensual, beautiful, healthy, alluring, slender |
| Weak | failure, loser, weak, timid, withdraw, follow, fragile, afraid, weakness, shy, lose, surrender, vulnerable, yield |
| Strong | strong, potent, succeed, loud, assert, leader, winner, dominant, command, confident, power, triumph, shout, bold |
| Family | loved, sisters, mother, reunited, estranged, aunt, relatives, grandchildren, godmother, kin, grand- sons, sons, son, parents, stepmother, childless, paramour, nieces, children, niece, father, twins, sister, fame, daughters, youngest, uncle, uncles, aunts, eldest, cousins, grandmother, children, loving, daughter, paternal, girls, nephews, friends, mothers, grandfather, cousin, maternal, married, nephew, wedding, grandson |
| Career | branch, managers, usurping, subsidiary, engineering, performs, fiscal, personnel, duties, offices, clerical, engineer, executive, functions, revenues, entity, competitive, competitor, employing, chairman, director, commissions, audit, promotion, professional, assistant, company, auditors, oversight, departments, comptroller, president, manager, operations, marketing, directors, share-holder, engineers, corporate, salaries, internal, management, salaried, corporation, revenue, salary, usurpation, managing, delegated, operating |
| Manual labour | sailor, bricklayer, server, butcher, gardener, cook, repairer, maid, guard, farmer, fisher, carpenter, paver, cleaner, cabinetmaker, barber, breeder, washer, miner, builder, baker, fisherman, plumber, labourer, servant |
| Non-manual labour | teacher, judge, manager, lawyer, director, mathematician, physician, medic, designer, bookkeeper, nurse, librarian, doctor, educator, auditor, clerk, midwife, translator, inspector, surgeon |
| Mental illness | sleep, pica, disorders, nightmare, personality, histrionic, stress, dependence, anxiety, terror, emotional, delusion, depression, panic, abuse, disorder, mania, hysteria |
| Physical illness | scurvy, sciatica, asthma, gangrene, gerd, cowpox, lice, rickets, malaria, epilepsy, sars, diphtheria, smallpox, bronchitis, thrush, leprosy, typhus, sids, watkins, measles, jaundice, shingles, cholera, boil, pneumonia, mumps, rheumatism, rabies, abscess, warts, plague, dysentery, syphilis, cancer, influenza, ulcers, tetanus |
| Crime | arrested, unreliable, detaineded, arrest, detain, murder, murdered, criminal, criminally, thug, theft, thief, mugger, mugging, suspicious, executed, illegal, unjust, jailed, jail, prison, arson, arsonist, kidnap, kidnapped, assaulted, assault, released, custody, police, sheriff, bailed, bail |
| lawfulness | loyal, charming, friendly, respectful, dutiful, grateful, amiable, honourable, honourably, good, faithfully, faithful, pleasant, praised, just, dignified, approving, approve, compliment, generous, faithful, intelligent, appreciative, delighted, appreciate |
| Attribute | Ratio of correctly classified entities | Ratio of incorrectly classified entities | Ratio of unable to classify |
| Non-whites | 0.89 | 0.036 | 0.07 |
| Whites | 0.75 | 0.18 | 0.07 |
| Males | 0.89 | 0.036 | 0.07 |
| Females | 0.79 | 0.21 | 0 |
| Word | Sentence |
| ingenious | This comprehensive piece of clockwork cost the ingenious and indefatig-able artist (one Jacob Lovelace, of Exeter,) 34 years’ labour. |
| elderly | y un away for upwards of 16 Months past; elderly NEGRO WOMAN hamed LOUISA, belongifg to the Estate of the late Ancup. |
| active | FOR SALE, STRONG active NEGRO GIRL, about 24 Years of Age, she is a good Cook, can W asu, [rron, and is well acquainted with Housework in general. |
| beautiful | and the young husband was hurried away, being scarcely permitted to take a parting kiss from his blooming and beautiful bride. |
| blind | Dick, of the Mundingo Country, blind mark, about 18 years of ane, says he belongs te the estate Of ee Nichole, dec. of Mantego bay. |
| sick | The young wife had snatched upa.; few of her own and her baby’s clothes; the husband, | Openiug Chorus, though sick, had attended to his duty to the last, and es | Song caped penniless with the clothes on his back. |
| free | A free black girl JOSEPHINE, detained by the Police as being diseased; Proprietors and Managers an the Country are kindly requested to have the said Josephine apprehended ‘and lodged in the Towa Prison, the usual reward will be paid |
| brave | From that moment the brave Lopez Lara was only occupied in devising means for delivering this notorious criminal into the hvids of justice. |
| Source | #Files | #Sentences |
| Caribbean Project | 7487 | 5224591 |
| Danish Royal Library | 5661 | 657618 |
| Total | 13148 | 5882209 |
| Period | Decade | #Issues | Total |
| International conflicts | 1710–1770 | 15 | |
| 1770s | 747 | 1886 | |
| and slave rebellions | 1780s | 283 | |
| 1790s | 841 | ||
| Revolutions and nation building | 1800s | 604 | 3790 |
| 1810s | 1347 | ||
| 1820s | 1839 | ||
| Abolishment of slavery | 1830s | 1838 | 7453 |
| 1840s | 1197 | ||
| 1850s | 1111 | ||
| 1860s | 1521 | ||
| 1870s | 1786 |
| #Entities | #Males | #Females | #Non-whites | #Non-white males | #Non-white females |
| 601 468 | 387 292 | 78 821 | 8 525 | 4 543 | 1 548 |
| Tokenisation | Embedding Size | Min Freq | Mean JS Top 20 | Correct Word in Top 5 (all words) | % Misspelling in vocabulary |
| BPE | 100 | 20 | 0.66 | 37.04 | 94.44 |
| 100 | 100 | 0.66 | 37.04 | 94.44 | |
| 300 | 20 | 0.63 | 40.74 | 94.44 | |
| 300 | 100 | 0.64 | 39.81 | 94.44 | |
| SpaCy | 100 | 20 | 0.59 | 63.89 | 74.07 |
| 100 | 100 | 0.65 | 48.15 | 56.48 | |
| 300 | 20 | 0.55 | 63.89 | 74.07 | |
| 300 | 100 | 0.61 | 50.00 | 56.48 |
| Subgroup | Wordlist |
| Males | husband, suitor, brother, boyhood, beau, salesman, daddy, man, spokesman, chairman, lad, mister, men, sperm, dad, gelding, gentle-man, boy, sir, horsemen, paternity, statesman, prince, sons, countryman, pa, suitors, stallion, fella, monks, fiance, chap, uncles, godfather, bulls, males, grandfather, penis, lions, nephew, monk, countrymen, grandsons, beards, schoolboy, councilmen, dads, fellow, colts, mr, king, father, fraternal, baritone, gentlemen, fathers, husbands, guy, semen, brotherhood, nephews, lion, lads, grandson, widower, bachelor, kings, male, son, brothers, uncle, brethren, boys, councilman, czar, beard, bull, salesmen, fraternity, dude, colt, john, he, himself, his |
| Females | sisters, queen, ladies, princess, witch, mother, nun, aunt, princes, house-wife, women, convent, gals, witches, stepmother, wife, granddaughter, mis, widows, nieces, studs, niece, actresses, wives, sister, dowry, hens, daughters, womb, monastery, ms, misses, mama, mrs, fillies, woman, aunts, girl, actress, wench, brides, grandmother, stud, lady, female, maid, gal, queens, hostess, daughter, grandmothers, girls, heiress, moms, maids, mistress, mothers, mom, mare, folly, maternal, bride, widow, goddess, diva, maiden, hen, housewives, heroine, nuns, females', she, herself, hers, her |
| Non-whites | negro, negros, creole, indian, negroes, colored, mulatto, mulattos, negresse, mundingo, brown, browns, african, congo, black, blacks, dark, creoles |
| Whites | (any entity that was not classified as Non-white) |
| Attribute | Wordlist |
| Males | husband, man, mister, gentleman, boy, sir, prince, countryman, fiance, godfather, grandfather, nephew, fellow, mr, king, father, guy, grandson, widower, bachelor, male, son, brother, uncle, brethren |
| Females | sister, queen, lady, witch, mother, aunt, princes, housewife, stepmother, wife, granddaughter, mis, niece, ms, misses, mrs, woman, girl, wench, bride, grandmother, female, maid, daughter, mistress, bride, widow, maiden |
| European countries | ireland, georgia, france, monaco, poland, cyprus, greece, hungary, norway, portugal, belgium, luxembourg, finland, albania, germany, netherlands, montenegro, scotland, spain, europe, russia, vatican, switzerland, lithuania, bulgaria, wales, ukraine, romania, denmark, england, italy, bosnia, turkey, malta, iceland, austria, croatia, sweden, macedonia |
| African countries | liberia, mozambique, gambia, ghana, morocco, chad, senegal, togo, algeria, egypt, benin, ethiopia, niger, madagascar, guinea, mauritius, africa, Mali, congo, angola |
| Caribbean countries | barbuda, bahamas, jamaica, dominica, haiti, antigua, Grenada, caribbean, barbados, cuba, trinidad, dominican, nevis, kitts, lucia, croix, tobago, grenadines, puerto, rico |
| Target | Wordlist |
| Appearance | apt, discerning, judicious, imaginative, inquiring, intelligent, inquisitive, wise, shrewd, logical, astute, intuitive, precocious, analytical, smart, ingenious, reflective, inventive, venerable, genius, brilliant, clever, thoughtful |
| Intelligence | bold, strong, muscular, thin, veluptuous, blushing, athletic, gorgeous, handsome, homely, feeble, fashionable, attractive, weak, plump, ugly, slim, stout, pretty, fat, sensual, beautiful, healthy, alluring, slender |
| Weak | failure, loser, weak, timid, withdraw, follow, fragile, afraid, weakness, shy, lose, surrender, vulnerable, yield |
| Strong | strong, potent, succeed, loud, assert, leader, winner, dominant, command, confident, power, triumph, shout, bold |
| Family | loved, sisters, mother, reunited, estranged, aunt, relatives, grandchildren, godmother, kin, grand- sons, sons, son, parents, stepmother, childless, paramour, nieces, children, niece, father, twins, sister, fame, daughters, youngest, uncle, uncles, aunts, eldest, cousins, grandmother, children, loving, daughter, paternal, girls, nephews, friends, mothers, grandfather, cousin, maternal, married, nephew, wedding, grandson |
| Career | branch, managers, usurping, subsidiary, engineering, performs, fiscal, personnel, duties, offices, clerical, engineer, executive, functions, revenues, entity, competitive, competitor, employing, chairman, director, commissions, audit, promotion, professional, assistant, company, auditors, oversight, departments, comptroller, president, manager, operations, marketing, directors, share-holder, engineers, corporate, salaries, internal, management, salaried, corporation, revenue, salary, usurpation, managing, delegated, operating |
| Manual labour | sailor, bricklayer, server, butcher, gardener, cook, repairer, maid, guard, farmer, fisher, carpenter, paver, cleaner, cabinetmaker, barber, breeder, washer, miner, builder, baker, fisherman, plumber, labourer, servant |
| Non-manual labour | teacher, judge, manager, lawyer, director, mathematician, physician, medic, designer, bookkeeper, nurse, librarian, doctor, educator, auditor, clerk, midwife, translator, inspector, surgeon |
| Mental illness | sleep, pica, disorders, nightmare, personality, histrionic, stress, dependence, anxiety, terror, emotional, delusion, depression, panic, abuse, disorder, mania, hysteria |
| Physical illness | scurvy, sciatica, asthma, gangrene, gerd, cowpox, lice, rickets, malaria, epilepsy, sars, diphtheria, smallpox, bronchitis, thrush, leprosy, typhus, sids, watkins, measles, jaundice, shingles, cholera, boil, pneumonia, mumps, rheumatism, rabies, abscess, warts, plague, dysentery, syphilis, cancer, influenza, ulcers, tetanus |
| Crime | arrested, unreliable, detaineded, arrest, detain, murder, murdered, criminal, criminally, thug, theft, thief, mugger, mugging, suspicious, executed, illegal, unjust, jailed, jail, prison, arson, arsonist, kidnap, kidnapped, assaulted, assault, released, custody, police, sheriff, bailed, bail |
| lawfulness | loyal, charming, friendly, respectful, dutiful, grateful, amiable, honourable, honourably, good, faithfully, faithful, pleasant, praised, just, dignified, approving, approve, compliment, generous, faithful, intelligent, appreciative, delighted, appreciate |
| Attribute | Ratio of correctly classified entities | Ratio of incorrectly classified entities | Ratio of unable to classify |
| Non-whites | 0.89 | 0.036 | 0.07 |
| Whites | 0.75 | 0.18 | 0.07 |
| Males | 0.89 | 0.036 | 0.07 |
| Females | 0.79 | 0.21 | 0 |
| Word | Sentence |
| ingenious | This comprehensive piece of clockwork cost the ingenious and indefatig-able artist (one Jacob Lovelace, of Exeter,) 34 years’ labour. |
| elderly | y un away for upwards of 16 Months past; elderly NEGRO WOMAN hamed LOUISA, belongifg to the Estate of the late Ancup. |
| active | FOR SALE, STRONG active NEGRO GIRL, about 24 Years of Age, she is a good Cook, can W asu, [rron, and is well acquainted with Housework in general. |
| beautiful | and the young husband was hurried away, being scarcely permitted to take a parting kiss from his blooming and beautiful bride. |
| blind | Dick, of the Mundingo Country, blind mark, about 18 years of ane, says he belongs te the estate Of ee Nichole, dec. of Mantego bay. |
| sick | The young wife had snatched upa.; few of her own and her baby’s clothes; the husband, | Openiug Chorus, though sick, had attended to his duty to the last, and es | Song caped penniless with the clothes on his back. |
| free | A free black girl JOSEPHINE, detained by the Police as being diseased; Proprietors and Managers an the Country are kindly requested to have the said Josephine apprehended ‘and lodged in the Towa Prison, the usual reward will be paid |
| brave | From that moment the brave Lopez Lara was only occupied in devising means for delivering this notorious criminal into the hvids of justice. |
| Source | #Files | #Sentences |
| Caribbean Project | 7487 | 5224591 |
| Danish Royal Library | 5661 | 657618 |
| Total | 13148 | 5882209 |
| Period | Decade | #Issues | Total |
| International conflicts | 1710–1770 | 15 | |
| 1770s | 747 | 1886 | |
| and slave rebellions | 1780s | 283 | |
| 1790s | 841 | ||
| Revolutions and nation building | 1800s | 604 | 3790 |
| 1810s | 1347 | ||
| 1820s | 1839 | ||
| Abolishment of slavery | 1830s | 1838 | 7453 |
| 1840s | 1197 | ||
| 1850s | 1111 | ||
| 1860s | 1521 | ||
| 1870s | 1786 |
| #Entities | #Males | #Females | #Non-whites | #Non-white males | #Non-white females |
| 601 468 | 387 292 | 78 821 | 8 525 | 4 543 | 1 548 |
| Tokenisation | Embedding Size | Min Freq | Mean JS Top 20 | Correct Word in Top 5 (all words) | % Misspelling in vocabulary |
| BPE | 100 | 20 | 0.66 | 37.04 | 94.44 |
| 100 | 100 | 0.66 | 37.04 | 94.44 | |
| 300 | 20 | 0.63 | 40.74 | 94.44 | |
| 300 | 100 | 0.64 | 39.81 | 94.44 | |
| SpaCy | 100 | 20 | 0.59 | 63.89 | 74.07 |
| 100 | 100 | 0.65 | 48.15 | 56.48 | |
| 300 | 20 | 0.55 | 63.89 | 74.07 | |
| 300 | 100 | 0.61 | 50.00 | 56.48 |
| Subgroup | Wordlist |
| Males | husband, suitor, brother, boyhood, beau, salesman, daddy, man, spokesman, chairman, lad, mister, men, sperm, dad, gelding, gentle-man, boy, sir, horsemen, paternity, statesman, prince, sons, countryman, pa, suitors, stallion, fella, monks, fiance, chap, uncles, godfather, bulls, males, grandfather, penis, lions, nephew, monk, countrymen, grandsons, beards, schoolboy, councilmen, dads, fellow, colts, mr, king, father, fraternal, baritone, gentlemen, fathers, husbands, guy, semen, brotherhood, nephews, lion, lads, grandson, widower, bachelor, kings, male, son, brothers, uncle, brethren, boys, councilman, czar, beard, bull, salesmen, fraternity, dude, colt, john, he, himself, his |
| Females | sisters, queen, ladies, princess, witch, mother, nun, aunt, princes, house-wife, women, convent, gals, witches, stepmother, wife, granddaughter, mis, widows, nieces, studs, niece, actresses, wives, sister, dowry, hens, daughters, womb, monastery, ms, misses, mama, mrs, fillies, woman, aunts, girl, actress, wench, brides, grandmother, stud, lady, female, maid, gal, queens, hostess, daughter, grandmothers, girls, heiress, moms, maids, mistress, mothers, mom, mare, folly, maternal, bride, widow, goddess, diva, maiden, hen, housewives, heroine, nuns, females', she, herself, hers, her |
| Non-whites | negro, negros, creole, indian, negroes, colored, mulatto, mulattos, negresse, mundingo, brown, browns, african, congo, black, blacks, dark, creoles |
| Whites | (any entity that was not classified as Non-white) |
| Attribute | Wordlist |
| Males | husband, man, mister, gentleman, boy, sir, prince, countryman, fiance, godfather, grandfather, nephew, fellow, mr, king, father, guy, grandson, widower, bachelor, male, son, brother, uncle, brethren |
| Females | sister, queen, lady, witch, mother, aunt, princes, housewife, stepmother, wife, granddaughter, mis, niece, ms, misses, mrs, woman, girl, wench, bride, grandmother, female, maid, daughter, mistress, bride, widow, maiden |
| European countries | ireland, georgia, france, monaco, poland, cyprus, greece, hungary, norway, portugal, belgium, luxembourg, finland, albania, germany, netherlands, montenegro, scotland, spain, europe, russia, vatican, switzerland, lithuania, bulgaria, wales, ukraine, romania, denmark, england, italy, bosnia, turkey, malta, iceland, austria, croatia, sweden, macedonia |
| African countries | liberia, mozambique, gambia, ghana, morocco, chad, senegal, togo, algeria, egypt, benin, ethiopia, niger, madagascar, guinea, mauritius, africa, Mali, congo, angola |
| Caribbean countries | barbuda, bahamas, jamaica, dominica, haiti, antigua, Grenada, caribbean, barbados, cuba, trinidad, dominican, nevis, kitts, lucia, croix, tobago, grenadines, puerto, rico |
| Target | Wordlist |
| Appearance | apt, discerning, judicious, imaginative, inquiring, intelligent, inquisitive, wise, shrewd, logical, astute, intuitive, precocious, analytical, smart, ingenious, reflective, inventive, venerable, genius, brilliant, clever, thoughtful |
| Intelligence | bold, strong, muscular, thin, veluptuous, blushing, athletic, gorgeous, handsome, homely, feeble, fashionable, attractive, weak, plump, ugly, slim, stout, pretty, fat, sensual, beautiful, healthy, alluring, slender |
| Weak | failure, loser, weak, timid, withdraw, follow, fragile, afraid, weakness, shy, lose, surrender, vulnerable, yield |
| Strong | strong, potent, succeed, loud, assert, leader, winner, dominant, command, confident, power, triumph, shout, bold |
| Family | loved, sisters, mother, reunited, estranged, aunt, relatives, grandchildren, godmother, kin, grand- sons, sons, son, parents, stepmother, childless, paramour, nieces, children, niece, father, twins, sister, fame, daughters, youngest, uncle, uncles, aunts, eldest, cousins, grandmother, children, loving, daughter, paternal, girls, nephews, friends, mothers, grandfather, cousin, maternal, married, nephew, wedding, grandson |
| Career | branch, managers, usurping, subsidiary, engineering, performs, fiscal, personnel, duties, offices, clerical, engineer, executive, functions, revenues, entity, competitive, competitor, employing, chairman, director, commissions, audit, promotion, professional, assistant, company, auditors, oversight, departments, comptroller, president, manager, operations, marketing, directors, share-holder, engineers, corporate, salaries, internal, management, salaried, corporation, revenue, salary, usurpation, managing, delegated, operating |
| Manual labour | sailor, bricklayer, server, butcher, gardener, cook, repairer, maid, guard, farmer, fisher, carpenter, paver, cleaner, cabinetmaker, barber, breeder, washer, miner, builder, baker, fisherman, plumber, labourer, servant |
| Non-manual labour | teacher, judge, manager, lawyer, director, mathematician, physician, medic, designer, bookkeeper, nurse, librarian, doctor, educator, auditor, clerk, midwife, translator, inspector, surgeon |
| Mental illness | sleep, pica, disorders, nightmare, personality, histrionic, stress, dependence, anxiety, terror, emotional, delusion, depression, panic, abuse, disorder, mania, hysteria |
| Physical illness | scurvy, sciatica, asthma, gangrene, gerd, cowpox, lice, rickets, malaria, epilepsy, sars, diphtheria, smallpox, bronchitis, thrush, leprosy, typhus, sids, watkins, measles, jaundice, shingles, cholera, boil, pneumonia, mumps, rheumatism, rabies, abscess, warts, plague, dysentery, syphilis, cancer, influenza, ulcers, tetanus |
| Crime | arrested, unreliable, detaineded, arrest, detain, murder, murdered, criminal, criminally, thug, theft, thief, mugger, mugging, suspicious, executed, illegal, unjust, jailed, jail, prison, arson, arsonist, kidnap, kidnapped, assaulted, assault, released, custody, police, sheriff, bailed, bail |
| lawfulness | loyal, charming, friendly, respectful, dutiful, grateful, amiable, honourable, honourably, good, faithfully, faithful, pleasant, praised, just, dignified, approving, approve, compliment, generous, faithful, intelligent, appreciative, delighted, appreciate |
| Attribute | Ratio of correctly classified entities | Ratio of incorrectly classified entities | Ratio of unable to classify |
| Non-whites | 0.89 | 0.036 | 0.07 |
| Whites | 0.75 | 0.18 | 0.07 |
| Males | 0.89 | 0.036 | 0.07 |
| Females | 0.79 | 0.21 | 0 |
| Word | Sentence |
| ingenious | This comprehensive piece of clockwork cost the ingenious and indefatig-able artist (one Jacob Lovelace, of Exeter,) 34 years’ labour. |
| elderly | y un away for upwards of 16 Months past; elderly NEGRO WOMAN hamed LOUISA, belongifg to the Estate of the late Ancup. |
| active | FOR SALE, STRONG active NEGRO GIRL, about 24 Years of Age, she is a good Cook, can W asu, [rron, and is well acquainted with Housework in general. |
| beautiful | and the young husband was hurried away, being scarcely permitted to take a parting kiss from his blooming and beautiful bride. |
| blind | Dick, of the Mundingo Country, blind mark, about 18 years of ane, says he belongs te the estate Of ee Nichole, dec. of Mantego bay. |
| sick | The young wife had snatched upa.; few of her own and her baby’s clothes; the husband, | Openiug Chorus, though sick, had attended to his duty to the last, and es | Song caped penniless with the clothes on his back. |
| free | A free black girl JOSEPHINE, detained by the Police as being diseased; Proprietors and Managers an the Country are kindly requested to have the said Josephine apprehended ‘and lodged in the Towa Prison, the usual reward will be paid |
| brave | From that moment the brave Lopez Lara was only occupied in devising means for delivering this notorious criminal into the hvids of justice. |
| Dataset | #D | #R | #W/#R | #T | #IA | #T/#I | POS #n-gram | DP #n-gram | #vocab |
| DP | 2 | 6037 | 16.49 | 9309 | 0 | 1.54 | 14.25/16.05/16.52 | 13.51/14.24/14.43 | 6512 |
| DW | 2 | 6009 | 16.46 | 10390 | 0 | 1.73 | 14.22/16.02/16.48 | 13.48/14.21/14.40 | 6484 |
| DX | 2 | 5989 | 16.43 | 10252 | 0 | 1.71 | 14.20/16.00/16.46 | 13.46/14.19/14.38 | 6467 |
| DMASTE | 8 | 7524 | 59.68 | 28233 | 11945 | 3.75 | 36.67/52.11/57.92 | 35.08/40.20/41.93 | 19226 |
| Domain | Electronics | Fashion | Beauty | Home | Book | Pet | Toy | Grocery |
| Train | 1395 | 851 | 535 | 1050 | 0 | 0 | 0 | 0 |
| Dev | 200 | 121 | 77 | 152 | 159 | 167 | 173 | 173 |
| Test | 399 | 245 | 154 | 301 | 325 | 340 | 354 | 353 |
| Term | Aspect terms | Opinion terms | Triplets |
| IAA | 0.666 | 0.664 | 0.593 |
| Method | Electronics | Beauty | Fashion | Home | Average |
| BMRC | 41.95±0.34 | 38.57±0.97 | 44.87±0.69 | 41.18±0.66 | 41.64 |
| BART-ABSA | 43.38±1.37 | 41.13±1.14 | 43.89±0.82 | 40.56±1.04 | 42.24 |
| GAS | 47.10±0.64 | 44.32±0.52 | 47.80±1.07 | 47.22±1.13 | 46.61 |
| Span-ASTE | 47.86±0.74 | 46.46±0.66 | 50.38±0.68 | 49.14±0.41 | 48.46 |
| Domain | BMRC | BART-ABSA | GAS | Span-ASTE | BMRC+AT | Span-ASTE+AT |
| Electronics→Book | 33.74±0.64 | 35.43±0.61 | 35.57±0.76 | 40.36±1.59 | 32.56±1.57 | 39.58±0.70 |
| Beauty→Book | 30.01±1.29 | 30.24±0.83 | 30.96±0.99 | 38.58±1.58 | 28.56±0.88 | 38.79±0.89 |
| Fashion→Book | 31.71±1.37 | 32.45±1.77 | 36.26±0.64 | 39.88±0.74 | 30.80±1.59 | 39.60±0.69 |
| Home→Book | 31.93±1.79 | 33.48±1.15 | 35.91±0.75 | 39.45±0.73 | 31.31±0.45 | 39.35±0.93 |
| Electronics→Grocery | 39.68±0.70 | 40.39±1.28 | 39.16±1.25 | 45.36±1.12 | 40.15±0.55 | 43.90±0.47 |
| Beauty→Grocery | 34.46±0.85 | 34.22±0.59 | 36.22±0.50 | 40.32±0.81 | 34.91±1.00 | 40.30±1.64 |
| Fashion→Grocery | 37.18±0.50 | 37.27±1.08 | 40.13±0.38 | 43.41±0.83 | 37.56±0.29 | 42.29±0.79 |
| Home→Grocery | 39.03±1.78 | 38.56±1.14 | 42.51±0.37 | 43.74±1.02 | 38.83±0.75 | 41.78±1.79 |
| Electronics→Toy | 42.39±0.88 | 42.49±1.13 | 43.55±0.52 | 47.23±0.38 | 41.38±1.20 | 47.43±0.67 |
| Beauty→Toy | 35.66±0.94 | 33.87±1.05 | 37.95±0.63 | 41.19±0.60 | 35.98±1.14 | 40.86±0.74 |
| Fashion→Toy | 41.13±1.12 | 40.08±1.15 | 42.78±0.36 | 46.83±0.94 | 40.74±1.20 | 46.09±0.87 |
| Home→Toy | 40.26±1.22 | 40.81±1.27 | 44.16±0.73 | 47.60±1.09 | 40.42±0.51 | 46.47±0.75 |
| Electronics→Pet | 37.39±0.60 | 36.88±1.07 | 38.17±0.89 | 41.04±0.48 | 36.70±0.43 | 40.09±0.77 |
| Beauty→Pet | 32.80±1.07 | 32.07±0.76 | 32.55±0.57 | 36.41±0.40 | 32.76±0.54 | 35.38±0.83 |
| Fashion→Pet | 35.97±0.84 | 34.92±0.85 | 36.13±0.74 | 40.57±0.57 | 36.07±0.81 | 38.78±0.81 |
| Home→Pet | 37.64±1.03 | 37.26±1.59 | 39.38±0.70 | 41.42±0.51 | 37.24±0.97 | 40.64±1.05 |
| Average | 36.31 | 36.28 | 38.21 | 42.09 | 36.00 | 41.33 |
| Domain | Book | Grocery | Toy | Pet | Average |
| B + E | 40.94±1.14 | 45.94±0.56 | 48.07±0.75 | 41.70±0.79 | 44.16 |
| E + F | 40.59±0.86 | 46.14±0.98 | 50.10±0.83 | 41.90±0.79 | 44.68 |
| E + H | 41.08±1.50 | 45.34±0.50 | 48.97±0.69 | 42.70±0.41 | 44.52 |
| B + F | 41.13±0.51 | 44.60±0.73 | 46.81±0.91 | 41.03±0.59 | 43.39 |
| B + H | 41.31±1.18 | 44.31±0.84 | 47.12±0.87 | 41.67±0.93 | 43.60 |
| F + H | 42.40±0.31 | 45.97±0.56 | 49.11±0.74 | 43.12±0.70 | 45.15 |
| B + E + F | 41.34±0.73 | 46.28±0.78 | 49.44±1.08 | 41.12±0.58 | 44.80 |
| B + E + H | 41.21±0.48 | 45.39±0.49 | 49.33±0.32 | 43.48±1.12 | 44.85 |
| E + F + H | 41.44±0.71 | 45.39±0.80 | 50.24±0.57 | 43.59±1.06 | 45.17 |
| B + F + H | 41.73±0.38 | 45.40±0.63 | 48.55±0.71 | 43.46±0.19 | 44.79 |
| ALL | 41.83±0.57 | 46.07±0.47 | 50.16±0.49 | 43.62±1.32 | 45.42 |
| Electronics | |
| Good case. It has minimum padding that makes the phone feel secure but too much to be burdensome. The card slots are convenient. | _card slots, convenient, POS); (case, Good, POS); (padding, minimum, POS). |
| They work very well. The connections are nice and tight and I loose no quality over the length of the cord. | (connections, nice, POS); (connections, tight, POS); (length of the cord, loose no quality over, POS); (work, very well, POS). |
| Beauty | |
| This smells so good exactly like the body lotion and spray mist. It is a soft feminine fragrance not too overpowering. I would highly recommend! | (NULL, would highly recommend, POS); (fragrance, not too overpowering, POS); (fragrance, soft feminine, POS); (smells, so good, POS). |
| Revision products are awesome. I switch away from this in the winter months since it's not as moisturizing as an emollient based cleanser. | (Revision, awesome, POS); (NULL, not as moisturizing, NEG). |
| Home | |
| I think i just started a new hobby, this Victorinox is pretty awesome. I might have to get a few more designs. | (Victorinox, pretty awesome, POS). |
| Too big and cumbersome to be useful. If design were down-sized a couple of inches, they would be perfect. Much too big for my salad bowls. | (NULL, Much too big, NEG); (NULL, Too big, NEG); (NULL, cumbersome, NEG). |
| Fashion | |
| the fabric is very, very thin. It should be underwear, not a regular shirt. fits too snug and is see through | (NULL, see through, NEG); (fabric, very thin, NEG); (fits, too snug, NEG). |
| What can I say that you don't already know? Classic Chucks. One of the best, most versatile, and most durable sneakers out there. | (Chucks, Classic, POS); (sneakers, One of the best, POS); (sneakers, most durable, POS); (sneakers, most versatile, POS). |
| Book | |
| It was a good book, haven't read it before bow. Very interesting, with good suspense and humor mixed in! 10/10 | (NULL, Very interesting, POS); (book, good, POS); (humor, good, POS); (suspense, good, POS). |
| Challenging to read but worth it. Persevere to the end of the book. Ask God to open your mind and heart. | (NULL, Challenging to read, NEU); (NULL, worth, POS). |
| Grocery | |
| These are the best Slim Jims every, I have tried them, but the taste, the texture of the Honey BBQ is the Greatest. | (Slim Jims, best, POS); (taste, Greatest, POS); (texture, Greatest, POS). |
| A good almond extract, and I like that it's organic. I took off one star for high price. Should not cost that much. | (almond extract, good, POS); (organic, like, POS); (price, high, NEG). |
| Pet | |
| This is a good prefilter. The clear plastic adapters are a little brittle so be careful when attaching to your filter input. | (clear plastic adapters, a little brittle, NEG); (prefilter, good, POS). |
| Bought these as cat treats. Have to break them up, but my cat goes crazy for them, and this tub is lasting forever. | (NULL, goes crazy for, POS); (tub, lasting forever, POS). |
| Toy | |
| Super simple dynamics, and a great game for friends, family, kids, etc. ! Easy to learn, and variety of play is good. | (NULL, Easy to learn, POS); (dynamics, Super, POS); (dynamics, simple, POS); (game, great, POS); (variety of play, good, POS). |
| My 1 year old loved this for his birthday. Such a fun, easy toy. I would buy this again and again. | (NULL, loved, POS); (toy, easy, POS); (toy, fun, POS). |
| α | 1 | 3 | 5 | 7 | 10 | 15 | 20 | 30 | 50 | 100 |
| E→K | 33.79 | 37.46 | 38.36 | 38.74 | 38.74 | 39.46 | 35.68 | 38.99 | 35.17 | 35.70 |
| B→K | 31.55 | 35.20 | 37.95 | 36.63 | 37.83 | 36.77 | 35.37 | 34.99 | 36.47 | 28.41 |
| F→K | 32.15 | 39.83 | 39.74 | 39.93 | 39.95 | 39.38 | 38.88 | 39.86 | 40.13 | 38.93 |
| H→K | 35.15 | 35.64 | 40.12 | 38.12 | 39.96 | 38.10 | 38.45 | 37.40 | 32.99 | 38.49 |
| E→G | 39.86 | 44.71 | 44.86 | 45.11 | 44.39 | 45.66 | 44.23 | 44.59 | 44.46 | 45.48 |
| B→G | 18.92 | 40.90 | 40.69 | 42.16 | 41.01 | 40.39 | 40.49 | 41.41 | 40.51 | 41.49 |
| F→G | 38.23 | 43.98 | 43.36 | 43.88 | 44.89 | 42.65 | 44.40 | 44.23 | 43.97 | 43.78 |
| H→G | 27.58 | 44.18 | 45.00 | 43.74 | 44.38 | 44.46 | 44.25 | 44.15 | 43.52 | 44.12 |
| E→T | 33.21 | 47.82 | 48.65 | 48.48 | 47.23 | 47.95 | 47.57 | 48.26 | 48.64 | 47.33 |
| B→T | 30.05 | 41.54 | 41.12 | 43.13 | 42.65 | 42.03 | 41.76 | 41.35 | 41.52 | 42.45 |
| F→T | 33.63 | 46.47 | 47.12 | 47.65 | 47.16 | 47.73 | 48.00 | 47.31 | 46.53 | 47.08 |
| H→T | 42.09 | 48.27 | 48.24 | 48.30 | 48.44 | 48.93 | 46.19 | 47.01 | 48.41 | 48.91 |
| E→P | 33.51 | 41.18 | 41.95 | 41.19 | 41.43 | 41.64 | 41.81 | 42.01 | 42.36 | 42.06 |
| B→P | 18.03 | 35.96 | 37.05 | 35.71 | 36.92 | 36.08 | 36.26 | 36.72 | 37.15 | 36.10 |
| F→P | 35.44 | 39.30 | 39.87 | 39.98 | 40.00 | 40.07 | 40.24 | 39.61 | 40.41 | 40.40 |
| H→P | 35.41 | 41.28 | 41.55 | 42.24 | 41.35 | 42.51 | 41.41 | 37.80 | 41.25 | 41.48 |
| AVE | 32.41 | 41.48 | 42.23 | 42.19 | 42.27 | 42.11 | 41.56 | 41.61 | 41.47 | 41.39 |
| α | 1 | 10 | 30 | 50 | 100 | 500 | 700 | 800 | 1000 | 1500 |
| E→K | 0.75 | 2.22 | 6.18 | 16.90 | 6.61 | 28.08 | 31.86 | 31.67 | 30.64 | 30.23 |
| B→K | 0.30 | 1.04 | 9.57 | 17.53 | 26.56 | 27.93 | 28.62 | 27.62 | 29.46 | 29.88 |
| F→K | 0.22 | 7.66 | 8.05 | 14.82 | 27.14 | 30.34 | 32.09 | 30.72 | 31.21 | 31.09 |
| H→K | 0.37 | 9.81 | 11.22 | 18.96 | 22.74 | 26.66 | 27.92 | 31.23 | 28.52 | 28.93 |
| E→G | 2.43 | 12.42 | 24.97 | 19.98 | 35.43 | 38.71 | 39.14 | 39.48 | 39.05 | 38.13 |
| B→G | 2.13 | 20.84 | 33.99 | 32.90 | 32.63 | 33.92 | 33.84 | 32.88 | 32.89 | 32.78 |
| F→G | 2.42 | 37.73 | 27.40 | 34.71 | 38.03 | 38.98 | 37.99 | 37.56 | 38.62 | 38.20 |
| H→G | 4.55 | 30.18 | 32.95 | 37.50 | 37.72 | 37.62 | 37.86 | 38.48 | 37.65 | 37.85 |
| E→T | 4.57 | 31.09 | 35.18 | 38.12 | 42.73 | 43.35 | 42.74 | 42.86 | 43.01 | 42.59 |
| B→T | 1.65 | 7.45 | 23.93 | 33.70 | 35.53 | 36.41 | 35.39 | 35.46 | 35.90 | 35.73 |
| F→T | 4.10 | 37.50 | 38.43 | 40.38 | 41.23 | 42.66 | 41.65 | 42.26 | 42.21 | 42.58 |
| H→T | 5.11 | 35.05 | 41.77 | 43.45 | 43.22 | 44.54 | 44.43 | 44.27 | 43.79 | 44.72 |
| E→P | 10.47 | 29.64 | 37.53 | 36.99 | 38.14 | 37.62 | 37.78 | 38.59 | 38.03 | 37.65 |
| B→P | 3.15 | 26.35 | 30.14 | 32.04 | 31.64 | 31.55 | 32.29 | 32.44 | 33.25 | 32.20 |
| F→P | 8.24 | 33.79 | 36.40 | 37.27 | 36.35 | 36.72 | 37.20 | 36.79 | 36.82 | 37.19 |
| H→P | 17.51 | 32.01 | 39.06 | 37.24 | 37.80 | 38.30 | 38.33 | 38.87 | 39.24 | 39.29 |
| AVE | 4.25 | 22.17 | 27.30 | 30.78 | 33.34 | 35.84 | 36.20 | 36.32 | 36.27 | 36.19 |
| Dataset | #D | #R | #W/#R | #T | #IA | #T/#I | POS #n-gram | DP #n-gram | #vocab |
| DP | 2 | 6037 | 16.49 | 9309 | 0 | 1.54 | 14.25/16.05/16.52 | 13.51/14.24/14.43 | 6512 |
| DW | 2 | 6009 | 16.46 | 10390 | 0 | 1.73 | 14.22/16.02/16.48 | 13.48/14.21/14.40 | 6484 |
| DX | 2 | 5989 | 16.43 | 10252 | 0 | 1.71 | 14.20/16.00/16.46 | 13.46/14.19/14.38 | 6467 |
| DMASTE | 8 | 7524 | 59.68 | 28233 | 11945 | 3.75 | 36.67/52.11/57.92 | 35.08/40.20/41.93 | 19226 |
| Domain | Electronics | Fashion | Beauty | Home | Book | Pet | Toy | Grocery |
| Train | 1395 | 851 | 535 | 1050 | 0 | 0 | 0 | 0 |
| Dev | 200 | 121 | 77 | 152 | 159 | 167 | 173 | 173 |
| Test | 399 | 245 | 154 | 301 | 325 | 340 | 354 | 353 |
| Term | Aspect terms | Opinion terms | Triplets |
| IAA | 0.666 | 0.664 | 0.593 |
| Method | Electronics | Beauty | Fashion | Home | Average |
| BMRC | 41.95±0.34 | 38.57±0.97 | 44.87±0.69 | 41.18±0.66 | 41.64 |
| BART-ABSA | 43.38±1.37 | 41.13±1.14 | 43.89±0.82 | 40.56±1.04 | 42.24 |
| GAS | 47.10±0.64 | 44.32±0.52 | 47.80±1.07 | 47.22±1.13 | 46.61 |
| Span-ASTE | 47.86±0.74 | 46.46±0.66 | 50.38±0.68 | 49.14±0.41 | 48.46 |
| Domain | BMRC | BART-ABSA | GAS | Span-ASTE | BMRC+AT | Span-ASTE+AT |
| Electronics→Book | 33.74±0.64 | 35.43±0.61 | 35.57±0.76 | 40.36±1.59 | 32.56±1.57 | 39.58±0.70 |
| Beauty→Book | 30.01±1.29 | 30.24±0.83 | 30.96±0.99 | 38.58±1.58 | 28.56±0.88 | 38.79±0.89 |
| Fashion→Book | 31.71±1.37 | 32.45±1.77 | 36.26±0.64 | 39.88±0.74 | 30.80±1.59 | 39.60±0.69 |
| Home→Book | 31.93±1.79 | 33.48±1.15 | 35.91±0.75 | 39.45±0.73 | 31.31±0.45 | 39.35±0.93 |
| Electronics→Grocery | 39.68±0.70 | 40.39±1.28 | 39.16±1.25 | 45.36±1.12 | 40.15±0.55 | 43.90±0.47 |
| Beauty→Grocery | 34.46±0.85 | 34.22±0.59 | 36.22±0.50 | 40.32±0.81 | 34.91±1.00 | 40.30±1.64 |
| Fashion→Grocery | 37.18±0.50 | 37.27±1.08 | 40.13±0.38 | 43.41±0.83 | 37.56±0.29 | 42.29±0.79 |
| Home→Grocery | 39.03±1.78 | 38.56±1.14 | 42.51±0.37 | 43.74±1.02 | 38.83±0.75 | 41.78±1.79 |
| Electronics→Toy | 42.39±0.88 | 42.49±1.13 | 43.55±0.52 | 47.23±0.38 | 41.38±1.20 | 47.43±0.67 |
| Beauty→Toy | 35.66±0.94 | 33.87±1.05 | 37.95±0.63 | 41.19±0.60 | 35.98±1.14 | 40.86±0.74 |
| Fashion→Toy | 41.13±1.12 | 40.08±1.15 | 42.78±0.36 | 46.83±0.94 | 40.74±1.20 | 46.09±0.87 |
| Home→Toy | 40.26±1.22 | 40.81±1.27 | 44.16±0.73 | 47.60±1.09 | 40.42±0.51 | 46.47±0.75 |
| Electronics→Pet | 37.39±0.60 | 36.88±1.07 | 38.17±0.89 | 41.04±0.48 | 36.70±0.43 | 40.09±0.77 |
| Beauty→Pet | 32.80±1.07 | 32.07±0.76 | 32.55±0.57 | 36.41±0.40 | 32.76±0.54 | 35.38±0.83 |
| Fashion→Pet | 35.97±0.84 | 34.92±0.85 | 36.13±0.74 | 40.57±0.57 | 36.07±0.81 | 38.78±0.81 |
| Home→Pet | 37.64±1.03 | 37.26±1.59 | 39.38±0.70 | 41.42±0.51 | 37.24±0.97 | 40.64±1.05 |
| Average | 36.31 | 36.28 | 38.21 | 42.09 | 36.00 | 41.33 |
| Domain | Book | Grocery | Toy | Pet | Average |
| B + E | 40.94±1.14 | 45.94±0.56 | 48.07±0.75 | 41.70±0.79 | 44.16 |
| E + F | 40.59±0.86 | 46.14±0.98 | 50.10±0.83 | 41.90±0.79 | 44.68 |
| E + H | 41.08±1.50 | 45.34±0.50 | 48.97±0.69 | 42.70±0.41 | 44.52 |
| B + F | 41.13±0.51 | 44.60±0.73 | 46.81±0.91 | 41.03±0.59 | 43.39 |
| B + H | 41.31±1.18 | 44.31±0.84 | 47.12±0.87 | 41.67±0.93 | 43.60 |
| F + H | 42.40±0.31 | 45.97±0.56 | 49.11±0.74 | 43.12±0.70 | 45.15 |
| B + E + F | 41.34±0.73 | 46.28±0.78 | 49.44±1.08 | 41.12±0.58 | 44.80 |
| B + E + H | 41.21±0.48 | 45.39±0.49 | 49.33±0.32 | 43.48±1.12 | 44.85 |
| E + F + H | 41.44±0.71 | 45.39±0.80 | 50.24±0.57 | 43.59±1.06 | 45.17 |
| B + F + H | 41.73±0.38 | 45.40±0.63 | 48.55±0.71 | 43.46±0.19 | 44.79 |
| ALL | 41.83±0.57 | 46.07±0.47 | 50.16±0.49 | 43.62±1.32 | 45.42 |
| Electronics | |
| Good case. It has minimum padding that makes the phone feel secure but too much to be burdensome. The card slots are convenient. | _card slots, convenient, POS); (case, Good, POS); (padding, minimum, POS). |
| They work very well. The connections are nice and tight and I loose no quality over the length of the cord. | (connections, nice, POS); (connections, tight, POS); (length of the cord, loose no quality over, POS); (work, very well, POS). |
| Beauty | |
| This smells so good exactly like the body lotion and spray mist. It is a soft feminine fragrance not too overpowering. I would highly recommend! | (NULL, would highly recommend, POS); (fragrance, not too overpowering, POS); (fragrance, soft feminine, POS); (smells, so good, POS). |
| Revision products are awesome. I switch away from this in the winter months since it's not as moisturizing as an emollient based cleanser. | (Revision, awesome, POS); (NULL, not as moisturizing, NEG). |
| Home | |
| I think i just started a new hobby, this Victorinox is pretty awesome. I might have to get a few more designs. | (Victorinox, pretty awesome, POS). |
| Too big and cumbersome to be useful. If design were down-sized a couple of inches, they would be perfect. Much too big for my salad bowls. | (NULL, Much too big, NEG); (NULL, Too big, NEG); (NULL, cumbersome, NEG). |
| Fashion | |
| the fabric is very, very thin. It should be underwear, not a regular shirt. fits too snug and is see through | (NULL, see through, NEG); (fabric, very thin, NEG); (fits, too snug, NEG). |
| What can I say that you don't already know? Classic Chucks. One of the best, most versatile, and most durable sneakers out there. | (Chucks, Classic, POS); (sneakers, One of the best, POS); (sneakers, most durable, POS); (sneakers, most versatile, POS). |
| Book | |
| It was a good book, haven't read it before bow. Very interesting, with good suspense and humor mixed in! 10/10 | (NULL, Very interesting, POS); (book, good, POS); (humor, good, POS); (suspense, good, POS). |
| Challenging to read but worth it. Persevere to the end of the book. Ask God to open your mind and heart. | (NULL, Challenging to read, NEU); (NULL, worth, POS). |
| Grocery | |
| These are the best Slim Jims every, I have tried them, but the taste, the texture of the Honey BBQ is the Greatest. | (Slim Jims, best, POS); (taste, Greatest, POS); (texture, Greatest, POS). |
| A good almond extract, and I like that it's organic. I took off one star for high price. Should not cost that much. | (almond extract, good, POS); (organic, like, POS); (price, high, NEG). |
| Pet | |
| This is a good prefilter. The clear plastic adapters are a little brittle so be careful when attaching to your filter input. | (clear plastic adapters, a little brittle, NEG); (prefilter, good, POS). |
| Bought these as cat treats. Have to break them up, but my cat goes crazy for them, and this tub is lasting forever. | (NULL, goes crazy for, POS); (tub, lasting forever, POS). |
| Toy | |
| Super simple dynamics, and a great game for friends, family, kids, etc. ! Easy to learn, and variety of play is good. | (NULL, Easy to learn, POS); (dynamics, Super, POS); (dynamics, simple, POS); (game, great, POS); (variety of play, good, POS). |
| My 1 year old loved this for his birthday. Such a fun, easy toy. I would buy this again and again. | (NULL, loved, POS); (toy, easy, POS); (toy, fun, POS). |
| α | 1 | 3 | 5 | 7 | 10 | 15 | 20 | 30 | 50 | 100 |
| E→K | 33.79 | 37.46 | 38.36 | 38.74 | 38.74 | 39.46 | 35.68 | 38.99 | 35.17 | 35.70 |
| B→K | 31.55 | 35.20 | 37.95 | 36.63 | 37.83 | 36.77 | 35.37 | 34.99 | 36.47 | 28.41 |
| F→K | 32.15 | 39.83 | 39.74 | 39.93 | 39.95 | 39.38 | 38.88 | 39.86 | 40.13 | 38.93 |
| H→K | 35.15 | 35.64 | 40.12 | 38.12 | 39.96 | 38.10 | 38.45 | 37.40 | 32.99 | 38.49 |
| E→G | 39.86 | 44.71 | 44.86 | 45.11 | 44.39 | 45.66 | 44.23 | 44.59 | 44.46 | 45.48 |
| B→G | 18.92 | 40.90 | 40.69 | 42.16 | 41.01 | 40.39 | 40.49 | 41.41 | 40.51 | 41.49 |
| F→G | 38.23 | 43.98 | 43.36 | 43.88 | 44.89 | 42.65 | 44.40 | 44.23 | 43.97 | 43.78 |
| H→G | 27.58 | 44.18 | 45.00 | 43.74 | 44.38 | 44.46 | 44.25 | 44.15 | 43.52 | 44.12 |
| E→T | 33.21 | 47.82 | 48.65 | 48.48 | 47.23 | 47.95 | 47.57 | 48.26 | 48.64 | 47.33 |
| B→T | 30.05 | 41.54 | 41.12 | 43.13 | 42.65 | 42.03 | 41.76 | 41.35 | 41.52 | 42.45 |
| F→T | 33.63 | 46.47 | 47.12 | 47.65 | 47.16 | 47.73 | 48.00 | 47.31 | 46.53 | 47.08 |
| H→T | 42.09 | 48.27 | 48.24 | 48.30 | 48.44 | 48.93 | 46.19 | 47.01 | 48.41 | 48.91 |
| E→P | 33.51 | 41.18 | 41.95 | 41.19 | 41.43 | 41.64 | 41.81 | 42.01 | 42.36 | 42.06 |
| B→P | 18.03 | 35.96 | 37.05 | 35.71 | 36.92 | 36.08 | 36.26 | 36.72 | 37.15 | 36.10 |
| F→P | 35.44 | 39.30 | 39.87 | 39.98 | 40.00 | 40.07 | 40.24 | 39.61 | 40.41 | 40.40 |
| H→P | 35.41 | 41.28 | 41.55 | 42.24 | 41.35 | 42.51 | 41.41 | 37.80 | 41.25 | 41.48 |
| AVE | 32.41 | 41.48 | 42.23 | 42.19 | 42.27 | 42.11 | 41.56 | 41.61 | 41.47 | 41.39 |
| α | 1 | 10 | 30 | 50 | 100 | 500 | 700 | 800 | 1000 | 1500 |
| E→K | 0.75 | 2.22 | 6.18 | 16.90 | 6.61 | 28.08 | 31.86 | 31.67 | 30.64 | 30.23 |
| B→K | 0.30 | 1.04 | 9.57 | 17.53 | 26.56 | 27.93 | 28.62 | 27.62 | 29.46 | 29.88 |
| F→K | 0.22 | 7.66 | 8.05 | 14.82 | 27.14 | 30.34 | 32.09 | 30.72 | 31.21 | 31.09 |
| H→K | 0.37 | 9.81 | 11.22 | 18.96 | 22.74 | 26.66 | 27.92 | 31.23 | 28.52 | 28.93 |
| E→G | 2.43 | 12.42 | 24.97 | 19.98 | 35.43 | 38.71 | 39.14 | 39.48 | 39.05 | 38.13 |
| B→G | 2.13 | 20.84 | 33.99 | 32.90 | 32.63 | 33.92 | 33.84 | 32.88 | 32.89 | 32.78 |
| F→G | 2.42 | 37.73 | 27.40 | 34.71 | 38.03 | 38.98 | 37.99 | 37.56 | 38.62 | 38.20 |
| H→G | 4.55 | 30.18 | 32.95 | 37.50 | 37.72 | 37.62 | 37.86 | 38.48 | 37.65 | 37.85 |
| E→T | 4.57 | 31.09 | 35.18 | 38.12 | 42.73 | 43.35 | 42.74 | 42.86 | 43.01 | 42.59 |
| B→T | 1.65 | 7.45 | 23.93 | 33.70 | 35.53 | 36.41 | 35.39 | 35.46 | 35.90 | 35.73 |
| F→T | 4.10 | 37.50 | 38.43 | 40.38 | 41.23 | 42.66 | 41.65 | 42.26 | 42.21 | 42.58 |
| H→T | 5.11 | 35.05 | 41.77 | 43.45 | 43.22 | 44.54 | 44.43 | 44.27 | 43.79 | 44.72 |
| E→P | 10.47 | 29.64 | 37.53 | 36.99 | 38.14 | 37.62 | 37.78 | 38.59 | 38.03 | 37.65 |
| B→P | 3.15 | 26.35 | 30.14 | 32.04 | 31.64 | 31.55 | 32.29 | 32.44 | 33.25 | 32.20 |
| F→P | 8.24 | 33.79 | 36.40 | 37.27 | 36.35 | 36.72 | 37.20 | 36.79 | 36.82 | 37.19 |
| H→P | 17.51 | 32.01 | 39.06 | 37.24 | 37.80 | 38.30 | 38.33 | 38.87 | 39.24 | 39.29 |
| AVE | 4.25 | 22.17 | 27.30 | 30.78 | 33.34 | 35.84 | 36.20 | 36.32 | 36.27 | 36.19 |
| Dataset | #D | #R | #W/#R | #T | #IA | #T/#I | POS #n-gram | DP #n-gram | #vocab |
| DP | 2 | 6037 | 16.49 | 9309 | 0 | 1.54 | 14.25/16.05/16.52 | 13.51/14.24/14.43 | 6512 |
| DW | 2 | 6009 | 16.46 | 10390 | 0 | 1.73 | 14.22/16.02/16.48 | 13.48/14.21/14.40 | 6484 |
| DX | 2 | 5989 | 16.43 | 10252 | 0 | 1.71 | 14.20/16.00/16.46 | 13.46/14.19/14.38 | 6467 |
| DMASTE | 8 | 7524 | 59.68 | 28233 | 11945 | 3.75 | 36.67/52.11/57.92 | 35.08/40.20/41.93 | 19226 |
| Domain | Electronics | Fashion | Beauty | Home | Book | Pet | Toy | Grocery |
| Train | 1395 | 851 | 535 | 1050 | 0 | 0 | 0 | 0 |
| Dev | 200 | 121 | 77 | 152 | 159 | 167 | 173 | 173 |
| Test | 399 | 245 | 154 | 301 | 325 | 340 | 354 | 353 |
| Term | Aspect terms | Opinion terms | Triplets |
| IAA | 0.666 | 0.664 | 0.593 |
| Method | Electronics | Beauty | Fashion | Home | Average |
| BMRC | 41.95±0.34 | 38.57±0.97 | 44.87±0.69 | 41.18±0.66 | 41.64 |
| BART-ABSA | 43.38±1.37 | 41.13±1.14 | 43.89±0.82 | 40.56±1.04 | 42.24 |
| GAS | 47.10±0.64 | 44.32±0.52 | 47.80±1.07 | 47.22±1.13 | 46.61 |
| Span-ASTE | 47.86±0.74 | 46.46±0.66 | 50.38±0.68 | 49.14±0.41 | 48.46 |
| Domain | BMRC | BART-ABSA | GAS | Span-ASTE | BMRC+AT | Span-ASTE+AT |
| Electronics→Book | 33.74±0.64 | 35.43±0.61 | 35.57±0.76 | 40.36±1.59 | 32.56±1.57 | 39.58±0.70 |
| Beauty→Book | 30.01±1.29 | 30.24±0.83 | 30.96±0.99 | 38.58±1.58 | 28.56±0.88 | 38.79±0.89 |
| Fashion→Book | 31.71±1.37 | 32.45±1.77 | 36.26±0.64 | 39.88±0.74 | 30.80±1.59 | 39.60±0.69 |
| Home→Book | 31.93±1.79 | 33.48±1.15 | 35.91±0.75 | 39.45±0.73 | 31.31±0.45 | 39.35±0.93 |
| Electronics→Grocery | 39.68±0.70 | 40.39±1.28 | 39.16±1.25 | 45.36±1.12 | 40.15±0.55 | 43.90±0.47 |
| Beauty→Grocery | 34.46±0.85 | 34.22±0.59 | 36.22±0.50 | 40.32±0.81 | 34.91±1.00 | 40.30±1.64 |
| Fashion→Grocery | 37.18±0.50 | 37.27±1.08 | 40.13±0.38 | 43.41±0.83 | 37.56±0.29 | 42.29±0.79 |
| Home→Grocery | 39.03±1.78 | 38.56±1.14 | 42.51±0.37 | 43.74±1.02 | 38.83±0.75 | 41.78±1.79 |
| Electronics→Toy | 42.39±0.88 | 42.49±1.13 | 43.55±0.52 | 47.23±0.38 | 41.38±1.20 | 47.43±0.67 |
| Beauty→Toy | 35.66±0.94 | 33.87±1.05 | 37.95±0.63 | 41.19±0.60 | 35.98±1.14 | 40.86±0.74 |
| Fashion→Toy | 41.13±1.12 | 40.08±1.15 | 42.78±0.36 | 46.83±0.94 | 40.74±1.20 | 46.09±0.87 |
| Home→Toy | 40.26±1.22 | 40.81±1.27 | 44.16±0.73 | 47.60±1.09 | 40.42±0.51 | 46.47±0.75 |
| Electronics→Pet | 37.39±0.60 | 36.88±1.07 | 38.17±0.89 | 41.04±0.48 | 36.70±0.43 | 40.09±0.77 |
| Beauty→Pet | 32.80±1.07 | 32.07±0.76 | 32.55±0.57 | 36.41±0.40 | 32.76±0.54 | 35.38±0.83 |
| Fashion→Pet | 35.97±0.84 | 34.92±0.85 | 36.13±0.74 | 40.57±0.57 | 36.07±0.81 | 38.78±0.81 |
| Home→Pet | 37.64±1.03 | 37.26±1.59 | 39.38±0.70 | 41.42±0.51 | 37.24±0.97 | 40.64±1.05 |
| Average | 36.31 | 36.28 | 38.21 | 42.09 | 36.00 | 41.33 |
| Domain | Book | Grocery | Toy | Pet | Average |
| B + E | 40.94±1.14 | 45.94±0.56 | 48.07±0.75 | 41.70±0.79 | 44.16 |
| E + F | 40.59±0.86 | 46.14±0.98 | 50.10±0.83 | 41.90±0.79 | 44.68 |
| E + H | 41.08±1.50 | 45.34±0.50 | 48.97±0.69 | 42.70±0.41 | 44.52 |
| B + F | 41.13±0.51 | 44.60±0.73 | 46.81±0.91 | 41.03±0.59 | 43.39 |
| B + H | 41.31±1.18 | 44.31±0.84 | 47.12±0.87 | 41.67±0.93 | 43.60 |
| F + H | 42.40±0.31 | 45.97±0.56 | 49.11±0.74 | 43.12±0.70 | 45.15 |
| B + E + F | 41.34±0.73 | 46.28±0.78 | 49.44±1.08 | 41.12±0.58 | 44.80 |
| B + E + H | 41.21±0.48 | 45.39±0.49 | 49.33±0.32 | 43.48±1.12 | 44.85 |
| E + F + H | 41.44±0.71 | 45.39±0.80 | 50.24±0.57 | 43.59±1.06 | 45.17 |
| B + F + H | 41.73±0.38 | 45.40±0.63 | 48.55±0.71 | 43.46±0.19 | 44.79 |
| ALL | 41.83±0.57 | 46.07±0.47 | 50.16±0.49 | 43.62±1.32 | 45.42 |
| Electronics | |
| Good case. It has minimum padding that makes the phone feel secure but too much to be burdensome. The card slots are convenient. | _card slots, convenient, POS); (case, Good, POS); (padding, minimum, POS). |
| They work very well. The connections are nice and tight and I loose no quality over the length of the cord. | (connections, nice, POS); (connections, tight, POS); (length of the cord, loose no quality over, POS); (work, very well, POS). |
| Beauty | |
| This smells so good exactly like the body lotion and spray mist. It is a soft feminine fragrance not too overpowering. I would highly recommend! | (NULL, would highly recommend, POS); (fragrance, not too overpowering, POS); (fragrance, soft feminine, POS); (smells, so good, POS). |
| Revision products are awesome. I switch away from this in the winter months since it's not as moisturizing as an emollient based cleanser. | (Revision, awesome, POS); (NULL, not as moisturizing, NEG). |
| Home | |
| I think i just started a new hobby, this Victorinox is pretty awesome. I might have to get a few more designs. | (Victorinox, pretty awesome, POS). |
| Too big and cumbersome to be useful. If design were down-sized a couple of inches, they would be perfect. Much too big for my salad bowls. | (NULL, Much too big, NEG); (NULL, Too big, NEG); (NULL, cumbersome, NEG). |
| Fashion | |
| the fabric is very, very thin. It should be underwear, not a regular shirt. fits too snug and is see through | (NULL, see through, NEG); (fabric, very thin, NEG); (fits, too snug, NEG). |
| What can I say that you don't already know? Classic Chucks. One of the best, most versatile, and most durable sneakers out there. | (Chucks, Classic, POS); (sneakers, One of the best, POS); (sneakers, most durable, POS); (sneakers, most versatile, POS). |
| Book | |
| It was a good book, haven't read it before bow. Very interesting, with good suspense and humor mixed in! 10/10 | (NULL, Very interesting, POS); (book, good, POS); (humor, good, POS); (suspense, good, POS). |
| Challenging to read but worth it. Persevere to the end of the book. Ask God to open your mind and heart. | (NULL, Challenging to read, NEU); (NULL, worth, POS). |
| Grocery | |
| These are the best Slim Jims every, I have tried them, but the taste, the texture of the Honey BBQ is the Greatest. | (Slim Jims, best, POS); (taste, Greatest, POS); (texture, Greatest, POS). |
| A good almond extract, and I like that it's organic. I took off one star for high price. Should not cost that much. | (almond extract, good, POS); (organic, like, POS); (price, high, NEG). |
| Pet | |
| This is a good prefilter. The clear plastic adapters are a little brittle so be careful when attaching to your filter input. | (clear plastic adapters, a little brittle, NEG); (prefilter, good, POS). |
| Bought these as cat treats. Have to break them up, but my cat goes crazy for them, and this tub is lasting forever. | (NULL, goes crazy for, POS); (tub, lasting forever, POS). |
| Toy | |
| Super simple dynamics, and a great game for friends, family, kids, etc. ! Easy to learn, and variety of play is good. | (NULL, Easy to learn, POS); (dynamics, Super, POS); (dynamics, simple, POS); (game, great, POS); (variety of play, good, POS). |
| My 1 year old loved this for his birthday. Such a fun, easy toy. I would buy this again and again. | (NULL, loved, POS); (toy, easy, POS); (toy, fun, POS). |
| α | 1 | 3 | 5 | 7 | 10 | 15 | 20 | 30 | 50 | 100 |
| E→K | 33.79 | 37.46 | 38.36 | 38.74 | 38.74 | 39.46 | 35.68 | 38.99 | 35.17 | 35.70 |
| B→K | 31.55 | 35.20 | 37.95 | 36.63 | 37.83 | 36.77 | 35.37 | 34.99 | 36.47 | 28.41 |
| F→K | 32.15 | 39.83 | 39.74 | 39.93 | 39.95 | 39.38 | 38.88 | 39.86 | 40.13 | 38.93 |
| H→K | 35.15 | 35.64 | 40.12 | 38.12 | 39.96 | 38.10 | 38.45 | 37.40 | 32.99 | 38.49 |
| E→G | 39.86 | 44.71 | 44.86 | 45.11 | 44.39 | 45.66 | 44.23 | 44.59 | 44.46 | 45.48 |
| B→G | 18.92 | 40.90 | 40.69 | 42.16 | 41.01 | 40.39 | 40.49 | 41.41 | 40.51 | 41.49 |
| F→G | 38.23 | 43.98 | 43.36 | 43.88 | 44.89 | 42.65 | 44.40 | 44.23 | 43.97 | 43.78 |
| H→G | 27.58 | 44.18 | 45.00 | 43.74 | 44.38 | 44.46 | 44.25 | 44.15 | 43.52 | 44.12 |
| E→T | 33.21 | 47.82 | 48.65 | 48.48 | 47.23 | 47.95 | 47.57 | 48.26 | 48.64 | 47.33 |
| B→T | 30.05 | 41.54 | 41.12 | 43.13 | 42.65 | 42.03 | 41.76 | 41.35 | 41.52 | 42.45 |
| F→T | 33.63 | 46.47 | 47.12 | 47.65 | 47.16 | 47.73 | 48.00 | 47.31 | 46.53 | 47.08 |
| H→T | 42.09 | 48.27 | 48.24 | 48.30 | 48.44 | 48.93 | 46.19 | 47.01 | 48.41 | 48.91 |
| E→P | 33.51 | 41.18 | 41.95 | 41.19 | 41.43 | 41.64 | 41.81 | 42.01 | 42.36 | 42.06 |
| B→P | 18.03 | 35.96 | 37.05 | 35.71 | 36.92 | 36.08 | 36.26 | 36.72 | 37.15 | 36.10 |
| F→P | 35.44 | 39.30 | 39.87 | 39.98 | 40.00 | 40.07 | 40.24 | 39.61 | 40.41 | 40.40 |
| H→P | 35.41 | 41.28 | 41.55 | 42.24 | 41.35 | 42.51 | 41.41 | 37.80 | 41.25 | 41.48 |
| AVE | 32.41 | 41.48 | 42.23 | 42.19 | 42.27 | 42.11 | 41.56 | 41.61 | 41.47 | 41.39 |
| α | 1 | 10 | 30 | 50 | 100 | 500 | 700 | 800 | 1000 | 1500 |
| E→K | 0.75 | 2.22 | 6.18 | 16.90 | 6.61 | 28.08 | 31.86 | 31.67 | 30.64 | 30.23 |
| B→K | 0.30 | 1.04 | 9.57 | 17.53 | 26.56 | 27.93 | 28.62 | 27.62 | 29.46 | 29.88 |
| F→K | 0.22 | 7.66 | 8.05 | 14.82 | 27.14 | 30.34 | 32.09 | 30.72 | 31.21 | 31.09 |
| H→K | 0.37 | 9.81 | 11.22 | 18.96 | 22.74 | 26.66 | 27.92 | 31.23 | 28.52 | 28.93 |
| E→G | 2.43 | 12.42 | 24.97 | 19.98 | 35.43 | 38.71 | 39.14 | 39.48 | 39.05 | 38.13 |
| B→G | 2.13 | 20.84 | 33.99 | 32.90 | 32.63 | 33.92 | 33.84 | 32.88 | 32.89 | 32.78 |
| F→G | 2.42 | 37.73 | 27.40 | 34.71 | 38.03 | 38.98 | 37.99 | 37.56 | 38.62 | 38.20 |
| H→G | 4.55 | 30.18 | 32.95 | 37.50 | 37.72 | 37.62 | 37.86 | 38.48 | 37.65 | 37.85 |
| E→T | 4.57 | 31.09 | 35.18 | 38.12 | 42.73 | 43.35 | 42.74 | 42.86 | 43.01 | 42.59 |
| B→T | 1.65 | 7.45 | 23.93 | 33.70 | 35.53 | 36.41 | 35.39 | 35.46 | 35.90 | 35.73 |
| F→T | 4.10 | 37.50 | 38.43 | 40.38 | 41.23 | 42.66 | 41.65 | 42.26 | 42.21 | 42.58 |
| H→T | 5.11 | 35.05 | 41.77 | 43.45 | 43.22 | 44.54 | 44.43 | 44.27 | 43.79 | 44.72 |
| E→P | 10.47 | 29.64 | 37.53 | 36.99 | 38.14 | 37.62 | 37.78 | 38.59 | 38.03 | 37.65 |
| B→P | 3.15 | 26.35 | 30.14 | 32.04 | 31.64 | 31.55 | 32.29 | 32.44 | 33.25 | 32.20 |
| F→P | 8.24 | 33.79 | 36.40 | 37.27 | 36.35 | 36.72 | 37.20 | 36.79 | 36.82 | 37.19 |
| H→P | 17.51 | 32.01 | 39.06 | 37.24 | 37.80 | 38.30 | 38.33 | 38.87 | 39.24 | 39.29 |
| AVE | 4.25 | 22.17 | 27.30 | 30.78 | 33.34 | 35.84 | 36.20 | 36.32 | 36.27 | 36.19 |
| Dataset | #D | #R | #W/#R | #T | #IA | #T/#I | POS #n-gram | DP #n-gram | #vocab |
| DP | 2 | 6037 | 16.49 | 9309 | 0 | 1.54 | 14.25/16.05/16.52 | 13.51/14.24/14.43 | 6512 |
| DW | 2 | 6009 | 16.46 | 10390 | 0 | 1.73 | 14.22/16.02/16.48 | 13.48/14.21/14.40 | 6484 |
| DX | 2 | 5989 | 16.43 | 10252 | 0 | 1.71 | 14.20/16.00/16.46 | 13.46/14.19/14.38 | 6467 |
| DMASTE | 8 | 7524 | 59.68 | 28233 | 11945 | 3.75 | 36.67/52.11/57.92 | 35.08/40.20/41.93 | 19226 |
| Domain | Electronics | Fashion | Beauty | Home | Book | Pet | Toy | Grocery |
| Train | 1395 | 851 | 535 | 1050 | 0 | 0 | 0 | 0 |
| Dev | 200 | 121 | 77 | 152 | 159 | 167 | 173 | 173 |
| Test | 399 | 245 | 154 | 301 | 325 | 340 | 354 | 353 |
| Term | Aspect terms | Opinion terms | Triplets |
| IAA | 0.666 | 0.664 | 0.593 |
| Method | Electronics | Beauty | Fashion | Home | Average |
| BMRC | 41.95±0.34 | 38.57±0.97 | 44.87±0.69 | 41.18±0.66 | 41.64 |
| BART-ABSA | 43.38±1.37 | 41.13±1.14 | 43.89±0.82 | 40.56±1.04 | 42.24 |
| GAS | 47.10±0.64 | 44.32±0.52 | 47.80±1.07 | 47.22±1.13 | 46.61 |
| Span-ASTE | 47.86±0.74 | 46.46±0.66 | 50.38±0.68 | 49.14±0.41 | 48.46 |
| Domain | BMRC | BART-ABSA | GAS | Span-ASTE | BMRC+AT | Span-ASTE+AT |
| Electronics→Book | 33.74±0.64 | 35.43±0.61 | 35.57±0.76 | 40.36±1.59 | 32.56±1.57 | 39.58±0.70 |
| Beauty→Book | 30.01±1.29 | 30.24±0.83 | 30.96±0.99 | 38.58±1.58 | 28.56±0.88 | 38.79±0.89 |
| Fashion→Book | 31.71±1.37 | 32.45±1.77 | 36.26±0.64 | 39.88±0.74 | 30.80±1.59 | 39.60±0.69 |
| Home→Book | 31.93±1.79 | 33.48±1.15 | 35.91±0.75 | 39.45±0.73 | 31.31±0.45 | 39.35±0.93 |
| Electronics→Grocery | 39.68±0.70 | 40.39±1.28 | 39.16±1.25 | 45.36±1.12 | 40.15±0.55 | 43.90±0.47 |
| Beauty→Grocery | 34.46±0.85 | 34.22±0.59 | 36.22±0.50 | 40.32±0.81 | 34.91±1.00 | 40.30±1.64 |
| Fashion→Grocery | 37.18±0.50 | 37.27±1.08 | 40.13±0.38 | 43.41±0.83 | 37.56±0.29 | 42.29±0.79 |
| Home→Grocery | 39.03±1.78 | 38.56±1.14 | 42.51±0.37 | 43.74±1.02 | 38.83±0.75 | 41.78±1.79 |
| Electronics→Toy | 42.39±0.88 | 42.49±1.13 | 43.55±0.52 | 47.23±0.38 | 41.38±1.20 | 47.43±0.67 |
| Beauty→Toy | 35.66±0.94 | 33.87±1.05 | 37.95±0.63 | 41.19±0.60 | 35.98±1.14 | 40.86±0.74 |
| Fashion→Toy | 41.13±1.12 | 40.08±1.15 | 42.78±0.36 | 46.83±0.94 | 40.74±1.20 | 46.09±0.87 |
| Home→Toy | 40.26±1.22 | 40.81±1.27 | 44.16±0.73 | 47.60±1.09 | 40.42±0.51 | 46.47±0.75 |
| Electronics→Pet | 37.39±0.60 | 36.88±1.07 | 38.17±0.89 | 41.04±0.48 | 36.70±0.43 | 40.09±0.77 |
| Beauty→Pet | 32.80±1.07 | 32.07±0.76 | 32.55±0.57 | 36.41±0.40 | 32.76±0.54 | 35.38±0.83 |
| Fashion→Pet | 35.97±0.84 | 34.92±0.85 | 36.13±0.74 | 40.57±0.57 | 36.07±0.81 | 38.78±0.81 |
| Home→Pet | 37.64±1.03 | 37.26±1.59 | 39.38±0.70 | 41.42±0.51 | 37.24±0.97 | 40.64±1.05 |
| Average | 36.31 | 36.28 | 38.21 | 42.09 | 36.00 | 41.33 |
| Domain | Book | Grocery | Toy | Pet | Average |
| B + E | 40.94±1.14 | 45.94±0.56 | 48.07±0.75 | 41.70±0.79 | 44.16 |
| E + F | 40.59±0.86 | 46.14±0.98 | 50.10±0.83 | 41.90±0.79 | 44.68 |
| E + H | 41.08±1.50 | 45.34±0.50 | 48.97±0.69 | 42.70±0.41 | 44.52 |
| B + F | 41.13±0.51 | 44.60±0.73 | 46.81±0.91 | 41.03±0.59 | 43.39 |
| B + H | 41.31±1.18 | 44.31±0.84 | 47.12±0.87 | 41.67±0.93 | 43.60 |
| F + H | 42.40±0.31 | 45.97±0.56 | 49.11±0.74 | 43.12±0.70 | 45.15 |
| B + E + F | 41.34±0.73 | 46.28±0.78 | 49.44±1.08 | 41.12±0.58 | 44.80 |
| B + E + H | 41.21±0.48 | 45.39±0.49 | 49.33±0.32 | 43.48±1.12 | 44.85 |
| E + F + H | 41.44±0.71 | 45.39±0.80 | 50.24±0.57 | 43.59±1.06 | 45.17 |
| B + F + H | 41.73±0.38 | 45.40±0.63 | 48.55±0.71 | 43.46±0.19 | 44.79 |
| ALL | 41.83±0.57 | 46.07±0.47 | 50.16±0.49 | 43.62±1.32 | 45.42 |
| Electronics | |
| Good case. It has minimum padding that makes the phone feel secure but too much to be burdensome. The card slots are convenient. | _card slots, convenient, POS); (case, Good, POS); (padding, minimum, POS). |
| They work very well. The connections are nice and tight and I loose no quality over the length of the cord. | (connections, nice, POS); (connections, tight, POS); (length of the cord, loose no quality over, POS); (work, very well, POS). |
| Beauty | |
| This smells so good exactly like the body lotion and spray mist. It is a soft feminine fragrance not too overpowering. I would highly recommend! | (NULL, would highly recommend, POS); (fragrance, not too overpowering, POS); (fragrance, soft feminine, POS); (smells, so good, POS). |
| Revision products are awesome. I switch away from this in the winter months since it's not as moisturizing as an emollient based cleanser. | (Revision, awesome, POS); (NULL, not as moisturizing, NEG). |
| Home | |
| I think i just started a new hobby, this Victorinox is pretty awesome. I might have to get a few more designs. | (Victorinox, pretty awesome, POS). |
| Too big and cumbersome to be useful. If design were down-sized a couple of inches, they would be perfect. Much too big for my salad bowls. | (NULL, Much too big, NEG); (NULL, Too big, NEG); (NULL, cumbersome, NEG). |
| Fashion | |
| the fabric is very, very thin. It should be underwear, not a regular shirt. fits too snug and is see through | (NULL, see through, NEG); (fabric, very thin, NEG); (fits, too snug, NEG). |
| What can I say that you don't already know? Classic Chucks. One of the best, most versatile, and most durable sneakers out there. | (Chucks, Classic, POS); (sneakers, One of the best, POS); (sneakers, most durable, POS); (sneakers, most versatile, POS). |
| Book | |
| It was a good book, haven't read it before bow. Very interesting, with good suspense and humor mixed in! 10/10 | (NULL, Very interesting, POS); (book, good, POS); (humor, good, POS); (suspense, good, POS). |
| Challenging to read but worth it. Persevere to the end of the book. Ask God to open your mind and heart. | (NULL, Challenging to read, NEU); (NULL, worth, POS). |
| Grocery | |
| These are the best Slim Jims every, I have tried them, but the taste, the texture of the Honey BBQ is the Greatest. | (Slim Jims, best, POS); (taste, Greatest, POS); (texture, Greatest, POS). |
| A good almond extract, and I like that it's organic. I took off one star for high price. Should not cost that much. | (almond extract, good, POS); (organic, like, POS); (price, high, NEG). |
| Pet | |
| This is a good prefilter. The clear plastic adapters are a little brittle so be careful when attaching to your filter input. | (clear plastic adapters, a little brittle, NEG); (prefilter, good, POS). |
| Bought these as cat treats. Have to break them up, but my cat goes crazy for them, and this tub is lasting forever. | (NULL, goes crazy for, POS); (tub, lasting forever, POS). |
| Toy | |
| Super simple dynamics, and a great game for friends, family, kids, etc. ! Easy to learn, and variety of play is good. | (NULL, Easy to learn, POS); (dynamics, Super, POS); (dynamics, simple, POS); (game, great, POS); (variety of play, good, POS). |
| My 1 year old loved this for his birthday. Such a fun, easy toy. I would buy this again and again. | (NULL, loved, POS); (toy, easy, POS); (toy, fun, POS). |
| α | 1 | 3 | 5 | 7 | 10 | 15 | 20 | 30 | 50 | 100 |
| E→K | 33.79 | 37.46 | 38.36 | 38.74 | 38.74 | 39.46 | 35.68 | 38.99 | 35.17 | 35.70 |
| B→K | 31.55 | 35.20 | 37.95 | 36.63 | 37.83 | 36.77 | 35.37 | 34.99 | 36.47 | 28.41 |
| F→K | 32.15 | 39.83 | 39.74 | 39.93 | 39.95 | 39.38 | 38.88 | 39.86 | 40.13 | 38.93 |
| H→K | 35.15 | 35.64 | 40.12 | 38.12 | 39.96 | 38.10 | 38.45 | 37.40 | 32.99 | 38.49 |
| E→G | 39.86 | 44.71 | 44.86 | 45.11 | 44.39 | 45.66 | 44.23 | 44.59 | 44.46 | 45.48 |
| B→G | 18.92 | 40.90 | 40.69 | 42.16 | 41.01 | 40.39 | 40.49 | 41.41 | 40.51 | 41.49 |
| F→G | 38.23 | 43.98 | 43.36 | 43.88 | 44.89 | 42.65 | 44.40 | 44.23 | 43.97 | 43.78 |
| H→G | 27.58 | 44.18 | 45.00 | 43.74 | 44.38 | 44.46 | 44.25 | 44.15 | 43.52 | 44.12 |
| E→T | 33.21 | 47.82 | 48.65 | 48.48 | 47.23 | 47.95 | 47.57 | 48.26 | 48.64 | 47.33 |
| B→T | 30.05 | 41.54 | 41.12 | 43.13 | 42.65 | 42.03 | 41.76 | 41.35 | 41.52 | 42.45 |
| F→T | 33.63 | 46.47 | 47.12 | 47.65 | 47.16 | 47.73 | 48.00 | 47.31 | 46.53 | 47.08 |
| H→T | 42.09 | 48.27 | 48.24 | 48.30 | 48.44 | 48.93 | 46.19 | 47.01 | 48.41 | 48.91 |
| E→P | 33.51 | 41.18 | 41.95 | 41.19 | 41.43 | 41.64 | 41.81 | 42.01 | 42.36 | 42.06 |
| B→P | 18.03 | 35.96 | 37.05 | 35.71 | 36.92 | 36.08 | 36.26 | 36.72 | 37.15 | 36.10 |
| F→P | 35.44 | 39.30 | 39.87 | 39.98 | 40.00 | 40.07 | 40.24 | 39.61 | 40.41 | 40.40 |
| H→P | 35.41 | 41.28 | 41.55 | 42.24 | 41.35 | 42.51 | 41.41 | 37.80 | 41.25 | 41.48 |
| AVE | 32.41 | 41.48 | 42.23 | 42.19 | 42.27 | 42.11 | 41.56 | 41.61 | 41.47 | 41.39 |
| α | 1 | 10 | 30 | 50 | 100 | 500 | 700 | 800 | 1000 | 1500 |
| E→K | 0.75 | 2.22 | 6.18 | 16.90 | 6.61 | 28.08 | 31.86 | 31.67 | 30.64 | 30.23 |
| B→K | 0.30 | 1.04 | 9.57 | 17.53 | 26.56 | 27.93 | 28.62 | 27.62 | 29.46 | 29.88 |
| F→K | 0.22 | 7.66 | 8.05 | 14.82 | 27.14 | 30.34 | 32.09 | 30.72 | 31.21 | 31.09 |
| H→K | 0.37 | 9.81 | 11.22 | 18.96 | 22.74 | 26.66 | 27.92 | 31.23 | 28.52 | 28.93 |
| E→G | 2.43 | 12.42 | 24.97 | 19.98 | 35.43 | 38.71 | 39.14 | 39.48 | 39.05 | 38.13 |
| B→G | 2.13 | 20.84 | 33.99 | 32.90 | 32.63 | 33.92 | 33.84 | 32.88 | 32.89 | 32.78 |
| F→G | 2.42 | 37.73 | 27.40 | 34.71 | 38.03 | 38.98 | 37.99 | 37.56 | 38.62 | 38.20 |
| H→G | 4.55 | 30.18 | 32.95 | 37.50 | 37.72 | 37.62 | 37.86 | 38.48 | 37.65 | 37.85 |
| E→T | 4.57 | 31.09 | 35.18 | 38.12 | 42.73 | 43.35 | 42.74 | 42.86 | 43.01 | 42.59 |
| B→T | 1.65 | 7.45 | 23.93 | 33.70 | 35.53 | 36.41 | 35.39 | 35.46 | 35.90 | 35.73 |
| F→T | 4.10 | 37.50 | 38.43 | 40.38 | 41.23 | 42.66 | 41.65 | 42.26 | 42.21 | 42.58 |
| H→T | 5.11 | 35.05 | 41.77 | 43.45 | 43.22 | 44.54 | 44.43 | 44.27 | 43.79 | 44.72 |
| E→P | 10.47 | 29.64 | 37.53 | 36.99 | 38.14 | 37.62 | 37.78 | 38.59 | 38.03 | 37.65 |
| B→P | 3.15 | 26.35 | 30.14 | 32.04 | 31.64 | 31.55 | 32.29 | 32.44 | 33.25 | 32.20 |
| F→P | 8.24 | 33.79 | 36.40 | 37.27 | 36.35 | 36.72 | 37.20 | 36.79 | 36.82 | 37.19 |
| H→P | 17.51 | 32.01 | 39.06 | 37.24 | 37.80 | 38.30 | 38.33 | 38.87 | 39.24 | 39.29 |
| AVE | 4.25 | 22.17 | 27.30 | 30.78 | 33.34 | 35.84 | 36.20 | 36.32 | 36.27 | 36.19 |
| Utterance (gold label) | \\( \\hat{p}_{\\text{min-max}} \\), \\( \\sigma_{m}(\\hat{p}) \\) | Label predictions over 50 runs |
| funny joke (general) | [0.98-0.99], 0.003 (low) | general:50 |
| start house cleanup (IOT) | [0.002-0.97], 0.17 (high) | lists:26, IOT:6, general:6, play:5, news:3, social:1, calendar:1 |
| search for gluten free menus (cooking) | [0.002-0.693], 0.06 (low) | lists:28, take-away:18, social:1, music:1, cooking:1, play:1 |
| Attribute | MASSIVE | CLINC150 |
| Source | AmazonAlexa AI | DiaoGLUE |
| Domains | 18 | - |
| Intentss | 60 | 150 |
| TrainHoldout(Unseen) | 11,5142974 | 15,0003,000 |
| Balanced? | No. | Yes. 100per intent |
| Classification task | Domain | Intent |
| Massive | Clink150 | |||||
| Methods | Accuracy(%) | ΔLEm(%) ↑ | % of Eb | Accuracy(%) | ΔLEm(%) ↑ | % of Eb |
| Control baseline | 90.6 ± 0.6 | - | - | 95.1 ± 0.8 | - | - |
| Ensemble baseline (Eb) | 91.3 ± 0.5 | 34.5 | - | 95.4 ± 0.6 | 31.1 | - |
| L2 Regularization | 90.3 ± 0.5 | -2.3 | -7 | 94.9 ± 0.7 | -0.6 | -2 |
| SWA | 91.0 ± 0.5 | 17.6 | 51 | 95.2 ± 0.7 | 7.3 | 23 |
| Label Smoothing | 90.8 ± 0.5 | 5.7 | 17 | 95.2 ± 0.8 | 6.1 | 20 |
| TGTSS (Ours) | 91.3 ± 0.6 | 31.4 | 91 | 95.3 ± 0.8 | 26.7 | 86 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Label Smoothing | 37.9 | 40.5 |
| Ensemble baseline | 55.5 | 61.7 |
| TGTSS (Ours) | 53.4 | 55.9 |
| ΔLEm (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Control baseline | - | - |
| Ensemble baseline | 27.4 | 24.1 |
| L2 Reg. | 3.8 | 4.2 |
| SWA | 11.3 | 4.3 |
| Label Smoothing | 5.4 | 8.1 |
| TGTSS (Ours) | 26 | 22.4 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Clinc150 |
| Label Smoothing | 14.9 | 20.7 |
| Ensemble baseline | 36.4 | 40.7 |
| TGTSS (Ours) | 31.5 | 33.6 |
| Utterance (gold label) | \\( \\hat{p}_{\\text{min-max}} \\), \\( \\sigma_{m}(\\hat{p}) \\) | Label predictions over 50 runs |
| funny joke (general) | [0.98-0.99], 0.003 (low) | general:50 |
| start house cleanup (IOT) | [0.002-0.97], 0.17 (high) | lists:26, IOT:6, general:6, play:5, news:3, social:1, calendar:1 |
| search for gluten free menus (cooking) | [0.002-0.693], 0.06 (low) | lists:28, take-away:18, social:1, music:1, cooking:1, play:1 |
| Attribute | MASSIVE | CLINC150 |
| Source | AmazonAlexa AI | DiaoGLUE |
| Domains | 18 | - |
| Intentss | 60 | 150 |
| TrainHoldout(Unseen) | 11,5142974 | 15,0003,000 |
| Balanced? | No. | Yes. 100per intent |
| Classification task | Domain | Intent |
| Massive | Clink150 | |||||
| Methods | Accuracy(%) | ΔLEm(%) ↑ | % of Eb | Accuracy(%) | ΔLEm(%) ↑ | % of Eb |
| Control baseline | 90.6 ± 0.6 | - | - | 95.1 ± 0.8 | - | - |
| Ensemble baseline (Eb) | 91.3 ± 0.5 | 34.5 | - | 95.4 ± 0.6 | 31.1 | - |
| L2 Regularization | 90.3 ± 0.5 | -2.3 | -7 | 94.9 ± 0.7 | -0.6 | -2 |
| SWA | 91.0 ± 0.5 | 17.6 | 51 | 95.2 ± 0.7 | 7.3 | 23 |
| Label Smoothing | 90.8 ± 0.5 | 5.7 | 17 | 95.2 ± 0.8 | 6.1 | 20 |
| TGTSS (Ours) | 91.3 ± 0.6 | 31.4 | 91 | 95.3 ± 0.8 | 26.7 | 86 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Label Smoothing | 37.9 | 40.5 |
| Ensemble baseline | 55.5 | 61.7 |
| TGTSS (Ours) | 53.4 | 55.9 |
| ΔLEm (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Control baseline | - | - |
| Ensemble baseline | 27.4 | 24.1 |
| L2 Reg. | 3.8 | 4.2 |
| SWA | 11.3 | 4.3 |
| Label Smoothing | 5.4 | 8.1 |
| TGTSS (Ours) | 26 | 22.4 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Clinc150 |
| Label Smoothing | 14.9 | 20.7 |
| Ensemble baseline | 36.4 | 40.7 |
| TGTSS (Ours) | 31.5 | 33.6 |
| Utterance (gold label) | \( \hat{p}_{\text{min-max}} \), \( \sigma_{m}(\hat{p}) \) | Label predictions over 50 runs |
| funny joke (general) | [0.98-0.99], 0.003 (low) | general:50 |
| start house cleanup (IOT) | [0.002-0.97], 0.17 (high) | lists:26, IOT:6, general:6, play:5, news:3, social:1, calendar:1 |
| search for gluten free menus (cooking) | [0.002-0.693], 0.06 (low) | lists:28, take-away:18, social:1, music:1, cooking:1, play:1 |
| Attribute | MASSIVE | CLINC150 |
| Source | AmazonAlexa AI | DiaoGLUE |
| Domains | 18 | - |
| Intentss | 60 | 150 |
| TrainHoldout(Unseen) | 11,5142974 | 15,0003,000 |
| Balanced? | No. | Yes. 100per intent |
| Classification task | Domain | Intent |
| Massive | Clink150 | |||||
| Methods | Accuracy(%) | ΔLEm(%) ↑ | % of Eb | Accuracy(%) | ΔLEm(%) ↑ | % of Eb |
| Control baseline | 90.6 ± 0.6 | - | - | 95.1 ± 0.8 | - | - |
| Ensemble baseline (Eb) | 91.3 ± 0.5 | 34.5 | - | 95.4 ± 0.6 | 31.1 | - |
| L2 Regularization | 90.3 ± 0.5 | -2.3 | -7 | 94.9 ± 0.7 | -0.6 | -2 |
| SWA | 91.0 ± 0.5 | 17.6 | 51 | 95.2 ± 0.7 | 7.3 | 23 |
| Label Smoothing | 90.8 ± 0.5 | 5.7 | 17 | 95.2 ± 0.8 | 6.1 | 20 |
| TGTSS (Ours) | 91.3 ± 0.6 | 31.4 | 91 | 95.3 ± 0.8 | 26.7 | 86 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Label Smoothing | 37.9 | 40.5 |
| Ensemble baseline | 55.5 | 61.7 |
| TGTSS (Ours) | 53.4 | 55.9 |
| ΔLEm (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Control baseline | - | - |
| Ensemble baseline | 27.4 | 24.1 |
| L2 Reg. | 3.8 | 4.2 |
| SWA | 11.3 | 4.3 |
| Label Smoothing | 5.4 | 8.1 |
| TGTSS (Ours) | 26 | 22.4 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Clinc150 |
| Label Smoothing | 14.9 | 20.7 |
| Ensemble baseline | 36.4 | 40.7 |
| TGTSS (Ours) | 31.5 | 33.6 |
| Utterance (gold label) | \\( \\hat{p}_{\\text{min-max}} \\), \\( \\sigma_{m}(\\hat{p}) \\) | Label predictions over 50 runs |
| funny joke (general) | [0.98-0.99], 0.003 (low) | general:50 |
| start house cleanup (IOT) | [0.002-0.97], 0.17 (high) | lists:26, IOT:6, general:6, play:5, news:3, social:1, calendar:1 |
| search for gluten free menus (cooking) | [0.002-0.693], 0.06 (low) | lists:28, take-away:18, social:1, music:1, cooking:1, play:1 |
| Attribute | MASSIVE | CLINC150 |
| Source | AmazonAlexa AI | DiaoGLUE |
| Domains | 18 | - |
| Intentss | 60 | 150 |
| TrainHoldout(Unseen) | 11,5142974 | 15,0003,000 |
| Balanced? | No. | Yes. 100per intent |
| Classification task | Domain | Intent |
| Massive | Clink150 | |||||
| Methods | Accuracy(%) | ΔLEm(%) ↑ | % of Eb | Accuracy(%) | ΔLEm(%) ↑ | % of Eb |
| Control baseline | 90.6 ± 0.6 | - | - | 95.1 ± 0.8 | - | - |
| Ensemble baseline (Eb) | 91.3 ± 0.5 | 34.5 | - | 95.4 ± 0.6 | 31.1 | - |
| L2 Regularization | 90.3 ± 0.5 | -2.3 | -7 | 94.9 ± 0.7 | -0.6 | -2 |
| SWA | 91.0 ± 0.5 | 17.6 | 51 | 95.2 ± 0.7 | 7.3 | 23 |
| Label Smoothing | 90.8 ± 0.5 | 5.7 | 17 | 95.2 ± 0.8 | 6.1 | 20 |
| TGTSS (Ours) | 91.3 ± 0.6 | 31.4 | 91 | 95.3 ± 0.8 | 26.7 | 86 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Label Smoothing | 37.9 | 40.5 |
| Ensemble baseline | 55.5 | 61.7 |
| TGTSS (Ours) | 53.4 | 55.9 |
| ΔLEm (%) ↑ | ||
| Methods | Massive | Cinc150 |
| Control baseline | - | - |
| Ensemble baseline | 27.4 | 24.1 |
| L2 Reg. | 3.8 | 4.2 |
| SWA | 11.3 | 4.3 |
| Label Smoothing | 5.4 | 8.1 |
| TGTSS (Ours) | 26 | 22.4 |
| ΔLEs (%) ↑ | ||
| Methods | Massive | Clinc150 |
| Label Smoothing | 14.9 | 20.7 |
| Ensemble baseline | 36.4 | 40.7 |
| TGTSS (Ours) | 31.5 | 33.6 |
| Socio-affective engagement | Examples | Count |
| Appreciation | "I really like the way you edited my goals. Thank you!" | 343 |
| Positive sentiment | "I feel that all the strategies are good ones and THEY WORK!!" | 357 |
| Negative sentiment | "I have now lost 2 replies! This will be my last try for tonight." | 25 |
| Cancer-related experience | "I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again. | 262 |
| Personal information | "Yea. My daughter and son-in-law came to visit. We had a great "Thanksgiving dinner. | 385 |
| Interest in communication | "Please let me know if you need any other information." | 161 |
| Vocatives | The patient addresses the study nurse by the name. | 612 |
| Greetings | "Sincerely," "Good morning", "With God's grace on us both," | 956 |
| Total Socio-affective spans | 3101 | |
| Cognitive engagement | Examples | Count |
| Intervention information | "It's hard to get up. I try to sleep late (until 8 or 8:30) and get in to work by 10, although I'd rather sleep until 10. I force myself to get up." | 1224 |
| Information-seeking question | "Are there other foods which make sores worse?" | 177 |
| Clarification question | "Are you talking about a based on a 1 to 10 type of thing?" | 14 |
| Acknowledgment | "I have received and read your email." | 50 |
| Agreement | "The care plan looks good. I don't have anything to add." | 298 |
| Disagreement | "He won't read the material, his mind is made up." | 13 |
| Initiative taking | "I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we." | 17 |
| Total Cognitive engagement spans | 1793 |
| # Patients | # Patient Turns | # Nurse Turns | # Unique Tokens |
| 68 | 1K | 1.1K | 29K |
| # Socio-affective spans | # Cognitive spans | ||
| 3101 | 1793 |
| Mean of f (SD) | Median of f (IQR) | |
| Socio-affective engagement categories | ||
| Positive sentiment | 3.63 (5.12) | 2 (4.25) |
| Negative sentiment | 0.35 (0.99) | 0 (0) |
| Appreciation | 3.96 (4.27) | 3 (5.25) |
| Cancer-related experience | 2.76 (2.97) | 2 (3.25) |
| Personal information | 4.38 (3.84) | 3 (4) |
| Vocatives | 8.6 (9.11) | 6.5(10.5) |
| Interest in communication | 2.24 (2.73) | 1 (3) |
| Greetings | 8.81(7.48) | 7 (8) |
| Cognitive engagement categories | ||
| Intervention information | 12.41 (9.36) | 10 (12.25) |
| Acknowledgement | 0.74 (1.19) | 0 (1) |
| Information-seeking question | 2.26 (3) | 1 (3) |
| Clarification question | 0.21 (0.59) | 0 (0) |
| Agreement | 3.38 (3.9) | 2 (6) |
| Disagreement | 0.19 (0.4) | 0 (0) |
| Initiative taking | 0.22 (0.59) | 0 (0) |
| Engagement category | Freq-ρ (p) |
| Socio-affective engagement category | |
| Positive sentiment | .35 (.008) |
| Appreciation | .25 (.007) |
| Cancer-related experience | .34 (.01) |
| Personal information | .34 (.01) |
| Vocatives | .33 (.01) |
| Interest in communication | .32 (.01) |
| Greetings | .27 (.04) |
| Cognitive engagement category | |
| Intervention information | .38 (.003) |
| Information-seeking question | .26 (.05) |
| Agreement | .34 (.009) |
| Initiative taking | .26 (.047) |
| Socio-Affective | |||
| Model | P | R | F1 |
| SVM-W | 81.1 (0.0) | 39.8 (0.0) | 40.7 (0.0) |
| SVM-C | 84.7 (0.0) | 59.7 (0.0) | 65.6 (0.0) |
| Bio-BERT | 75.4 (2.1) | 78.0 (1.1) | 76.4 (1.7) |
| BERT | 75.1 (1.1) | 76.6 (0.5) | 75.6 (0.6) |
| XLNet | 77.4 (1.9) | 78.7 (1.6) | 77.7 (1.4) |
| RoBerta | 78.6 (1.3) | 79.6 (1.2) | 78.8 (1.0) |
| Cognitive | |||
| Model | P | R | F1 |
| SVM-W | 62.1 (0.0) | 58.9 (0.0) | 57.6 (0.0) |
| SVM-C | 72.1 (0.0) | 73.5 (0.0) | 72.5 (0.0) |
| Bio-BERT | 78.0 (2.1) | 79.3 (0.4) | 78.5 (1.3) |
| BERT | 77.5 (0.9) | 78.3 (0.6) | 77.6 (0.6) |
| XLNet | 61.9 (0.8) | 63.9 (2.3) | 62.8 (0.7) |
| RoBerta | 61.5 (2.3) | 66.9 (0.2) | 63.8 (1.3) |
| Step | Intervention element | Description |
| 1 | The nurse initiates the message board by introducing herself and explaining how the intervention works on the message boards. The nurse then invited the patient to talk about herself and her experience with ovarian cancer and symptoms in general. | |
| 2 | Representational assessment | The nurse proceeds to ask a series of protocolized questions in the first few messages. This is for both the patient and the nurse to get a full understanding of the patient's experience with each symptom. As the patient thinks about the questions and writes responses, she may find that she sees some things more clearly or perhaps differently. A few examples of the protocolized questions include: \n· What does your fatigue feel like and how severe is it? \n· How has this symptom affected your life? Are you unable to do anything because of it? \n· How does your fatigue affect you emotionally? |
| 3 | Identifying and exploring Gaps, Errors, and Confusions | Throughout interactions with the participant, the nurse keeps attending to evidence of any confusion, concerns, or misconceptions about symptom management that the patient states or suggests during her message board posts. Any of these constitutes a barrier to effective symptom management for participants. |
| 4 | Creating conditions for conceptual change | If any concerns/ misconceptions/gaps are identified, the nurse will (1) discuss the relationship between identified concerns/confusion/misconceptions and consequences of poor symptoms that the patient talked about during representational assessment and (2) provide information to address/counter patient's concerns, gaps, or misconceptions. |
| 5 | Introducing replacement information | The nurse will then guide the patient to read the relevant clinical practice guidelines and provide relevant information to the patient to choose and use different strategies that fit well into their life and needs for managing symptoms. |
| 6 | Summary | The nurse will summarize information and the ways she believes it can benefit the patient. For example, increase comfort and less interference with life. |
| 7 | Goal Setting and Planning | The nurse will encourage the patient to work together on developing symptom goals and specific strategies to reach those goals. |
| 8 | Goal and Strategy Review | After 2 weeks, the nurse will ask protocolized questions to work with the patient to evaluate strategies, reinforce success, and outline modifications as needed. The nurse will also discuss any barriers that occur in the implementation of strategies and work with the patient to identify new or different strategies that could be integrated into her life. A few examples of the protocolized questions include: \n· Were you able to use the above strategies? \n· If not, what things prevented you from doing so? \n· If yes, how well did the strategies (name/list strategies) work in helping you reach your goal? |
| Socio-affective engagement | Description | Example |
| Positive sentiment | 1. Participant explicitly expresses positive emotions or feedback on the intervention design, symptom management progress (e.g., the effectiveness of strategies).2. *Very emotional greetings and goodbyes can fall under this code provided there is a clearly expressed strong emotional connection (intensity), for example \"Oh. It's so nice to meet you!\" | 1. Participant: \"Other than that, I feel that all the strategies are good ones and THEY WORK!!\"2. Participant: \"I feel relieved now that I know many other people have similar problems. That makes me feel normal \\(⑤\\) \"3. Participant: \"I will look forward to your advice on the issues!\" |
| Negative sentiment | Participant explicitly expresses negative emotions or feedback on the intervention design or content, the symptom management progress (e.g., the effectiveness of strategies). | Participant: \"We have been having computer problems. After writing part of a draft yesterday and half of one today, we contacted Jim from your group for tech support. Both messages to you were spontaneously deleted!!!!!!!\" |
| Appreciation | Participant explicitly appreciate and recognize the nurse's contributions and support. | Participant: \"I really appreciate the time you take to discuss on my symptoms. You have much insight and your recommendations are very helpful.\" |
| Cancer-related experience | 1. Participant shows her vulnerability (e.g., fear, guilt, traumatic cancer experience).2. Participant opens to the nurse about her feelings and beliefs of cancer trajectory and symptom management.3. Participant discusses with nurse about the cancer-related issues that she finds hard to discuss with others. | 1. Participant: \"I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again.\"2. Participant: \"Honestly I feel like a prisoner to it. I have to plan for it weekly, bug the nurse on my HCT to get me appointments, I feel it constantly everyday, I can't wear some of my clothes anymore or I fear I can't get an appointment and hit a weekend and have to go 3 extra days is misery.\"3. Participant: \"It's really hard for me to tell my family about my pain. My husband doesn't understand how much pain I have.\" |
| Personal information | Participant shares personal information with the nurse (e.g., family, friends, pets, health care team, religion) | Participant: \"Yea. My daughter and son-in-law came to visit. We had a great Thanksgiving dinner. Thanks for asking.\" |
| Interest in communication | 1. Participant invites nurse to ask for more information or clarification if needed.2. Participant explicitly expresses interest in further communication in general. | 1. Participant: \"Please let me know if you need any other information.\"2. Participant: \"I will update you on the treatment next Wednesday.\" |
| Vocatives | Participant addresses the study name by the name. | \"Names of study nurses\" |
| Greetings | 1. Participant starts posts with a greeting (e.g., Dear, Hello)\n2. Participant signs off the email (e.g., yours, ..., sincerely, ...., “thanks, ...) \n3. Participant shows benevolence, kindness, and social respect to the nurse. | 1. Participant: “Good morning.”\n2. Participant: “Take care, [name]”\n3. Participant: “May He bless your service in helping women to survive with less pain and life to gain.” |
| Cognitive engagement | Description | Example |
| Intervention information | 1. Participant answers nurse questions seeking information to complete intervention activities.\n2. Participant answers nurse questions seeking clarification about something participant said. The goal is to reduce misunderstanding.\n3. Participant reports completing intervention tasks assigned/suggested by the nurse.\n4. Participant provides information or updates unprompted (not asked by the nurse). For example, proposing new strategies, technical issues of the website.\n5. Participant updates the nurse about her cancer treatment that impact her symptom management. | 1. Nurse: “How does it affect you emotionally?”\nParticipant: “I probably get cranky when I am tired, but my dear husband doesn’t complain.”\n2. Nurse: “What do you mean you don’t like to take pills?” (Participant mentioned that she doesn’t like taking pills for symptoms).\nParticipant: “I have to take many prescribed medications. I don’t want to take more that I have to”.\n3. Participant: “I did read the guide you suggested. I guess I can start walking around the block”.\n4. Participant: “I just went to see the dentist and walk with Jim after that.” (Context: seeing a dentist is a strategy participant chose to manage her mouth sore).\n5. Participant: “I was completely drained after chemo last Wed. I didn’t have energy to do anything.” |
| Clarification question | Participant asks nurse questions seeking clarification about something the nurse said. The goal is to reduce misunderstanding. | 1. Participant: “my nutritionist told me I should eat more fruits and vegetable. But the guide says I should not have raw fruits/vegetables if I have diarrhea. Can I still have salads sometimes?”\n2. Participant: “taking pain meds won’t cause addiction, even matter if I take it every day?” |
| Information-seeking question | 1. Participant proposes a question to seek for answers. The specific indicators contain “?”, “who”, “what”, “where”, “when”, “why”, “how”, “how-much/many” or other key question symbols.\n2. Participant seeks for advice/suggestions.\n3. Participant asks the nurse for a favor. | 1. Participant: “I checked out the diet website. Do you know what’s the regimen?”\n2. Participant: “I look forward to your answer to my questions about your list.”\n3. Participant: “It would be wonderful if you could add the goals and strategies as I’m getting ready to go out of town and am really busy preparing for that and trying to pace myself.” |
| Acknowledgement | Participant acknowledges/affirms the nurse suggestions, ideas, summary, questions. | Participant: “I did read your e-mail last night before going to bed. I didn’t answer because I was tired, but I did keep thinking about your questions.” |
| Agreement | 1. Participant explicitly shares the same opinion with the nurse on the topics (e.g., suggestions on symptom management strategies, barriers to good symptom management, reading resources), including restating or paraphrasing or summarizing what the nurse has said. The specific indicators contain “I agree”, “I think so”, etc.\n2. Participant agrees to the plans or do what the nurse suggests doing, including stating that she will try what the nurse has said (being open to different/new experience). The specific indicators contain “I will...”, “Okay”, etc.\n3. Participant approves what the nurse does or suggests. | 1. Participant: “In just the few days of doing yoga I have become aware that my posture is not good. I think maybe that is affecting the way I walk as you suggested when you asked how I walk.”\n2. Nurse: “I was thinking what might make the most sense is to do your 2-week follow-up for Sleep before moving on to headaches. How’s that sound to you? I will have the Sleep Follow-Up posted up there (on web page) under the Sleep topic heading.”\n3. Participant: “okay, I will take a look at the Sleep Follow-up.”\n4. Participant: “The plan looks fine. I don’t have anything to add.” |
| Disagreement | 1. Participant explicitly disagrees with nurse on a certain topic. The specific indicators contain \"I disagree\", \"I don't think so\", etc. \n2. Participant disagrees to do what the nurse suggests doing, including explaining why they don't want to or can't try what the nurse suggested. The specific indicators contain \"I don't...\", \"doesn't work\", etc. | 1. Participant: \"What do you mean? I don't think this is helping me. I was hoping that the program can help me solve the problem.\" \n2. Nurse: \"here are other similar activities that you might also like, such as Tai Chi. Many community centers offer both of these. Look for classes for people over 50... they often start out a little slower and easier. Don't push too hard or fast when you start out.\" \nParticipant: \"I tried Tai Chi two years ago at the senior center. My sister does it and recommended it. I did not get the feeling of renewal from it that she did. Actually, I did not get much from it at all. I went twice a week for two months.\" |
| Initiative taking | 1. Participant anticipates the later steps of the intervention. For example, when the nurse asks questions about patient's symptom representation (element 1 Representation Assessment), the patient both describes the symptoms AND proceeds to describe the strategies she wants to try during the intervention (element 6 Goal setting and Planning). \n2. Participant changes the course of the intervention by taking initiative in changing symptoms to work with the nurse before the nurse's prompts. \n3. Participant offers suggestions to future study design | 1. Nurse: \"Can you tell me about your hot flashes? You can describe a typical day.\" \nParticipant: \"It usually happen at night. I have to get up several times because it is too hot. In looking at the Symptom Care Guides, there are four strategies I would like to work on: 1. Keep a symptom diary. (I have printed a template. I will keep it on the end table where I sit to watch TV, crochet, read, etc.)2. Yoga. I will look on the TV an see if I can find a yoga class. Vitamin E. I will call and get approval from my HCT.4. Paced Respirations (abdominal breathing).\" \nNurse: \"Wow, you're jumping right into this process! That's wonderful. We are in the process of updating our [Resource Library] for Hot Flashes and will add some Yoga web sites, soon. Before we go too much further, though, I do have a few more questions for you.... Don't I always © This is so we can both get the full picture. Can you tell me a little more about how your hot flashes feel?\" \n2. Participant: \"I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we.\" \n3. Participant: \"Oh just remembered an idea: I wonder if a message box would be helpful to add to the questionnaires? The patient could type out any comments/explanations in answering the questions. Just wondering in the event of any chemo and/or life changes that occurred during this research time that impacted the current questionnaire.\" |
| Socio-affective engagement | Examples | Count |
| Appreciation | "I really like the way you edited my goals. Thank you!" | 343 |
| Positive sentiment | "I feel that all the strategies are good ones and THEY WORK!!" | 357 |
| Negative sentiment | "I have now lost 2 replies! This will be my last try for tonight." | 25 |
| Cancer-related experience | "I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again. | 262 |
| Personal information | "Yea. My daughter and son-in-law came to visit. We had a great "Thanksgiving dinner. | 385 |
| Interest in communication | "Please let me know if you need any other information." | 161 |
| Vocatives | The patient addresses the study nurse by the name. | 612 |
| Greetings | "Sincerely," "Good morning", "With God's grace on us both," | 956 |
| Total Socio-affective spans | 3101 | |
| Cognitive engagement | Examples | Count |
| Intervention information | "It's hard to get up. I try to sleep late (until 8 or 8:30) and get in to work by 10, although I'd rather sleep until 10. I force myself to get up." | 1224 |
| Information-seeking question | "Are there other foods which make sores worse?" | 177 |
| Clarification question | "Are you talking about a based on a 1 to 10 type of thing?" | 14 |
| Acknowledgment | "I have received and read your email." | 50 |
| Agreement | "The care plan looks good. I don't have anything to add." | 298 |
| Disagreement | "He won't read the material, his mind is made up." | 13 |
| Initiative taking | "I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we." | 17 |
| Total Cognitive engagement spans | 1793 |
| # Patients | # Patient Turns | # Nurse Turns | # Unique Tokens |
| 68 | 1K | 1.1K | 29K |
| # Socio-affective spans | # Cognitive spans | ||
| 3101 | 1793 |
| Mean of f (SD) | Median of f (IQR) | |
| Socio-affective engagement categories | ||
| Positive sentiment | 3.63 (5.12) | 2 (4.25) |
| Negative sentiment | 0.35 (0.99) | 0 (0) |
| Appreciation | 3.96 (4.27) | 3 (5.25) |
| Cancer-related experience | 2.76 (2.97) | 2 (3.25) |
| Personal information | 4.38 (3.84) | 3 (4) |
| Vocatives | 8.6 (9.11) | 6.5(10.5) |
| Interest in communication | 2.24 (2.73) | 1 (3) |
| Greetings | 8.81(7.48) | 7 (8) |
| Cognitive engagement categories | ||
| Intervention information | 12.41 (9.36) | 10 (12.25) |
| Acknowledgement | 0.74 (1.19) | 0 (1) |
| Information-seeking question | 2.26 (3) | 1 (3) |
| Clarification question | 0.21 (0.59) | 0 (0) |
| Agreement | 3.38 (3.9) | 2 (6) |
| Disagreement | 0.19 (0.4) | 0 (0) |
| Initiative taking | 0.22 (0.59) | 0 (0) |
| Engagement category | Freq-ρ (p) |
| Socio-affective engagement category | |
| Positive sentiment | .35 (.008) |
| Appreciation | .25 (.007) |
| Cancer-related experience | .34 (.01) |
| Personal information | .34 (.01) |
| Vocatives | .33 (.01) |
| Interest in communication | .32 (.01) |
| Greetings | .27 (.04) |
| Cognitive engagement category | |
| Intervention information | .38 (.003) |
| Information-seeking question | .26 (.05) |
| Agreement | .34 (.009) |
| Initiative taking | .26 (.047) |
| Socio-Affective | |||
| Model | P | R | F1 |
| SVM-W | 81.1 (0.0) | 39.8 (0.0) | 40.7 (0.0) |
| SVM-C | 84.7 (0.0) | 59.7 (0.0) | 65.6 (0.0) |
| Bio-BERT | 75.4 (2.1) | 78.0 (1.1) | 76.4 (1.7) |
| BERT | 75.1 (1.1) | 76.6 (0.5) | 75.6 (0.6) |
| XLNet | 77.4 (1.9) | 78.7 (1.6) | 77.7 (1.4) |
| RoBerta | 78.6 (1.3) | 79.6 (1.2) | 78.8 (1.0) |
| Cognitive | |||
| Model | P | R | F1 |
| SVM-W | 62.1 (0.0) | 58.9 (0.0) | 57.6 (0.0) |
| SVM-C | 72.1 (0.0) | 73.5 (0.0) | 72.5 (0.0) |
| Bio-BERT | 78.0 (2.1) | 79.3 (0.4) | 78.5 (1.3) |
| BERT | 77.5 (0.9) | 78.3 (0.6) | 77.6 (0.6) |
| XLNet | 61.9 (0.8) | 63.9 (2.3) | 62.8 (0.7) |
| RoBerta | 61.5 (2.3) | 66.9 (0.2) | 63.8 (1.3) |
| Step | Intervention element | Description |
| 1 | The nurse initiates the message board by introducing herself and explaining how the intervention works on the message boards. The nurse then invited the patient to talk about herself and her experience with ovarian cancer and symptoms in general. | |
| 2 | Representational assessment | The nurse proceeds to ask a series of protocolized questions in the first few messages. This is for both the patient and the nurse to get a full understanding of the patient's experience with each symptom. As the patient thinks about the questions and writes responses, she may find that she sees some things more clearly or perhaps differently. A few examples of the protocolized questions include: \n· What does your fatigue feel like and how severe is it? \n· How has this symptom affected your life? Are you unable to do anything because of it? \n· How does your fatigue affect you emotionally? |
| 3 | Identifying and exploring Gaps, Errors, and Confusions | Throughout interactions with the participant, the nurse keeps attending to evidence of any confusion, concerns, or misconceptions about symptom management that the patient states or suggests during her message board posts. Any of these constitutes a barrier to effective symptom management for participants. |
| 4 | Creating conditions for conceptual change | If any concerns/ misconceptions/gaps are identified, the nurse will (1) discuss the relationship between identified concerns/confusion/misconceptions and consequences of poor symptoms that the patient talked about during representational assessment and (2) provide information to address/counter patient's concerns, gaps, or misconceptions. |
| 5 | Introducing replacement information | The nurse will then guide the patient to read the relevant clinical practice guidelines and provide relevant information to the patient to choose and use different strategies that fit well into their life and needs for managing symptoms. |
| 6 | Summary | The nurse will summarize information and the ways she believes it can benefit the patient. For example, increase comfort and less interference with life. |
| 7 | Goal Setting and Planning | The nurse will encourage the patient to work together on developing symptom goals and specific strategies to reach those goals. |
| 8 | Goal and Strategy Review | After 2 weeks, the nurse will ask protocolized questions to work with the patient to evaluate strategies, reinforce success, and outline modifications as needed. The nurse will also discuss any barriers that occur in the implementation of strategies and work with the patient to identify new or different strategies that could be integrated into her life. A few examples of the protocolized questions include: \n· Were you able to use the above strategies? \n· If not, what things prevented you from doing so? \n· If yes, how well did the strategies (name/list strategies) work in helping you reach your goal? |
| Socio-affective engagement | Description | Example |
| Positive sentiment | 1. Participant explicitly expresses positive emotions or feedback on the intervention design, symptom management progress (e.g., the effectiveness of strategies).2. *Very emotional greetings and goodbyes can fall under this code provided there is a clearly expressed strong emotional connection (intensity), for example "Oh. It's so nice to meet you!" | 1. Participant: "Other than that, I feel that all the strategies are good ones and THEY WORK!!"2. Participant: "I feel relieved now that I know many other people have similar problems. That makes me feel normal \\(⑤\\) "3. Participant: "I will look forward to your advice on the issues!" |
| Negative sentiment | Participant explicitly expresses negative emotions or feedback on the intervention design or content, the symptom management progress (e.g., the effectiveness of strategies). | Participant: "We have been having computer problems. After writing part of a draft yesterday and half of one today, we contacted Jim from your group for tech support. Both messages to you were spontaneously deleted!!!!!!!" |
| Appreciation | Participant explicitly appreciate and recognize the nurse's contributions and support. | Participant: "I really appreciate the time you take to discuss on my symptoms. You have much insight and your recommendations are very helpful." |
| Cancer-related experience | 1. Participant shows her vulnerability (e.g., fear, guilt, traumatic cancer experience).2. Participant opens to the nurse about her feelings and beliefs of cancer trajectory and symptom management.3. Participant discusses with nurse about the cancer-related issues that she finds hard to discuss with others. | 1. Participant: "I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again."2. Participant: "Honestly I feel like a prisoner to it. I have to plan for it weekly, bug the nurse on my HCT to get me appointments, I feel it constantly everyday, I can't wear some of my clothes anymore or I fear I can't get an appointment and hit a weekend and have to go 3 extra days is misery."3. Participant: "It's really hard for me to tell my family about my pain. My husband doesn't understand how much pain I have." |
| Personal information | Participant shares personal information with the nurse (e.g., family, friends, pets, health care team, religion) | Participant: "Yea. My daughter and son-in-law came to visit. We had a great Thanksgiving dinner. Thanks for asking." |
| Interest in communication | 1. Participant invites nurse to ask for more information or clarification if needed.2. Participant explicitly expresses interest in further communication in general. | 1. Participant: "Please let me know if you need any other information."2. Participant: "I will update you on the treatment next Wednesday." |
| Vocatives | Participant addresses the study name by the name. | "Names of study nurses" |
| Greetings | 1. Participant starts posts with a greeting (e.g., Dear, Hello)\n2. Participant signs off the email (e.g., yours, ..., sincerely, ...., “thanks, ...) \n3. Participant shows benevolence, kindness, and social respect to the nurse. | 1. Participant: “Good morning.”\n2. Participant: “Take care, [name]”\n3. Participant: “May He bless your service in helping women to survive with less pain and life to gain.” |
| Cognitive engagement | Description | Example |
| Intervention information | 1. Participant answers nurse questions seeking information to complete intervention activities.\n2. Participant answers nurse questions seeking clarification about something participant said. The goal is to reduce misunderstanding.\n3. Participant reports completing intervention tasks assigned/suggested by the nurse.\n4. Participant provides information or updates unprompted (not asked by the nurse). For example, proposing new strategies, technical issues of the website.\n5. Participant updates the nurse about her cancer treatment that impact her symptom management. | 1. Nurse: “How does it affect you emotionally?”\nParticipant: “I probably get cranky when I am tired, but my dear husband doesn’t complain.”\n2. Nurse: “What do you mean you don’t like to take pills?” (Participant mentioned that she doesn’t like taking pills for symptoms).\nParticipant: “I have to take many prescribed medications. I don’t want to take more that I have to”.\n3. Participant: “I did read the guide you suggested. I guess I can start walking around the block”.\n4. Participant: “I just went to see the dentist and walk with Jim after that.” (Context: seeing a dentist is a strategy participant chose to manage her mouth sore).\n5. Participant: “I was completely drained after chemo last Wed. I didn’t have energy to do anything.” |
| Clarification question | Participant asks nurse questions seeking clarification about something the nurse said. The goal is to reduce misunderstanding. | 1. Participant: “my nutritionist told me I should eat more fruits and vegetable. But the guide says I should not have raw fruits/vegetables if I have diarrhea. Can I still have salads sometimes?”\n2. Participant: “taking pain meds won’t cause addiction, even matter if I take it every day?” |
| Information-seeking question | 1. Participant proposes a question to seek for answers. The specific indicators contain “?”, “who”, “what”, “where”, “when”, “why”, “how”, “how-much/many” or other key question symbols.\n2. Participant seeks for advice/suggestions.\n3. Participant asks the nurse for a favor. | 1. Participant: “I checked out the diet website. Do you know what’s the regimen?”\n2. Participant: “I look forward to your answer to my questions about your list.”\n3. Participant: “It would be wonderful if you could add the goals and strategies as I’m getting ready to go out of town and am really busy preparing for that and trying to pace myself.” |
| Acknowledgement | Participant acknowledges/affirms the nurse suggestions, ideas, summary, questions. | Participant: “I did read your e-mail last night before going to bed. I didn’t answer because I was tired, but I did keep thinking about your questions.” |
| Agreement | 1. Participant explicitly shares the same opinion with the nurse on the topics (e.g., suggestions on symptom management strategies, barriers to good symptom management, reading resources), including restating or paraphrasing or summarizing what the nurse has said. The specific indicators contain “I agree”, “I think so”, etc.\n2. Participant agrees to the plans or do what the nurse suggests doing, including stating that she will try what the nurse has said (being open to different/new experience). The specific indicators contain “I will...”, “Okay”, etc.\n3. Participant approves what the nurse does or suggests. | 1. Participant: “In just the few days of doing yoga I have become aware that my posture is not good. I think maybe that is affecting the way I walk as you suggested when you asked how I walk.”\n2. Nurse: “I was thinking what might make the most sense is to do your 2-week follow-up for Sleep before moving on to headaches. How’s that sound to you? I will have the Sleep Follow-Up posted up there (on web page) under the Sleep topic heading.”\n3. Participant: “okay, I will take a look at the Sleep Follow-up.”\n4. Participant: “The plan looks fine. I don’t have anything to add.” |
| Disagreement | 1. Participant explicitly disagrees with nurse on a certain topic. The specific indicators contain "I disagree", "I don't think so", etc. \n2. Participant disagrees to do what the nurse suggests doing, including explaining why they don't want to or can't try what the nurse suggested. The specific indicators contain "I don't...", "doesn't work", etc. | 1. Participant: "What do you mean? I don't think this is helping me. I was hoping that the program can help me solve the problem." \n2. Nurse: "here are other similar activities that you might also like, such as Tai Chi. Many community centers offer both of these. Look for classes for people over 50... they often start out a little slower and easier. Don't push too hard or fast when you start out." \nParticipant: "I tried Tai Chi two years ago at the senior center. My sister does it and recommended it. I did not get the feeling of renewal from it that she did. Actually, I did not get much from it at all. I went twice a week for two months." |
| Initiative taking | 1. Participant anticipates the later steps of the intervention. For example, when the nurse asks questions about patient's symptom representation (element 1 Representation Assessment), the patient both describes the symptoms AND proceeds to describe the strategies she wants to try during the intervention (element 6 Goal setting and Planning). \n2. Participant changes the course of the intervention by taking initiative in changing symptoms to work with the nurse before the nurse's prompts. \n3. Participant offers suggestions to future study design | 1. Nurse: "Can you tell me about your hot flashes? You can describe a typical day." \nParticipant: "It usually happen at night. I have to get up several times because it is too hot. In looking at the Symptom Care Guides, there are four strategies I would like to work on: 1. Keep a symptom diary. (I have printed a template. I will keep it on the end table where I sit to watch TV, crochet, read, etc.)2. Yoga. I will look on the TV an see if I can find a yoga class. Vitamin E. I will call and get approval from my HCT.4. Paced Respirations (abdominal breathing)." \nNurse: "Wow, you're jumping right into this process! That's wonderful. We are in the process of updating our [Resource Library] for Hot Flashes and will add some Yoga web sites, soon. Before we go too much further, though, I do have a few more questions for you.... Don't I always © This is so we can both get the full picture. Can you tell me a little more about how your hot flashes feel?" \n2. Participant: "I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we." \n3. Participant: "Oh just remembered an idea: I wonder if a message box would be helpful to add to the questionnaires? The patient could type out any comments/explanations in answering the questions. Just wondering in the event of any chemo and/or life changes that occurred during this research time that impacted the current questionnaire." |
| Socio-affective engagement | Examples | Count |
| Appreciation | "I really like the way you edited my goals. Thank you!" | 343 |
| Positive sentiment | "I feel that all the strategies are good ones and THEY WORK!!" | 357 |
| Negative sentiment | "I have now lost 2 replies! This will be my last try for tonight." | 25 |
| Cancer-related experience | "I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again. | 262 |
| Personal information | "Yea. My daughter and son-in-law came to visit. We had a great "Thanksgiving dinner. | 385 |
| Interest in communication | "Please let me know if you need any other information." | 161 |
| Vocatives | The patient addresses the study nurse by the name. | 612 |
| Greetings | "Sincerely," "Good morning", "With God's grace on us both," | 956 |
| Total Socio-affective spans | 3101 | |
| Cognitive engagement | Examples | Count |
| Intervention information | "It's hard to get up. I try to sleep late (until 8 or 8:30) and get in to work by 10, although I'd rather sleep until 10. I force myself to get up." | 1224 |
| Information-seeking question | "Are there other foods which make sores worse?" | 177 |
| Clarification question | "Are you talking about a based on a 1 to 10 type of thing?" | 14 |
| Acknowledgment | "I have received and read your email." | 50 |
| Agreement | "The care plan looks good. I don't have anything to add." | 298 |
| Disagreement | "He won't read the material, his mind is made up." | 13 |
| Initiative taking | "I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we." | 17 |
| Total Cognitive engagement spans | 1793 | |
| # Patients | # Patient Turns | # Nurse Turns | # Unique Tokens |
| 68 | 1K | 1.1K | 29K |
| # Socio-affective spans | # Cognitive spans | ||
| 3101 | 1793 | ||
| Mean of f (SD) | Median of f (IQR) | |
| Socio-affective engagement categories | ||
| Positive sentiment | 3.63 (5.12) | 2 (4.25) |
| Negative sentiment | 0.35 (0.99) | 0 (0) |
| Appreciation | 3.96 (4.27) | 3 (5.25) |
| Cancer-related experience | 2.76 (2.97) | 2 (3.25) |
| Personal information | 4.38 (3.84) | 3 (4) |
| Vocatives | 8.6 (9.11) | 6.5(10.5) |
| Interest in communication | 2.24 (2.73) | 1 (3) |
| Greetings | 8.81(7.48) | 7 (8) |
| Cognitive engagement categories | ||
| Intervention information | 12.41 (9.36) | 10 (12.25) |
| Acknowledgement | 0.74 (1.19) | 0 (1) |
| Information-seeking question | 2.26 (3) | 1 (3) |
| Clarification question | 0.21 (0.59) | 0 (0) |
| Agreement | 3.38 (3.9) | 2 (6) |
| Disagreement | 0.19 (0.4) | 0 (0) |
| Initiative taking | 0.22 (0.59) | 0 (0) |
| Engagement category | Freq-ρ (p) |
| Socio-affective engagement category | |
| Positive sentiment | .35 (.008) |
| Appreciation | .25 (.007) |
| Cancer-related experience | .34 (.01) |
| Personal information | .34 (.01) |
| Vocatives | .33 (.01) |
| Interest in communication | .32 (.01) |
| Greetings | .27 (.04) |
| Cognitive engagement category | |
| Intervention information | .38 (.003) |
| Information-seeking question | .26 (.05) |
| Agreement | .34 (.009) |
| Initiative taking | .26 (.047) |
| Socio-Affective | |||
| Model | P | R | F1 |
| SVM-W | 81.1 (0.0) | 39.8 (0.0) | 40.7 (0.0) |
| SVM-C | 84.7 (0.0) | 59.7 (0.0) | 65.6 (0.0) |
| Bio-BERT | 75.4 (2.1) | 78.0 (1.1) | 76.4 (1.7) |
| BERT | 75.1 (1.1) | 76.6 (0.5) | 75.6 (0.6) |
| XLNet | 77.4 (1.9) | 78.7 (1.6) | 77.7 (1.4) |
| RoBerta | 78.6 (1.3) | 79.6 (1.2) | 78.8 (1.0) |
| Cognitive | |||
| Model | P | R | F1 |
| SVM-W | 62.1 (0.0) | 58.9 (0.0) | 57.6 (0.0) |
| SVM-C | 72.1 (0.0) | 73.5 (0.0) | 72.5 (0.0) |
| Bio-BERT | 78.0 (2.1) | 79.3 (0.4) | 78.5 (1.3) |
| BERT | 77.5 (0.9) | 78.3 (0.6) | 77.6 (0.6) |
| XLNet | 61.9 (0.8) | 63.9 (2.3) | 62.8 (0.7) |
| RoBerta | 61.5 (2.3) | 66.9 (0.2) | 63.8 (1.3) |
| Step | Intervention element | Description |
| 1 | The nurse initiates the message board by introducing herself and explaining how the intervention works on the message boards. The nurse then invited the patient to talk about herself and her experience with ovarian cancer and symptoms in general. | |
| 2 | Representational assessment | The nurse proceeds to ask a series of protocolized questions in the first few messages. This is for both the patient and the nurse to get a full understanding of the patient's experience with each symptom. As the patient thinks about the questions and writes responses, she may find that she sees some things more clearly or perhaps differently. A few examples of the protocolized questions include: +· What does your fatigue feel like and how severe is it? +· How has this symptom affected your life? Are you unable to do anything because of it? +· How does your fatigue affect you emotionally? |
| 3 | Identifying and exploring Gaps, Errors, and Confusions | Throughout interactions with the participant, the nurse keeps attending to evidence of any confusion, concerns, or misconceptions about symptom management that the patient states or suggests during her message board posts. Any of these constitutes a barrier to effective symptom management for participants. |
| 4 | Creating conditions for conceptual change | If any concerns/ misconceptions/gaps are identified, the nurse will (1) discuss the relationship between identified concerns/confusion/misconceptions and consequences of poor symptoms that the patient talked about during representational assessment and (2) provide information to address/counter patient's concerns, gaps, or misconceptions. |
| 5 | Introducing replacement information | The nurse will then guide the patient to read the relevant clinical practice guidelines and provide relevant information to the patient to choose and use different strategies that fit well into their life and needs for managing symptoms. |
| 6 | Summary | The nurse will summarize information and the ways she believes it can benefit the patient. For example, increase comfort and less interference with life. |
| 7 | Goal Setting and Planning | The nurse will encourage the patient to work together on developing symptom goals and specific strategies to reach those goals. |
| 8 | Goal and Strategy Review | After 2 weeks, the nurse will ask protocolized questions to work with the patient to evaluate strategies, reinforce success, and outline modifications as needed. The nurse will also discuss any barriers that occur in the implementation of strategies and work with the patient to identify new or different strategies that could be integrated into her life. A few examples of the protocolized questions include: +· Were you able to use the above strategies? +· If not, what things prevented you from doing so? +· If yes, how well did the strategies (name/list strategies) work in helping you reach your goal? |
| Socio-affective engagement | Description | Example |
| Positive sentiment | 1. Participant explicitly expresses positive emotions or feedback on the intervention design, symptom management progress (e.g., the effectiveness of strategies).2. *Very emotional greetings and goodbyes can fall under this code provided there is a clearly expressed strong emotional connection (intensity), for example "Oh. It's so nice to meet you!" | 1. Participant: "Other than that, I feel that all the strategies are good ones and THEY WORK!!"2. Participant: "I feel relieved now that I know many other people have similar problems. That makes me feel normal \(⑤\) "3. Participant: "I will look forward to your advice on the issues!" |
| Negative sentiment | Participant explicitly expresses negative emotions or feedback on the intervention design or content, the symptom management progress (e.g., the effectiveness of strategies). | Participant: "We have been having computer problems. After writing part of a draft yesterday and half of one today, we contacted Jim from your group for tech support. Both messages to you were spontaneously deleted!!!!!!!" |
| Appreciation | Participant explicitly appreciate and recognize the nurse's contributions and support. | Participant: "I really appreciate the time you take to discuss on my symptoms. You have much insight and your recommendations are very helpful." |
| Cancer-related experience | 1. Participant shows her vulnerability (e.g., fear, guilt, traumatic cancer experience).2. Participant opens to the nurse about her feelings and beliefs of cancer trajectory and symptom management.3. Participant discusses with nurse about the cancer-related issues that she finds hard to discuss with others. | 1. Participant: "I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again."2. Participant: "Honestly I feel like a prisoner to it. I have to plan for it weekly, bug the nurse on my HCT to get me appointments, I feel it constantly everyday, I can't wear some of my clothes anymore or I fear I can't get an appointment and hit a weekend and have to go 3 extra days is misery."3. Participant: "It's really hard for me to tell my family about my pain. My husband doesn't understand how much pain I have." |
| Personal information | Participant shares personal information with the nurse (e.g., family, friends, pets, health care team, religion) | Participant: "Yea. My daughter and son-in-law came to visit. We had a great Thanksgiving dinner. Thanks for asking." |
| Interest in communication | 1. Participant invites nurse to ask for more information or clarification if needed.2. Participant explicitly expresses interest in further communication in general. | 1. Participant: "Please let me know if you need any other information."2. Participant: "I will update you on the treatment next Wednesday." |
| Vocatives | Participant addresses the study name by the name. | "Names of study nurses" |
| Greetings | 1. Participant starts posts with a greeting (e.g., Dear, Hello) +2. Participant signs off the email (e.g., yours, ..., sincerely, ...., “thanks, ...) +3. Participant shows benevolence, kindness, and social respect to the nurse. | 1. Participant: “Good morning.” +2. Participant: “Take care, [name]” +3. Participant: “May He bless your service in helping women to survive with less pain and life to gain.” |
| Cognitive engagement | Description | Example |
| Intervention information | 1. Participant answers nurse questions seeking information to complete intervention activities. +2. Participant answers nurse questions seeking clarification about something participant said. The goal is to reduce misunderstanding. +3. Participant reports completing intervention tasks assigned/suggested by the nurse. +4. Participant provides information or updates unprompted (not asked by the nurse). For example, proposing new strategies, technical issues of the website. +5. Participant updates the nurse about her cancer treatment that impact her symptom management. | 1. Nurse: “How does it affect you emotionally?” +Participant: “I probably get cranky when I am tired, but my dear husband doesn’t complain.” +2. Nurse: “What do you mean you don’t like to take pills?” (Participant mentioned that she doesn’t like taking pills for symptoms). +Participant: “I have to take many prescribed medications. I don’t want to take more that I have to”. +3. Participant: “I did read the guide you suggested. I guess I can start walking around the block”. +4. Participant: “I just went to see the dentist and walk with Jim after that.” (Context: seeing a dentist is a strategy participant chose to manage her mouth sore). +5. Participant: “I was completely drained after chemo last Wed. I didn’t have energy to do anything.” |
| Clarification question | Participant asks nurse questions seeking clarification about something the nurse said. The goal is to reduce misunderstanding. | 1. Participant: “my nutritionist told me I should eat more fruits and vegetable. But the guide says I should not have raw fruits/vegetables if I have diarrhea. Can I still have salads sometimes?” +2. Participant: “taking pain meds won’t cause addiction, even matter if I take it every day?” |
| Information-seeking question | 1. Participant proposes a question to seek for answers. The specific indicators contain “?”, “who”, “what”, “where”, “when”, “why”, “how”, “how-much/many” or other key question symbols. +2. Participant seeks for advice/suggestions. +3. Participant asks the nurse for a favor. | 1. Participant: “I checked out the diet website. Do you know what’s the regimen?” +2. Participant: “I look forward to your answer to my questions about your list.” +3. Participant: “It would be wonderful if you could add the goals and strategies as I’m getting ready to go out of town and am really busy preparing for that and trying to pace myself.” |
| Acknowledgement | Participant acknowledges/affirms the nurse suggestions, ideas, summary, questions. | Participant: “I did read your e-mail last night before going to bed. I didn’t answer because I was tired, but I did keep thinking about your questions.” |
| Agreement | 1. Participant explicitly shares the same opinion with the nurse on the topics (e.g., suggestions on symptom management strategies, barriers to good symptom management, reading resources), including restating or paraphrasing or summarizing what the nurse has said. The specific indicators contain “I agree”, “I think so”, etc. +2. Participant agrees to the plans or do what the nurse suggests doing, including stating that she will try what the nurse has said (being open to different/new experience). The specific indicators contain “I will...”, “Okay”, etc. +3. Participant approves what the nurse does or suggests. | 1. Participant: “In just the few days of doing yoga I have become aware that my posture is not good. I think maybe that is affecting the way I walk as you suggested when you asked how I walk.” +2. Nurse: “I was thinking what might make the most sense is to do your 2-week follow-up for Sleep before moving on to headaches. How’s that sound to you? I will have the Sleep Follow-Up posted up there (on web page) under the Sleep topic heading.” +3. Participant: “okay, I will take a look at the Sleep Follow-up.” +4. Participant: “The plan looks fine. I don’t have anything to add.” |
| Disagreement | 1. Participant explicitly disagrees with nurse on a certain topic. The specific indicators contain "I disagree", "I don't think so", etc. +2. Participant disagrees to do what the nurse suggests doing, including explaining why they don't want to or can't try what the nurse suggested. The specific indicators contain "I don't...", "doesn't work", etc. | 1. Participant: "What do you mean? I don't think this is helping me. I was hoping that the program can help me solve the problem." +2. Nurse: "here are other similar activities that you might also like, such as Tai Chi. Many community centers offer both of these. Look for classes for people over 50... they often start out a little slower and easier. Don't push too hard or fast when you start out." +Participant: "I tried Tai Chi two years ago at the senior center. My sister does it and recommended it. I did not get the feeling of renewal from it that she did. Actually, I did not get much from it at all. I went twice a week for two months." |
| Initiative taking | 1. Participant anticipates the later steps of the intervention. For example, when the nurse asks questions about patient's symptom representation (element 1 Representation Assessment), the patient both describes the symptoms AND proceeds to describe the strategies she wants to try during the intervention (element 6 Goal setting and Planning). +2. Participant changes the course of the intervention by taking initiative in changing symptoms to work with the nurse before the nurse's prompts. +3. Participant offers suggestions to future study design | 1. Nurse: "Can you tell me about your hot flashes? You can describe a typical day." +Participant: "It usually happen at night. I have to get up several times because it is too hot. In looking at the Symptom Care Guides, there are four strategies I would like to work on: 1. Keep a symptom diary. (I have printed a template. I will keep it on the end table where I sit to watch TV, crochet, read, etc.)2. Yoga. I will look on the TV an see if I can find a yoga class. Vitamin E. I will call and get approval from my HCT.4. Paced Respirations (abdominal breathing)." +Nurse: "Wow, you're jumping right into this process! That's wonderful. We are in the process of updating our [Resource Library] for Hot Flashes and will add some Yoga web sites, soon. Before we go too much further, though, I do have a few more questions for you.... Don't I always © This is so we can both get the full picture. Can you tell me a little more about how your hot flashes feel?" +2. Participant: "I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we." +3. Participant: "Oh just remembered an idea: I wonder if a message box would be helpful to add to the questionnaires? The patient could type out any comments/explanations in answering the questions. Just wondering in the event of any chemo and/or life changes that occurred during this research time that impacted the current questionnaire." |
| Socio-affective engagement | Examples | Count |
| Appreciation | "I really like the way you edited my goals. Thank you!" | 343 |
| Positive sentiment | "I feel that all the strategies are good ones and THEY WORK!!" | 357 |
| Negative sentiment | "I have now lost 2 replies! This will be my last try for tonight." | 25 |
| Cancer-related experience | "I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again. | 262 |
| Personal information | "Yea. My daughter and son-in-law came to visit. We had a great "Thanksgiving dinner. | 385 |
| Interest in communication | "Please let me know if you need any other information." | 161 |
| Vocatives | The patient addresses the study nurse by the name. | 612 |
| Greetings | "Sincerely," "Good morning", "With God's grace on us both," | 956 |
| Total Socio-affective spans | 3101 | |
| Cognitive engagement | Examples | Count |
| Intervention information | "It's hard to get up. I try to sleep late (until 8 or 8:30) and get in to work by 10, although I'd rather sleep until 10. I force myself to get up." | 1224 |
| Information-seeking question | "Are there other foods which make sores worse?" | 177 |
| Clarification question | "Are you talking about a based on a 1 to 10 type of thing?" | 14 |
| Acknowledgment | "I have received and read your email." | 50 |
| Agreement | "The care plan looks good. I don't have anything to add." | 298 |
| Disagreement | "He won't read the material, his mind is made up." | 13 |
| Initiative taking | "I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we." | 17 |
| Total Cognitive engagement spans | 1793 |
| # Patients | # Patient Turns | # Nurse Turns | # Unique Tokens |
| 68 | 1K | 1.1K | 29K |
| # Socio-affective spans | # Cognitive spans | ||
| 3101 | 1793 |
| Mean of f (SD) | Median of f (IQR) | |
| Socio-affective engagement categories | ||
| Positive sentiment | 3.63 (5.12) | 2 (4.25) |
| Negative sentiment | 0.35 (0.99) | 0 (0) |
| Appreciation | 3.96 (4.27) | 3 (5.25) |
| Cancer-related experience | 2.76 (2.97) | 2 (3.25) |
| Personal information | 4.38 (3.84) | 3 (4) |
| Vocatives | 8.6 (9.11) | 6.5(10.5) |
| Interest in communication | 2.24 (2.73) | 1 (3) |
| Greetings | 8.81(7.48) | 7 (8) |
| Cognitive engagement categories | ||
| Intervention information | 12.41 (9.36) | 10 (12.25) |
| Acknowledgement | 0.74 (1.19) | 0 (1) |
| Information-seeking question | 2.26 (3) | 1 (3) |
| Clarification question | 0.21 (0.59) | 0 (0) |
| Agreement | 3.38 (3.9) | 2 (6) |
| Disagreement | 0.19 (0.4) | 0 (0) |
| Initiative taking | 0.22 (0.59) | 0 (0) |
| Engagement category | Freq-ρ (p) |
| Socio-affective engagement category | |
| Positive sentiment | .35 (.008) |
| Appreciation | .25 (.007) |
| Cancer-related experience | .34 (.01) |
| Personal information | .34 (.01) |
| Vocatives | .33 (.01) |
| Interest in communication | .32 (.01) |
| Greetings | .27 (.04) |
| Cognitive engagement category | |
| Intervention information | .38 (.003) |
| Information-seeking question | .26 (.05) |
| Agreement | .34 (.009) |
| Initiative taking | .26 (.047) |
| Socio-Affective | |||
| Model | P | R | F1 |
| SVM-W | 81.1 (0.0) | 39.8 (0.0) | 40.7 (0.0) |
| SVM-C | 84.7 (0.0) | 59.7 (0.0) | 65.6 (0.0) |
| Bio-BERT | 75.4 (2.1) | 78.0 (1.1) | 76.4 (1.7) |
| BERT | 75.1 (1.1) | 76.6 (0.5) | 75.6 (0.6) |
| XLNet | 77.4 (1.9) | 78.7 (1.6) | 77.7 (1.4) |
| RoBerta | 78.6 (1.3) | 79.6 (1.2) | 78.8 (1.0) |
| Cognitive | |||
| Model | P | R | F1 |
| SVM-W | 62.1 (0.0) | 58.9 (0.0) | 57.6 (0.0) |
| SVM-C | 72.1 (0.0) | 73.5 (0.0) | 72.5 (0.0) |
| Bio-BERT | 78.0 (2.1) | 79.3 (0.4) | 78.5 (1.3) |
| BERT | 77.5 (0.9) | 78.3 (0.6) | 77.6 (0.6) |
| XLNet | 61.9 (0.8) | 63.9 (2.3) | 62.8 (0.7) |
| RoBerta | 61.5 (2.3) | 66.9 (0.2) | 63.8 (1.3) |
| Step | Intervention element | Description |
| 1 | The nurse initiates the message board by introducing herself and explaining how the intervention works on the message boards. The nurse then invited the patient to talk about herself and her experience with ovarian cancer and symptoms in general. | |
| 2 | Representational assessment | The nurse proceeds to ask a series of protocolized questions in the first few messages. This is for both the patient and the nurse to get a full understanding of the patient's experience with each symptom. As the patient thinks about the questions and writes responses, she may find that she sees some things more clearly or perhaps differently. A few examples of the protocolized questions include: \n· What does your fatigue feel like and how severe is it? \n· How has this symptom affected your life? Are you unable to do anything because of it? \n· How does your fatigue affect you emotionally? |
| 3 | Identifying and exploring Gaps, Errors, and Confusions | Throughout interactions with the participant, the nurse keeps attending to evidence of any confusion, concerns, or misconceptions about symptom management that the patient states or suggests during her message board posts. Any of these constitutes a barrier to effective symptom management for participants. |
| 4 | Creating conditions for conceptual change | If any concerns/ misconceptions/gaps are identified, the nurse will (1) discuss the relationship between identified concerns/confusion/misconceptions and consequences of poor symptoms that the patient talked about during representational assessment and (2) provide information to address/counter patient's concerns, gaps, or misconceptions. |
| 5 | Introducing replacement information | The nurse will then guide the patient to read the relevant clinical practice guidelines and provide relevant information to the patient to choose and use different strategies that fit well into their life and needs for managing symptoms. |
| 6 | Summary | The nurse will summarize information and the ways she believes it can benefit the patient. For example, increase comfort and less interference with life. |
| 7 | Goal Setting and Planning | The nurse will encourage the patient to work together on developing symptom goals and specific strategies to reach those goals. |
| 8 | Goal and Strategy Review | After 2 weeks, the nurse will ask protocolized questions to work with the patient to evaluate strategies, reinforce success, and outline modifications as needed. The nurse will also discuss any barriers that occur in the implementation of strategies and work with the patient to identify new or different strategies that could be integrated into her life. A few examples of the protocolized questions include: \n· Were you able to use the above strategies? \n· If not, what things prevented you from doing so? \n· If yes, how well did the strategies (name/list strategies) work in helping you reach your goal? |
| Socio-affective engagement | Description | Example |
| Positive sentiment | 1. Participant explicitly expresses positive emotions or feedback on the intervention design, symptom management progress (e.g., the effectiveness of strategies).2. *Very emotional greetings and goodbyes can fall under this code provided there is a clearly expressed strong emotional connection (intensity), for example \"Oh. It's so nice to meet you!\" | 1. Participant: \"Other than that, I feel that all the strategies are good ones and THEY WORK!!\"2. Participant: \"I feel relieved now that I know many other people have similar problems. That makes me feel normal \\(⑤\\) \"3. Participant: \"I will look forward to your advice on the issues!\" |
| Negative sentiment | Participant explicitly expresses negative emotions or feedback on the intervention design or content, the symptom management progress (e.g., the effectiveness of strategies). | Participant: \"We have been having computer problems. After writing part of a draft yesterday and half of one today, we contacted Jim from your group for tech support. Both messages to you were spontaneously deleted!!!!!!!\" |
| Appreciation | Participant explicitly appreciate and recognize the nurse's contributions and support. | Participant: \"I really appreciate the time you take to discuss on my symptoms. You have much insight and your recommendations are very helpful.\" |
| Cancer-related experience | 1. Participant shows her vulnerability (e.g., fear, guilt, traumatic cancer experience).2. Participant opens to the nurse about her feelings and beliefs of cancer trajectory and symptom management.3. Participant discusses with nurse about the cancer-related issues that she finds hard to discuss with others. | 1. Participant: \"I have caught myself cycling through emotional fatigue times where I'm completely frustrated with myself and life and want to give up and then have some good news or an outing and everything is back to fine again.\"2. Participant: \"Honestly I feel like a prisoner to it. I have to plan for it weekly, bug the nurse on my HCT to get me appointments, I feel it constantly everyday, I can't wear some of my clothes anymore or I fear I can't get an appointment and hit a weekend and have to go 3 extra days is misery.\"3. Participant: \"It's really hard for me to tell my family about my pain. My husband doesn't understand how much pain I have.\" |
| Personal information | Participant shares personal information with the nurse (e.g., family, friends, pets, health care team, religion) | Participant: \"Yea. My daughter and son-in-law came to visit. We had a great Thanksgiving dinner. Thanks for asking.\" |
| Interest in communication | 1. Participant invites nurse to ask for more information or clarification if needed.2. Participant explicitly expresses interest in further communication in general. | 1. Participant: \"Please let me know if you need any other information.\"2. Participant: \"I will update you on the treatment next Wednesday.\" |
| Vocatives | Participant addresses the study name by the name. | \"Names of study nurses\" |
| Greetings | 1. Participant starts posts with a greeting (e.g., Dear, Hello)\n2. Participant signs off the email (e.g., yours, ..., sincerely, ...., “thanks, ...) \n3. Participant shows benevolence, kindness, and social respect to the nurse. | 1. Participant: “Good morning.”\n2. Participant: “Take care, [name]”\n3. Participant: “May He bless your service in helping women to survive with less pain and life to gain.” |
| Cognitive engagement | Description | Example |
| Intervention information | 1. Participant answers nurse questions seeking information to complete intervention activities.\n2. Participant answers nurse questions seeking clarification about something participant said. The goal is to reduce misunderstanding.\n3. Participant reports completing intervention tasks assigned/suggested by the nurse.\n4. Participant provides information or updates unprompted (not asked by the nurse). For example, proposing new strategies, technical issues of the website.\n5. Participant updates the nurse about her cancer treatment that impact her symptom management. | 1. Nurse: “How does it affect you emotionally?”\nParticipant: “I probably get cranky when I am tired, but my dear husband doesn’t complain.”\n2. Nurse: “What do you mean you don’t like to take pills?” (Participant mentioned that she doesn’t like taking pills for symptoms).\nParticipant: “I have to take many prescribed medications. I don’t want to take more that I have to”.\n3. Participant: “I did read the guide you suggested. I guess I can start walking around the block”.\n4. Participant: “I just went to see the dentist and walk with Jim after that.” (Context: seeing a dentist is a strategy participant chose to manage her mouth sore).\n5. Participant: “I was completely drained after chemo last Wed. I didn’t have energy to do anything.” |
| Clarification question | Participant asks nurse questions seeking clarification about something the nurse said. The goal is to reduce misunderstanding. | 1. Participant: “my nutritionist told me I should eat more fruits and vegetable. But the guide says I should not have raw fruits/vegetables if I have diarrhea. Can I still have salads sometimes?”\n2. Participant: “taking pain meds won’t cause addiction, even matter if I take it every day?” |
| Information-seeking question | 1. Participant proposes a question to seek for answers. The specific indicators contain “?”, “who”, “what”, “where”, “when”, “why”, “how”, “how-much/many” or other key question symbols.\n2. Participant seeks for advice/suggestions.\n3. Participant asks the nurse for a favor. | 1. Participant: “I checked out the diet website. Do you know what’s the regimen?”\n2. Participant: “I look forward to your answer to my questions about your list.”\n3. Participant: “It would be wonderful if you could add the goals and strategies as I’m getting ready to go out of town and am really busy preparing for that and trying to pace myself.” |
| Acknowledgement | Participant acknowledges/affirms the nurse suggestions, ideas, summary, questions. | Participant: “I did read your e-mail last night before going to bed. I didn’t answer because I was tired, but I did keep thinking about your questions.” |
| Agreement | 1. Participant explicitly shares the same opinion with the nurse on the topics (e.g., suggestions on symptom management strategies, barriers to good symptom management, reading resources), including restating or paraphrasing or summarizing what the nurse has said. The specific indicators contain “I agree”, “I think so”, etc.\n2. Participant agrees to the plans or do what the nurse suggests doing, including stating that she will try what the nurse has said (being open to different/new experience). The specific indicators contain “I will...”, “Okay”, etc.\n3. Participant approves what the nurse does or suggests. | 1. Participant: “In just the few days of doing yoga I have become aware that my posture is not good. I think maybe that is affecting the way I walk as you suggested when you asked how I walk.”\n2. Nurse: “I was thinking what might make the most sense is to do your 2-week follow-up for Sleep before moving on to headaches. How’s that sound to you? I will have the Sleep Follow-Up posted up there (on web page) under the Sleep topic heading.”\n3. Participant: “okay, I will take a look at the Sleep Follow-up.”\n4. Participant: “The plan looks fine. I don’t have anything to add.” |
| Disagreement | 1. Participant explicitly disagrees with nurse on a certain topic. The specific indicators contain \"I disagree\", \"I don't think so\", etc. \n2. Participant disagrees to do what the nurse suggests doing, including explaining why they don't want to or can't try what the nurse suggested. The specific indicators contain \"I don't...\", \"doesn't work\", etc. | 1. Participant: \"What do you mean? I don't think this is helping me. I was hoping that the program can help me solve the problem.\" \n2. Nurse: \"here are other similar activities that you might also like, such as Tai Chi. Many community centers offer both of these. Look for classes for people over 50... they often start out a little slower and easier. Don't push too hard or fast when you start out.\" \nParticipant: \"I tried Tai Chi two years ago at the senior center. My sister does it and recommended it. I did not get the feeling of renewal from it that she did. Actually, I did not get much from it at all. I went twice a week for two months.\" |
| Initiative taking | 1. Participant anticipates the later steps of the intervention. For example, when the nurse asks questions about patient's symptom representation (element 1 Representation Assessment), the patient both describes the symptoms AND proceeds to describe the strategies she wants to try during the intervention (element 6 Goal setting and Planning). \n2. Participant changes the course of the intervention by taking initiative in changing symptoms to work with the nurse before the nurse's prompts. \n3. Participant offers suggestions to future study design | 1. Nurse: \"Can you tell me about your hot flashes? You can describe a typical day.\" \nParticipant: \"It usually happen at night. I have to get up several times because it is too hot. In looking at the Symptom Care Guides, there are four strategies I would like to work on: 1. Keep a symptom diary. (I have printed a template. I will keep it on the end table where I sit to watch TV, crochet, read, etc.)2. Yoga. I will look on the TV an see if I can find a yoga class. Vitamin E. I will call and get approval from my HCT.4. Paced Respirations (abdominal breathing).\" \nNurse: \"Wow, you're jumping right into this process! That's wonderful. We are in the process of updating our [Resource Library] for Hot Flashes and will add some Yoga web sites, soon. Before we go too much further, though, I do have a few more questions for you.... Don't I always © This is so we can both get the full picture. Can you tell me a little more about how your hot flashes feel?\" \n2. Participant: \"I think we can start with the sleep disturbance first. I am working on my eating now by myself, so that will take a while to see the affects of that, so let's move on to the next one, shall we.\" \n3. Participant: \"Oh just remembered an idea: I wonder if a message box would be helpful to add to the questionnaires? The patient could type out any comments/explanations in answering the questions. Just wondering in the event of any chemo and/or life changes that occurred during this research time that impacted the current questionnaire.\" |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 | E-P | E-R | E-F1 | |
| w/o Pre-training | Seq2Seq | 28.55 | 22.85 | 15.45 | 25.61 | 11.24 | 16.79 | 10.44 | 12.88 |
| Seq2Seq-Entity | 29.13 | 23.22 | 15.66 | 25.79 | 11.42 | 23.79 | 15.89 | 19.06 | |
| HRED | 31.61 | 25.22 | 17.05 | 24.17 | 9.79 | 15.56 | 10.12 | 12.26 | |
| HRED-Entity | 32.84 | 26.12 | 17.63 | 24.26 | 9.76 | 21.75 | 15.33 | 17.98 | |
| VRBot | 29.69 | 23.90 | 16.34 | 24.69 | 11.23 | 18.67 | 9.72 | 12.78 | |
| w/ Pre-trained LM | GPT-2 | 35.27 | 28.19 | 19.16 | 28.74 | 13.61 | 18.29 | 14.45 | 16.14 |
| GPT-2-Entity | 34.56 | 27.56 | 18.71 | 28.78 | 13.62 | 21.27 | 17.10 | 18.96 | |
| BART | 34.94 | 27.99 | 19.06 | 29.03 | 14.40 | 19.97 | 14.29 | 16.66 | |
| BART-Entity | 34.14 | 27.19 | 18.42 | 28.52 | 13.67 | 23.49 | 16.90 | 19.66 | |
| DFMED | 42.56† | 33.34† | 22.53† | 29.31 | 14.21 | 22.48 | 22.84† | 22.66† | |
| w/o Act Flow | 36.79 | 29.18 | 19.81 | 29.45 | 14.26 | 22.73 | 21.70 | 22.20 | |
| w/o Entity Flow | 42.14 | 32.83 | 21.95 | 29.26 | 13.73 | 16.86 | 22.06 | 19.11 | |
| w/o Interweaving | 42.35 | 33.02 | 22.19 | 29.02 | 14.11 | 22.14 | 20.62 | 21.34 |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 |
| Seq2Seq | 23.52 | 18.56 | 12.13 | 23.56 | 8.67 |
| HRED | 26.75 | 21.08 | 13.91 | 22.93 | 7.80 |
| VRBot | 30.04 | 23.76 | 16.36 | 18.71 | 7.28 |
| GPT-2 | 33.76 | 26.58 | 17.82 | 26.80 | 10.56 |
| BART | 33.62 | 26.43 | 17.64 | 27.91 | 11.43 |
| DFMED | 40.20† | 30.97† | 20.76† | 28.28† | 11.54† |
| w/o Act Flow | 35.47 | 28.11 | 18.78 | 27.97 | 11.45 |
| w/o Entity Flow | 39.14 | 29.92 | 19.73 | 27.17 | 10.47 |
| w/o Interweaving | 39.34 | 30.45 | 20.38 | 28.03 | 11.39 |
| Methods | FLU | KC | EQ |
| BART | 3.82 | 1.86 | 3.06 |
| BART-Entity | 3.85 | 2.03 | 3.35 |
| DFMED | 4.00 | 2.14 | 3.61 |
| Gold | 4.12 | 3.97 | 4.35 |
| Methods | MedDG | KaMed | ||||
| Weighted-F1 | R@20 | B-4 | Weighted-F1 | R@20 | B-4 | |
| DFMED | 62.83 | 56.53 | 22.53 | 55.81 | 52.23 | 20.76 |
| w/o Flow Modeling | 62.09 | 54.74 | 22.11 | 55.16 | 51.31 | 20.21 |
| w/o Interweaving | 62.13 | 54.98 | 22.19 | 55.17 | 51.44 | 20.38 |
| w/o Entity attends to Act | 62.37 | 55.75 | 22.34 | 55.43 | 51.84 | 20.62 |
| w/o Act attends to Entity | 62.43 | 55.91 | 22.40 | 55.62 | 52.10 | 20.67 |
| Dialogue Acts | MedDG | KaMed |
| Inquire | 25.49% | 20.95% |
| Make a diagnosis | 6.72% | 8.64% |
| Prescribe medications | 10.12% | 13.74% |
| State a required medical test | 4.25% | 8.29% |
| Provide daily precautions | 7.51% | 5.29% |
| Inform | 29.91% | 30.04% |
| Chitchat | 15.98% | 13.04% |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 | E-P | E-R | E-F1 | |
| w/o Pre-training | Seq2Seq | 28.55 | 22.85 | 15.45 | 25.61 | 11.24 | 16.79 | 10.44 | 12.88 |
| Seq2Seq-Entity | 29.13 | 23.22 | 15.66 | 25.79 | 11.42 | 23.79 | 15.89 | 19.06 | |
| HRED | 31.61 | 25.22 | 17.05 | 24.17 | 9.79 | 15.56 | 10.12 | 12.26 | |
| HRED-Entity | 32.84 | 26.12 | 17.63 | 24.26 | 9.76 | 21.75 | 15.33 | 17.98 | |
| VRBot | 29.69 | 23.90 | 16.34 | 24.69 | 11.23 | 18.67 | 9.72 | 12.78 | |
| w/ Pre-trained LM | GPT-2 | 35.27 | 28.19 | 19.16 | 28.74 | 13.61 | 18.29 | 14.45 | 16.14 |
| GPT-2-Entity | 34.56 | 27.56 | 18.71 | 28.78 | 13.62 | 21.27 | 17.10 | 18.96 | |
| BART | 34.94 | 27.99 | 19.06 | 29.03 | 14.40 | 19.97 | 14.29 | 16.66 | |
| BART-Entity | 34.14 | 27.19 | 18.42 | 28.52 | 13.67 | 23.49 | 16.90 | 19.66 | |
| DFMED | 42.56† | 33.34† | 22.53† | 29.31 | 14.21 | 22.48 | 22.84† | 22.66† | |
| w/o Act Flow | 36.79 | 29.18 | 19.81 | 29.45 | 14.26 | 22.73 | 21.70 | 22.20 | |
| w/o Entity Flow | 42.14 | 32.83 | 21.95 | 29.26 | 13.73 | 16.86 | 22.06 | 19.11 | |
| w/o Interweaving | 42.35 | 33.02 | 22.19 | 29.02 | 14.11 | 22.14 | 20.62 | 21.34 |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 |
| Seq2Seq | 23.52 | 18.56 | 12.13 | 23.56 | 8.67 |
| HRED | 26.75 | 21.08 | 13.91 | 22.93 | 7.80 |
| VRBot | 30.04 | 23.76 | 16.36 | 18.71 | 7.28 |
| GPT-2 | 33.76 | 26.58 | 17.82 | 26.80 | 10.56 |
| BART | 33.62 | 26.43 | 17.64 | 27.91 | 11.43 |
| DFMED | 40.20† | 30.97† | 20.76† | 28.28† | 11.54† |
| w/o Act Flow | 35.47 | 28.11 | 18.78 | 27.97 | 11.45 |
| w/o Entity Flow | 39.14 | 29.92 | 19.73 | 27.17 | 10.47 |
| w/o Interweaving | 39.34 | 30.45 | 20.38 | 28.03 | 11.39 |
| Methods | FLU | KC | EQ |
| BART | 3.82 | 1.86 | 3.06 |
| BART-Entity | 3.85 | 2.03 | 3.35 |
| DFMED | 4.00 | 2.14 | 3.61 |
| Gold | 4.12 | 3.97 | 4.35 |
| Methods | MedDG | KaMed | ||||
| Weighted-F1 | R@20 | B-4 | Weighted-F1 | R@20 | B-4 | |
| DFMED | 62.83 | 56.53 | 22.53 | 55.81 | 52.23 | 20.76 |
| w/o Flow Modeling | 62.09 | 54.74 | 22.11 | 55.16 | 51.31 | 20.21 |
| w/o Interweaving | 62.13 | 54.98 | 22.19 | 55.17 | 51.44 | 20.38 |
| w/o Entity attends to Act | 62.37 | 55.75 | 22.34 | 55.43 | 51.84 | 20.62 |
| w/o Act attends to Entity | 62.43 | 55.91 | 22.40 | 55.62 | 52.10 | 20.67 |
| Dialogue Acts | MedDG | KaMed |
| Inquire | 25.49% | 20.95% |
| Make a diagnosis | 6.72% | 8.64% |
| Prescribe medications | 10.12% | 13.74% |
| State a required medical test | 4.25% | 8.29% |
| Provide daily precautions | 7.51% | 5.29% |
| Inform | 29.91% | 30.04% |
| Chitchat | 15.98% | 13.04% |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 | E-P | E-R | E-F1 | |
| w/o Pre-training | Seq2Seq | 28.55 | 22.85 | 15.45 | 25.61 | 11.24 | 16.79 | 10.44 | 12.88 |
| Seq2Seq-Entity | 29.13 | 23.22 | 15.66 | 25.79 | 11.42 | 23.79 | 15.89 | 19.06 | |
| HRED | 31.61 | 25.22 | 17.05 | 24.17 | 9.79 | 15.56 | 10.12 | 12.26 | |
| HRED-Entity | 32.84 | 26.12 | 17.63 | 24.26 | 9.76 | 21.75 | 15.33 | 17.98 | |
| VRBot | 29.69 | 23.90 | 16.34 | 24.69 | 11.23 | 18.67 | 9.72 | 12.78 | |
| w/ Pre-trained LM | GPT-2 | 35.27 | 28.19 | 19.16 | 28.74 | 13.61 | 18.29 | 14.45 | 16.14 |
| GPT-2-Entity | 34.56 | 27.56 | 18.71 | 28.78 | 13.62 | 21.27 | 17.10 | 18.96 | |
| BART | 34.94 | 27.99 | 19.06 | 29.03 | 14.40 | 19.97 | 14.29 | 16.66 | |
| BART-Entity | 34.14 | 27.19 | 18.42 | 28.52 | 13.67 | 23.49 | 16.90 | 19.66 | |
| DFMED | 42.56† | 33.34† | 22.53† | 29.31 | 14.21 | 22.48 | 22.84† | 22.66† | |
| w/o Act Flow | 36.79 | 29.18 | 19.81 | 29.45 | 14.26 | 22.73 | 21.70 | 22.20 | |
| w/o Entity Flow | 42.14 | 32.83 | 21.95 | 29.26 | 13.73 | 16.86 | 22.06 | 19.11 | |
| w/o Interweaving | 42.35 | 33.02 | 22.19 | 29.02 | 14.11 | 22.14 | 20.62 | 21.34 | |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 |
| Seq2Seq | 23.52 | 18.56 | 12.13 | 23.56 | 8.67 |
| HRED | 26.75 | 21.08 | 13.91 | 22.93 | 7.80 |
| VRBot | 30.04 | 23.76 | 16.36 | 18.71 | 7.28 |
| GPT-2 | 33.76 | 26.58 | 17.82 | 26.80 | 10.56 |
| BART | 33.62 | 26.43 | 17.64 | 27.91 | 11.43 |
| DFMED | 40.20† | 30.97† | 20.76† | 28.28† | 11.54† |
| w/o Act Flow | 35.47 | 28.11 | 18.78 | 27.97 | 11.45 |
| w/o Entity Flow | 39.14 | 29.92 | 19.73 | 27.17 | 10.47 |
| w/o Interweaving | 39.34 | 30.45 | 20.38 | 28.03 | 11.39 |
| Methods | FLU | KC | EQ |
| BART | 3.82 | 1.86 | 3.06 |
| BART-Entity | 3.85 | 2.03 | 3.35 |
| DFMED | 4.00 | 2.14 | 3.61 |
| Gold | 4.12 | 3.97 | 4.35 |
| Methods | MedDG | KaMed | ||||
| Weighted-F1 | R@20 | B-4 | Weighted-F1 | R@20 | B-4 | |
| DFMED | 62.83 | 56.53 | 22.53 | 55.81 | 52.23 | 20.76 |
| w/o Flow Modeling | 62.09 | 54.74 | 22.11 | 55.16 | 51.31 | 20.21 |
| w/o Interweaving | 62.13 | 54.98 | 22.19 | 55.17 | 51.44 | 20.38 |
| w/o Entity attends to Act | 62.37 | 55.75 | 22.34 | 55.43 | 51.84 | 20.62 |
| w/o Act attends to Entity | 62.43 | 55.91 | 22.40 | 55.62 | 52.10 | 20.67 |
| Dialogue Acts | MedDG | KaMed |
| Inquire | 25.49% | 20.95% |
| Make a diagnosis | 6.72% | 8.64% |
| Prescribe medications | 10.12% | 13.74% |
| State a required medical test | 4.25% | 8.29% |
| Provide daily precautions | 7.51% | 5.29% |
| Inform | 29.91% | 30.04% |
| Chitchat | 15.98% | 13.04% |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 | E-P | E-R | E-F1 | |
| w/o Pre-training | Seq2Seq | 28.55 | 22.85 | 15.45 | 25.61 | 11.24 | 16.79 | 10.44 | 12.88 |
| Seq2Seq-Entity | 29.13 | 23.22 | 15.66 | 25.79 | 11.42 | 23.79 | 15.89 | 19.06 | |
| HRED | 31.61 | 25.22 | 17.05 | 24.17 | 9.79 | 15.56 | 10.12 | 12.26 | |
| HRED-Entity | 32.84 | 26.12 | 17.63 | 24.26 | 9.76 | 21.75 | 15.33 | 17.98 | |
| VRBot | 29.69 | 23.90 | 16.34 | 24.69 | 11.23 | 18.67 | 9.72 | 12.78 | |
| w/ Pre-trained LM | GPT-2 | 35.27 | 28.19 | 19.16 | 28.74 | 13.61 | 18.29 | 14.45 | 16.14 |
| GPT-2-Entity | 34.56 | 27.56 | 18.71 | 28.78 | 13.62 | 21.27 | 17.10 | 18.96 | |
| BART | 34.94 | 27.99 | 19.06 | 29.03 | 14.40 | 19.97 | 14.29 | 16.66 | |
| BART-Entity | 34.14 | 27.19 | 18.42 | 28.52 | 13.67 | 23.49 | 16.90 | 19.66 | |
| DFMED | 42.56† | 33.34† | 22.53† | 29.31 | 14.21 | 22.48 | 22.84† | 22.66† | |
| w/o Act Flow | 36.79 | 29.18 | 19.81 | 29.45 | 14.26 | 22.73 | 21.70 | 22.20 | |
| w/o Entity Flow | 42.14 | 32.83 | 21.95 | 29.26 | 13.73 | 16.86 | 22.06 | 19.11 | |
| w/o Interweaving | 42.35 | 33.02 | 22.19 | 29.02 | 14.11 | 22.14 | 20.62 | 21.34 |
| Methods | B-1 | B-2 | B-4 | R-1 | R-2 |
| Seq2Seq | 23.52 | 18.56 | 12.13 | 23.56 | 8.67 |
| HRED | 26.75 | 21.08 | 13.91 | 22.93 | 7.80 |
| VRBot | 30.04 | 23.76 | 16.36 | 18.71 | 7.28 |
| GPT-2 | 33.76 | 26.58 | 17.82 | 26.80 | 10.56 |
| BART | 33.62 | 26.43 | 17.64 | 27.91 | 11.43 |
| DFMED | 40.20† | 30.97† | 20.76† | 28.28† | 11.54† |
| w/o Act Flow | 35.47 | 28.11 | 18.78 | 27.97 | 11.45 |
| w/o Entity Flow | 39.14 | 29.92 | 19.73 | 27.17 | 10.47 |
| w/o Interweaving | 39.34 | 30.45 | 20.38 | 28.03 | 11.39 |
| Methods | FLU | KC | EQ |
| BART | 3.82 | 1.86 | 3.06 |
| BART-Entity | 3.85 | 2.03 | 3.35 |
| DFMED | 4.00 | 2.14 | 3.61 |
| Gold | 4.12 | 3.97 | 4.35 |
| Methods | MedDG | KaMed | ||||
| Weighted-F1 | R@20 | B-4 | Weighted-F1 | R@20 | B-4 | |
| DFMED | 62.83 | 56.53 | 22.53 | 55.81 | 52.23 | 20.76 |
| w/o Flow Modeling | 62.09 | 54.74 | 22.11 | 55.16 | 51.31 | 20.21 |
| w/o Interweaving | 62.13 | 54.98 | 22.19 | 55.17 | 51.44 | 20.38 |
| w/o Entity attends to Act | 62.37 | 55.75 | 22.34 | 55.43 | 51.84 | 20.62 |
| w/o Act attends to Entity | 62.43 | 55.91 | 22.40 | 55.62 | 52.10 | 20.67 |
| Dialogue Acts | MedDG | KaMed |
| Inquire | 25.49% | 20.95% |
| Make a diagnosis | 6.72% | 8.64% |
| Prescribe medications | 10.12% | 13.74% |
| State a required medical test | 4.25% | 8.29% |
| Provide daily precautions | 7.51% | 5.29% |
| Inform | 29.91% | 30.04% |
| Chitchat | 15.98% | 13.04% |
| Dataset | #Samples |
| AG News (Other Categories) | 60,000 |
| NewsCatcher | 60,000 |
| AG News Oracle Data | 60,000 |
| Twitter Mental Health | 20,000 |
| Twitter Offensive Language | 25,000 |
| Twitter Oracle Data | 150,000 |
| Wikipedia Oracle Data | 100,000 |
| False Positive Rate | News | Wikipedia | |||||||
| 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | |
| Likelihood Ratio Attacks: | |||||||||
| Base Reference Model | 4.24% | 0.91% | 0.16% | 5.66% | 0.98% | 0.22% | 1.21% | 0.12% | 0.01% |
| Candidate Reference Model 1 | 4.91% | 0.95% | 0.15% | 6.49% | 1.10% | 0.24% | |||
| Candidate Reference Model 2 | 4.76% | 0.92% | 0.15% | 6.61% | 1.19% | 0.25% | |||
| Oracle Reference Model* | 18.90% | 3.76% | 0.16% | 13.90% | 1.59% | 0.28% | 11.70% | 3.70% | 0.12% |
| Reference-Free Attacks: | |||||||||
| LOSS Attack | 3.50% | 0.10% | 0.01% | 2.08% | 0.11% | 0.02% | 1.06% | 0.11% | 0.01% |
| Neighbour Attack (Ours) | 8.29% | 1.73% | 0.29% | 7.35% | 1.43% | 0.28% | 2.32% | 0.27% | 0.10% |
| News | Wiki | ||
| LiRA: | |||
| Base Reference Model | 0.76 | 0.75 | 0.54 |
| Candidate Reference 1 | 0.78 | 0.81 | |
| Candidate Reference 2 | 0.75 | 0.77 | |
| Oracle Reference* | 0.94 | 0.89 | 0.90 |
| Other Attacks: | |||
| LOSS Attack | 0.64 | 0.60 | 0.52 |
| Neighbour Attack | 0.79 | 0.77 | 0.62 |
| #Neighbours | 5 | 10 | 25 | 50 | 100 |
| News: | |||||
| 1% FPR | 2.98% | 4.57% | 6.65% | 8.19% | 8.29% |
| 0.1% FPR | 0.53% | 0.79% | 1.43% | 1.50% | 1.73% |
| 0.01% FPR | 0.05% | 0.07% | 0.18% | 0.23% | 0.29% |
| Twitter: | |||||
| 1% FPR | 3.93% | 4.88% | 6.21% | 6.63% | 7.35% |
| 0.1% FPR | 0.57% | 0.62% | 1.01% | 1.34% | 1.43% |
| 0.01% FPR | 0.05% | 0.07% | 0.10% | 0.23% | 0.28% |
| Wikipedia: | |||||
| 1% FPR | 1.57% | 1.81% | 2.02% | 2.17% | 2.32% |
| 0.1% FPR | 0.16% | 0.21% | 0.23% | 0.26% | 0.27% |
| 0.01% FPR | 0.05% | 0.08% | 0.09% | 0.10% | 0.10% |
| #Word Replacements | 1 | 2 | 3 |
| News: | |||
| 1% FPR | 8.29% | 4.09% | 4.18% |
| 0.1% FPR | 1.73% | 0.85% | 0.94% |
| 0.01% FPR | 0.29% | 0.23% | 0.21% |
| Twitter: | |||
| 1% FPR | 7.35% | 4.86% | 4.37% |
| 0.1% FPR | 1.43% | 0.74% | 0.72% |
| 0.01% FPR | 0.28% | 0.14% | 0.11% |
| Wikipedia: | |||
| 1% FPR | 2.32% | 1.76% | 1.44% |
| 0.1% FPR | 0.27% | 0.23% | 0.17% |
| 0.01% FPR | 0.10% | 0.07% | 0.03% |
| ε = 5 | ε = 10 | ε = ∞ | |
| TPR @ 1% FPR | 1.29% | 1.52% | 8.29% |
| TPR @ 0.1% FPR | 0.09% | 0.13% | 1.73% |
| TPR @ 0.01% FPR | 0.01% | 0.01% | 0.29% |
| Dataset | #Samples |
| AG News (Other Categories) | 60,000 |
| NewsCatcher | 60,000 |
| AG News Oracle Data | 60,000 |
| Twitter Mental Health | 20,000 |
| Twitter Offensive Language | 25,000 |
| Twitter Oracle Data | 150,000 |
| Wikipedia Oracle Data | 100,000 |
| False Positive Rate | News | Wikipedia | |||||||
| 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | |
| Likelihood Ratio Attacks: | |||||||||
| Base Reference Model | 4.24% | 0.91% | 0.16% | 5.66% | 0.98% | 0.22% | 1.21% | 0.12% | 0.01% |
| Candidate Reference Model 1 | 4.91% | 0.95% | 0.15% | 6.49% | 1.10% | 0.24% | |||
| Candidate Reference Model 2 | 4.76% | 0.92% | 0.15% | 6.61% | 1.19% | 0.25% | |||
| Oracle Reference Model* | 18.90% | 3.76% | 0.16% | 13.90% | 1.59% | 0.28% | 11.70% | 3.70% | 0.12% |
| Reference-Free Attacks: | |||||||||
| LOSS Attack | 3.50% | 0.10% | 0.01% | 2.08% | 0.11% | 0.02% | 1.06% | 0.11% | 0.01% |
| Neighbour Attack (Ours) | 8.29% | 1.73% | 0.29% | 7.35% | 1.43% | 0.28% | 2.32% | 0.27% | 0.10% |
| News | Wiki | ||
| LiRA: | |||
| Base Reference Model | 0.76 | 0.75 | 0.54 |
| Candidate Reference 1 | 0.78 | 0.81 | |
| Candidate Reference 2 | 0.75 | 0.77 | |
| Oracle Reference* | 0.94 | 0.89 | 0.90 |
| Other Attacks: | |||
| LOSS Attack | 0.64 | 0.60 | 0.52 |
| Neighbour Attack | 0.79 | 0.77 | 0.62 |
| #Neighbours | 5 | 10 | 25 | 50 | 100 |
| News: | |||||
| 1% FPR | 2.98% | 4.57% | 6.65% | 8.19% | 8.29% |
| 0.1% FPR | 0.53% | 0.79% | 1.43% | 1.50% | 1.73% |
| 0.01% FPR | 0.05% | 0.07% | 0.18% | 0.23% | 0.29% |
| Twitter: | |||||
| 1% FPR | 3.93% | 4.88% | 6.21% | 6.63% | 7.35% |
| 0.1% FPR | 0.57% | 0.62% | 1.01% | 1.34% | 1.43% |
| 0.01% FPR | 0.05% | 0.07% | 0.10% | 0.23% | 0.28% |
| Wikipedia: | |||||
| 1% FPR | 1.57% | 1.81% | 2.02% | 2.17% | 2.32% |
| 0.1% FPR | 0.16% | 0.21% | 0.23% | 0.26% | 0.27% |
| 0.01% FPR | 0.05% | 0.08% | 0.09% | 0.10% | 0.10% |
| #Word Replacements | 1 | 2 | 3 |
| News: | |||
| 1% FPR | 8.29% | 4.09% | 4.18% |
| 0.1% FPR | 1.73% | 0.85% | 0.94% |
| 0.01% FPR | 0.29% | 0.23% | 0.21% |
| Twitter: | |||
| 1% FPR | 7.35% | 4.86% | 4.37% |
| 0.1% FPR | 1.43% | 0.74% | 0.72% |
| 0.01% FPR | 0.28% | 0.14% | 0.11% |
| Wikipedia: | |||
| 1% FPR | 2.32% | 1.76% | 1.44% |
| 0.1% FPR | 0.27% | 0.23% | 0.17% |
| 0.01% FPR | 0.10% | 0.07% | 0.03% |
| ε = 5 | ε = 10 | ε = ∞ | |
| TPR @ 1% FPR | 1.29% | 1.52% | 8.29% |
| TPR @ 0.1% FPR | 0.09% | 0.13% | 1.73% |
| TPR @ 0.01% FPR | 0.01% | 0.01% | 0.29% |
| Dataset | #Samples |
| AG News (Other Categories) | 60,000 |
| NewsCatcher | 60,000 |
| AG News Oracle Data | 60,000 |
| Twitter Mental Health | 20,000 |
| Twitter Offensive Language | 25,000 |
| Twitter Oracle Data | 150,000 |
| Wikipedia Oracle Data | 100,000 |
| False Positive Rate | News | Wikipedia | |||||||
| 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | |
| Likelihood Ratio Attacks: | |||||||||
| Base Reference Model | 4.24% | 0.91% | 0.16% | 5.66% | 0.98% | 0.22% | 1.21% | 0.12% | 0.01% |
| Candidate Reference Model 1 | 4.91% | 0.95% | 0.15% | 6.49% | 1.10% | 0.24% | |||
| Candidate Reference Model 2 | 4.76% | 0.92% | 0.15% | 6.61% | 1.19% | 0.25% | |||
| Oracle Reference Model* | 18.90% | 3.76% | 0.16% | 13.90% | 1.59% | 0.28% | 11.70% | 3.70% | 0.12% |
| Reference-Free Attacks: | |||||||||
| LOSS Attack | 3.50% | 0.10% | 0.01% | 2.08% | 0.11% | 0.02% | 1.06% | 0.11% | 0.01% |
| Neighbour Attack (Ours) | 8.29% | 1.73% | 0.29% | 7.35% | 1.43% | 0.28% | 2.32% | 0.27% | 0.10% |
| News | Wiki | ||
| LiRA: | |||
| Base Reference Model | 0.76 | 0.75 | 0.54 |
| Candidate Reference 1 | 0.78 | 0.81 | |
| Candidate Reference 2 | 0.75 | 0.77 | |
| Oracle Reference* | 0.94 | 0.89 | 0.90 |
| Other Attacks: | |||
| LOSS Attack | 0.64 | 0.60 | 0.52 |
| Neighbour Attack | 0.79 | 0.77 | 0.62 |
| #Neighbours | 5 | 10 | 25 | 50 | 100 |
| News: | |||||
| 1% FPR | 2.98% | 4.57% | 6.65% | 8.19% | 8.29% |
| 0.1% FPR | 0.53% | 0.79% | 1.43% | 1.50% | 1.73% |
| 0.01% FPR | 0.05% | 0.07% | 0.18% | 0.23% | 0.29% |
| Twitter: | |||||
| 1% FPR | 3.93% | 4.88% | 6.21% | 6.63% | 7.35% |
| 0.1% FPR | 0.57% | 0.62% | 1.01% | 1.34% | 1.43% |
| 0.01% FPR | 0.05% | 0.07% | 0.10% | 0.23% | 0.28% |
| Wikipedia: | |||||
| 1% FPR | 1.57% | 1.81% | 2.02% | 2.17% | 2.32% |
| 0.1% FPR | 0.16% | 0.21% | 0.23% | 0.26% | 0.27% |
| 0.01% FPR | 0.05% | 0.08% | 0.09% | 0.10% | 0.10% |
| #Word Replacements | 1 | 2 | 3 |
| News: | |||
| 1% FPR | 8.29% | 4.09% | 4.18% |
| 0.1% FPR | 1.73% | 0.85% | 0.94% |
| 0.01% FPR | 0.29% | 0.23% | 0.21% |
| Twitter: | |||
| 1% FPR | 7.35% | 4.86% | 4.37% |
| 0.1% FPR | 1.43% | 0.74% | 0.72% |
| 0.01% FPR | 0.28% | 0.14% | 0.11% |
| Wikipedia: | |||
| 1% FPR | 2.32% | 1.76% | 1.44% |
| 0.1% FPR | 0.27% | 0.23% | 0.17% |
| 0.01% FPR | 0.10% | 0.07% | 0.03% |
| ε = 5 | ε = 10 | ε = ∞ | |
| TPR @ 1% FPR | 1.29% | 1.52% | 8.29% |
| TPR @ 0.1% FPR | 0.09% | 0.13% | 1.73% |
| TPR @ 0.01% FPR | 0.01% | 0.01% | 0.29% |
| Dataset | #Samples |
| AG News (Other Categories) | 60,000 |
| NewsCatcher | 60,000 |
| AG News Oracle Data | 60,000 |
| Twitter Mental Health | 20,000 |
| Twitter Offensive Language | 25,000 |
| Twitter Oracle Data | 150,000 |
| Wikipedia Oracle Data | 100,000 |
| False Positive Rate | News | Wikipedia | |||||||
| 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | 1% | 0.1% | 0.01% | |
| Likelihood Ratio Attacks: | |||||||||
| Base Reference Model | 4.24% | 0.91% | 0.16% | 5.66% | 0.98% | 0.22% | 1.21% | 0.12% | 0.01% |
| Candidate Reference Model 1 | 4.91% | 0.95% | 0.15% | 6.49% | 1.10% | 0.24% | |||
| Candidate Reference Model 2 | 4.76% | 0.92% | 0.15% | 6.61% | 1.19% | 0.25% | |||
| Oracle Reference Model* | 18.90% | 3.76% | 0.16% | 13.90% | 1.59% | 0.28% | 11.70% | 3.70% | 0.12% |
| Reference-Free Attacks: | |||||||||
| LOSS Attack | 3.50% | 0.10% | 0.01% | 2.08% | 0.11% | 0.02% | 1.06% | 0.11% | 0.01% |
| Neighbour Attack (Ours) | 8.29% | 1.73% | 0.29% | 7.35% | 1.43% | 0.28% | 2.32% | 0.27% | 0.10% |
| News | Wiki | ||
| LiRA: | |||
| Base Reference Model | 0.76 | 0.75 | 0.54 |
| Candidate Reference 1 | 0.78 | 0.81 | |
| Candidate Reference 2 | 0.75 | 0.77 | |
| Oracle Reference* | 0.94 | 0.89 | 0.90 |
| Other Attacks: | |||
| LOSS Attack | 0.64 | 0.60 | 0.52 |
| Neighbour Attack | 0.79 | 0.77 | 0.62 |
| #Neighbours | 5 | 10 | 25 | 50 | 100 |
| News: | |||||
| 1% FPR | 2.98% | 4.57% | 6.65% | 8.19% | 8.29% |
| 0.1% FPR | 0.53% | 0.79% | 1.43% | 1.50% | 1.73% |
| 0.01% FPR | 0.05% | 0.07% | 0.18% | 0.23% | 0.29% |
| Twitter: | |||||
| 1% FPR | 3.93% | 4.88% | 6.21% | 6.63% | 7.35% |
| 0.1% FPR | 0.57% | 0.62% | 1.01% | 1.34% | 1.43% |
| 0.01% FPR | 0.05% | 0.07% | 0.10% | 0.23% | 0.28% |
| Wikipedia: | |||||
| 1% FPR | 1.57% | 1.81% | 2.02% | 2.17% | 2.32% |
| 0.1% FPR | 0.16% | 0.21% | 0.23% | 0.26% | 0.27% |
| 0.01% FPR | 0.05% | 0.08% | 0.09% | 0.10% | 0.10% |
| #Word Replacements | 1 | 2 | 3 |
| News: | |||
| 1% FPR | 8.29% | 4.09% | 4.18% |
| 0.1% FPR | 1.73% | 0.85% | 0.94% |
| 0.01% FPR | 0.29% | 0.23% | 0.21% |
| Twitter: | |||
| 1% FPR | 7.35% | 4.86% | 4.37% |
| 0.1% FPR | 1.43% | 0.74% | 0.72% |
| 0.01% FPR | 0.28% | 0.14% | 0.11% |
| Wikipedia: | |||
| 1% FPR | 2.32% | 1.76% | 1.44% |
| 0.1% FPR | 0.27% | 0.23% | 0.17% |
| 0.01% FPR | 0.10% | 0.07% | 0.03% |
| ε = 5 | ε = 10 | ε = ∞ | |
| TPR @ 1% FPR | 1.29% | 1.52% | 8.29% |
| TPR @ 0.1% FPR | 0.09% | 0.13% | 1.73% |
| TPR @ 0.01% FPR | 0.01% | 0.01% | 0.29% |
| Category | Dataset | Type | #Train | #Test | L | |Y| |
| Extractive QA | SQuAD (Rajpurkar et al., 2016) | Open QA | 86,588 | 10,507 | 10 / 120 | - |
| BioASQ (Tsatsaronis et al., 2015) | Open QA | 24,559 | 1,504 | 10 / 200 | - | |
| QASC (Khot et al., 2020) | Multi-choice QA | 8,134 | 926 | 8 / 18 | - | |
| Knowledge-intensive QA | TriviaQA (Joshi et al., 2017) | Open QA | 61,688 | 7,785 | 13 / 677 | - |
| TextbookQA (Kembhavi et al., 2017) | Open QA | 15,154 | 1,503 | 10 / 581 | - | |
| Classification | TREC (Voorhees and Tice, 2000) | Question class. | 5,452 | 500 | 10 | 6 |
| MRPC (Dolan and Brockett, 2005) | Paraphrase class. | 3,668 | 408 | 22 / 21 | 2 | |
| MNLI (Williams et al., 2018) | NLI | 392,702 | 9,815 | 22 / 11 | 3 | |
| MNLI-mm (ibid.) | NLI | 392,702 | 9,832 | 22 / 11 | 3 | |
| QNLI (Wang et al., 2018) | NLI | 104,743 | 5,463 | 11 / 30 | 3 |
| TREC | MNLI | MNLI-mm | QNLI | MRPC | Avg. | |
| Majority | 18.8 | 32.7 | 33.3 | 49.5 | 81.2 | 43.1 |
| RoBERTa | *88.82.1 | *45.86.4 | *47.86.8 | *60.26.5 | 76.62.5 | 63.8 |
| iPET | *85.04.1 | 71.21.7 | 71.82.6 | *70.36.2 | 70.44.7 | 73.7 |
| LM-BFF | *89.41.7 | 70.71.3 | *72.01.2 | *69.21.9 | *78.13.4 | 75.9 |
| FewshotQA | 91.02.0 | *47.96.3 | *46.15.9 | *61.06.4 | *67.64.8 | 62.7 |
| FewshotQA-m | 92.41.4 | *50.11.0 | *50.62.5 | *71.82.1 | 74.03.7 | 67.8 |
| RAG | *81.12.0 | *62.40.9 | *61.81.2 | *74.91.5 | 70.23.3 | 70.1 |
| RAG-m | *87.81.7 | *70.01.4 | 69.11.4 | *83.21.5 | 74.92.8 | 77.0 |
| Ours | 91.71.3 | 72.91.7 | 69.61.4 | 84.41.8 | 73.42.5 | 78.4 |
| Model | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| FewshotQA | 68.9 | 63.0 | 82.6 | 65.2 | 37.7 |
| FewshotQA-m | 76.6 | 63.4 | 85.9 | 65.9 | 38.2 |
| RAG-m | 80.0 | 62.9 | 88.9 | 66.6 | 27.7 |
| Ours | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Ours (oracle) | 93.5 | 94.2 | 99.1 | 80.7 | 83.2 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 1.8 | 1.5 | 1.2 | 1.8 | 2.3 |
| DPR (Wiki) | 11.5 | 1.8 | 15.7 | 4.9 | 24.3 |
| DPR (PAQ) | 16.9 | 1.5 | 26.1 | 29.3 | 24.0 |
| Contriever | 14.1 | 7.3 | 28.0 | 27.9 | 24.3 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 74.3 | 61.8 | 88.7 | 56.5 | 29.6 |
| DPR (Wiki) | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| DPR (PAQ) | 78.8 | 63.5 | 86.8 | 57.6 | 33.5 |
| Contriever | 81.1 | 62.5 | 88.7 | 58.9 | 32.4 |
| Memory | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| All tasks | 83.5 | 63.2 | 89.2 | 61.4 | 36.8 |
| Semantically similar tasks | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Format name | Format | Example | F1 |
| Options in question, generate letter | question: q? {aA, ..., aH} \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: i | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: D | 15.6 |
| Options in question, generate answer | question: q? {aA, ..., aH} \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 39.4 |
| Options in context, generate answer | question: q? \\n answer: [MASK] \\n context: {aA, ..., aH}. c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? \\n answer: [MASK] \\n context: (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker. A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 82.6 |
| No options, generate answer | question: q? \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 49.8 |
| Category | Dataset | Type | #Train | #Test | L | |Y| |
| Extractive QA | SQuAD (Rajpurkar et al., 2016) | Open QA | 86,588 | 10,507 | 10 / 120 | - |
| BioASQ (Tsatsaronis et al., 2015) | Open QA | 24,559 | 1,504 | 10 / 200 | - | |
| QASC (Khot et al., 2020) | Multi-choice QA | 8,134 | 926 | 8 / 18 | - | |
| Knowledge-intensive QA | TriviaQA (Joshi et al., 2017) | Open QA | 61,688 | 7,785 | 13 / 677 | - |
| TextbookQA (Kembhavi et al., 2017) | Open QA | 15,154 | 1,503 | 10 / 581 | - | |
| Classification | TREC (Voorhees and Tice, 2000) | Question class. | 5,452 | 500 | 10 | 6 |
| MRPC (Dolan and Brockett, 2005) | Paraphrase class. | 3,668 | 408 | 22 / 21 | 2 | |
| MNLI (Williams et al., 2018) | NLI | 392,702 | 9,815 | 22 / 11 | 3 | |
| MNLI-mm (ibid.) | NLI | 392,702 | 9,832 | 22 / 11 | 3 | |
| QNLI (Wang et al., 2018) | NLI | 104,743 | 5,463 | 11 / 30 | 3 |
| TREC | MNLI | MNLI-mm | QNLI | MRPC | Avg. | |
| Majority | 18.8 | 32.7 | 33.3 | 49.5 | 81.2 | 43.1 |
| RoBERTa | *88.82.1 | *45.86.4 | *47.86.8 | *60.26.5 | 76.62.5 | 63.8 |
| iPET | *85.04.1 | 71.21.7 | 71.82.6 | *70.36.2 | 70.44.7 | 73.7 |
| LM-BFF | *89.41.7 | 70.71.3 | *72.01.2 | *69.21.9 | *78.13.4 | 75.9 |
| FewshotQA | 91.02.0 | *47.96.3 | *46.15.9 | *61.06.4 | *67.64.8 | 62.7 |
| FewshotQA-m | 92.41.4 | *50.11.0 | *50.62.5 | *71.82.1 | 74.03.7 | 67.8 |
| RAG | *81.12.0 | *62.40.9 | *61.81.2 | *74.91.5 | 70.23.3 | 70.1 |
| RAG-m | *87.81.7 | *70.01.4 | 69.11.4 | *83.21.5 | 74.92.8 | 77.0 |
| Ours | 91.71.3 | 72.91.7 | 69.61.4 | 84.41.8 | 73.42.5 | 78.4 |
| Model | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| FewshotQA | 68.9 | 63.0 | 82.6 | 65.2 | 37.7 |
| FewshotQA-m | 76.6 | 63.4 | 85.9 | 65.9 | 38.2 |
| RAG-m | 80.0 | 62.9 | 88.9 | 66.6 | 27.7 |
| Ours | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Ours (oracle) | 93.5 | 94.2 | 99.1 | 80.7 | 83.2 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 1.8 | 1.5 | 1.2 | 1.8 | 2.3 |
| DPR (Wiki) | 11.5 | 1.8 | 15.7 | 4.9 | 24.3 |
| DPR (PAQ) | 16.9 | 1.5 | 26.1 | 29.3 | 24.0 |
| Contriever | 14.1 | 7.3 | 28.0 | 27.9 | 24.3 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 74.3 | 61.8 | 88.7 | 56.5 | 29.6 |
| DPR (Wiki) | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| DPR (PAQ) | 78.8 | 63.5 | 86.8 | 57.6 | 33.5 |
| Contriever | 81.1 | 62.5 | 88.7 | 58.9 | 32.4 |
| Memory | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| All tasks | 83.5 | 63.2 | 89.2 | 61.4 | 36.8 |
| Semantically similar tasks | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Format name | Format | Example | F1 |
| Options in question, generate letter | question: q? {aA, ..., aH} \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: i | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: D | 15.6 |
| Options in question, generate answer | question: q? {aA, ..., aH} \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 39.4 |
| Options in context, generate answer | question: q? \\n answer: [MASK] \\n context: {aA, ..., aH}. c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? \\n answer: [MASK] \\n context: (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker. A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 82.6 |
| No options, generate answer | question: q? \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 49.8 |
| Category | Dataset | Type | #Train | #Test | L | |Y| |
| Extractive QA | SQuAD (Rajpurkar et al., 2016) | Open QA | 86,588 | 10,507 | 10 / 120 | - |
| BioASQ (Tsatsaronis et al., 2015) | Open QA | 24,559 | 1,504 | 10 / 200 | - | |
| QASC (Khot et al., 2020) | Multi-choice QA | 8,134 | 926 | 8 / 18 | - | |
| Knowledge-intensive QA | TriviaQA (Joshi et al., 2017) | Open QA | 61,688 | 7,785 | 13 / 677 | - |
| TextbookQA (Kembhavi et al., 2017) | Open QA | 15,154 | 1,503 | 10 / 581 | - | |
| Classification | TREC (Voorhees and Tice, 2000) | Question class. | 5,452 | 500 | 10 | 6 |
| MRPC (Dolan and Brockett, 2005) | Paraphrase class. | 3,668 | 408 | 22 / 21 | 2 | |
| MNLI (Williams et al., 2018) | NLI | 392,702 | 9,815 | 22 / 11 | 3 | |
| MNLI-mm (ibid.) | NLI | 392,702 | 9,832 | 22 / 11 | 3 | |
| QNLI (Wang et al., 2018) | NLI | 104,743 | 5,463 | 11 / 30 | 3 |
| TREC | MNLI | MNLI-mm | QNLI | MRPC | Avg. | |
| Majority | 18.8 | 32.7 | 33.3 | 49.5 | 81.2 | 43.1 |
| RoBERTa | *88.82.1 | *45.86.4 | *47.86.8 | *60.26.5 | 76.62.5 | 63.8 |
| iPET | *85.04.1 | 71.21.7 | 71.82.6 | *70.36.2 | 70.44.7 | 73.7 |
| LM-BFF | *89.41.7 | 70.71.3 | *72.01.2 | *69.21.9 | *78.13.4 | 75.9 |
| FewshotQA | 91.02.0 | *47.96.3 | *46.15.9 | *61.06.4 | *67.64.8 | 62.7 |
| FewshotQA-m | 92.41.4 | *50.11.0 | *50.62.5 | *71.82.1 | 74.03.7 | 67.8 |
| RAG | *81.12.0 | *62.40.9 | *61.81.2 | *74.91.5 | 70.23.3 | 70.1 |
| RAG-m | *87.81.7 | *70.01.4 | 69.11.4 | *83.21.5 | 74.92.8 | 77.0 |
| Ours | 91.71.3 | 72.91.7 | 69.61.4 | 84.41.8 | 73.42.5 | 78.4 |
| Model | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| FewshotQA | 68.9 | 63.0 | 82.6 | 65.2 | 37.7 |
| FewshotQA-m | 76.6 | 63.4 | 85.9 | 65.9 | 38.2 |
| RAG-m | 80.0 | 62.9 | 88.9 | 66.6 | 27.7 |
| Ours | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Ours (oracle) | 93.5 | 94.2 | 99.1 | 80.7 | 83.2 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 1.8 | 1.5 | 1.2 | 1.8 | 2.3 |
| DPR (Wiki) | 11.5 | 1.8 | 15.7 | 4.9 | 24.3 |
| DPR (PAQ) | 16.9 | 1.5 | 26.1 | 29.3 | 24.0 |
| Contriever | 14.1 | 7.3 | 28.0 | 27.9 | 24.3 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 74.3 | 61.8 | 88.7 | 56.5 | 29.6 |
| DPR (Wiki) | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| DPR (PAQ) | 78.8 | 63.5 | 86.8 | 57.6 | 33.5 |
| Contriever | 81.1 | 62.5 | 88.7 | 58.9 | 32.4 |
| Memory | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| All tasks | 83.5 | 63.2 | 89.2 | 61.4 | 36.8 |
| Semantically similar tasks | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Format name | Format | Example | F1 |
| Options in question, generate letter | question: q? {aA, ..., aH} \n answer: [MASK] \n context: c1. c2. ⇒ question: q? \n answer: i | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \n answer: [MASK] \n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \n answer: D | 15.6 |
| Options in question, generate answer | question: q? {aA, ..., aH} \n answer: [MASK] \n context: c1. c2. ⇒ question: q? \n answer: a | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \n answer: [MASK] \n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \n answer: Helps it survive | 39.4 |
| Options in context, generate answer | question: q? \n answer: [MASK] \n context: {aA, ..., aH}. c1. c2. ⇒ question: q? \n answer: a | question: What does sunlight do for a plant? \n answer: [MASK] \n context: (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker. A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \n answer: Helps it survive | 82.6 |
| No options, generate answer | question: q? \n answer: [MASK] \n context: c1. c2. ⇒ question: q? \n answer: a | question: What does sunlight do for a plant? \n answer: [MASK] \n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \n answer: Helps it survive | 49.8 |
| Category | Dataset | Type | #Train | #Test | L | |Y| |
| Extractive QA | SQuAD (Rajpurkar et al., 2016) | Open QA | 86,588 | 10,507 | 10 / 120 | - |
| BioASQ (Tsatsaronis et al., 2015) | Open QA | 24,559 | 1,504 | 10 / 200 | - | |
| QASC (Khot et al., 2020) | Multi-choice QA | 8,134 | 926 | 8 / 18 | - | |
| Knowledge-intensive QA | TriviaQA (Joshi et al., 2017) | Open QA | 61,688 | 7,785 | 13 / 677 | - |
| TextbookQA (Kembhavi et al., 2017) | Open QA | 15,154 | 1,503 | 10 / 581 | - | |
| Classification | TREC (Voorhees and Tice, 2000) | Question class. | 5,452 | 500 | 10 | 6 |
| MRPC (Dolan and Brockett, 2005) | Paraphrase class. | 3,668 | 408 | 22 / 21 | 2 | |
| MNLI (Williams et al., 2018) | NLI | 392,702 | 9,815 | 22 / 11 | 3 | |
| MNLI-mm (ibid.) | NLI | 392,702 | 9,832 | 22 / 11 | 3 | |
| QNLI (Wang et al., 2018) | NLI | 104,743 | 5,463 | 11 / 30 | 3 |
| TREC | MNLI | MNLI-mm | QNLI | MRPC | Avg. | |
| Majority | 18.8 | 32.7 | 33.3 | 49.5 | 81.2 | 43.1 |
| RoBERTa | *88.82.1 | *45.86.4 | *47.86.8 | *60.26.5 | 76.62.5 | 63.8 |
| iPET | *85.04.1 | 71.21.7 | 71.82.6 | *70.36.2 | 70.44.7 | 73.7 |
| LM-BFF | *89.41.7 | 70.71.3 | *72.01.2 | *69.21.9 | *78.13.4 | 75.9 |
| FewshotQA | 91.02.0 | *47.96.3 | *46.15.9 | *61.06.4 | *67.64.8 | 62.7 |
| FewshotQA-m | 92.41.4 | *50.11.0 | *50.62.5 | *71.82.1 | 74.03.7 | 67.8 |
| RAG | *81.12.0 | *62.40.9 | *61.81.2 | *74.91.5 | 70.23.3 | 70.1 |
| RAG-m | *87.81.7 | *70.01.4 | 69.11.4 | *83.21.5 | 74.92.8 | 77.0 |
| Ours | 91.71.3 | 72.91.7 | 69.61.4 | 84.41.8 | 73.42.5 | 78.4 |
| Model | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| FewshotQA | 68.9 | 63.0 | 82.6 | 65.2 | 37.7 |
| FewshotQA-m | 76.6 | 63.4 | 85.9 | 65.9 | 38.2 |
| RAG-m | 80.0 | 62.9 | 88.9 | 66.6 | 27.7 |
| Ours | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Ours (oracle) | 93.5 | 94.2 | 99.1 | 80.7 | 83.2 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 1.8 | 1.5 | 1.2 | 1.8 | 2.3 |
| DPR (Wiki) | 11.5 | 1.8 | 15.7 | 4.9 | 24.3 |
| DPR (PAQ) | 16.9 | 1.5 | 26.1 | 29.3 | 24.0 |
| Contriever | 14.1 | 7.3 | 28.0 | 27.9 | 24.3 |
| Retriever | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| Random | 74.3 | 61.8 | 88.7 | 56.5 | 29.6 |
| DPR (Wiki) | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| DPR (PAQ) | 78.8 | 63.5 | 86.8 | 57.6 | 33.5 |
| Contriever | 81.1 | 62.5 | 88.7 | 58.9 | 32.4 |
| Memory | SQuAD | BioASQ | QASC | TriviaQA | TbQA |
| All tasks | 83.5 | 63.2 | 89.2 | 61.4 | 36.8 |
| Semantically similar tasks | 83.9 | 64.7 | 89.2 | 62.9 | 37.2 |
| Format name | Format | Example | F1 |
| Options in question, generate letter | question: q? {aA, ..., aH} \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: i | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: D | 15.6 |
| Options in question, generate answer | question: q? {aA, ..., aH} \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 39.4 |
| Options in context, generate answer | question: q? \\n answer: [MASK] \\n context: {aA, ..., aH}. c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? \\n answer: [MASK] \\n context: (A) during the day (B) kills it (C) it can be seen (D) Helps it survive (E) Helps it drink water (F) It gets heated up (G) adding heat (H) Makes the color darker. A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 82.6 |
| No options, generate answer | question: q? \\n answer: [MASK] \\n context: c1. c2. ⇒ question: q? \\n answer: a | question: What does sunlight do for a plant? \\n answer: [MASK] \\n context: A plant requires food for survival. All plants require sunlight to make their food. ⇒ question: ... \\n answer: Helps it survive | 49.8 |
| Availability bias score (%) | |||||
| Training size | 10-shot | 100-shot | 1,000-shot | 10,000-shot | 25,296 |
| Negative | 26.3 (2.1) | 77.7 (2.1) | 39.7 (3.9) | 47.0 (3.6) | 52.0 (2.8) |
| Mechanism | 20.0 (0.0) | 20.0 (0.0) | 13.7 (0.5) | 17.3 (1.2) | 16.7 (1.3) |
| Advice | 20.0 (0.0) | 18.3 (1.7) | 8.3 (2.6) | 7.0 (2.4) | 8.3 (2.6) |
| Effect | 33.7 (2.1) | 20.0 (0.0) | 16.3 (2.1) | 12.7 (2.5) | 11.7 (2.4) |
| Int | 20.0 (0.0) | 19.3 (0.5) | 1.3 (0.5) | 10.0 (0.8) | 15.3 (1.2) |
| Test data | # pairs | F1 |
| Paraphrase, including invalid paraphrases | 300 | 44.6 |
| Original sentences of the above | 300 | 10.2 |
| Paraphrase, excluding invalid paraphrases | 208 | 55.7 |
| Original sentences of the above | 208 | 9.0 |
| \\(F_{\\text{FEP}}\\) | #Ori. wrong | #Pp. correct | \\(\\Delta\\) |
| [0.99, 1.00) | 78 | 77 | 98.7 |
| [0.97, 0.99) | 23 | 12 | 52.2 |
| [0.95, 0.97) | 35 | 27 | 77.1 |
| [0.00, 1.00) | 141 | 119 | 84.4 |
| Example ID | Original Sentence | GPT-3 Paraphrase |
| 1 | The concomitant use of @DRUG$ with other @DRUG$ or with other agents which produce dry mouth, constipation, somnolence (drowsiness), and/or other anticholinergic-like effects may increase the frequency and/or severity of such effects. | The simultaneous use of @DRUG$ with other @DRUG$ medications or agents which can cause dry mouth, constipation, drowsiness, and/or other anticholinergic-like side effects may make these effects more common and/or more severe. |
| 2 | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs, such as amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$, when indicated, requires careful monitoring. | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs requires careful monitoring. Drugs that may require special monitoring include amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$. |
| Availability bias score (%) | |||||
| Training size | 10-shot | 100-shot | 1,000-shot | 10,000-shot | 25,296 |
| Negative | 26.3 (2.1) | 77.7 (2.1) | 39.7 (3.9) | 47.0 (3.6) | 52.0 (2.8) |
| Mechanism | 20.0 (0.0) | 20.0 (0.0) | 13.7 (0.5) | 17.3 (1.2) | 16.7 (1.3) |
| Advice | 20.0 (0.0) | 18.3 (1.7) | 8.3 (2.6) | 7.0 (2.4) | 8.3 (2.6) |
| Effect | 33.7 (2.1) | 20.0 (0.0) | 16.3 (2.1) | 12.7 (2.5) | 11.7 (2.4) |
| Int | 20.0 (0.0) | 19.3 (0.5) | 1.3 (0.5) | 10.0 (0.8) | 15.3 (1.2) |
| Test data | # pairs | F1 |
| Paraphrase, including invalid paraphrases | 300 | 44.6 |
| Original sentences of the above | 300 | 10.2 |
| Paraphrase, excluding invalid paraphrases | 208 | 55.7 |
| Original sentences of the above | 208 | 9.0 |
| \\(F_{\\text{FEP}}\\) | #Ori. wrong | #Pp. correct | \\(\\Delta\\) |
| [0.99, 1.00) | 78 | 77 | 98.7 |
| [0.97, 0.99) | 23 | 12 | 52.2 |
| [0.95, 0.97) | 35 | 27 | 77.1 |
| [0.00, 1.00) | 141 | 119 | 84.4 |
| Example ID | Original Sentence | GPT-3 Paraphrase |
| 1 | The concomitant use of @DRUG$ with other @DRUG$ or with other agents which produce dry mouth, constipation, somnolence (drowsiness), and/or other anticholinergic-like effects may increase the frequency and/or severity of such effects. | The simultaneous use of @DRUG$ with other @DRUG$ medications or agents which can cause dry mouth, constipation, drowsiness, and/or other anticholinergic-like side effects may make these effects more common and/or more severe. |
| 2 | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs, such as amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$, when indicated, requires careful monitoring. | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs requires careful monitoring. Drugs that may require special monitoring include amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$. |
| Availability bias score (%) | |||||
| Training size | 10-shot | 100-shot | 1,000-shot | 10,000-shot | 25,296 |
| Negative | 26.3 (2.1) | 77.7 (2.1) | 39.7 (3.9) | 47.0 (3.6) | 52.0 (2.8) |
| Mechanism | 20.0 (0.0) | 20.0 (0.0) | 13.7 (0.5) | 17.3 (1.2) | 16.7 (1.3) |
| Advice | 20.0 (0.0) | 18.3 (1.7) | 8.3 (2.6) | 7.0 (2.4) | 8.3 (2.6) |
| Effect | 33.7 (2.1) | 20.0 (0.0) | 16.3 (2.1) | 12.7 (2.5) | 11.7 (2.4) |
| Int | 20.0 (0.0) | 19.3 (0.5) | 1.3 (0.5) | 10.0 (0.8) | 15.3 (1.2) |
| Test data | # pairs | F1 |
| Paraphrase, including invalid paraphrases | 300 | 44.6 |
| Original sentences of the above | 300 | 10.2 |
| Paraphrase, excluding invalid paraphrases | 208 | 55.7 |
| Original sentences of the above | 208 | 9.0 |
| \(F_{\text{FEP}}\) | #Ori. wrong | #Pp. correct | \(\Delta\) |
| [0.99, 1.00) | 78 | 77 | 98.7 |
| [0.97, 0.99) | 23 | 12 | 52.2 |
| [0.95, 0.97) | 35 | 27 | 77.1 |
| [0.00, 1.00) | 141 | 119 | 84.4 |
| Example ID | Original Sentence | GPT-3 Paraphrase |
| 1 | The concomitant use of @DRUG$ with other @DRUG$ or with other agents which produce dry mouth, constipation, somnolence (drowsiness), and/or other anticholinergic-like effects may increase the frequency and/or severity of such effects. | The simultaneous use of @DRUG$ with other @DRUG$ medications or agents which can cause dry mouth, constipation, drowsiness, and/or other anticholinergic-like side effects may make these effects more common and/or more severe. |
| 2 | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs, such as amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$, when indicated, requires careful monitoring. | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs requires careful monitoring. Drugs that may require special monitoring include amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$. |
| Availability bias score (%) | |||||
| Training size | 10-shot | 100-shot | 1,000-shot | 10,000-shot | 25,296 |
| Negative | 26.3 (2.1) | 77.7 (2.1) | 39.7 (3.9) | 47.0 (3.6) | 52.0 (2.8) |
| Mechanism | 20.0 (0.0) | 20.0 (0.0) | 13.7 (0.5) | 17.3 (1.2) | 16.7 (1.3) |
| Advice | 20.0 (0.0) | 18.3 (1.7) | 8.3 (2.6) | 7.0 (2.4) | 8.3 (2.6) |
| Effect | 33.7 (2.1) | 20.0 (0.0) | 16.3 (2.1) | 12.7 (2.5) | 11.7 (2.4) |
| Int | 20.0 (0.0) | 19.3 (0.5) | 1.3 (0.5) | 10.0 (0.8) | 15.3 (1.2) |
| Test data | # pairs | F1 |
| Paraphrase, including invalid paraphrases | 300 | 44.6 |
| Original sentences of the above | 300 | 10.2 |
| Paraphrase, excluding invalid paraphrases | 208 | 55.7 |
| Original sentences of the above | 208 | 9.0 |
| \\(F_{\\text{FEP}}\\) | #Ori. wrong | #Pp. correct | \\(\\Delta\\) |
| [0.99, 1.00) | 78 | 77 | 98.7 |
| [0.97, 0.99) | 23 | 12 | 52.2 |
| [0.95, 0.97) | 35 | 27 | 77.1 |
| [0.00, 1.00) | 141 | 119 | 84.4 |
| Example ID | Original Sentence | GPT-3 Paraphrase |
| 1 | The concomitant use of @DRUG$ with other @DRUG$ or with other agents which produce dry mouth, constipation, somnolence (drowsiness), and/or other anticholinergic-like effects may increase the frequency and/or severity of such effects. | The simultaneous use of @DRUG$ with other @DRUG$ medications or agents which can cause dry mouth, constipation, drowsiness, and/or other anticholinergic-like side effects may make these effects more common and/or more severe. |
| 2 | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs, such as amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$, when indicated, requires careful monitoring. | Concurrent and/or sequential systemic or topical use of other potentially neurotoxic and/or nephrotoxic drugs requires careful monitoring. Drugs that may require special monitoring include amphotericin B, aminoglycosides, bacitracin, polymyxin B, @DRUG$, viomycin, or @DRUG$. |
| GB | Language Family | Word Order | Syn. Dist. | Phylo. Dist. | |
| ar | 28.0 | Afro-Asiatic: Semitic | SVO/VSO | 0.57 | 1.00 |
| bg | 58.0 | Indo-European: Slavic | SVO | 0.48 | 0.86 |
| de | 67.0 | Indo-European: Germanic | SVO/soV | 0.42 | 0.43 |
| el | 47.0 | Indo-European: Greek | SVO/VSO | 0.52 | 0.83 |
| en | 301.0 | Indo-European: Germanic | SVO | 0.00 | 0.00 |
| es | 54.0 | Indo-European: Romance | SVO | 0.40 | 0.90 |
| fr | 57.0 | Indo-European: Romance | SVO | 0.46 | 0.90 |
| hi | 21.0 | Indo-European: Indic | SOV | 0.59 | 0.90 |
| ru | 279.0 | Indo-European: Slavic | SVO | 0.49 | 0.90 |
| sw | 1.7 | Niger-Congo: Bantu | SVO | 0.57 | 1.00 |
| th | 72.0 | Tai-Kadai: Kam-Tai | SVO | 0.56 | 1.00 |
| tr | 21.0 | Altaic: Turkic | SOV | 0.70 | 1.00 |
| ur | 5.7 | Indo-European: Indic | SOV | 0.67 | 0.90 |
| vi | 138.0 | Austro-Asiatic: Viet-Muong | SVO | 0.57 | 1.00 |
| zh | 47.0 | Sino-Tibetan: Chinese | SVO | 0.57 | 1.00 |
| Train cost | XNLI (acc) | |||||||||||||||||
| EFLOPs | days | ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 70.2 | 78.4 | 76.2 | 76.1 | 79.2 | 78.9 | 65.6 | 72.5 | 68.2 | 70.1 | 67.1 | 60.9 | 72.1 | 72.4 | 72.0 |
| BL.Small | 21.1 | 8.1 | 65.4 | 71.1 | 68.0 | 69.1 | 71.3 | 71.0 | 61.8 | 66.2 | 63.8 | 64.9 | 64.4 | 57.9 | 65.7 | 67.6 | 66.3 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 70.1 | 77.8 | 75.7 | 75.5 | 78.5 | 78.2 | 63.9 | 72.2 | 67.5 | 70.2 | 64.8 | 59.2 | 70.8 | 72.0 | 71.2 |
| MINJOINT | 21.1 | 8.1 | 69.3 | 77.6 | 75.0 | 74.7 | 78.4 | 77.7 | 62.4 | 71.7 | 66.9 | 68.8 | 63.7 | 58.1 | 69.2 | 70.8 | 70.3 |
| Train cost | MLQA (F1) | PAWS-X (acc) | ||||||||||||||
| EFLOPs | days | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |||
| Standard | BL_BASE | 54.1 | 20.9 | 51.2 | 61.0 | 66.6 | 48.5 | 57.3 | 56.8 | 56.9 | 84.7 | 85.8 | 86.0 | 77.3 | 83.4 | |
| BL.Small | 21.1 | 8.1 | 46.0 | 54.6 | 59.8 | 43.1 | 53.2 | 50.8 | 51.2 | 74.2 | 76.4 | 76.6 | 70.8 | 74.5 | ||
| Proposed | MINIPOST | 29.3 | 11.3 | 50.8 | 60.4 | 66.7 | 48.9 | 56.6 | 56.7 | 56.7 | 83.8 | 86.2 | 86.3 | 76.0 | 83.1 | |
| MINJOINT | 21.1 | 8.1 | 50.8 | 60.1 | 66.0 | 46.4 | 57.2 | 55.6 | 56.0 | 84.4 | 85.1 | 86.7 | 77.9 | 83.5 |
| XNLI | MLQA | PAWS-X | |||||||||||||||||||||||||
| ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |
| BL_BASE | 2.5 | 1.2 | 1.1 | 1.5 | 0.8 | 0.8 | 3.2 | 1.8 | 1.4 | 1.8 | 2.3 | 8.3 | 1.0 | 1.7 | 2.1 | 2.6 | 1.1 | 0.8 | 3.4 | 1.2 | 1.8 | 1.8 | 0.7 | 0.6 | 0.6 | 0.7 | 0.6 |
| MINIPOST | 1.7 | 0.9 | 0.8 | 0.9 | 0.4 | 0.4 | 2.5 | 1.3 | 1.1 | 1.3 | 3.0 | 6.5 | 0.8 | 1.1 | 1.6 | 1.7 | 0.8 | 0.5 | 1.7 | 0.9 | 1.2 | 1.1 | 0.3 | 0.3 | 0.3 | 0.4 | 0.3 |
| MINJOINT | 1.0 | 0.5 | 0.5 | 0.6 | 0.3 | 0.3 | 5.9 | 0.6 | 0.6 | 0.7 | 5.3 | 5.4 | 0.6 | 0.8 | 1.6 | 1.0 | 0.6 | 0.4 | 5.3 | 0.5 | 0.9 | 1.5 | 0.2 | 0.2 | 0.2 | 0.2 | 0.2 |
| Layer: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 | 0.5 | 0.7 | 1.0 | 1.1 | 1.4 |
| MLQA | 0.7 | 0.6 | 0.6 | 1.1 | 1.0 | 1.4 | |
| PAWS | 0.2 | 0.2 | 0.4 | 0.5 | 0.7 | 1.0 | |
| ar | XNLI | ∞ | 1.0 | 0.9 | 1.3 | 1.8 | 2.4 |
| MLQA | ∞ | 1.0 | 1.1 | 2.2 | 2.1 | 2.5 | |
| tr | XNLI | ∞ | 5.3 | 1.5 | 1.5 | 1.5 | 1.6 |
| BL_BASE | 86.4 |
| BL.Small | 79.6 |
| MINIJOINT (full) | 86.2 |
| MINIJOINT (mini-model) | 79.2 |
| XQuAD | ||||||||||
| ar | de | el | es | hi | ru | th | tr | vi | zh | |
| BL_BASE | 2.3 | 1.1 | 1.7 | 0.8 | 3.3 | 1.9 | 2.1 | 2.2 | 1.2 | 1.5 |
| MINIPOST | 1.4 | 0.8 | 1.0 | 0.4 | 1.3 | 1.3 | 1.4 | 1.2 | 0.8 | 1.0 |
| MINIJOINT | 0.8 | 0.4 | 0.6 | 0.5 | 3.1 | 0.6 | 0.6 | ∞ | 0.5 | 0.6 |
| Figure: | Fig. 4 | A2(a) | A2(b) | A2(c) | A3(a) | A3(b) | |
| N= | 2 | 66.2 | 51.5 | 83.8 | 54.1 | 58.3 | 46.8 |
| 4 | 69.3 | 55.4 | 84.4 | 59.0 | 63.7 | 53.2 | |
| 6 | 70.3 | 56.2 | 83.7 | 59.8 | 65.6 | 56.0 | |
| 8 | 70.6 | 54.8 | 83.3 | 58.6 | 66.7 | 55.0 | |
| 10 | 71.4 | 55.8 | 82.9 | 58.8 | 67.9 | 53.2 | |
| 12 | 71.1 | 56.0 | 81.5 | 60.3 | 68.5 | 57.4 | |
| BL_Base | 71.2 | 56.1 | 84.7 | 59.6 | 67.1 | 57.2 |
| Train cost | XQuAD (acc) | |||||||||||||
| EFLOPs | days | ar | de | el | es | hi | ru | th | tr | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 52.4 | 69.2 | 70.0 | 74.8 | 53.8 | 68.7 | 57.1 | 57.2 | 65.9 | 53.7 | 62.3 |
| BL.Small | 21.1 | 8.1 | 48.2 | 61.2 | 63.3 | 65.8 | 46.2 | 61.0 | 50.4 | 52.1 | 60.3 | 46.2 | 55.5 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 52.4 | 68.8 | 69.9 | 75.1 | 55.3 | 68.2 | 56.9 | 58.6 | 65.7 | 53.5 | 62.4 |
| MINJOINT | 21.1 | 8.1 | 53.8 | 70.1 | 69.9 | 73.7 | 51.8 | 68.6 | 56.5 | 53.2 | 64.9 | 52.5 | 61.5 |
| Layers: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 (0.2 - 0.4) | 0.5 (0.3 - 0.7) | 0.7 (0.5 - 0.9) | 1.0 (0.6 - 1.2) | 1.1 (0.7 - 1.4) | 1.4 (0.8 - 1.7) |
| XQUAD | 0.4 (0.2 - 0.4) | 0.4 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 0.9 (0.6 - 1.2) | 0.7 (0.7 - 1.4) | 1.3 (0.8 - 1.7) | |
| MLQA | 0.7 (0.6 - 0.8) | 0.6 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 1.1 (0.6 - 1.2) | 1.0 (0.7 - 1.4) | 1.4 (0.8 - 1.7) | |
| PAWS | 0.2 (0.0 - 0.2) | 0.2 (0.2 - 0.3) | 0.4 (0.0 - 0.5) | 0.5 (0.0 - 0.6) | 0.7 (0.0 - 0.7) | 1.0 (0.8 - 1.7) | |
| ar | XNLI | ∞ | 1.0 (0.7 - 1.0) | 0.9 (0.9 - 1.4) | 1.3 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) |
| XQUAD | ∞ | 0.8 (0.7 - 1.0) | 0.9 (0.5 - 0.9) | 1.6 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) | |
| MLQA | ∞ | 1.0 (1.0 - 1.3) | 1.1 (0.9 - 1.4) | 2.2 (1.7 - 2.3) | 2.1 (1.4 - 2.1) | 2.5 (1.7 - 2.5) | |
| tr | XNLI | ∞ | 5.3 (5.2 - 5.5) | 1.5 (1.4 - 1.8) | 1.5 (1.2 - 1.7) | 1.5 (1.4 - 2.1) | 1.6 (0.8 - 1.7) |
| XQUAD | ∞ | ∞ | 1.3 (0.9 - 1.4) | 3.0 (2.9 - 3.5) | ∞ | 2.7 (2.5 - 3.3) |
| XNLI | MLQA | |
| Run #1 | 67.7 | 51.4 |
| Run #2 | 69.3 | 51.1 |
| Run #3 | 68.7 | 52.7 |
| Main run | 69.6 | 49.6 |
| GB | Language Family | Word Order | Syn. Dist. | Phylo. Dist. | |
| ar | 28.0 | Afro-Asiatic: Semitic | SVO/VSO | 0.57 | 1.00 |
| bg | 58.0 | Indo-European: Slavic | SVO | 0.48 | 0.86 |
| de | 67.0 | Indo-European: Germanic | SVO/soV | 0.42 | 0.43 |
| el | 47.0 | Indo-European: Greek | SVO/VSO | 0.52 | 0.83 |
| en | 301.0 | Indo-European: Germanic | SVO | 0.00 | 0.00 |
| es | 54.0 | Indo-European: Romance | SVO | 0.40 | 0.90 |
| fr | 57.0 | Indo-European: Romance | SVO | 0.46 | 0.90 |
| hi | 21.0 | Indo-European: Indic | SOV | 0.59 | 0.90 |
| ru | 279.0 | Indo-European: Slavic | SVO | 0.49 | 0.90 |
| sw | 1.7 | Niger-Congo: Bantu | SVO | 0.57 | 1.00 |
| th | 72.0 | Tai-Kadai: Kam-Tai | SVO | 0.56 | 1.00 |
| tr | 21.0 | Altaic: Turkic | SOV | 0.70 | 1.00 |
| ur | 5.7 | Indo-European: Indic | SOV | 0.67 | 0.90 |
| vi | 138.0 | Austro-Asiatic: Viet-Muong | SVO | 0.57 | 1.00 |
| zh | 47.0 | Sino-Tibetan: Chinese | SVO | 0.57 | 1.00 |
| Train cost | XNLI (acc) | |||||||||||||||||
| EFLOPs | days | ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 70.2 | 78.4 | 76.2 | 76.1 | 79.2 | 78.9 | 65.6 | 72.5 | 68.2 | 70.1 | 67.1 | 60.9 | 72.1 | 72.4 | 72.0 |
| BL.Small | 21.1 | 8.1 | 65.4 | 71.1 | 68.0 | 69.1 | 71.3 | 71.0 | 61.8 | 66.2 | 63.8 | 64.9 | 64.4 | 57.9 | 65.7 | 67.6 | 66.3 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 70.1 | 77.8 | 75.7 | 75.5 | 78.5 | 78.2 | 63.9 | 72.2 | 67.5 | 70.2 | 64.8 | 59.2 | 70.8 | 72.0 | 71.2 |
| MINJOINT | 21.1 | 8.1 | 69.3 | 77.6 | 75.0 | 74.7 | 78.4 | 77.7 | 62.4 | 71.7 | 66.9 | 68.8 | 63.7 | 58.1 | 69.2 | 70.8 | 70.3 |
| Train cost | MLQA (F1) | PAWS-X (acc) | ||||||||||||||
| EFLOPs | days | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |||
| Standard | BL_BASE | 54.1 | 20.9 | 51.2 | 61.0 | 66.6 | 48.5 | 57.3 | 56.8 | 56.9 | 84.7 | 85.8 | 86.0 | 77.3 | 83.4 | |
| BL.Small | 21.1 | 8.1 | 46.0 | 54.6 | 59.8 | 43.1 | 53.2 | 50.8 | 51.2 | 74.2 | 76.4 | 76.6 | 70.8 | 74.5 | ||
| Proposed | MINIPOST | 29.3 | 11.3 | 50.8 | 60.4 | 66.7 | 48.9 | 56.6 | 56.7 | 56.7 | 83.8 | 86.2 | 86.3 | 76.0 | 83.1 | |
| MINJOINT | 21.1 | 8.1 | 50.8 | 60.1 | 66.0 | 46.4 | 57.2 | 55.6 | 56.0 | 84.4 | 85.1 | 86.7 | 77.9 | 83.5 |
| XNLI | MLQA | PAWS-X | |||||||||||||||||||||||||
| ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |
| BL_BASE | 2.5 | 1.2 | 1.1 | 1.5 | 0.8 | 0.8 | 3.2 | 1.8 | 1.4 | 1.8 | 2.3 | 8.3 | 1.0 | 1.7 | 2.1 | 2.6 | 1.1 | 0.8 | 3.4 | 1.2 | 1.8 | 1.8 | 0.7 | 0.6 | 0.6 | 0.7 | 0.6 |
| MINIPOST | 1.7 | 0.9 | 0.8 | 0.9 | 0.4 | 0.4 | 2.5 | 1.3 | 1.1 | 1.3 | 3.0 | 6.5 | 0.8 | 1.1 | 1.6 | 1.7 | 0.8 | 0.5 | 1.7 | 0.9 | 1.2 | 1.1 | 0.3 | 0.3 | 0.3 | 0.4 | 0.3 |
| MINJOINT | 1.0 | 0.5 | 0.5 | 0.6 | 0.3 | 0.3 | 5.9 | 0.6 | 0.6 | 0.7 | 5.3 | 5.4 | 0.6 | 0.8 | 1.6 | 1.0 | 0.6 | 0.4 | 5.3 | 0.5 | 0.9 | 1.5 | 0.2 | 0.2 | 0.2 | 0.2 | 0.2 |
| Layer: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 | 0.5 | 0.7 | 1.0 | 1.1 | 1.4 |
| MLQA | 0.7 | 0.6 | 0.6 | 1.1 | 1.0 | 1.4 | |
| PAWS | 0.2 | 0.2 | 0.4 | 0.5 | 0.7 | 1.0 | |
| ar | XNLI | ∞ | 1.0 | 0.9 | 1.3 | 1.8 | 2.4 |
| MLQA | ∞ | 1.0 | 1.1 | 2.2 | 2.1 | 2.5 | |
| tr | XNLI | ∞ | 5.3 | 1.5 | 1.5 | 1.5 | 1.6 |
| BL_BASE | 86.4 |
| BL.Small | 79.6 |
| MINIJOINT (full) | 86.2 |
| MINIJOINT (mini-model) | 79.2 |
| XQuAD | ||||||||||
| ar | de | el | es | hi | ru | th | tr | vi | zh | |
| BL_BASE | 2.3 | 1.1 | 1.7 | 0.8 | 3.3 | 1.9 | 2.1 | 2.2 | 1.2 | 1.5 |
| MINIPOST | 1.4 | 0.8 | 1.0 | 0.4 | 1.3 | 1.3 | 1.4 | 1.2 | 0.8 | 1.0 |
| MINIJOINT | 0.8 | 0.4 | 0.6 | 0.5 | 3.1 | 0.6 | 0.6 | ∞ | 0.5 | 0.6 |
| Figure: | Fig. 4 | A2(a) | A2(b) | A2(c) | A3(a) | A3(b) | |
| N= | 2 | 66.2 | 51.5 | 83.8 | 54.1 | 58.3 | 46.8 |
| 4 | 69.3 | 55.4 | 84.4 | 59.0 | 63.7 | 53.2 | |
| 6 | 70.3 | 56.2 | 83.7 | 59.8 | 65.6 | 56.0 | |
| 8 | 70.6 | 54.8 | 83.3 | 58.6 | 66.7 | 55.0 | |
| 10 | 71.4 | 55.8 | 82.9 | 58.8 | 67.9 | 53.2 | |
| 12 | 71.1 | 56.0 | 81.5 | 60.3 | 68.5 | 57.4 | |
| BL_Base | 71.2 | 56.1 | 84.7 | 59.6 | 67.1 | 57.2 |
| Train cost | XQuAD (acc) | |||||||||||||
| EFLOPs | days | ar | de | el | es | hi | ru | th | tr | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 52.4 | 69.2 | 70.0 | 74.8 | 53.8 | 68.7 | 57.1 | 57.2 | 65.9 | 53.7 | 62.3 |
| BL.Small | 21.1 | 8.1 | 48.2 | 61.2 | 63.3 | 65.8 | 46.2 | 61.0 | 50.4 | 52.1 | 60.3 | 46.2 | 55.5 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 52.4 | 68.8 | 69.9 | 75.1 | 55.3 | 68.2 | 56.9 | 58.6 | 65.7 | 53.5 | 62.4 |
| MINJOINT | 21.1 | 8.1 | 53.8 | 70.1 | 69.9 | 73.7 | 51.8 | 68.6 | 56.5 | 53.2 | 64.9 | 52.5 | 61.5 |
| Layers: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 (0.2 - 0.4) | 0.5 (0.3 - 0.7) | 0.7 (0.5 - 0.9) | 1.0 (0.6 - 1.2) | 1.1 (0.7 - 1.4) | 1.4 (0.8 - 1.7) |
| XQUAD | 0.4 (0.2 - 0.4) | 0.4 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 0.9 (0.6 - 1.2) | 0.7 (0.7 - 1.4) | 1.3 (0.8 - 1.7) | |
| MLQA | 0.7 (0.6 - 0.8) | 0.6 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 1.1 (0.6 - 1.2) | 1.0 (0.7 - 1.4) | 1.4 (0.8 - 1.7) | |
| PAWS | 0.2 (0.0 - 0.2) | 0.2 (0.2 - 0.3) | 0.4 (0.0 - 0.5) | 0.5 (0.0 - 0.6) | 0.7 (0.0 - 0.7) | 1.0 (0.8 - 1.7) | |
| ar | XNLI | ∞ | 1.0 (0.7 - 1.0) | 0.9 (0.9 - 1.4) | 1.3 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) |
| XQUAD | ∞ | 0.8 (0.7 - 1.0) | 0.9 (0.5 - 0.9) | 1.6 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) | |
| MLQA | ∞ | 1.0 (1.0 - 1.3) | 1.1 (0.9 - 1.4) | 2.2 (1.7 - 2.3) | 2.1 (1.4 - 2.1) | 2.5 (1.7 - 2.5) | |
| tr | XNLI | ∞ | 5.3 (5.2 - 5.5) | 1.5 (1.4 - 1.8) | 1.5 (1.2 - 1.7) | 1.5 (1.4 - 2.1) | 1.6 (0.8 - 1.7) |
| XQUAD | ∞ | ∞ | 1.3 (0.9 - 1.4) | 3.0 (2.9 - 3.5) | ∞ | 2.7 (2.5 - 3.3) |
| XNLI | MLQA | |
| Run #1 | 67.7 | 51.4 |
| Run #2 | 69.3 | 51.1 |
| Run #3 | 68.7 | 52.7 |
| Main run | 69.6 | 49.6 |
| GB | Language Family | Word Order | Syn. Dist. | Phylo. Dist. | |
| ar | 28.0 | Afro-Asiatic: Semitic | SVO/VSO | 0.57 | 1.00 |
| bg | 58.0 | Indo-European: Slavic | SVO | 0.48 | 0.86 |
| de | 67.0 | Indo-European: Germanic | SVO/soV | 0.42 | 0.43 |
| el | 47.0 | Indo-European: Greek | SVO/VSO | 0.52 | 0.83 |
| en | 301.0 | Indo-European: Germanic | SVO | 0.00 | 0.00 |
| es | 54.0 | Indo-European: Romance | SVO | 0.40 | 0.90 |
| fr | 57.0 | Indo-European: Romance | SVO | 0.46 | 0.90 |
| hi | 21.0 | Indo-European: Indic | SOV | 0.59 | 0.90 |
| ru | 279.0 | Indo-European: Slavic | SVO | 0.49 | 0.90 |
| sw | 1.7 | Niger-Congo: Bantu | SVO | 0.57 | 1.00 |
| th | 72.0 | Tai-Kadai: Kam-Tai | SVO | 0.56 | 1.00 |
| tr | 21.0 | Altaic: Turkic | SOV | 0.70 | 1.00 |
| ur | 5.7 | Indo-European: Indic | SOV | 0.67 | 0.90 |
| vi | 138.0 | Austro-Asiatic: Viet-Muong | SVO | 0.57 | 1.00 |
| zh | 47.0 | Sino-Tibetan: Chinese | SVO | 0.57 | 1.00 |
| Train cost | XNLI (acc) | |||||||||||||||||
| EFLOPs | days | ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 70.2 | 78.4 | 76.2 | 76.1 | 79.2 | 78.9 | 65.6 | 72.5 | 68.2 | 70.1 | 67.1 | 60.9 | 72.1 | 72.4 | 72.0 |
| BL.Small | 21.1 | 8.1 | 65.4 | 71.1 | 68.0 | 69.1 | 71.3 | 71.0 | 61.8 | 66.2 | 63.8 | 64.9 | 64.4 | 57.9 | 65.7 | 67.6 | 66.3 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 70.1 | 77.8 | 75.7 | 75.5 | 78.5 | 78.2 | 63.9 | 72.2 | 67.5 | 70.2 | 64.8 | 59.2 | 70.8 | 72.0 | 71.2 |
| MINJOINT | 21.1 | 8.1 | 69.3 | 77.6 | 75.0 | 74.7 | 78.4 | 77.7 | 62.4 | 71.7 | 66.9 | 68.8 | 63.7 | 58.1 | 69.2 | 70.8 | 70.3 | |
| Train cost | MLQA (F1) | PAWS-X (acc) | ||||||||||||||
| EFLOPs | days | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |||
| Standard | BL_BASE | 54.1 | 20.9 | 51.2 | 61.0 | 66.6 | 48.5 | 57.3 | 56.8 | 56.9 | 84.7 | 85.8 | 86.0 | 77.3 | 83.4 | |
| BL.Small | 21.1 | 8.1 | 46.0 | 54.6 | 59.8 | 43.1 | 53.2 | 50.8 | 51.2 | 74.2 | 76.4 | 76.6 | 70.8 | 74.5 | ||
| Proposed | MINIPOST | 29.3 | 11.3 | 50.8 | 60.4 | 66.7 | 48.9 | 56.6 | 56.7 | 56.7 | 83.8 | 86.2 | 86.3 | 76.0 | 83.1 | |
| MINJOINT | 21.1 | 8.1 | 50.8 | 60.1 | 66.0 | 46.4 | 57.2 | 55.6 | 56.0 | 84.4 | 85.1 | 86.7 | 77.9 | 83.5 | ||
| XNLI | MLQA | PAWS-X | |||||||||||||||||||||||||||||
| ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |||||
| BL_BASE | 2.5 | 1.2 | 1.1 | 1.5 | 0.8 | 0.8 | 3.2 | 1.8 | 1.4 | 1.8 | 2.3 | 8.3 | 1.0 | 1.7 | 2.1 | 2.6 | 1.1 | 0.8 | 3.4 | 1.2 | 1.8 | 1.8 | 0.7 | 0.6 | 0.6 | 0.7 | 0.6 | ||||
| MINIPOST | 1.7 | 0.9 | 0.8 | 0.9 | 0.4 | 0.4 | 2.5 | 1.3 | 1.1 | 1.3 | 3.0 | 6.5 | 0.8 | 1.1 | 1.6 | 1.7 | 0.8 | 0.5 | 1.7 | 0.9 | 1.2 | 1.1 | 0.3 | 0.3 | 0.3 | 0.4 | 0.3 | ||||
| MINJOINT | 1.0 | 0.5 | 0.5 | 0.6 | 0.3 | 0.3 | 5.9 | 0.6 | 0.6 | 0.7 | 5.3 | 5.4 | 0.6 | 0.8 | 1.6 | 1.0 | 0.6 | 0.4 | 5.3 | 0.5 | 0.9 | 1.5 | 0.2 | 0.2 | 0.2 | 0.2 | 0.2 | ||||
| Layer: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 | 0.5 | 0.7 | 1.0 | 1.1 | 1.4 |
| MLQA | 0.7 | 0.6 | 0.6 | 1.1 | 1.0 | 1.4 | |
| PAWS | 0.2 | 0.2 | 0.4 | 0.5 | 0.7 | 1.0 | |
| ar | XNLI | ∞ | 1.0 | 0.9 | 1.3 | 1.8 | 2.4 |
| MLQA | ∞ | 1.0 | 1.1 | 2.2 | 2.1 | 2.5 | |
| tr | XNLI | ∞ | 5.3 | 1.5 | 1.5 | 1.5 | 1.6 |
| BL_BASE | 86.4 |
| BL.Small | 79.6 |
| MINIJOINT (full) | 86.2 |
| MINIJOINT (mini-model) | 79.2 |
| XQuAD | ||||||||||
| ar | de | el | es | hi | ru | th | tr | vi | zh | |
| BL_BASE | 2.3 | 1.1 | 1.7 | 0.8 | 3.3 | 1.9 | 2.1 | 2.2 | 1.2 | 1.5 |
| MINIPOST | 1.4 | 0.8 | 1.0 | 0.4 | 1.3 | 1.3 | 1.4 | 1.2 | 0.8 | 1.0 |
| MINIJOINT | 0.8 | 0.4 | 0.6 | 0.5 | 3.1 | 0.6 | 0.6 | ∞ | 0.5 | 0.6 |
| Figure: | Fig. 4 | A2(a) | A2(b) | A2(c) | A3(a) | A3(b) | |
| N= | 2 | 66.2 | 51.5 | 83.8 | 54.1 | 58.3 | 46.8 |
| 4 | 69.3 | 55.4 | 84.4 | 59.0 | 63.7 | 53.2 | |
| 6 | 70.3 | 56.2 | 83.7 | 59.8 | 65.6 | 56.0 | |
| 8 | 70.6 | 54.8 | 83.3 | 58.6 | 66.7 | 55.0 | |
| 10 | 71.4 | 55.8 | 82.9 | 58.8 | 67.9 | 53.2 | |
| 12 | 71.1 | 56.0 | 81.5 | 60.3 | 68.5 | 57.4 | |
| BL_Base | 71.2 | 56.1 | 84.7 | 59.6 | 67.1 | 57.2 |
| Train cost | XQuAD (acc) | |||||||||||||
| EFLOPs | days | ar | de | el | es | hi | ru | th | tr | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 52.4 | 69.2 | 70.0 | 74.8 | 53.8 | 68.7 | 57.1 | 57.2 | 65.9 | 53.7 | 62.3 |
| BL.Small | 21.1 | 8.1 | 48.2 | 61.2 | 63.3 | 65.8 | 46.2 | 61.0 | 50.4 | 52.1 | 60.3 | 46.2 | 55.5 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 52.4 | 68.8 | 69.9 | 75.1 | 55.3 | 68.2 | 56.9 | 58.6 | 65.7 | 53.5 | 62.4 |
| MINJOINT | 21.1 | 8.1 | 53.8 | 70.1 | 69.9 | 73.7 | 51.8 | 68.6 | 56.5 | 53.2 | 64.9 | 52.5 | 61.5 | |
| Layers: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 (0.2 - 0.4) | 0.5 (0.3 - 0.7) | 0.7 (0.5 - 0.9) | 1.0 (0.6 - 1.2) | 1.1 (0.7 - 1.4) | 1.4 (0.8 - 1.7) |
| XQUAD | 0.4 (0.2 - 0.4) | 0.4 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 0.9 (0.6 - 1.2) | 0.7 (0.7 - 1.4) | 1.3 (0.8 - 1.7) | |
| MLQA | 0.7 (0.6 - 0.8) | 0.6 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 1.1 (0.6 - 1.2) | 1.0 (0.7 - 1.4) | 1.4 (0.8 - 1.7) | |
| PAWS | 0.2 (0.0 - 0.2) | 0.2 (0.2 - 0.3) | 0.4 (0.0 - 0.5) | 0.5 (0.0 - 0.6) | 0.7 (0.0 - 0.7) | 1.0 (0.8 - 1.7) | |
| ar | XNLI | ∞ | 1.0 (0.7 - 1.0) | 0.9 (0.9 - 1.4) | 1.3 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) |
| XQUAD | ∞ | 0.8 (0.7 - 1.0) | 0.9 (0.5 - 0.9) | 1.6 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) | |
| MLQA | ∞ | 1.0 (1.0 - 1.3) | 1.1 (0.9 - 1.4) | 2.2 (1.7 - 2.3) | 2.1 (1.4 - 2.1) | 2.5 (1.7 - 2.5) | |
| tr | XNLI | ∞ | 5.3 (5.2 - 5.5) | 1.5 (1.4 - 1.8) | 1.5 (1.2 - 1.7) | 1.5 (1.4 - 2.1) | 1.6 (0.8 - 1.7) |
| XQUAD | ∞ | ∞ | 1.3 (0.9 - 1.4) | 3.0 (2.9 - 3.5) | ∞ | 2.7 (2.5 - 3.3) |
| XNLI | MLQA | |
| Run #1 | 67.7 | 51.4 |
| Run #2 | 69.3 | 51.1 |
| Run #3 | 68.7 | 52.7 |
| Main run | 69.6 | 49.6 |
| GB | Language Family | Word Order | Syn. Dist. | Phylo. Dist. | |
| ar | 28.0 | Afro-Asiatic: Semitic | SVO/VSO | 0.57 | 1.00 |
| bg | 58.0 | Indo-European: Slavic | SVO | 0.48 | 0.86 |
| de | 67.0 | Indo-European: Germanic | SVO/soV | 0.42 | 0.43 |
| el | 47.0 | Indo-European: Greek | SVO/VSO | 0.52 | 0.83 |
| en | 301.0 | Indo-European: Germanic | SVO | 0.00 | 0.00 |
| es | 54.0 | Indo-European: Romance | SVO | 0.40 | 0.90 |
| fr | 57.0 | Indo-European: Romance | SVO | 0.46 | 0.90 |
| hi | 21.0 | Indo-European: Indic | SOV | 0.59 | 0.90 |
| ru | 279.0 | Indo-European: Slavic | SVO | 0.49 | 0.90 |
| sw | 1.7 | Niger-Congo: Bantu | SVO | 0.57 | 1.00 |
| th | 72.0 | Tai-Kadai: Kam-Tai | SVO | 0.56 | 1.00 |
| tr | 21.0 | Altaic: Turkic | SOV | 0.70 | 1.00 |
| ur | 5.7 | Indo-European: Indic | SOV | 0.67 | 0.90 |
| vi | 138.0 | Austro-Asiatic: Viet-Muong | SVO | 0.57 | 1.00 |
| zh | 47.0 | Sino-Tibetan: Chinese | SVO | 0.57 | 1.00 |
| Train cost | XNLI (acc) | |||||||||||||||||
| EFLOPs | days | ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 70.2 | 78.4 | 76.2 | 76.1 | 79.2 | 78.9 | 65.6 | 72.5 | 68.2 | 70.1 | 67.1 | 60.9 | 72.1 | 72.4 | 72.0 |
| BL.Small | 21.1 | 8.1 | 65.4 | 71.1 | 68.0 | 69.1 | 71.3 | 71.0 | 61.8 | 66.2 | 63.8 | 64.9 | 64.4 | 57.9 | 65.7 | 67.6 | 66.3 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 70.1 | 77.8 | 75.7 | 75.5 | 78.5 | 78.2 | 63.9 | 72.2 | 67.5 | 70.2 | 64.8 | 59.2 | 70.8 | 72.0 | 71.2 |
| MINJOINT | 21.1 | 8.1 | 69.3 | 77.6 | 75.0 | 74.7 | 78.4 | 77.7 | 62.4 | 71.7 | 66.9 | 68.8 | 63.7 | 58.1 | 69.2 | 70.8 | 70.3 |
| Train cost | MLQA (F1) | PAWS-X (acc) | ||||||||||||||
| EFLOPs | days | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |||
| Standard | BL_BASE | 54.1 | 20.9 | 51.2 | 61.0 | 66.6 | 48.5 | 57.3 | 56.8 | 56.9 | 84.7 | 85.8 | 86.0 | 77.3 | 83.4 | |
| BL.Small | 21.1 | 8.1 | 46.0 | 54.6 | 59.8 | 43.1 | 53.2 | 50.8 | 51.2 | 74.2 | 76.4 | 76.6 | 70.8 | 74.5 | ||
| Proposed | MINIPOST | 29.3 | 11.3 | 50.8 | 60.4 | 66.7 | 48.9 | 56.6 | 56.7 | 56.7 | 83.8 | 86.2 | 86.3 | 76.0 | 83.1 | |
| MINJOINT | 21.1 | 8.1 | 50.8 | 60.1 | 66.0 | 46.4 | 57.2 | 55.6 | 56.0 | 84.4 | 85.1 | 86.7 | 77.9 | 83.5 |
| XNLI | MLQA | PAWS-X | |||||||||||||||||||||||||
| ar | bg | de | el | es | fr | hi | ru | sw | th | tr | ur | vi | zh | avg | ar | de | es | hi | vi | zh | avg | de | fr | es | zh | avg | |
| BL_BASE | 2.5 | 1.2 | 1.1 | 1.5 | 0.8 | 0.8 | 3.2 | 1.8 | 1.4 | 1.8 | 2.3 | 8.3 | 1.0 | 1.7 | 2.1 | 2.6 | 1.1 | 0.8 | 3.4 | 1.2 | 1.8 | 1.8 | 0.7 | 0.6 | 0.6 | 0.7 | 0.6 |
| MINIPOST | 1.7 | 0.9 | 0.8 | 0.9 | 0.4 | 0.4 | 2.5 | 1.3 | 1.1 | 1.3 | 3.0 | 6.5 | 0.8 | 1.1 | 1.6 | 1.7 | 0.8 | 0.5 | 1.7 | 0.9 | 1.2 | 1.1 | 0.3 | 0.3 | 0.3 | 0.4 | 0.3 |
| MINJOINT | 1.0 | 0.5 | 0.5 | 0.6 | 0.3 | 0.3 | 5.9 | 0.6 | 0.6 | 0.7 | 5.3 | 5.4 | 0.6 | 0.8 | 1.6 | 1.0 | 0.6 | 0.4 | 5.3 | 0.5 | 0.9 | 1.5 | 0.2 | 0.2 | 0.2 | 0.2 | 0.2 |
| Layer: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 | 0.5 | 0.7 | 1.0 | 1.1 | 1.4 |
| MLQA | 0.7 | 0.6 | 0.6 | 1.1 | 1.0 | 1.4 | |
| PAWS | 0.2 | 0.2 | 0.4 | 0.5 | 0.7 | 1.0 | |
| ar | XNLI | ∞ | 1.0 | 0.9 | 1.3 | 1.8 | 2.4 |
| MLQA | ∞ | 1.0 | 1.1 | 2.2 | 2.1 | 2.5 | |
| tr | XNLI | ∞ | 5.3 | 1.5 | 1.5 | 1.5 | 1.6 |
| BL_BASE | 86.4 |
| BL.Small | 79.6 |
| MINIJOINT (full) | 86.2 |
| MINIJOINT (mini-model) | 79.2 |
| XQuAD | ||||||||||
| ar | de | el | es | hi | ru | th | tr | vi | zh | |
| BL_BASE | 2.3 | 1.1 | 1.7 | 0.8 | 3.3 | 1.9 | 2.1 | 2.2 | 1.2 | 1.5 |
| MINIPOST | 1.4 | 0.8 | 1.0 | 0.4 | 1.3 | 1.3 | 1.4 | 1.2 | 0.8 | 1.0 |
| MINIJOINT | 0.8 | 0.4 | 0.6 | 0.5 | 3.1 | 0.6 | 0.6 | ∞ | 0.5 | 0.6 |
| Figure: | Fig. 4 | A2(a) | A2(b) | A2(c) | A3(a) | A3(b) | |
| N= | 2 | 66.2 | 51.5 | 83.8 | 54.1 | 58.3 | 46.8 |
| 4 | 69.3 | 55.4 | 84.4 | 59.0 | 63.7 | 53.2 | |
| 6 | 70.3 | 56.2 | 83.7 | 59.8 | 65.6 | 56.0 | |
| 8 | 70.6 | 54.8 | 83.3 | 58.6 | 66.7 | 55.0 | |
| 10 | 71.4 | 55.8 | 82.9 | 58.8 | 67.9 | 53.2 | |
| 12 | 71.1 | 56.0 | 81.5 | 60.3 | 68.5 | 57.4 | |
| BL_Base | 71.2 | 56.1 | 84.7 | 59.6 | 67.1 | 57.2 |
| Train cost | XQuAD (acc) | |||||||||||||
| EFLOPs | days | ar | de | el | es | hi | ru | th | tr | vi | zh | avg | ||
| Standard | BL_BASE | 54.1 | 20.9 | 52.4 | 69.2 | 70.0 | 74.8 | 53.8 | 68.7 | 57.1 | 57.2 | 65.9 | 53.7 | 62.3 |
| BL.Small | 21.1 | 8.1 | 48.2 | 61.2 | 63.3 | 65.8 | 46.2 | 61.0 | 50.4 | 52.1 | 60.3 | 46.2 | 55.5 | |
| Proposed | MINIPOST | 29.3 | 11.3 | 52.4 | 68.8 | 69.9 | 75.1 | 55.3 | 68.2 | 56.9 | 58.6 | 65.7 | 53.5 | 62.4 |
| MINJOINT | 21.1 | 8.1 | 53.8 | 70.1 | 69.9 | 73.7 | 51.8 | 68.6 | 56.5 | 53.2 | 64.9 | 52.5 | 61.5 |
| Layers: | 2 | 4 | 6 | 8 | 10 | 12 | |
| de | XNLI | 0.4 (0.2 - 0.4) | 0.5 (0.3 - 0.7) | 0.7 (0.5 - 0.9) | 1.0 (0.6 - 1.2) | 1.1 (0.7 - 1.4) | 1.4 (0.8 - 1.7) |
| XQUAD | 0.4 (0.2 - 0.4) | 0.4 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 0.9 (0.6 - 1.2) | 0.7 (0.7 - 1.4) | 1.3 (0.8 - 1.7) | |
| MLQA | 0.7 (0.6 - 0.8) | 0.6 (0.3 - 0.7) | 0.6 (0.5 - 0.9) | 1.1 (0.6 - 1.2) | 1.0 (0.7 - 1.4) | 1.4 (0.8 - 1.7) | |
| PAWS | 0.2 (0.0 - 0.2) | 0.2 (0.2 - 0.3) | 0.4 (0.0 - 0.5) | 0.5 (0.0 - 0.6) | 0.7 (0.0 - 0.7) | 1.0 (0.8 - 1.7) | |
| ar | XNLI | ∞ | 1.0 (0.7 - 1.0) | 0.9 (0.9 - 1.4) | 1.3 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) |
| XQUAD | ∞ | 0.8 (0.7 - 1.0) | 0.9 (0.5 - 0.9) | 1.6 (1.2 - 1.7) | 1.8 (1.4 - 2.1) | 2.4 (1.7 - 2.5) | |
| MLQA | ∞ | 1.0 (1.0 - 1.3) | 1.1 (0.9 - 1.4) | 2.2 (1.7 - 2.3) | 2.1 (1.4 - 2.1) | 2.5 (1.7 - 2.5) | |
| tr | XNLI | ∞ | 5.3 (5.2 - 5.5) | 1.5 (1.4 - 1.8) | 1.5 (1.2 - 1.7) | 1.5 (1.4 - 2.1) | 1.6 (0.8 - 1.7) |
| XQUAD | ∞ | ∞ | 1.3 (0.9 - 1.4) | 3.0 (2.9 - 3.5) | ∞ | 2.7 (2.5 - 3.3) |
| XNLI | MLQA | |
| Run #1 | 67.7 | 51.4 |
| Run #2 | 69.3 | 51.1 |
| Run #3 | 68.7 | 52.7 |
| Main run | 69.6 | 49.6 |
| Metric | Acquisition Method | Superior |
| Acc | Statistics | ↑ |
| ASR | Statistics | ↑ |
| Dis | Levenshtein | ↓ |
| Flu | GPT2LM | ↓ |
| Sem | USE | ↑ |
| WMR | Statistics | ↓ |
| Err | Ginger | ↓ |
| Dataset | Victim Model | ||
| Bert | BertEntity | PCNN | |
| Wiki80 | 86.21% | 87.09% | 67.14% |
| TACRED | 87.57% | 88.45% | 80.64% |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 38.11% | 23.86 | 628.49 | 84.58% | 43.45% | 5.20 | 19.78% | 32.64 | 587.08 | 86.16% | 36.26% | 5.77 |
| TB | 34.00% | 20.00 | 938.21 | 75.95% | 78.52% | 8.55 | 17.40% | 28.45 | 810.19 | 78.95% | 77.42% | 10.90 |
| Gen | 14.51% | 20.10 | 504.52 | 89.75% | 37.68% | 5.01 | 8.11% | 25.23 | 344.63 | 91.92% | 28.82% | 5.57 |
| HF | 13.15% | 34.10 | 927.54 | 82.71% | 47.39% | 5.05 | 4.08% | 36.89 | 556.78 | 86.90% | 33.60% | 5.40 |
| NV | 30.83% | 17.66 | 339.16 | 91.19% | 12.06% | 0.57 | 15.65% | 24.40 | 282.95 | 91.33% | 13.18% | 0.74 |
| NVR | 31.14% | 18.74 | 353.37 | 91.05% | 13.05% | 0.60 | 15.73% | 25.47 | 315.08 | 91.24% | 13.92% | 0.79 |
| NVJ | 32.31% | 19.82 | 357.27 | 90.36% | 14.39% | 0.60 | 15.92% | 26.77 | 294.09 | 90.61% | 16.26% | 0.77 |
| NVJR | 32.57% | 20.89 | 372.01 | 90.23% | 15.42% | 0.63 | 16.03% | 28.00 | 325.82 | 90.45% | 17.01% | 0.81 |
| Dataset | Method | |||
| Gen | HF | PWWS | TB | |
| Wiki80 | NOUN: 51.7% | NOUN: 35.9% | NOUN: 61.5% | NOUN: 55.7% |
| VERB: 26.6% | PREP: 21.5% | VERB: 17.7% | VERB: 18.6% | |
| ADJ: 15.0% | VERB: 20.8% | ADJ: 17.6% | ADJ: 16.8% | |
| ADV: 4.0% | ADJ: 10.7% | ADV: 1.9% | ADV: 3.0% | |
| TACRED | NOUN: 53.2% | NOUN: 40.7% | NOUN: 58.4% | NOUN: 55.8% |
| VERB: 28.2% | VERB: 19.3% | VERB: 19.8% | VERB: 20.6% | |
| ADJ: 12.1% | PREP: 16.6% | ADJ: 17.4% | ADJ: 15.8% | |
| ADV: 4.6% | ADJ: 11.8% | ADV: 2.5% | ADV: 3.1% |
| Dataset | Victim Model | Dependency Parsing | ||||||
| Similarity | Overlap | Relation | ||||||
| UAS↑ | LAS↑ | Root | Head | Tail | Head | Tail | ||
| Wiki80 | Bert | 93.88% | 92.69% | 67.41% | 56.22% | 53.86% | nmod, nsubj, obl | obl, nmod, compound |
| BertEntity | 94.05% | 92.91% | 70.11% | 57.61% | 58.98% | nmod, nsubj, appos | obl, nmod, compound | |
| PCNN | 94.71% | 93.42% | 67.60% | 51.87% | 55.61% | nsubj, nmod, obl | obl, compound, nmod | |
| TACRED | Bert | 93.12% | 91.64% | 62.40% | 47.69% | 43.79% | nsubj, nmod, nmod:poss | nsubj, obl, nmod |
| BertEntity | 92.90% | 91.45% | 61.76% | 52.24% | 46.59% | nsubj, nmod, nmod:poss | obl, nsubj, nmod | |
| PCNN | 93.16% | 91.84% | 62.18% | 41.97% | 42.49% | nsubj, nmod:poss, nmod | nsubj, amod, nmod |
| Result | Method | ||||
| NV | PWWS | Gen | TB | HF | |
| Average score | 3.52 | 3.05 | 3.03 | 2.75 | 2.30 |
| Score >= 3 | 78% | 67% | 67% | 55% | 35% |
| Dataset | Victim Model (Vα) | Target Model | ||
| Bert | BertEntity | PCNN | ||
| Wiki80 | Bert(1439) | - | 43.71% | 57.61% |
| BertEntity(1392) | 48.35% | - | 57.61% | |
| PCNN(1284) | 14.80% | 14.95% | - | |
| TACRED | Bert(3508) | - | 23.66% | 40.76% |
| BertEntity(2050) | 54.49% | - | 53.56% | |
| PCNN(772) | 15.41% | 15.80% | - |
| Attack Method | Word-Level | Character/Word-Level | |||
| PWWS | Gen | Ours | TB | HF | |
| Perturbation Options | Sub | Sub | Sub | Ins, Del, Swap, Sub-C, Sub-W | Flip, Ins, Del, Sub |
| Importance Ranking | Required | Not Required | Not Required | Required | Not Required |
| Extra Requirements | GloVe vectors, Counter-fitting method, Google 1 billion words language model | / | Pre-trained GloVe model | One-Hot input representation, Word Embedding |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 34.40% | 24.21 | 648.04 | 83.95% | 44.15% | 5.23 | 29.34% | 32.96 | 512.49 | 86.25% | 36.27% | 5.93 |
| TB | 30.10% | 19.87 | 936.60 | 76.57% | 73.98% | 8.20 | 25.97% | 27.56 | 781.14 | 81.75% | 70.14% | 9.86 |
| Gen | 13.94% | 20.70 | 516.27 | 89.62% | 37.97% | 5.10 | 13.50% | 26.97 | 368.46 | 91.42% | 29.48% | 5.77 |
| HF | 12.55% | 35.17 | 975.18 | 81.00% | 48.34% | 5.09 | 5.88% | 38.31 | 567.73 | 86.17% | 34.28% | 5.53 |
| NV | 29.81% | 17.85 | 342.37 | 91.00% | 12.14% | 0.59 | 25.83% | 25.46 | 275.21 | 91.13% | 13.41% | 0.81 |
| NVR | 29.97% | 18.94 | 356.95 | 90.83% | 13.21% | 0.62 | 25.99% | 26.63 | 286.01 | 91.04% | 14.15% | 0.84 |
| NVJ | 31.09% | 20.02 | 362.64 | 90.09% | 14.14% | 0.62 | 26.07% | 27.58 | 285.36 | 90.49% | 16.72% | 0.82 |
| NVJR | 31.19% | 21.07 | 378.48 | 89.93% | 15.26% | 0.65 | 26.27% | 28.82 | 296.72 | 90.38% | 17.49% | 0.85 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 32.79% | 24.59 | 660.85 | 83.90% | 44.44% | 5.24 | 19.37% | 30.91 | 542.86 | 86.31% | 36.04% | 5.79 |
| TB | 28.85% | 20.25 | 986.84 | 74.87% | 77.06% | 8.33 | 16.64% | 29.10 | 860.45 | 78.47% | 74.99% | 10.60 |
| Gen | 12.12% | 20.85 | 525.69 | 89.34% | 38.07% | 5.00 | 7.22% | 26.00 | 364.15 | 91.28% | 29.28% | 5.74 |
| HF | 10.99% | 34.91 | 993.70 | 80.95% | 48.44% | 5.09 | 4.29% | 42.19 | 734.48 | 84.44% | 36.76% | 5.35 |
| NV | 28.54% | 18.13 | 357.90 | 90.67% | 12.48% | 0.55 | 14.94% | 24.15 | 275.77 | 91.41% | 13.44% | 0.76 |
| NVR | 28.81% | 19.12 | 371.86 | 90.50% | 13.55% | 0.59 | 15.00% | 25.20 | 287.31 | 91.22% | 14.46% | 0.80 |
| NVJ | 29.59% | 20.35 | 372.73 | 89.82% | 14.69% | 0.58 | 15.37% | 26.77 | 286.80 | 90.74% | 16.11% | 0.79 |
| NVJR | 29.83% | 21.45 | 387.00 | 89.67% | 15.90% | 0.62 | 15.41% | 27.73 | 297.15 | 90.57% | 16.92% | 0.83 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 47.15% | 22.78 | 576.58 | 85.90% | 41.77% | 5.12 | 10.63% | 34.05 | 705.90 | 85.92% | 36.48% | 5.59 |
| TB | 43.06% | 19.89 | 891.19 | 76.42% | 84.52% | 9.12 | 9.58% | 28.70 | 788.97 | 76.63% | 87.12% | 12.24 |
| Gen | 17.47% | 18.74 | 471.60 | 90.30% | 37.00% | 4.92 | 3.60% | 22.71 | 301.27 | 93.05% | 27.71% | 5.21 |
| HF | 15.90% | 32.23 | 813.75 | 86.17% | 45.40% | 4.97 | 2.06% | 30.18 | 368.13 | 90.09% | 29.76% | 5.33 |
| NV | 34.15% | 17.01 | 317.22 | 91.91% | 11.56% | 0.57 | 6.17% | 23.60 | 297.86 | 91.46% | 12.68% | 0.66 |
| NVR | 34.63% | 18.16 | 331.29 | 91.82% | 12.39% | 0.59 | 6.21% | 24.59 | 371.93 | 91.45% | 13.15% | 0.73 |
| NVJ | 36.25% | 19.08 | 336.43 | 91.18% | 14.35% | 0.59 | 6.31% | 25.97 | 310.11 | 90.59% | 15.94% | 0.70 |
| NVJR | 36.70% | 20.16 | 350.55 | 91.08% | 15.10% | 0.61 | 6.42% | 27.44 | 383.60 | 90.40% | 16.62% | 0.76 |
| Instruction\nBelow are six short snippets of text. The first one is the REFERENCE TEXT, and the other five are generated by different methods, which do not appear in any fixed order.\nPlease read TEXT1 - TEXT5, then select a score for them in the drop-down box at the beginning of each TEXT.\nNOTE: The total score is 5 points, according to the following scoring rules:\n1. TEXT does NOT change much from the REFERENCE TEXT. (0 - 2 points)\n2. TEXT is good in terms of grammatically and fluency. (0 - 2 points)\n3. TEXT is likely to be written by a human (and not by a machine). (0 - 1 point)\nFor example, if one sentence is good at 1. and 2., but bad at 3., you can give it a score of 2, 2, 0, which means the total score is 4.\nYou can give different sentences with the same score. |
| Test Items\nQ1:\nREFERENCE TEXT: Rosa Hernandez recalled how Piedra smiled and leaned forward the day he became her dentist. |
| [4] TEXT1: rosa guzman reminding how piedra giggled and leaned forward the day he became her dentist.\n[4] TEXT2: rosa hernandez recalled how piedra smiled and leenad forward the day he became her dentist.\n[3] TEXT3: rosa hernandez recalled how piedra smiled and leaned forward the day he get her dentist.\n[3] TEXT4: rosas gonzalo recalled why piedra whispered nor whispered eagerly the jour he became her dentist.\n[4] TEXT5: Rosa Hernandez knew how Piedra beamed and leaned forward the day he became her dentist. |
| Example 1 (generated against the Bert model) | |
| Original\n(Relation = contains administrative territorial entity) | Situated on the German border and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| PWWS\n(Prediction = located on terrain feature) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (turn the island of corsica). |
| TB\n(Prediction = has part) | situated on te german frontiers and not far from te german town of karlsruhe, it is te easternmost comune in metropolitan france (excluding te island of corsica). |
| Gen\n(Prediction = has part) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (excluded the island of corsica). |
| HF\n(Prediction = has part) | implanted during the german border nor never considerably for the german township from karlsruhe, he makes the easternmost township in metropolitan france (excluding the island of corsica). |
| NV\n(Prediction = has part) | localized on the German margin and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| Example 2 (generated against the BertEntity model) | |
| Original\n(Relation = after a work by) | It is loosely based on the novel "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| PWWS\n(Prediction = screenwriter) | it is loosely based on the novel "t night watch" by sergei lukyanenko, and is the foremost part of a duology, followed by "day watch". |
| TB\n(Prediction = screenwriter) | it is loosely bases on the novel "te night watch" by sergei lukyanenko, and is the frist pat of a duologj, followed by "day watch". |
| Gen\n(Prediction = screenwriter) | it is loosely based on the newer "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| HF\n(Prediction = screenwriter) | he makes freely based on the novel "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| NVR\n(Prediction = screenwriter) | It is slackly based on the romance "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| Example 1 (generated against the BertEntity model) | |
| Original\n(Relation = per:title) | Benjamin Chertoff - 25-year-old cousin of Michael Chertoff ; senior researcher " for Popular Mechanics ' hit piece on 9-11 Truth Movement |
| PWWS\n(Prediction = NA) | benjamin chertoff - 25 - year - old first of michael chertoff ; senior researcher " f popular mechanics ' hit piece on 9 - 11 truth movement |
| TB\n(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; sineor researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| Gen\n(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; elders researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| HF\n(Prediction = NA) | benjamin chertoff - 25 - sunny - immemorial kinsman of michael chertoff ; senior researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| NVJ\n(Prediction = NA) | Benjamin Chertoff - 25-year-old cousin-german of Michael Chertoff ; precedential researcher " for best-selling Mechanics ' striking piece on 9-11 Truth Movement |
| Example 2 (generated against the PCNN model) | |
| Original\n(Relation = org:top_members/employees) | Clayton also was hands-on as he helped archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| PWWS\n(Prediction = NA) | clayton also was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , theater of literary manuscripts at the Huntington library in san marino , Calif. |
| TB\n(Prediction = NA) | clyton aso ws h ands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in sn marino , Calif. |
| Gen\n(Prediction = NA) | clayton also was hands - on as he aided archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| HF\n(Prediction = NA) | clay additionally was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| NVJR\n(Prediction = NA) | Clayton likewise was hands-on as he aided archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| Metric | Acquisition Method | Superior |
| Acc | Statistics | ↑ |
| ASR | Statistics | ↑ |
| Dis | Levenshtein | ↓ |
| Flu | GPT2LM | ↓ |
| Sem | USE | ↑ |
| WMR | Statistics | ↓ |
| Err | Ginger | ↓ |
| Dataset | Victim Model | ||
| Bert | BertEntity | PCNN | |
| Wiki80 | 86.21% | 87.09% | 67.14% |
| TACRED | 87.57% | 88.45% | 80.64% |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 38.11% | 23.86 | 628.49 | 84.58% | 43.45% | 5.20 | 19.78% | 32.64 | 587.08 | 86.16% | 36.26% | 5.77 |
| TB | 34.00% | 20.00 | 938.21 | 75.95% | 78.52% | 8.55 | 17.40% | 28.45 | 810.19 | 78.95% | 77.42% | 10.90 |
| Gen | 14.51% | 20.10 | 504.52 | 89.75% | 37.68% | 5.01 | 8.11% | 25.23 | 344.63 | 91.92% | 28.82% | 5.57 |
| HF | 13.15% | 34.10 | 927.54 | 82.71% | 47.39% | 5.05 | 4.08% | 36.89 | 556.78 | 86.90% | 33.60% | 5.40 |
| NV | 30.83% | 17.66 | 339.16 | 91.19% | 12.06% | 0.57 | 15.65% | 24.40 | 282.95 | 91.33% | 13.18% | 0.74 |
| NVR | 31.14% | 18.74 | 353.37 | 91.05% | 13.05% | 0.60 | 15.73% | 25.47 | 315.08 | 91.24% | 13.92% | 0.79 |
| NVJ | 32.31% | 19.82 | 357.27 | 90.36% | 14.39% | 0.60 | 15.92% | 26.77 | 294.09 | 90.61% | 16.26% | 0.77 |
| NVJR | 32.57% | 20.89 | 372.01 | 90.23% | 15.42% | 0.63 | 16.03% | 28.00 | 325.82 | 90.45% | 17.01% | 0.81 |
| Dataset | Method | |||
| Gen | HF | PWWS | TB | |
| Wiki80 | NOUN: 51.7% | NOUN: 35.9% | NOUN: 61.5% | NOUN: 55.7% |
| VERB: 26.6% | PREP: 21.5% | VERB: 17.7% | VERB: 18.6% | |
| ADJ: 15.0% | VERB: 20.8% | ADJ: 17.6% | ADJ: 16.8% | |
| ADV: 4.0% | ADJ: 10.7% | ADV: 1.9% | ADV: 3.0% | |
| TACRED | NOUN: 53.2% | NOUN: 40.7% | NOUN: 58.4% | NOUN: 55.8% |
| VERB: 28.2% | VERB: 19.3% | VERB: 19.8% | VERB: 20.6% | |
| ADJ: 12.1% | PREP: 16.6% | ADJ: 17.4% | ADJ: 15.8% | |
| ADV: 4.6% | ADJ: 11.8% | ADV: 2.5% | ADV: 3.1% |
| Dataset | Victim Model | Dependency Parsing | ||||||
| Similarity | Overlap | Relation | ||||||
| UAS↑ | LAS↑ | Root | Head | Tail | Head | Tail | ||
| Wiki80 | Bert | 93.88% | 92.69% | 67.41% | 56.22% | 53.86% | nmod, nsubj, obl | obl, nmod, compound |
| BertEntity | 94.05% | 92.91% | 70.11% | 57.61% | 58.98% | nmod, nsubj, appos | obl, nmod, compound | |
| PCNN | 94.71% | 93.42% | 67.60% | 51.87% | 55.61% | nsubj, nmod, obl | obl, compound, nmod | |
| TACRED | Bert | 93.12% | 91.64% | 62.40% | 47.69% | 43.79% | nsubj, nmod, nmod:poss | nsubj, obl, nmod |
| BertEntity | 92.90% | 91.45% | 61.76% | 52.24% | 46.59% | nsubj, nmod, nmod:poss | obl, nsubj, nmod | |
| PCNN | 93.16% | 91.84% | 62.18% | 41.97% | 42.49% | nsubj, nmod:poss, nmod | nsubj, amod, nmod |
| Result | Method | ||||
| NV | PWWS | Gen | TB | HF | |
| Average score | 3.52 | 3.05 | 3.03 | 2.75 | 2.30 |
| Score >= 3 | 78% | 67% | 67% | 55% | 35% |
| Dataset | Victim Model (Vα) | Target Model | ||
| Bert | BertEntity | PCNN | ||
| Wiki80 | Bert(1439) | - | 43.71% | 57.61% |
| BertEntity(1392) | 48.35% | - | 57.61% | |
| PCNN(1284) | 14.80% | 14.95% | - | |
| TACRED | Bert(3508) | - | 23.66% | 40.76% |
| BertEntity(2050) | 54.49% | - | 53.56% | |
| PCNN(772) | 15.41% | 15.80% | - |
| Attack Method | Word-Level | Character/Word-Level | |||
| PWWS | Gen | Ours | TB | HF | |
| Perturbation Options | Sub | Sub | Sub | Ins, Del, Swap, Sub-C, Sub-W | Flip, Ins, Del, Sub |
| Importance Ranking | Required | Not Required | Not Required | Required | Not Required |
| Extra Requirements | GloVe vectors, Counter-fitting method, Google 1 billion words language model | / | Pre-trained GloVe model | One-Hot input representation, Word Embedding |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 34.40% | 24.21 | 648.04 | 83.95% | 44.15% | 5.23 | 29.34% | 32.96 | 512.49 | 86.25% | 36.27% | 5.93 |
| TB | 30.10% | 19.87 | 936.60 | 76.57% | 73.98% | 8.20 | 25.97% | 27.56 | 781.14 | 81.75% | 70.14% | 9.86 |
| Gen | 13.94% | 20.70 | 516.27 | 89.62% | 37.97% | 5.10 | 13.50% | 26.97 | 368.46 | 91.42% | 29.48% | 5.77 |
| HF | 12.55% | 35.17 | 975.18 | 81.00% | 48.34% | 5.09 | 5.88% | 38.31 | 567.73 | 86.17% | 34.28% | 5.53 |
| NV | 29.81% | 17.85 | 342.37 | 91.00% | 12.14% | 0.59 | 25.83% | 25.46 | 275.21 | 91.13% | 13.41% | 0.81 |
| NVR | 29.97% | 18.94 | 356.95 | 90.83% | 13.21% | 0.62 | 25.99% | 26.63 | 286.01 | 91.04% | 14.15% | 0.84 |
| NVJ | 31.09% | 20.02 | 362.64 | 90.09% | 14.14% | 0.62 | 26.07% | 27.58 | 285.36 | 90.49% | 16.72% | 0.82 |
| NVJR | 31.19% | 21.07 | 378.48 | 89.93% | 15.26% | 0.65 | 26.27% | 28.82 | 296.72 | 90.38% | 17.49% | 0.85 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 32.79% | 24.59 | 660.85 | 83.90% | 44.44% | 5.24 | 19.37% | 30.91 | 542.86 | 86.31% | 36.04% | 5.79 |
| TB | 28.85% | 20.25 | 986.84 | 74.87% | 77.06% | 8.33 | 16.64% | 29.10 | 860.45 | 78.47% | 74.99% | 10.60 |
| Gen | 12.12% | 20.85 | 525.69 | 89.34% | 38.07% | 5.00 | 7.22% | 26.00 | 364.15 | 91.28% | 29.28% | 5.74 |
| HF | 10.99% | 34.91 | 993.70 | 80.95% | 48.44% | 5.09 | 4.29% | 42.19 | 734.48 | 84.44% | 36.76% | 5.35 |
| NV | 28.54% | 18.13 | 357.90 | 90.67% | 12.48% | 0.55 | 14.94% | 24.15 | 275.77 | 91.41% | 13.44% | 0.76 |
| NVR | 28.81% | 19.12 | 371.86 | 90.50% | 13.55% | 0.59 | 15.00% | 25.20 | 287.31 | 91.22% | 14.46% | 0.80 |
| NVJ | 29.59% | 20.35 | 372.73 | 89.82% | 14.69% | 0.58 | 15.37% | 26.77 | 286.80 | 90.74% | 16.11% | 0.79 |
| NVJR | 29.83% | 21.45 | 387.00 | 89.67% | 15.90% | 0.62 | 15.41% | 27.73 | 297.15 | 90.57% | 16.92% | 0.83 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 47.15% | 22.78 | 576.58 | 85.90% | 41.77% | 5.12 | 10.63% | 34.05 | 705.90 | 85.92% | 36.48% | 5.59 |
| TB | 43.06% | 19.89 | 891.19 | 76.42% | 84.52% | 9.12 | 9.58% | 28.70 | 788.97 | 76.63% | 87.12% | 12.24 |
| Gen | 17.47% | 18.74 | 471.60 | 90.30% | 37.00% | 4.92 | 3.60% | 22.71 | 301.27 | 93.05% | 27.71% | 5.21 |
| HF | 15.90% | 32.23 | 813.75 | 86.17% | 45.40% | 4.97 | 2.06% | 30.18 | 368.13 | 90.09% | 29.76% | 5.33 |
| NV | 34.15% | 17.01 | 317.22 | 91.91% | 11.56% | 0.57 | 6.17% | 23.60 | 297.86 | 91.46% | 12.68% | 0.66 |
| NVR | 34.63% | 18.16 | 331.29 | 91.82% | 12.39% | 0.59 | 6.21% | 24.59 | 371.93 | 91.45% | 13.15% | 0.73 |
| NVJ | 36.25% | 19.08 | 336.43 | 91.18% | 14.35% | 0.59 | 6.31% | 25.97 | 310.11 | 90.59% | 15.94% | 0.70 |
| NVJR | 36.70% | 20.16 | 350.55 | 91.08% | 15.10% | 0.61 | 6.42% | 27.44 | 383.60 | 90.40% | 16.62% | 0.76 |
| Instruction\nBelow are six short snippets of text. The first one is the REFERENCE TEXT, and the other five are generated by different methods, which do not appear in any fixed order.\nPlease read TEXT1 - TEXT5, then select a score for them in the drop-down box at the beginning of each TEXT.\nNOTE: The total score is 5 points, according to the following scoring rules:\n1. TEXT does NOT change much from the REFERENCE TEXT. (0 - 2 points)\n2. TEXT is good in terms of grammatically and fluency. (0 - 2 points)\n3. TEXT is likely to be written by a human (and not by a machine). (0 - 1 point)\nFor example, if one sentence is good at 1. and 2., but bad at 3., you can give it a score of 2, 2, 0, which means the total score is 4.\nYou can give different sentences with the same score. |
| Test Items\nQ1:\nREFERENCE TEXT: Rosa Hernandez recalled how Piedra smiled and leaned forward the day he became her dentist. |
| [4] TEXT1: rosa guzman reminding how piedra giggled and leaned forward the day he became her dentist.\n[4] TEXT2: rosa hernandez recalled how piedra smiled and leenad forward the day he became her dentist.\n[3] TEXT3: rosa hernandez recalled how piedra smiled and leaned forward the day he get her dentist.\n[3] TEXT4: rosas gonzalo recalled why piedra whispered nor whispered eagerly the jour he became her dentist.\n[4] TEXT5: Rosa Hernandez knew how Piedra beamed and leaned forward the day he became her dentist. |
| Example 1 (generated against the Bert model) | |
| Original\n(Relation = contains administrative territorial entity) | Situated on the German border and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| PWWS\n(Prediction = located on terrain feature) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (turn the island of corsica). |
| TB\n(Prediction = has part) | situated on te german frontiers and not far from te german town of karlsruhe, it is te easternmost comune in metropolitan france (excluding te island of corsica). |
| Gen\n(Prediction = has part) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (excluded the island of corsica). |
| HF\n(Prediction = has part) | implanted during the german border nor never considerably for the german township from karlsruhe, he makes the easternmost township in metropolitan france (excluding the island of corsica). |
| NV\n(Prediction = has part) | localized on the German margin and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| Example 2 (generated against the BertEntity model) | |
| Original\n(Relation = after a work by) | It is loosely based on the novel "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| PWWS\n(Prediction = screenwriter) | it is loosely based on the novel "t night watch" by sergei lukyanenko, and is the foremost part of a duology, followed by "day watch". |
| TB\n(Prediction = screenwriter) | it is loosely bases on the novel "te night watch" by sergei lukyanenko, and is the frist pat of a duologj, followed by "day watch". |
| Gen\n(Prediction = screenwriter) | it is loosely based on the newer "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| HF\n(Prediction = screenwriter) | he makes freely based on the novel "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| NVR\n(Prediction = screenwriter) | It is slackly based on the romance "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| Example 1 (generated against the BertEntity model) | |
| Original\n(Relation = per:title) | Benjamin Chertoff - 25-year-old cousin of Michael Chertoff ; senior researcher " for Popular Mechanics ' hit piece on 9-11 Truth Movement |
| PWWS\n(Prediction = NA) | benjamin chertoff - 25 - year - old first of michael chertoff ; senior researcher " f popular mechanics ' hit piece on 9 - 11 truth movement |
| TB\n(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; sineor researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| Gen\n(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; elders researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| HF\n(Prediction = NA) | benjamin chertoff - 25 - sunny - immemorial kinsman of michael chertoff ; senior researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| NVJ\n(Prediction = NA) | Benjamin Chertoff - 25-year-old cousin-german of Michael Chertoff ; precedential researcher " for best-selling Mechanics ' striking piece on 9-11 Truth Movement |
| Example 2 (generated against the PCNN model) | |
| Original\n(Relation = org:top_members/employees) | Clayton also was hands-on as he helped archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| PWWS\n(Prediction = NA) | clayton also was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , theater of literary manuscripts at the Huntington library in san marino , Calif. |
| TB\n(Prediction = NA) | clyton aso ws h ands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in sn marino , Calif. |
| Gen\n(Prediction = NA) | clayton also was hands - on as he aided archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| HF\n(Prediction = NA) | clay additionally was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| NVJR\n(Prediction = NA) | Clayton likewise was hands-on as he aided archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| Metric | Acquisition Method | Superior |
| Acc | Statistics | ↑ |
| ASR | Statistics | ↑ |
| Dis | Levenshtein | ↓ |
| Flu | GPT2LM | ↓ |
| Sem | USE | ↑ |
| WMR | Statistics | ↓ |
| Err | Ginger | ↓ |
| Dataset | Victim Model | ||
| Bert | BertEntity | PCNN | |
| Wiki80 | 86.21% | 87.09% | 67.14% |
| TACRED | 87.57% | 88.45% | 80.64% |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 38.11% | 23.86 | 628.49 | 84.58% | 43.45% | 5.20 | 19.78% | 32.64 | 587.08 | 86.16% | 36.26% | 5.77 |
| TB | 34.00% | 20.00 | 938.21 | 75.95% | 78.52% | 8.55 | 17.40% | 28.45 | 810.19 | 78.95% | 77.42% | 10.90 |
| Gen | 14.51% | 20.10 | 504.52 | 89.75% | 37.68% | 5.01 | 8.11% | 25.23 | 344.63 | 91.92% | 28.82% | 5.57 |
| HF | 13.15% | 34.10 | 927.54 | 82.71% | 47.39% | 5.05 | 4.08% | 36.89 | 556.78 | 86.90% | 33.60% | 5.40 |
| NV | 30.83% | 17.66 | 339.16 | 91.19% | 12.06% | 0.57 | 15.65% | 24.40 | 282.95 | 91.33% | 13.18% | 0.74 |
| NVR | 31.14% | 18.74 | 353.37 | 91.05% | 13.05% | 0.60 | 15.73% | 25.47 | 315.08 | 91.24% | 13.92% | 0.79 |
| NVJ | 32.31% | 19.82 | 357.27 | 90.36% | 14.39% | 0.60 | 15.92% | 26.77 | 294.09 | 90.61% | 16.26% | 0.77 |
| NVJR | 32.57% | 20.89 | 372.01 | 90.23% | 15.42% | 0.63 | 16.03% | 28.00 | 325.82 | 90.45% | 17.01% | 0.81 |
| Dataset | Method | |||
| Gen | HF | PWWS | TB | |
| Wiki80 | NOUN: 51.7% | NOUN: 35.9% | NOUN: 61.5% | NOUN: 55.7% |
| VERB: 26.6% | PREP: 21.5% | VERB: 17.7% | VERB: 18.6% | |
| ADJ: 15.0% | VERB: 20.8% | ADJ: 17.6% | ADJ: 16.8% | |
| ADV: 4.0% | ADJ: 10.7% | ADV: 1.9% | ADV: 3.0% | |
| TACRED | NOUN: 53.2% | NOUN: 40.7% | NOUN: 58.4% | NOUN: 55.8% |
| VERB: 28.2% | VERB: 19.3% | VERB: 19.8% | VERB: 20.6% | |
| ADJ: 12.1% | PREP: 16.6% | ADJ: 17.4% | ADJ: 15.8% | |
| ADV: 4.6% | ADJ: 11.8% | ADV: 2.5% | ADV: 3.1% | |
| Dataset | Victim Model | Dependency Parsing | ||||||
| Similarity | Overlap | Relation | ||||||
| UAS↑ | LAS↑ | Root | Head | Tail | Head | Tail | ||
| Wiki80 | Bert | 93.88% | 92.69% | 67.41% | 56.22% | 53.86% | nmod, nsubj, obl | obl, nmod, compound |
| BertEntity | 94.05% | 92.91% | 70.11% | 57.61% | 58.98% | nmod, nsubj, appos | obl, nmod, compound | |
| PCNN | 94.71% | 93.42% | 67.60% | 51.87% | 55.61% | nsubj, nmod, obl | obl, compound, nmod | |
| TACRED | Bert | 93.12% | 91.64% | 62.40% | 47.69% | 43.79% | nsubj, nmod, nmod:poss | nsubj, obl, nmod |
| BertEntity | 92.90% | 91.45% | 61.76% | 52.24% | 46.59% | nsubj, nmod, nmod:poss | obl, nsubj, nmod | |
| PCNN | 93.16% | 91.84% | 62.18% | 41.97% | 42.49% | nsubj, nmod:poss, nmod | nsubj, amod, nmod | |
| Result | Method | ||||
| NV | PWWS | Gen | TB | HF | |
| Average score | 3.52 | 3.05 | 3.03 | 2.75 | 2.30 |
| Score >= 3 | 78% | 67% | 67% | 55% | 35% |
| Dataset | Victim Model (Vα) | Target Model | ||
| Bert | BertEntity | PCNN | ||
| Wiki80 | Bert(1439) | - | 43.71% | 57.61% |
| BertEntity(1392) | 48.35% | - | 57.61% | |
| PCNN(1284) | 14.80% | 14.95% | - | |
| TACRED | Bert(3508) | - | 23.66% | 40.76% |
| BertEntity(2050) | 54.49% | - | 53.56% | |
| PCNN(772) | 15.41% | 15.80% | - | |
| Attack Method | Word-Level | Character/Word-Level | |||
| PWWS | Gen | Ours | TB | HF | |
| Perturbation Options | Sub | Sub | Sub | Ins, Del, Swap, Sub-C, Sub-W | Flip, Ins, Del, Sub |
| Importance Ranking | Required | Not Required | Not Required | Required | Not Required |
| Extra Requirements | GloVe vectors, Counter-fitting method, Google 1 billion words language model | / | Pre-trained GloVe model | One-Hot input representation, Word Embedding | |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 34.40% | 24.21 | 648.04 | 83.95% | 44.15% | 5.23 | 29.34% | 32.96 | 512.49 | 86.25% | 36.27% | 5.93 |
| TB | 30.10% | 19.87 | 936.60 | 76.57% | 73.98% | 8.20 | 25.97% | 27.56 | 781.14 | 81.75% | 70.14% | 9.86 |
| Gen | 13.94% | 20.70 | 516.27 | 89.62% | 37.97% | 5.10 | 13.50% | 26.97 | 368.46 | 91.42% | 29.48% | 5.77 |
| HF | 12.55% | 35.17 | 975.18 | 81.00% | 48.34% | 5.09 | 5.88% | 38.31 | 567.73 | 86.17% | 34.28% | 5.53 |
| NV | 29.81% | 17.85 | 342.37 | 91.00% | 12.14% | 0.59 | 25.83% | 25.46 | 275.21 | 91.13% | 13.41% | 0.81 |
| NVR | 29.97% | 18.94 | 356.95 | 90.83% | 13.21% | 0.62 | 25.99% | 26.63 | 286.01 | 91.04% | 14.15% | 0.84 |
| NVJ | 31.09% | 20.02 | 362.64 | 90.09% | 14.14% | 0.62 | 26.07% | 27.58 | 285.36 | 90.49% | 16.72% | 0.82 |
| NVJR | 31.19% | 21.07 | 378.48 | 89.93% | 15.26% | 0.65 | 26.27% | 28.82 | 296.72 | 90.38% | 17.49% | 0.85 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 32.79% | 24.59 | 660.85 | 83.90% | 44.44% | 5.24 | 19.37% | 30.91 | 542.86 | 86.31% | 36.04% | 5.79 |
| TB | 28.85% | 20.25 | 986.84 | 74.87% | 77.06% | 8.33 | 16.64% | 29.10 | 860.45 | 78.47% | 74.99% | 10.60 |
| Gen | 12.12% | 20.85 | 525.69 | 89.34% | 38.07% | 5.00 | 7.22% | 26.00 | 364.15 | 91.28% | 29.28% | 5.74 |
| HF | 10.99% | 34.91 | 993.70 | 80.95% | 48.44% | 5.09 | 4.29% | 42.19 | 734.48 | 84.44% | 36.76% | 5.35 |
| NV | 28.54% | 18.13 | 357.90 | 90.67% | 12.48% | 0.55 | 14.94% | 24.15 | 275.77 | 91.41% | 13.44% | 0.76 |
| NVR | 28.81% | 19.12 | 371.86 | 90.50% | 13.55% | 0.59 | 15.00% | 25.20 | 287.31 | 91.22% | 14.46% | 0.80 |
| NVJ | 29.59% | 20.35 | 372.73 | 89.82% | 14.69% | 0.58 | 15.37% | 26.77 | 286.80 | 90.74% | 16.11% | 0.79 |
| NVJR | 29.83% | 21.45 | 387.00 | 89.67% | 15.90% | 0.62 | 15.41% | 27.73 | 297.15 | 90.57% | 16.92% | 0.83 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 47.15% | 22.78 | 576.58 | 85.90% | 41.77% | 5.12 | 10.63% | 34.05 | 705.90 | 85.92% | 36.48% | 5.59 |
| TB | 43.06% | 19.89 | 891.19 | 76.42% | 84.52% | 9.12 | 9.58% | 28.70 | 788.97 | 76.63% | 87.12% | 12.24 |
| Gen | 17.47% | 18.74 | 471.60 | 90.30% | 37.00% | 4.92 | 3.60% | 22.71 | 301.27 | 93.05% | 27.71% | 5.21 |
| HF | 15.90% | 32.23 | 813.75 | 86.17% | 45.40% | 4.97 | 2.06% | 30.18 | 368.13 | 90.09% | 29.76% | 5.33 |
| NV | 34.15% | 17.01 | 317.22 | 91.91% | 11.56% | 0.57 | 6.17% | 23.60 | 297.86 | 91.46% | 12.68% | 0.66 |
| NVR | 34.63% | 18.16 | 331.29 | 91.82% | 12.39% | 0.59 | 6.21% | 24.59 | 371.93 | 91.45% | 13.15% | 0.73 |
| NVJ | 36.25% | 19.08 | 336.43 | 91.18% | 14.35% | 0.59 | 6.31% | 25.97 | 310.11 | 90.59% | 15.94% | 0.70 |
| NVJR | 36.70% | 20.16 | 350.55 | 91.08% | 15.10% | 0.61 | 6.42% | 27.44 | 383.60 | 90.40% | 16.62% | 0.76 |
| Instruction +Below are six short snippets of text. The first one is the REFERENCE TEXT, and the other five are generated by different methods, which do not appear in any fixed order. +Please read TEXT1 - TEXT5, then select a score for them in the drop-down box at the beginning of each TEXT. +NOTE: The total score is 5 points, according to the following scoring rules: +1. TEXT does NOT change much from the REFERENCE TEXT. (0 - 2 points) +2. TEXT is good in terms of grammatically and fluency. (0 - 2 points) +3. TEXT is likely to be written by a human (and not by a machine). (0 - 1 point) +For example, if one sentence is good at 1. and 2., but bad at 3., you can give it a score of 2, 2, 0, which means the total score is 4. +You can give different sentences with the same score. | |
| Test Items +Q1: +REFERENCE TEXT: Rosa Hernandez recalled how Piedra smiled and leaned forward the day he became her dentist. | |
| [4] TEXT1: rosa guzman reminding how piedra giggled and leaned forward the day he became her dentist. +[4] TEXT2: rosa hernandez recalled how piedra smiled and leenad forward the day he became her dentist. +[3] TEXT3: rosa hernandez recalled how piedra smiled and leaned forward the day he get her dentist. +[3] TEXT4: rosas gonzalo recalled why piedra whispered nor whispered eagerly the jour he became her dentist. +[4] TEXT5: Rosa Hernandez knew how Piedra beamed and leaned forward the day he became her dentist. |
| Example 1 (generated against the Bert model) | |
| Original +(Relation = contains administrative territorial entity) | Situated on the German border and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| PWWS +(Prediction = located on terrain feature) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (turn the island of corsica). |
| TB +(Prediction = has part) | situated on te german frontiers and not far from te german town of karlsruhe, it is te easternmost comune in metropolitan france (excluding te island of corsica). |
| Gen +(Prediction = has part) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (excluded the island of corsica). |
| HF +(Prediction = has part) | implanted during the german border nor never considerably for the german township from karlsruhe, he makes the easternmost township in metropolitan france (excluding the island of corsica). |
| NV +(Prediction = has part) | localized on the German margin and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| Example 2 (generated against the BertEntity model) | |
| Original +(Relation = after a work by) | It is loosely based on the novel "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| PWWS +(Prediction = screenwriter) | it is loosely based on the novel "t night watch" by sergei lukyanenko, and is the foremost part of a duology, followed by "day watch". |
| TB +(Prediction = screenwriter) | it is loosely bases on the novel "te night watch" by sergei lukyanenko, and is the frist pat of a duologj, followed by "day watch". |
| Gen +(Prediction = screenwriter) | it is loosely based on the newer "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| HF +(Prediction = screenwriter) | he makes freely based on the novel "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| NVR +(Prediction = screenwriter) | It is slackly based on the romance "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| Example 1 (generated against the BertEntity model) | |
| Original +(Relation = per:title) | Benjamin Chertoff - 25-year-old cousin of Michael Chertoff ; senior researcher " for Popular Mechanics ' hit piece on 9-11 Truth Movement |
| PWWS +(Prediction = NA) | benjamin chertoff - 25 - year - old first of michael chertoff ; senior researcher " f popular mechanics ' hit piece on 9 - 11 truth movement |
| TB +(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; sineor researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| Gen +(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; elders researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| HF +(Prediction = NA) | benjamin chertoff - 25 - sunny - immemorial kinsman of michael chertoff ; senior researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| NVJ +(Prediction = NA) | Benjamin Chertoff - 25-year-old cousin-german of Michael Chertoff ; precedential researcher " for best-selling Mechanics ' striking piece on 9-11 Truth Movement |
| Example 2 (generated against the PCNN model) | |
| Original +(Relation = org:top_members/employees) | Clayton also was hands-on as he helped archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| PWWS +(Prediction = NA) | clayton also was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , theater of literary manuscripts at the Huntington library in san marino , Calif. |
| TB +(Prediction = NA) | clyton aso ws h ands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in sn marino , Calif. |
| Gen +(Prediction = NA) | clayton also was hands - on as he aided archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| HF +(Prediction = NA) | clay additionally was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| NVJR +(Prediction = NA) | Clayton likewise was hands-on as he aided archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| Metric | Acquisition Method | Superior |
| Acc | Statistics | ↑ |
| ASR | Statistics | ↑ |
| Dis | Levenshtein | ↓ |
| Flu | GPT2LM | ↓ |
| Sem | USE | ↑ |
| WMR | Statistics | ↓ |
| Err | Ginger | ↓ |
| Dataset | Victim Model | ||
| Bert | BertEntity | PCNN | |
| Wiki80 | 86.21% | 87.09% | 67.14% |
| TACRED | 87.57% | 88.45% | 80.64% |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 38.11% | 23.86 | 628.49 | 84.58% | 43.45% | 5.20 | 19.78% | 32.64 | 587.08 | 86.16% | 36.26% | 5.77 |
| TB | 34.00% | 20.00 | 938.21 | 75.95% | 78.52% | 8.55 | 17.40% | 28.45 | 810.19 | 78.95% | 77.42% | 10.90 |
| Gen | 14.51% | 20.10 | 504.52 | 89.75% | 37.68% | 5.01 | 8.11% | 25.23 | 344.63 | 91.92% | 28.82% | 5.57 |
| HF | 13.15% | 34.10 | 927.54 | 82.71% | 47.39% | 5.05 | 4.08% | 36.89 | 556.78 | 86.90% | 33.60% | 5.40 |
| NV | 30.83% | 17.66 | 339.16 | 91.19% | 12.06% | 0.57 | 15.65% | 24.40 | 282.95 | 91.33% | 13.18% | 0.74 |
| NVR | 31.14% | 18.74 | 353.37 | 91.05% | 13.05% | 0.60 | 15.73% | 25.47 | 315.08 | 91.24% | 13.92% | 0.79 |
| NVJ | 32.31% | 19.82 | 357.27 | 90.36% | 14.39% | 0.60 | 15.92% | 26.77 | 294.09 | 90.61% | 16.26% | 0.77 |
| NVJR | 32.57% | 20.89 | 372.01 | 90.23% | 15.42% | 0.63 | 16.03% | 28.00 | 325.82 | 90.45% | 17.01% | 0.81 |
| Dataset | Method | |||
| Gen | HF | PWWS | TB | |
| Wiki80 | NOUN: 51.7% | NOUN: 35.9% | NOUN: 61.5% | NOUN: 55.7% |
| VERB: 26.6% | PREP: 21.5% | VERB: 17.7% | VERB: 18.6% | |
| ADJ: 15.0% | VERB: 20.8% | ADJ: 17.6% | ADJ: 16.8% | |
| ADV: 4.0% | ADJ: 10.7% | ADV: 1.9% | ADV: 3.0% | |
| TACRED | NOUN: 53.2% | NOUN: 40.7% | NOUN: 58.4% | NOUN: 55.8% |
| VERB: 28.2% | VERB: 19.3% | VERB: 19.8% | VERB: 20.6% | |
| ADJ: 12.1% | PREP: 16.6% | ADJ: 17.4% | ADJ: 15.8% | |
| ADV: 4.6% | ADJ: 11.8% | ADV: 2.5% | ADV: 3.1% |
| Dataset | Victim Model | Dependency Parsing | ||||||
| Similarity | Overlap | Relation | ||||||
| UAS↑ | LAS↑ | Root | Head | Tail | Head | Tail | ||
| Wiki80 | Bert | 93.88% | 92.69% | 67.41% | 56.22% | 53.86% | nmod, nsubj, obl | obl, nmod, compound |
| BertEntity | 94.05% | 92.91% | 70.11% | 57.61% | 58.98% | nmod, nsubj, appos | obl, nmod, compound | |
| PCNN | 94.71% | 93.42% | 67.60% | 51.87% | 55.61% | nsubj, nmod, obl | obl, compound, nmod | |
| TACRED | Bert | 93.12% | 91.64% | 62.40% | 47.69% | 43.79% | nsubj, nmod, nmod:poss | nsubj, obl, nmod |
| BertEntity | 92.90% | 91.45% | 61.76% | 52.24% | 46.59% | nsubj, nmod, nmod:poss | obl, nsubj, nmod | |
| PCNN | 93.16% | 91.84% | 62.18% | 41.97% | 42.49% | nsubj, nmod:poss, nmod | nsubj, amod, nmod |
| Result | Method | ||||
| NV | PWWS | Gen | TB | HF | |
| Average score | 3.52 | 3.05 | 3.03 | 2.75 | 2.30 |
| Score >= 3 | 78% | 67% | 67% | 55% | 35% |
| Dataset | Victim Model (Vα) | Target Model | ||
| Bert | BertEntity | PCNN | ||
| Wiki80 | Bert(1439) | - | 43.71% | 57.61% |
| BertEntity(1392) | 48.35% | - | 57.61% | |
| PCNN(1284) | 14.80% | 14.95% | - | |
| TACRED | Bert(3508) | - | 23.66% | 40.76% |
| BertEntity(2050) | 54.49% | - | 53.56% | |
| PCNN(772) | 15.41% | 15.80% | - |
| Attack Method | Word-Level | Character/Word-Level | |||
| PWWS | Gen | Ours | TB | HF | |
| Perturbation Options | Sub | Sub | Sub | Ins, Del, Swap, Sub-C, Sub-W | Flip, Ins, Del, Sub |
| Importance Ranking | Required | Not Required | Not Required | Required | Not Required |
| Extra Requirements | GloVe vectors, Counter-fitting method, Google 1 billion words language model | / | Pre-trained GloVe model | One-Hot input representation, Word Embedding |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 34.40% | 24.21 | 648.04 | 83.95% | 44.15% | 5.23 | 29.34% | 32.96 | 512.49 | 86.25% | 36.27% | 5.93 |
| TB | 30.10% | 19.87 | 936.60 | 76.57% | 73.98% | 8.20 | 25.97% | 27.56 | 781.14 | 81.75% | 70.14% | 9.86 |
| Gen | 13.94% | 20.70 | 516.27 | 89.62% | 37.97% | 5.10 | 13.50% | 26.97 | 368.46 | 91.42% | 29.48% | 5.77 |
| HF | 12.55% | 35.17 | 975.18 | 81.00% | 48.34% | 5.09 | 5.88% | 38.31 | 567.73 | 86.17% | 34.28% | 5.53 |
| NV | 29.81% | 17.85 | 342.37 | 91.00% | 12.14% | 0.59 | 25.83% | 25.46 | 275.21 | 91.13% | 13.41% | 0.81 |
| NVR | 29.97% | 18.94 | 356.95 | 90.83% | 13.21% | 0.62 | 25.99% | 26.63 | 286.01 | 91.04% | 14.15% | 0.84 |
| NVJ | 31.09% | 20.02 | 362.64 | 90.09% | 14.14% | 0.62 | 26.07% | 27.58 | 285.36 | 90.49% | 16.72% | 0.82 |
| NVJR | 31.19% | 21.07 | 378.48 | 89.93% | 15.26% | 0.65 | 26.27% | 28.82 | 296.72 | 90.38% | 17.49% | 0.85 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 32.79% | 24.59 | 660.85 | 83.90% | 44.44% | 5.24 | 19.37% | 30.91 | 542.86 | 86.31% | 36.04% | 5.79 |
| TB | 28.85% | 20.25 | 986.84 | 74.87% | 77.06% | 8.33 | 16.64% | 29.10 | 860.45 | 78.47% | 74.99% | 10.60 |
| Gen | 12.12% | 20.85 | 525.69 | 89.34% | 38.07% | 5.00 | 7.22% | 26.00 | 364.15 | 91.28% | 29.28% | 5.74 |
| HF | 10.99% | 34.91 | 993.70 | 80.95% | 48.44% | 5.09 | 4.29% | 42.19 | 734.48 | 84.44% | 36.76% | 5.35 |
| NV | 28.54% | 18.13 | 357.90 | 90.67% | 12.48% | 0.55 | 14.94% | 24.15 | 275.77 | 91.41% | 13.44% | 0.76 |
| NVR | 28.81% | 19.12 | 371.86 | 90.50% | 13.55% | 0.59 | 15.00% | 25.20 | 287.31 | 91.22% | 14.46% | 0.80 |
| NVJ | 29.59% | 20.35 | 372.73 | 89.82% | 14.69% | 0.58 | 15.37% | 26.77 | 286.80 | 90.74% | 16.11% | 0.79 |
| NVJR | 29.83% | 21.45 | 387.00 | 89.67% | 15.90% | 0.62 | 15.41% | 27.73 | 297.15 | 90.57% | 16.92% | 0.83 |
| Method | Wiki80 | TACRED | ||||||||||
| ASR | Dis | Flu | Sem | WMR | Err | ASR | Dis | Flu | Sem | WMR | Err | |
| PWWS | 47.15% | 22.78 | 576.58 | 85.90% | 41.77% | 5.12 | 10.63% | 34.05 | 705.90 | 85.92% | 36.48% | 5.59 |
| TB | 43.06% | 19.89 | 891.19 | 76.42% | 84.52% | 9.12 | 9.58% | 28.70 | 788.97 | 76.63% | 87.12% | 12.24 |
| Gen | 17.47% | 18.74 | 471.60 | 90.30% | 37.00% | 4.92 | 3.60% | 22.71 | 301.27 | 93.05% | 27.71% | 5.21 |
| HF | 15.90% | 32.23 | 813.75 | 86.17% | 45.40% | 4.97 | 2.06% | 30.18 | 368.13 | 90.09% | 29.76% | 5.33 |
| NV | 34.15% | 17.01 | 317.22 | 91.91% | 11.56% | 0.57 | 6.17% | 23.60 | 297.86 | 91.46% | 12.68% | 0.66 |
| NVR | 34.63% | 18.16 | 331.29 | 91.82% | 12.39% | 0.59 | 6.21% | 24.59 | 371.93 | 91.45% | 13.15% | 0.73 |
| NVJ | 36.25% | 19.08 | 336.43 | 91.18% | 14.35% | 0.59 | 6.31% | 25.97 | 310.11 | 90.59% | 15.94% | 0.70 |
| NVJR | 36.70% | 20.16 | 350.55 | 91.08% | 15.10% | 0.61 | 6.42% | 27.44 | 383.60 | 90.40% | 16.62% | 0.76 |
| Instruction\nBelow are six short snippets of text. The first one is the REFERENCE TEXT, and the other five are generated by different methods, which do not appear in any fixed order.\nPlease read TEXT1 - TEXT5, then select a score for them in the drop-down box at the beginning of each TEXT.\nNOTE: The total score is 5 points, according to the following scoring rules:\n1. TEXT does NOT change much from the REFERENCE TEXT. (0 - 2 points)\n2. TEXT is good in terms of grammatically and fluency. (0 - 2 points)\n3. TEXT is likely to be written by a human (and not by a machine). (0 - 1 point)\nFor example, if one sentence is good at 1. and 2., but bad at 3., you can give it a score of 2, 2, 0, which means the total score is 4.\nYou can give different sentences with the same score. |
| Test Items\nQ1:\nREFERENCE TEXT: Rosa Hernandez recalled how Piedra smiled and leaned forward the day he became her dentist. |
| [4] TEXT1: rosa guzman reminding how piedra giggled and leaned forward the day he became her dentist.\n[4] TEXT2: rosa hernandez recalled how piedra smiled and leenad forward the day he became her dentist.\n[3] TEXT3: rosa hernandez recalled how piedra smiled and leaned forward the day he get her dentist.\n[3] TEXT4: rosas gonzalo recalled why piedra whispered nor whispered eagerly the jour he became her dentist.\n[4] TEXT5: Rosa Hernandez knew how Piedra beamed and leaned forward the day he became her dentist. |
| Example 1 (generated against the Bert model) | |
| Original\n(Relation = contains administrative territorial entity) | Situated on the German border and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| PWWS\n(Prediction = located on terrain feature) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (turn the island of corsica). |
| TB\n(Prediction = has part) | situated on te german frontiers and not far from te german town of karlsruhe, it is te easternmost comune in metropolitan france (excluding te island of corsica). |
| Gen\n(Prediction = has part) | situated on the german border and not far from the german city of karlsruhe, it is the easternmost commune in metropolitan france (excluded the island of corsica). |
| HF\n(Prediction = has part) | implanted during the german border nor never considerably for the german township from karlsruhe, he makes the easternmost township in metropolitan france (excluding the island of corsica). |
| NV\n(Prediction = has part) | localized on the German margin and not far from the German city of Karlsruhe, it is the easternmost commune in Metropolitan France (excluding the island of Corsica). |
| Example 2 (generated against the BertEntity model) | |
| Original\n(Relation = after a work by) | It is loosely based on the novel "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| PWWS\n(Prediction = screenwriter) | it is loosely based on the novel "t night watch" by sergei lukyanenko, and is the foremost part of a duology, followed by "day watch". |
| TB\n(Prediction = screenwriter) | it is loosely bases on the novel "te night watch" by sergei lukyanenko, and is the frist pat of a duologj, followed by "day watch". |
| Gen\n(Prediction = screenwriter) | it is loosely based on the newer "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| HF\n(Prediction = screenwriter) | he makes freely based on the novel "the night watch" by sergei lukyanenko, and is the first part of a duology, followed by "day watch". |
| NVR\n(Prediction = screenwriter) | It is slackly based on the romance "The Night Watch" by Sergei Lukyanenko, and is the first part of a duology, followed by "Day Watch". |
| Example 1 (generated against the BertEntity model) | |
| Original\n(Relation = per:title) | Benjamin Chertoff - 25-year-old cousin of Michael Chertoff ; senior researcher " for Popular Mechanics ' hit piece on 9-11 Truth Movement |
| PWWS\n(Prediction = NA) | benjamin chertoff - 25 - year - old first of michael chertoff ; senior researcher " f popular mechanics ' hit piece on 9 - 11 truth movement |
| TB\n(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; sineor researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| Gen\n(Prediction = NA) | benjamin chertoff - 25 - year - old cousin of michael chertoff ; elders researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| HF\n(Prediction = NA) | benjamin chertoff - 25 - sunny - immemorial kinsman of michael chertoff ; senior researcher " for popular mechanics ' hit piece on 9 - 11 truth movement |
| NVJ\n(Prediction = NA) | Benjamin Chertoff - 25-year-old cousin-german of Michael Chertoff ; precedential researcher " for best-selling Mechanics ' striking piece on 9-11 Truth Movement |
| Example 2 (generated against the PCNN model) | |
| Original\n(Relation = org:top_members/employees) | Clayton also was hands-on as he helped archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| PWWS\n(Prediction = NA) | clayton also was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , theater of literary manuscripts at the Huntington library in san marino , Calif. |
| TB\n(Prediction = NA) | clyton aso ws h ands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in sn marino , Calif. |
| Gen\n(Prediction = NA) | clayton also was hands - on as he aided archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| HF\n(Prediction = NA) | clay additionally was hands - on as he helped archive of one of the largest collections of african - americana , " said sue hodson , director of literary manuscripts at the Huntington library in san marino , Calif. |
| NVJR\n(Prediction = NA) | Clayton likewise was hands-on as he aided archive of one of the largest collections of African-Americana , " said Sue Hodson , director of literary manuscripts at the Huntington Library in San Marino , Calif . |
| Training Sets | Doriginal | Drandom | Dnorept |
| Ratios (%) ↓ | 60.42 | 96.04 | 1.67 |
| Models | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 |
| Greedy | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 2.55 | 3.29 | 41.23 | 70.18 | 83.28 | 19.27 | 23.95 | 1.76 | 0.61 | 71.08 | 87.20 | 89.43 | 9.61 | 11.40 |
| UL | 3.20 | 7.16 | 33.91 | 61.90 | 76.89 | 25.13 | 31.90 | 2.01 | 1.63 | 59.43 | 81.63 | 85.89 | 11.66 | 14.30 |
| ScaleGrad | 4.61 | 7.66 | 29.82 | 50.69 | 66.14 | 36.96 | 47.34 | 2.87 | 11.17 | 52.29 | 69.53 | 76.16 | 18.16 | 24.40 |
| SELFCONT | 6.47 | 17.34 | 23.29 | 39.41 | 62.46 | 46.71 | 57.66 | 3.30 | 20.05 | 35.13 | 53.69 | 74.09 | 23.30 | 31.52 |
| Nucleus | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 20.66 | 21.09 | 19.40 | 30.22 | 48.11 | 71.92 | 84.75 | 18.68 | 88.54 | 20.95 | 32.53 | 48.87 | 60.38 | 81.55 |
| UL | 15.54 | 21.78 | 18.45 | 29.57 | 46.69 | 69.63 | 82.87 | 19.39 | 81.49 | 18.36 | 27.98 | 42.65 | 63.92 | 82.93 |
| ScaleGrad | 12.41 | 25.69 | 18.59 | 29.24 | 45.19 | 66.35 | 80.23 | 14.14 | 77.82 | 18.62 | 27.80 | 41.22 | 56.74 | 77.27 |
| SELFCONT | 19.02 | 34.37 | 16.45 | 26.47 | 45.10 | 72.02 | 84.78 | 19.86 | 89.84 | 17.56 | 26.98 | 43.39 | 63.33 | 83.51 |
| Ground Truth | 18.31 | 100 | 17.38 | 27.92 | 46.29 | 72.34 | 84.20 | 24.01 | 100 | 16.36 | 26.47 | 42.30 | 74.49 | 90.01 |
| Datasets | |Train| | |Validation| | |Test| | Avg. Len |
| Wikitext-103 | 201,632 | 448 | 480 | 512 |
| WritingPrompts | 272,600 | 15,620 | 15,138 | 439 |
| Training Steps | Cases | |
| 100 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. The first, and the " . The " The " <eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos></eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos> | |
| 1500 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of the building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the buildings, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building | He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building |
| 6000 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. She had begun in 1962, was built a series of New York Times in 1962 - 1938 after producing the Laidlay, and described as well as well as The Lasttunk's assistant ... ] named". She later explained : "[ 500 people who did not be turned off for me to me to me. She decided to his own song " As Rocham R. Nails , saying". She knew it 's script was going to make me to live film, and I mean it was through me to get it 't know what we 't know what we want to make me . On 17 November 1953 , and it was nominated for me. He had done what he wanted the BBC at a period at her following year because it appearance at the mid @-@ selling artist with her singing . She was nominated for the background scene in 1962 . She spent four days after getting part of the public appearances at the war appearance , California , and ] base . Aftermath Medie Glam artist of the summer May 1967 , New York Times , and the war as the Star Wars franchise . He began to use the National Association ( from his staff ; it was included Star ) , including the Lyds house east , and I was the West Virginia Tech back to the war and Mennon ; there . She developed by <eos> William Peninsular League <eos>=<eos> The script , 2004 . The script was named after the North America for the LAM passed . The script was commissioned to the American co @-@ person to produce producer ( present , taking place of the mid @-@ old , and the mid @-@ old @-@ old film , The Next Generation . The New York Times , having won the 4th birthday of the 4 , in the 4 million viewers . This was announced that it was cut of the media . The Elder Scrolls IV of the production , in East Coaster and The company entitled The Next Generation . <eos> For example , including the war , having performed on 6 , having released in East Coast Division . Upon its crew became a series of the produce |
| Training Set | Cases | |
| Doriginal | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has appeared in several films, including the television series The Bill, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. <eos>Boulter was born in London, England, on 23 May 1986. He is the third child of actor and actress Robert Boulter and his wife, Susan. He is the third of five children born to his wife Susan and their three children, Robert, Roberta, and Roberta. Robert Boulter's father, Robert Boulter, was a film director and producer. He was the first actor to be cast in a film role, and the first to be cast in a television series. He was also the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a tele-. tical series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was then cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a tele-. tical series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He were cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. Hewas cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He wasCast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast ina Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Sereas. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series | He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast ina television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a Television Series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cist in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the firsst actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series.Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast ina television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. |
| Models | Wikitext-103 | WritingPrompts |
| MLE | 0.9 | 0.9 |
| UL | 0.7 | 0.8 |
| ScaleGrad | 0.5 | 0.6 |
| SELFCONT | 0.6 | 0.7 |
| Models | Acc | Repetitive | Non-Repetitive | ||||
| P | R | F1 | P | R | F1 | ||
| MLE | 78.9 | 67.4 | 87.0 | 75.9 | 90.2 | 73.9 | 81.2 |
| UL | 80.5 | 70.9 | 83.0 | 76.5 | 88.2 | 78.9 | 83.3 |
| ScaleGrad | 81.6 | 75.9 | 76.1 | 76.0 | 85.2 | 85.0 | 85.1 |
| SELFCONT | 82.3 | 78.5 | 73.9 | 76.1 | 84.4 | 87.4 | 85.9 |
| Models | Cases (Greedy Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in the television series The Bill, The Bill 2, The Bill and The Bill 3. <eos>Boulter was born in London, England, and grew up in the United States. He studied at the University of California, Berkeley, and the University of California, Berkeley. He was a member of the British Academy of Film and Television Arts (BAFTA) in 2001. He has appeared in several television series, including The Bill, The Bill 2, The Bill 3, The Bill 4, The Bill 5, The Bill 6, The Bill 7, The Bill 8, The Bill 9, The Bill 10, The Bill 11, The Bill 12, The Bill 13, The Bill 14, The Bill 15, The Bill 16, The Bill 17, The Bill 18, The Bill 19, The Bill 20, The Bill 21, The Bill 22, The Bill 23, The Bill 24, The Bill 25, The Bill 26, The Bill 27, The Bill 28, The Bill 29, The Bill 30, The Bill 31, The Bill 32, The Bill 33, The Bill 34, The Bill 35, The Bill 36, The Bill 37, The Bill 39, The Bill 40, The Bill 41, The Bill 42, The Bill 43, The Bill 44, The Bill 45, The Bill 46, The Bill 47, The Bill 48, The Bill 49, The Bill 50, The Bill 51, The Bill 52, The Bill 53, The Bill 54, The Bill 55, The Bill 56, The Bill 57, The Bill 58, The Bill 59, The Bill 60, The Bill 61, The Bill 62, The Bill 63, The Bill 64, The Bill 65, The Bill 66, The Bill 67, The Bill 68, The Bill 69, The Bill 70, The Bill 71, The Bill 72, The Bill 73, The Bill 74, The Bill 75, The Bill 76, The Bill 77, The Bill 78, The Bill 79, The Bill 80, The Bill 81, The Bill 82, The Bill 83, The Bill 84, The Bill 85, The Bill 86, The Bill 87, The Bill 88, The Bill 89, The Bill 90, The Bill 91, The Bill 92, The Bill 93 |
| UL | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in several films, including the comedy @-@ drama The Man Who Would Be King (2002), the science fiction film The Man Who Would Be Super (2006), the science fiction film The Man Who Would Be Super (2008), the science fiction film The Man Who Would Be Super (2012), the science fiction film The Day After Tomorrow (2013), the science fiction film The Day After Tomorrow (2014), the science fiction film The Day After Tomorrow (2015), the science fiction film The Day After Tomorrow (2016), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2019), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), The science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science Fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction filmThe Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2020), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the sciencefiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (1999), The Man Who Would Never Be King 10 (2017), The Man Who Would Never Be King 11 (2017), The Man Who Would Never Be King 12 (2017), The Man Who Would Never Be King 13 (2017), The Man Who Would Never Be King 14 (2017), The Man Who Would Never Be King 15 (2017), The Man Who Would Never Be King 16 (2017), The Man Who Would Never Be King 17 (2017), The Man Who Would Never Be King 18 (2017), The Man Who Would Never Be King 19 (2017), The Man Who Would Never Be King 20 (2017), The Man Who Would Never Be King 21 (2017), The Man Who Would Never Be King 22 (2017), The Man Who Would Never Be King 23 (2017), The Man Who Would Never Be King 24 (2017), The Man Who Would Never Be King 25 (2017), The Man Who Would Never Be King 26 (2017), The Man Who Would Never Be King 27 (2017), The Man Who Would Never Be King 28 (2017), The Man Who Would Never Be King 29 (2017), The Man Who Would Never Be King 30 (2017), The Man Who Would Never Be King 31 (2017), The Man Who Would Never Be King 32 (2017), The Man Who Would Never Be King 33 (2017), The Man Who Would Never Be King 34 (2017), The Man Who Would Never Be King 35 (2017), The Man Who Woul |
| SELFCONT | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Boulter appeared in two episodes of the British comedy @-@ drama Oz in 2001, playing an American businessman who helps an Italian woman (Joanna Cassidy) escape from her abusive husband (Michael Sheen). He also played a character in the 2004 film Heat, which was directed by Michael Medvedev. Boulter appeared in two episodes of the British comedy @-@ drama Body of Proof in 2005, playing a retired police officer who is accused of murdering his wife (Christine Fay). He appeared in two episodes of the British television drama Body of Proof in 2006, playing a retired judge who defends his client, Simon Ashdown (Stephen Fry). Boulter also appeared in the Australian television series The X Factor in 2007, playing a retired judge who defends his client, Kieman Straker (Simon Ashdown). <eos>Boulter was nominated for a Golden Globe Award for Best Actor - Motion Picture Musical or Comedy, but lost to Hugh Grant for the role of Seamus Heaney. In 2010, he appeared in the British comedy @-@ drama The X Factor, playing a retired judge who is accused of murdering his wife (Christine Fay). He also appeared in the Australian television series The X Factor in 2010, playing a retired judge who defends his client, Simon Ashdown. In 2012, he appeared in the British comedy @-@ drama Body of Proof, playing a retired judge who is accused of murdering his wife (Joanna Cassidy). He also appeared in the Australian television series The X Factor in 2013, playing a retired judge who is accused of murdering his wife (Christine Fay). <eos>Boulter has appeared in a number of films, including the 2005 film Heat, and the 2008 comedy Body of Proof. He appeared in the British television series Heat 2, and the 2009 comedy Body of Proof. In 2011, he appeared in the Australian television series The X Factor. In 2012, he appeared in the British television series The X Factor. In 2013, he appeared in the British television series The X Factor. In 2014, he appeared in the Australian television series The X Factor. In 2015, he appeared in the British television series The X Factor. In 2016, he appeared in the Australian television series The X Factor. |
| Models | Cases (Nucleus Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Since 1994, he has appeared in a wide range of productions. He has played Crocodile Hunter, a domestic terrorist in A Change Is Gonna Come. His first feature film role was in a 1983 film, The Perfect Numbers, written by Ridley Scott. A science fiction parody of the British mafia, his next film role was the horror comedy The Imperfect People. The film was a box office hit and has been named one of the greatest movies ever made. <eos>Boulter portrayed a divorcee living in Chicago, Illinois, from the start of its development in 1986, when he met Hollywood agent Kevin De Matos. The two became very close and married on August 29, 1988. He has become a great fan of James Cameron's 1999 filmAvatar. In December 2011, he starred as a roadie who stands in a line at a restaurant. <eos><eos>= Early life== <eos> Robert Boulter was born in Dundee, Scotland, on November 6, 1961, and raised in Dundee, Scotland. He attended Eales College, Dundee. He received an MBE for his work at Eales as a schoolteacher in 1973, and graduated from Dundee University in 1974. After teaching history to three young teenagers, he acted in numerous short films. <eos><eos>= Career== <eos> He made his acting debut in 1976 with the short film The Quarryman, playing the character Andre Taylor in a variety of movies and television shows. Boulter has appeared on television and cinema advertisements as well. <eos> In 1982, Boulter co @-@ starred in the video game Dr. No. The following year he made his film debut in the video game El Story, played by Terry Gilliam. He also appeared in the 1984 video game Doctor Who, starring Matthew Broderick. <eos>Boulter met producer Shane Bobbitt in 1982, and the two became friends, and Broderick made Boulter his stand @-@ up comedy partner. On the strength of their relationship, Boulter starred in several feature film projects for the American television |
| UL | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He starred as the title character in 2008's A Charlie Brown Christmas, directed by Adam Mickiewicz. Boulter has also played a variety of other roles over the years, including as the titular character on <unk> USA in 2002, in the <unk> anthology series Blood's Wild in 2007, as the protagonist on the murder mystery A Beautiful Guy in 2009, and as the love interest for the titular character on Queen's College in 2012. Boulter has also played the supporting roles of Richard VIII and Queen Elizabeth in various media including television, film, and video games. He won the Academy Award for Best Actor for his performance in King George V. <eos> Born in Carlisle, Lancashire, Boulter grew up in Manchester. He has been involved in theatre since he was eight years old. His only film role came in 2000, when he played Jack Leppard in Harry Potter and the Goblet of Fire. A second appearance came in 2001, when he played the young Severus Snape in Romeo and Juliet. He has worked on television shows including Steven Soderbergh's From Russia, with Love, Watchmen, Dawson's Creek, Ealing Studios and the British Channel 7 documentary series Ghost Hunters. <eos>Boulter has made four appearances on television: seven as Colin Rudge in Star Trek: Voyager (1995-1997); 13 as Ian Frazier in The Sarah Jane Adventures (1997-1999); and 16 as Scott Reid in Michael Bay's Robin Hood: How the Legend Was Won (1999). He was nominated for a Laurence Olivier Award in 2001. <eos><eos>= Early life== <eos> Boulter was born on 12 April 1979 in Carlisle, Lancashire, England. His father is a retired pilot. He attended Elgin Grammar School, where he earned an academic degree. After leaving school, he worked in retail at an engineering firm in Manchester. <eos>Boulter attended High School, Carlisle before transferring to King's College, Cambridge. During this time, he became involved in theatre and became a student at Queen's College, Cambridge. |
| ScaleGrad | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. In 2004 he starred as the character Sherlock Holmes in the BBC medical drama Holby City. Boulter made his screen debut with the comedy series Chitty Chitty Bang Bang, which was broadcast in 2006. He then played Lord Stanley in The Lord of the Rings trilogy and in Pirates of the Caribbean: At World's End. In 2007 he portrayed David Marmion in the film adaptation of A Midsummer Night's Dream. <eos> He appeared in the British film Downton Abbey (2008), which won him a Golden Globe Award for Best Actor. He also starred in Steven Spielberg's fantasy film, The Wrestler, as the title character. <eos> In 2010, Boulter portrayed Leonardo DiCaprio in the directorial debut of Woody Allen's The Wrestling. The film tells the story of two men who live in Los Angeles, California, who have been feuding over money and power. The film received mixed reviews from critics. <eos><eos>= Early life== <eos> Boulter was born in Leicester, England, on 3 October 1963. His father, John, is a chemical engineer and designer. His mother, Mary (née <unk>), is a member of the Scottish Episcopal Church. He grew up in Loughborough, Leicester. He was educated at Eton College and then St Thomas' School, Northamptonshire. His older brother, Liam, plays football for Leicester City. <eos>Boulter started playing football when he was four years old. After a few years, he joined Brentford Town, where he played alongside Ronnie Brown. Boulter enjoyed the game and liked the fact that his father had taught him how to play football. When he turned twelve, he moved to Leicester City, but he left the club after one season because of disciplinary problems. Boulter did not attend Brentford's youth team, but took up playing football for them. During his time at Brentford, he played for several clubs including Chesterfield, Rochdale, Oldham Athletic, St James' Park, Scarborough and Lewes. He made his debut for the club aged 15 in 1971, an |
| SELFCONT | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He also appeared in a 2000 episode of Syfy's Geneva Live. <eos>Boulter is an accomplished box @-@ office actor and one of the best known box @-@ office draws in the history of British television. Boulter appeared in over 20 films and received many awards, including Four Academy Awards, including Best Actor and Best Supporting Actor, and the BAFTA Award for Best British Actor. He was nominated for six other BAFTAs, winning three, for his work on the television series and the 1997 film. He starred in The Bill in 2001 and again in 2002. In 2005, he appeared in The Gleason Room, the 2005 science fiction film about rediscovery of woolly alien relics, and in the 2006 biographical drama Brand New Eyes. In 2010, he starred in the stage production of Minor Threat and the 2007 psychological thriller Victoria's Secret. <eos>Boulter's stage and film career began with his performance in the 1997 romantic comedy Hamlet. In 2000, he was cast as Jonathan Simeone in the German @-@ language dramatisation of French novelist Raymond Lebowski's epic play, The Professionals. He took on the role of "Troy", an obsessive person who attempts to prove himself to courtiers. Although he enjoyed playing Troy, he took "enormous risks", in the words of theatre critic Graham McCann, who wrote that " there was nothing to lose in playing a man like Troy. " He co @-@ starred in The Professionals with Julianne Moore and Kim Novak. He portrayed the criminal Tammi Martineau in the 2004 biographical film Asterisk and appeared in several films and television shows. In 2005, he starred as Garth Snow in the Fox crime drama Dangerous Liaisons. <eos>Boulter is known for his film work in Hungary and abroad. He has also worked with Brandon Thomas and Sacha Baron Cohen. In 2011, he was nominated for a Laurence Olivier Award for Best Actor, with Olivier in the role of General Herculaneum. In 2012, he starred in The Phantom of the Opera, which opened at the BBC2 Leicester Square Theatre, with much of the stage cast from his earlier work |
| Training Sets | Doriginal | Drandom | Dnorept |
| Ratios (%) ↓ | 60.42 | 96.04 | 1.67 |
| Models | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 |
| Greedy | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 2.55 | 3.29 | 41.23 | 70.18 | 83.28 | 19.27 | 23.95 | 1.76 | 0.61 | 71.08 | 87.20 | 89.43 | 9.61 | 11.40 |
| UL | 3.20 | 7.16 | 33.91 | 61.90 | 76.89 | 25.13 | 31.90 | 2.01 | 1.63 | 59.43 | 81.63 | 85.89 | 11.66 | 14.30 |
| ScaleGrad | 4.61 | 7.66 | 29.82 | 50.69 | 66.14 | 36.96 | 47.34 | 2.87 | 11.17 | 52.29 | 69.53 | 76.16 | 18.16 | 24.40 |
| SELFCONT | 6.47 | 17.34 | 23.29 | 39.41 | 62.46 | 46.71 | 57.66 | 3.30 | 20.05 | 35.13 | 53.69 | 74.09 | 23.30 | 31.52 |
| Nucleus | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 20.66 | 21.09 | 19.40 | 30.22 | 48.11 | 71.92 | 84.75 | 18.68 | 88.54 | 20.95 | 32.53 | 48.87 | 60.38 | 81.55 |
| UL | 15.54 | 21.78 | 18.45 | 29.57 | 46.69 | 69.63 | 82.87 | 19.39 | 81.49 | 18.36 | 27.98 | 42.65 | 63.92 | 82.93 |
| ScaleGrad | 12.41 | 25.69 | 18.59 | 29.24 | 45.19 | 66.35 | 80.23 | 14.14 | 77.82 | 18.62 | 27.80 | 41.22 | 56.74 | 77.27 |
| SELFCONT | 19.02 | 34.37 | 16.45 | 26.47 | 45.10 | 72.02 | 84.78 | 19.86 | 89.84 | 17.56 | 26.98 | 43.39 | 63.33 | 83.51 |
| Ground Truth | 18.31 | 100 | 17.38 | 27.92 | 46.29 | 72.34 | 84.20 | 24.01 | 100 | 16.36 | 26.47 | 42.30 | 74.49 | 90.01 |
| Datasets | |Train| | |Validation| | |Test| | Avg. Len |
| Wikitext-103 | 201,632 | 448 | 480 | 512 |
| WritingPrompts | 272,600 | 15,620 | 15,138 | 439 |
| Training Steps | Cases | |
| 100 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. The first, and the " . The " The " <eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos></eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos> | |
| 1500 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of the building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the buildings, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building | He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building |
| 6000 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. She had begun in 1962, was built a series of New York Times in 1962 - 1938 after producing the Laidlay, and described as well as well as The Lasttunk's assistant ... ] named". She later explained : "[ 500 people who did not be turned off for me to me to me. She decided to his own song " As Rocham R. Nails , saying". She knew it 's script was going to make me to live film, and I mean it was through me to get it 't know what we 't know what we want to make me . On 17 November 1953 , and it was nominated for me. He had done what he wanted the BBC at a period at her following year because it appearance at the mid @-@ selling artist with her singing . She was nominated for the background scene in 1962 . She spent four days after getting part of the public appearances at the war appearance , California , and ] base . Aftermath Medie Glam artist of the summer May 1967 , New York Times , and the war as the Star Wars franchise . He began to use the National Association ( from his staff ; it was included Star ) , including the Lyds house east , and I was the West Virginia Tech back to the war and Mennon ; there . She developed by <eos> William Peninsular League <eos>=<eos> The script , 2004 . The script was named after the North America for the LAM passed . The script was commissioned to the American co @-@ person to produce producer ( present , taking place of the mid @-@ old , and the mid @-@ old @-@ old film , The Next Generation . The New York Times , having won the 4th birthday of the 4 , in the 4 million viewers . This was announced that it was cut of the media . The Elder Scrolls IV of the production , in East Coaster and The company entitled The Next Generation . <eos> For example , including the war , having performed on 6 , having released in East Coast Division . Upon its crew became a series of the produce |
| Training Set | Cases | |
| Doriginal | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has appeared in several films, including the television series The Bill, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. <eos>Boulter was born in London, England, on 23 May 1986. He is the third child of actor and actress Robert Boulter and his wife, Susan. He is the third of five children born to his wife Susan and their three children, Robert, Roberta, and Roberta. Robert Boulter's father, Robert Boulter, was a film director and producer. He was the first actor to be cast in a film role, and the first to be cast in a television series. He was also the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a tele-. tical series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was then cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a tele-. tical series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He were cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. Hewas cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He wasCast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast ina Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Sereas. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series | He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast ina television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a Television Series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cist in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the firsst actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series.Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast ina television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. |
| Models | Wikitext-103 | WritingPrompts |
| MLE | 0.9 | 0.9 |
| UL | 0.7 | 0.8 |
| ScaleGrad | 0.5 | 0.6 |
| SELFCONT | 0.6 | 0.7 |
| Models | Acc | Repetitive | Non-Repetitive | ||||
| P | R | F1 | P | R | F1 | ||
| MLE | 78.9 | 67.4 | 87.0 | 75.9 | 90.2 | 73.9 | 81.2 |
| UL | 80.5 | 70.9 | 83.0 | 76.5 | 88.2 | 78.9 | 83.3 |
| ScaleGrad | 81.6 | 75.9 | 76.1 | 76.0 | 85.2 | 85.0 | 85.1 |
| SELFCONT | 82.3 | 78.5 | 73.9 | 76.1 | 84.4 | 87.4 | 85.9 |
| Models | Cases (Greedy Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in the television series The Bill, The Bill 2, The Bill and The Bill 3. <eos>Boulter was born in London, England, and grew up in the United States. He studied at the University of California, Berkeley, and the University of California, Berkeley. He was a member of the British Academy of Film and Television Arts (BAFTA) in 2001. He has appeared in several television series, including The Bill, The Bill 2, The Bill 3, The Bill 4, The Bill 5, The Bill 6, The Bill 7, The Bill 8, The Bill 9, The Bill 10, The Bill 11, The Bill 12, The Bill 13, The Bill 14, The Bill 15, The Bill 16, The Bill 17, The Bill 18, The Bill 19, The Bill 20, The Bill 21, The Bill 22, The Bill 23, The Bill 24, The Bill 25, The Bill 26, The Bill 27, The Bill 28, The Bill 29, The Bill 30, The Bill 31, The Bill 32, The Bill 33, The Bill 34, The Bill 35, The Bill 36, The Bill 37, The Bill 39, The Bill 40, The Bill 41, The Bill 42, The Bill 43, The Bill 44, The Bill 45, The Bill 46, The Bill 47, The Bill 48, The Bill 49, The Bill 50, The Bill 51, The Bill 52, The Bill 53, The Bill 54, The Bill 55, The Bill 56, The Bill 57, The Bill 58, The Bill 59, The Bill 60, The Bill 61, The Bill 62, The Bill 63, The Bill 64, The Bill 65, The Bill 66, The Bill 67, The Bill 68, The Bill 69, The Bill 70, The Bill 71, The Bill 72, The Bill 73, The Bill 74, The Bill 75, The Bill 76, The Bill 77, The Bill 78, The Bill 79, The Bill 80, The Bill 81, The Bill 82, The Bill 83, The Bill 84, The Bill 85, The Bill 86, The Bill 87, The Bill 88, The Bill 89, The Bill 90, The Bill 91, The Bill 92, The Bill 93 |
| UL | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in several films, including the comedy @-@ drama The Man Who Would Be King (2002), the science fiction film The Man Who Would Be Super (2006), the science fiction film The Man Who Would Be Super (2008), the science fiction film The Man Who Would Be Super (2012), the science fiction film The Day After Tomorrow (2013), the science fiction film The Day After Tomorrow (2014), the science fiction film The Day After Tomorrow (2015), the science fiction film The Day After Tomorrow (2016), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2019), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), The science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science Fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction filmThe Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2020), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the sciencefiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (1999), The Man Who Would Never Be King 10 (2017), The Man Who Would Never Be King 11 (2017), The Man Who Would Never Be King 12 (2017), The Man Who Would Never Be King 13 (2017), The Man Who Would Never Be King 14 (2017), The Man Who Would Never Be King 15 (2017), The Man Who Would Never Be King 16 (2017), The Man Who Would Never Be King 17 (2017), The Man Who Would Never Be King 18 (2017), The Man Who Would Never Be King 19 (2017), The Man Who Would Never Be King 20 (2017), The Man Who Would Never Be King 21 (2017), The Man Who Would Never Be King 22 (2017), The Man Who Would Never Be King 23 (2017), The Man Who Would Never Be King 24 (2017), The Man Who Would Never Be King 25 (2017), The Man Who Would Never Be King 26 (2017), The Man Who Would Never Be King 27 (2017), The Man Who Would Never Be King 28 (2017), The Man Who Would Never Be King 29 (2017), The Man Who Would Never Be King 30 (2017), The Man Who Would Never Be King 31 (2017), The Man Who Would Never Be King 32 (2017), The Man Who Would Never Be King 33 (2017), The Man Who Would Never Be King 34 (2017), The Man Who Would Never Be King 35 (2017), The Man Who Woul |
| SELFCONT | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Boulter appeared in two episodes of the British comedy @-@ drama Oz in 2001, playing an American businessman who helps an Italian woman (Joanna Cassidy) escape from her abusive husband (Michael Sheen). He also played a character in the 2004 film Heat, which was directed by Michael Medvedev. Boulter appeared in two episodes of the British comedy @-@ drama Body of Proof in 2005, playing a retired police officer who is accused of murdering his wife (Christine Fay). He appeared in two episodes of the British television drama Body of Proof in 2006, playing a retired judge who defends his client, Simon Ashdown (Stephen Fry). Boulter also appeared in the Australian television series The X Factor in 2007, playing a retired judge who defends his client, Kieman Straker (Simon Ashdown). <eos>Boulter was nominated for a Golden Globe Award for Best Actor - Motion Picture Musical or Comedy, but lost to Hugh Grant for the role of Seamus Heaney. In 2010, he appeared in the British comedy @-@ drama The X Factor, playing a retired judge who is accused of murdering his wife (Christine Fay). He also appeared in the Australian television series The X Factor in 2010, playing a retired judge who defends his client, Simon Ashdown. In 2012, he appeared in the British comedy @-@ drama Body of Proof, playing a retired judge who is accused of murdering his wife (Joanna Cassidy). He also appeared in the Australian television series The X Factor in 2013, playing a retired judge who is accused of murdering his wife (Christine Fay). <eos>Boulter has appeared in a number of films, including the 2005 film Heat, and the 2008 comedy Body of Proof. He appeared in the British television series Heat 2, and the 2009 comedy Body of Proof. In 2011, he appeared in the Australian television series The X Factor. In 2012, he appeared in the British television series The X Factor. In 2013, he appeared in the British television series The X Factor. In 2014, he appeared in the Australian television series The X Factor. In 2015, he appeared in the British television series The X Factor. In 2016, he appeared in the Australian television series The X Factor. |
| Models | Cases (Nucleus Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Since 1994, he has appeared in a wide range of productions. He has played Crocodile Hunter, a domestic terrorist in A Change Is Gonna Come. His first feature film role was in a 1983 film, The Perfect Numbers, written by Ridley Scott. A science fiction parody of the British mafia, his next film role was the horror comedy The Imperfect People. The film was a box office hit and has been named one of the greatest movies ever made. <eos>Boulter portrayed a divorcee living in Chicago, Illinois, from the start of its development in 1986, when he met Hollywood agent Kevin De Matos. The two became very close and married on August 29, 1988. He has become a great fan of James Cameron's 1999 filmAvatar. In December 2011, he starred as a roadie who stands in a line at a restaurant. <eos><eos>= Early life== <eos> Robert Boulter was born in Dundee, Scotland, on November 6, 1961, and raised in Dundee, Scotland. He attended Eales College, Dundee. He received an MBE for his work at Eales as a schoolteacher in 1973, and graduated from Dundee University in 1974. After teaching history to three young teenagers, he acted in numerous short films. <eos><eos>= Career== <eos> He made his acting debut in 1976 with the short film The Quarryman, playing the character Andre Taylor in a variety of movies and television shows. Boulter has appeared on television and cinema advertisements as well. <eos> In 1982, Boulter co @-@ starred in the video game Dr. No. The following year he made his film debut in the video game El Story, played by Terry Gilliam. He also appeared in the 1984 video game Doctor Who, starring Matthew Broderick. <eos>Boulter met producer Shane Bobbitt in 1982, and the two became friends, and Broderick made Boulter his stand @-@ up comedy partner. On the strength of their relationship, Boulter starred in several feature film projects for the American television |
| UL | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He starred as the title character in 2008's A Charlie Brown Christmas, directed by Adam Mickiewicz. Boulter has also played a variety of other roles over the years, including as the titular character on <unk> USA in 2002, in the <unk> anthology series Blood's Wild in 2007, as the protagonist on the murder mystery A Beautiful Guy in 2009, and as the love interest for the titular character on Queen's College in 2012. Boulter has also played the supporting roles of Richard VIII and Queen Elizabeth in various media including television, film, and video games. He won the Academy Award for Best Actor for his performance in King George V. <eos> Born in Carlisle, Lancashire, Boulter grew up in Manchester. He has been involved in theatre since he was eight years old. His only film role came in 2000, when he played Jack Leppard in Harry Potter and the Goblet of Fire. A second appearance came in 2001, when he played the young Severus Snape in Romeo and Juliet. He has worked on television shows including Steven Soderbergh's From Russia, with Love, Watchmen, Dawson's Creek, Ealing Studios and the British Channel 7 documentary series Ghost Hunters. <eos>Boulter has made four appearances on television: seven as Colin Rudge in Star Trek: Voyager (1995-1997); 13 as Ian Frazier in The Sarah Jane Adventures (1997-1999); and 16 as Scott Reid in Michael Bay's Robin Hood: How the Legend Was Won (1999). He was nominated for a Laurence Olivier Award in 2001. <eos><eos>= Early life== <eos> Boulter was born on 12 April 1979 in Carlisle, Lancashire, England. His father is a retired pilot. He attended Elgin Grammar School, where he earned an academic degree. After leaving school, he worked in retail at an engineering firm in Manchester. <eos>Boulter attended High School, Carlisle before transferring to King's College, Cambridge. During this time, he became involved in theatre and became a student at Queen's College, Cambridge. |
| ScaleGrad | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. In 2004 he starred as the character Sherlock Holmes in the BBC medical drama Holby City. Boulter made his screen debut with the comedy series Chitty Chitty Bang Bang, which was broadcast in 2006. He then played Lord Stanley in The Lord of the Rings trilogy and in Pirates of the Caribbean: At World's End. In 2007 he portrayed David Marmion in the film adaptation of A Midsummer Night's Dream. <eos> He appeared in the British film Downton Abbey (2008), which won him a Golden Globe Award for Best Actor. He also starred in Steven Spielberg's fantasy film, The Wrestler, as the title character. <eos> In 2010, Boulter portrayed Leonardo DiCaprio in the directorial debut of Woody Allen's The Wrestling. The film tells the story of two men who live in Los Angeles, California, who have been feuding over money and power. The film received mixed reviews from critics. <eos><eos>= Early life== <eos> Boulter was born in Leicester, England, on 3 October 1963. His father, John, is a chemical engineer and designer. His mother, Mary (née <unk>), is a member of the Scottish Episcopal Church. He grew up in Loughborough, Leicester. He was educated at Eton College and then St Thomas' School, Northamptonshire. His older brother, Liam, plays football for Leicester City. <eos>Boulter started playing football when he was four years old. After a few years, he joined Brentford Town, where he played alongside Ronnie Brown. Boulter enjoyed the game and liked the fact that his father had taught him how to play football. When he turned twelve, he moved to Leicester City, but he left the club after one season because of disciplinary problems. Boulter did not attend Brentford's youth team, but took up playing football for them. During his time at Brentford, he played for several clubs including Chesterfield, Rochdale, Oldham Athletic, St James' Park, Scarborough and Lewes. He made his debut for the club aged 15 in 1971, an |
| SELFCONT | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He also appeared in a 2000 episode of Syfy's Geneva Live. <eos>Boulter is an accomplished box @-@ office actor and one of the best known box @-@ office draws in the history of British television. Boulter appeared in over 20 films and received many awards, including Four Academy Awards, including Best Actor and Best Supporting Actor, and the BAFTA Award for Best British Actor. He was nominated for six other BAFTAs, winning three, for his work on the television series and the 1997 film. He starred in The Bill in 2001 and again in 2002. In 2005, he appeared in The Gleason Room, the 2005 science fiction film about rediscovery of woolly alien relics, and in the 2006 biographical drama Brand New Eyes. In 2010, he starred in the stage production of Minor Threat and the 2007 psychological thriller Victoria's Secret. <eos>Boulter's stage and film career began with his performance in the 1997 romantic comedy Hamlet. In 2000, he was cast as Jonathan Simeone in the German @-@ language dramatisation of French novelist Raymond Lebowski's epic play, The Professionals. He took on the role of "Troy", an obsessive person who attempts to prove himself to courtiers. Although he enjoyed playing Troy, he took "enormous risks", in the words of theatre critic Graham McCann, who wrote that " there was nothing to lose in playing a man like Troy. " He co @-@ starred in The Professionals with Julianne Moore and Kim Novak. He portrayed the criminal Tammi Martineau in the 2004 biographical film Asterisk and appeared in several films and television shows. In 2005, he starred as Garth Snow in the Fox crime drama Dangerous Liaisons. <eos>Boulter is known for his film work in Hungary and abroad. He has also worked with Brandon Thomas and Sacha Baron Cohen. In 2011, he was nominated for a Laurence Olivier Award for Best Actor, with Olivier in the role of General Herculaneum. In 2012, he starred in The Phantom of the Opera, which opened at the BBC2 Leicester Square Theatre, with much of the stage cast from his earlier work |
| Training Sets | Doriginal | Drandom | Dnorept |
| Ratios (%) ↓ | 60.42 | 96.04 | 1.67 |
| Models | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 |
| Greedy | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 2.55 | 3.29 | 41.23 | 70.18 | 83.28 | 19.27 | 23.95 | 1.76 | 0.61 | 71.08 | 87.20 | 89.43 | 9.61 | 11.40 |
| UL | 3.20 | 7.16 | 33.91 | 61.90 | 76.89 | 25.13 | 31.90 | 2.01 | 1.63 | 59.43 | 81.63 | 85.89 | 11.66 | 14.30 |
| ScaleGrad | 4.61 | 7.66 | 29.82 | 50.69 | 66.14 | 36.96 | 47.34 | 2.87 | 11.17 | 52.29 | 69.53 | 76.16 | 18.16 | 24.40 |
| SELFCONT | 6.47 | 17.34 | 23.29 | 39.41 | 62.46 | 46.71 | 57.66 | 3.30 | 20.05 | 35.13 | 53.69 | 74.09 | 23.30 | 31.52 |
| Nucleus | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 20.66 | 21.09 | 19.40 | 30.22 | 48.11 | 71.92 | 84.75 | 18.68 | 88.54 | 20.95 | 32.53 | 48.87 | 60.38 | 81.55 |
| UL | 15.54 | 21.78 | 18.45 | 29.57 | 46.69 | 69.63 | 82.87 | 19.39 | 81.49 | 18.36 | 27.98 | 42.65 | 63.92 | 82.93 |
| ScaleGrad | 12.41 | 25.69 | 18.59 | 29.24 | 45.19 | 66.35 | 80.23 | 14.14 | 77.82 | 18.62 | 27.80 | 41.22 | 56.74 | 77.27 |
| SELFCONT | 19.02 | 34.37 | 16.45 | 26.47 | 45.10 | 72.02 | 84.78 | 19.86 | 89.84 | 17.56 | 26.98 | 43.39 | 63.33 | 83.51 |
| Ground Truth | 18.31 | 100 | 17.38 | 27.92 | 46.29 | 72.34 | 84.20 | 24.01 | 100 | 16.36 | 26.47 | 42.30 | 74.49 | 90.01 |
| Datasets | |Train| | |Validation| | |Test| | Avg. Len |
| Wikitext-103 | 201,632 | 448 | 480 | 512 |
| WritingPrompts | 272,600 | 15,620 | 15,138 | 439 |
| Training Steps | Cases | |
| 100 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. The first, and the " . The " The " <eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos></eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos> | |
| 1500 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of the building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the buildings, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building | He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building |
| 6000 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. She had begun in 1962, was built a series of New York Times in 1962 - 1938 after producing the Laidlay, and described as well as well as The Lasttunk's assistant ... ] named". She later explained : "[ 500 people who did not be turned off for me to me to me. She decided to his own song " As Rocham R. Nails , saying". She knew it 's script was going to make me to live film, and I mean it was through me to get it 't know what we 't know what we want to make me . On 17 November 1953 , and it was nominated for me. He had done what he wanted the BBC at a period at her following year because it appearance at the mid @-@ selling artist with her singing . She was nominated for the background scene in 1962 . She spent four days after getting part of the public appearances at the war appearance , California , and ] base . Aftermath Medie Glam artist of the summer May 1967 , New York Times , and the war as the Star Wars franchise . He began to use the National Association ( from his staff ; it was included Star ) , including the Lyds house east , and I was the West Virginia Tech back to the war and Mennon ; there . She developed by <eos> William Peninsular League <eos>=<eos> The script , 2004 . The script was named after the North America for the LAM passed . The script was commissioned to the American co @-@ person to produce producer ( present , taking place of the mid @-@ old , and the mid @-@ old @-@ old film , The Next Generation . The New York Times , having won the 4th birthday of the 4 , in the 4 million viewers . This was announced that it was cut of the media . The Elder Scrolls IV of the production , in East Coaster and The company entitled The Next Generation . <eos> For example , including the war , having performed on 6 , having released in East Coast Division . Upon its crew became a series of the produce |
| Training Set | Cases | |
| Doriginal | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has appeared in several films, including the television series The Bill, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. <eos>Boulter was born in London, England, on 23 May 1986. He is the third child of actor and actress Robert Boulter and his wife, Susan. He is the third of five children born to his wife Susan and their three children, Robert, Roberta, and Roberta. Robert Boulter's father, Robert Boulter, was a film director and producer. He was the first actor to be cast in a film role, and the first to be cast in a television series. He was also the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a tele-. tical series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was then cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a tele-. tical series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He were cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. Hewas cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He wasCast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast ina Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Sereas. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series | He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast ina television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a Television Series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cist in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the firsst actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series.Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast ina television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. |
| Models | Wikitext-103 | WritingPrompts |
| MLE | 0.9 | 0.9 |
| UL | 0.7 | 0.8 |
| ScaleGrad | 0.5 | 0.6 |
| SELFCONT | 0.6 | 0.7 |
| Models | Acc | Repetitive | Non-Repetitive | ||||
| P | R | F1 | P | R | F1 | ||
| MLE | 78.9 | 67.4 | 87.0 | 75.9 | 90.2 | 73.9 | 81.2 |
| UL | 80.5 | 70.9 | 83.0 | 76.5 | 88.2 | 78.9 | 83.3 |
| ScaleGrad | 81.6 | 75.9 | 76.1 | 76.0 | 85.2 | 85.0 | 85.1 |
| SELFCONT | 82.3 | 78.5 | 73.9 | 76.1 | 84.4 | 87.4 | 85.9 |
| Models | Cases (Greedy Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in the television series The Bill, The Bill 2, The Bill and The Bill 3. <eos>Boulter was born in London, England, and grew up in the United States. He studied at the University of California, Berkeley, and the University of California, Berkeley. He was a member of the British Academy of Film and Television Arts (BAFTA) in 2001. He has appeared in several television series, including The Bill, The Bill 2, The Bill 3, The Bill 4, The Bill 5, The Bill 6, The Bill 7, The Bill 8, The Bill 9, The Bill 10, The Bill 11, The Bill 12, The Bill 13, The Bill 14, The Bill 15, The Bill 16, The Bill 17, The Bill 18, The Bill 19, The Bill 20, The Bill 21, The Bill 22, The Bill 23, The Bill 24, The Bill 25, The Bill 26, The Bill 27, The Bill 28, The Bill 29, The Bill 30, The Bill 31, The Bill 32, The Bill 33, The Bill 34, The Bill 35, The Bill 36, The Bill 37, The Bill 39, The Bill 40, The Bill 41, The Bill 42, The Bill 43, The Bill 44, The Bill 45, The Bill 46, The Bill 47, The Bill 48, The Bill 49, The Bill 50, The Bill 51, The Bill 52, The Bill 53, The Bill 54, The Bill 55, The Bill 56, The Bill 57, The Bill 58, The Bill 59, The Bill 60, The Bill 61, The Bill 62, The Bill 63, The Bill 64, The Bill 65, The Bill 66, The Bill 67, The Bill 68, The Bill 69, The Bill 70, The Bill 71, The Bill 72, The Bill 73, The Bill 74, The Bill 75, The Bill 76, The Bill 77, The Bill 78, The Bill 79, The Bill 80, The Bill 81, The Bill 82, The Bill 83, The Bill 84, The Bill 85, The Bill 86, The Bill 87, The Bill 88, The Bill 89, The Bill 90, The Bill 91, The Bill 92, The Bill 93 |
| UL | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in several films, including the comedy @-@ drama The Man Who Would Be King (2002), the science fiction film The Man Who Would Be Super (2006), the science fiction film The Man Who Would Be Super (2008), the science fiction film The Man Who Would Be Super (2012), the science fiction film The Day After Tomorrow (2013), the science fiction film The Day After Tomorrow (2014), the science fiction film The Day After Tomorrow (2015), the science fiction film The Day After Tomorrow (2016), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2019), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), The science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science Fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction filmThe Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2020), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the sciencefiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (1999), The Man Who Would Never Be King 10 (2017), The Man Who Would Never Be King 11 (2017), The Man Who Would Never Be King 12 (2017), The Man Who Would Never Be King 13 (2017), The Man Who Would Never Be King 14 (2017), The Man Who Would Never Be King 15 (2017), The Man Who Would Never Be King 16 (2017), The Man Who Would Never Be King 17 (2017), The Man Who Would Never Be King 18 (2017), The Man Who Would Never Be King 19 (2017), The Man Who Would Never Be King 20 (2017), The Man Who Would Never Be King 21 (2017), The Man Who Would Never Be King 22 (2017), The Man Who Would Never Be King 23 (2017), The Man Who Would Never Be King 24 (2017), The Man Who Would Never Be King 25 (2017), The Man Who Would Never Be King 26 (2017), The Man Who Would Never Be King 27 (2017), The Man Who Would Never Be King 28 (2017), The Man Who Would Never Be King 29 (2017), The Man Who Would Never Be King 30 (2017), The Man Who Would Never Be King 31 (2017), The Man Who Would Never Be King 32 (2017), The Man Who Would Never Be King 33 (2017), The Man Who Would Never Be King 34 (2017), The Man Who Would Never Be King 35 (2017), The Man Who Woul |
| SELFCONT | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Boulter appeared in two episodes of the British comedy @-@ drama Oz in 2001, playing an American businessman who helps an Italian woman (Joanna Cassidy) escape from her abusive husband (Michael Sheen). He also played a character in the 2004 film Heat, which was directed by Michael Medvedev. Boulter appeared in two episodes of the British comedy @-@ drama Body of Proof in 2005, playing a retired police officer who is accused of murdering his wife (Christine Fay). He appeared in two episodes of the British television drama Body of Proof in 2006, playing a retired judge who defends his client, Simon Ashdown (Stephen Fry). Boulter also appeared in the Australian television series The X Factor in 2007, playing a retired judge who defends his client, Kieman Straker (Simon Ashdown). <eos>Boulter was nominated for a Golden Globe Award for Best Actor - Motion Picture Musical or Comedy, but lost to Hugh Grant for the role of Seamus Heaney. In 2010, he appeared in the British comedy @-@ drama The X Factor, playing a retired judge who is accused of murdering his wife (Christine Fay). He also appeared in the Australian television series The X Factor in 2010, playing a retired judge who defends his client, Simon Ashdown. In 2012, he appeared in the British comedy @-@ drama Body of Proof, playing a retired judge who is accused of murdering his wife (Joanna Cassidy). He also appeared in the Australian television series The X Factor in 2013, playing a retired judge who is accused of murdering his wife (Christine Fay). <eos>Boulter has appeared in a number of films, including the 2005 film Heat, and the 2008 comedy Body of Proof. He appeared in the British television series Heat 2, and the 2009 comedy Body of Proof. In 2011, he appeared in the Australian television series The X Factor. In 2012, he appeared in the British television series The X Factor. In 2013, he appeared in the British television series The X Factor. In 2014, he appeared in the Australian television series The X Factor. In 2015, he appeared in the British television series The X Factor. In 2016, he appeared in the Australian television series The X Factor. |
| Models | Cases (Nucleus Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Since 1994, he has appeared in a wide range of productions. He has played Crocodile Hunter, a domestic terrorist in A Change Is Gonna Come. His first feature film role was in a 1983 film, The Perfect Numbers, written by Ridley Scott. A science fiction parody of the British mafia, his next film role was the horror comedy The Imperfect People. The film was a box office hit and has been named one of the greatest movies ever made. <eos>Boulter portrayed a divorcee living in Chicago, Illinois, from the start of its development in 1986, when he met Hollywood agent Kevin De Matos. The two became very close and married on August 29, 1988. He has become a great fan of James Cameron's 1999 filmAvatar. In December 2011, he starred as a roadie who stands in a line at a restaurant. <eos><eos>= Early life== <eos> Robert Boulter was born in Dundee, Scotland, on November 6, 1961, and raised in Dundee, Scotland. He attended Eales College, Dundee. He received an MBE for his work at Eales as a schoolteacher in 1973, and graduated from Dundee University in 1974. After teaching history to three young teenagers, he acted in numerous short films. <eos><eos>= Career== <eos> He made his acting debut in 1976 with the short film The Quarryman, playing the character Andre Taylor in a variety of movies and television shows. Boulter has appeared on television and cinema advertisements as well. <eos> In 1982, Boulter co @-@ starred in the video game Dr. No. The following year he made his film debut in the video game El Story, played by Terry Gilliam. He also appeared in the 1984 video game Doctor Who, starring Matthew Broderick. <eos>Boulter met producer Shane Bobbitt in 1982, and the two became friends, and Broderick made Boulter his stand @-@ up comedy partner. On the strength of their relationship, Boulter starred in several feature film projects for the American television |
| UL | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He starred as the title character in 2008's A Charlie Brown Christmas, directed by Adam Mickiewicz. Boulter has also played a variety of other roles over the years, including as the titular character on <unk> USA in 2002, in the <unk> anthology series Blood's Wild in 2007, as the protagonist on the murder mystery A Beautiful Guy in 2009, and as the love interest for the titular character on Queen's College in 2012. Boulter has also played the supporting roles of Richard VIII and Queen Elizabeth in various media including television, film, and video games. He won the Academy Award for Best Actor for his performance in King George V. <eos> Born in Carlisle, Lancashire, Boulter grew up in Manchester. He has been involved in theatre since he was eight years old. His only film role came in 2000, when he played Jack Leppard in Harry Potter and the Goblet of Fire. A second appearance came in 2001, when he played the young Severus Snape in Romeo and Juliet. He has worked on television shows including Steven Soderbergh's From Russia, with Love, Watchmen, Dawson's Creek, Ealing Studios and the British Channel 7 documentary series Ghost Hunters. <eos>Boulter has made four appearances on television: seven as Colin Rudge in Star Trek: Voyager (1995-1997); 13 as Ian Frazier in The Sarah Jane Adventures (1997-1999); and 16 as Scott Reid in Michael Bay's Robin Hood: How the Legend Was Won (1999). He was nominated for a Laurence Olivier Award in 2001. <eos><eos>= Early life== <eos> Boulter was born on 12 April 1979 in Carlisle, Lancashire, England. His father is a retired pilot. He attended Elgin Grammar School, where he earned an academic degree. After leaving school, he worked in retail at an engineering firm in Manchester. <eos>Boulter attended High School, Carlisle before transferring to King's College, Cambridge. During this time, he became involved in theatre and became a student at Queen's College, Cambridge. |
| ScaleGrad | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. In 2004 he starred as the character Sherlock Holmes in the BBC medical drama Holby City. Boulter made his screen debut with the comedy series Chitty Chitty Bang Bang, which was broadcast in 2006. He then played Lord Stanley in The Lord of the Rings trilogy and in Pirates of the Caribbean: At World's End. In 2007 he portrayed David Marmion in the film adaptation of A Midsummer Night's Dream. <eos> He appeared in the British film Downton Abbey (2008), which won him a Golden Globe Award for Best Actor. He also starred in Steven Spielberg's fantasy film, The Wrestler, as the title character. <eos> In 2010, Boulter portrayed Leonardo DiCaprio in the directorial debut of Woody Allen's The Wrestling. The film tells the story of two men who live in Los Angeles, California, who have been feuding over money and power. The film received mixed reviews from critics. <eos><eos>= Early life== <eos> Boulter was born in Leicester, England, on 3 October 1963. His father, John, is a chemical engineer and designer. His mother, Mary (née <unk>), is a member of the Scottish Episcopal Church. He grew up in Loughborough, Leicester. He was educated at Eton College and then St Thomas' School, Northamptonshire. His older brother, Liam, plays football for Leicester City. <eos>Boulter started playing football when he was four years old. After a few years, he joined Brentford Town, where he played alongside Ronnie Brown. Boulter enjoyed the game and liked the fact that his father had taught him how to play football. When he turned twelve, he moved to Leicester City, but he left the club after one season because of disciplinary problems. Boulter did not attend Brentford's youth team, but took up playing football for them. During his time at Brentford, he played for several clubs including Chesterfield, Rochdale, Oldham Athletic, St James' Park, Scarborough and Lewes. He made his debut for the club aged 15 in 1971, an |
| SELFCONT | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He also appeared in a 2000 episode of Syfy's Geneva Live. <eos>Boulter is an accomplished box @-@ office actor and one of the best known box @-@ office draws in the history of British television. Boulter appeared in over 20 films and received many awards, including Four Academy Awards, including Best Actor and Best Supporting Actor, and the BAFTA Award for Best British Actor. He was nominated for six other BAFTAs, winning three, for his work on the television series and the 1997 film. He starred in The Bill in 2001 and again in 2002. In 2005, he appeared in The Gleason Room, the 2005 science fiction film about rediscovery of woolly alien relics, and in the 2006 biographical drama Brand New Eyes. In 2010, he starred in the stage production of Minor Threat and the 2007 psychological thriller Victoria's Secret. <eos>Boulter's stage and film career began with his performance in the 1997 romantic comedy Hamlet. In 2000, he was cast as Jonathan Simeone in the German @-@ language dramatisation of French novelist Raymond Lebowski's epic play, The Professionals. He took on the role of "Troy", an obsessive person who attempts to prove himself to courtiers. Although he enjoyed playing Troy, he took "enormous risks", in the words of theatre critic Graham McCann, who wrote that " there was nothing to lose in playing a man like Troy. " He co @-@ starred in The Professionals with Julianne Moore and Kim Novak. He portrayed the criminal Tammi Martineau in the 2004 biographical film Asterisk and appeared in several films and television shows. In 2005, he starred as Garth Snow in the Fox crime drama Dangerous Liaisons. <eos>Boulter is known for his film work in Hungary and abroad. He has also worked with Brandon Thomas and Sacha Baron Cohen. In 2011, he was nominated for a Laurence Olivier Award for Best Actor, with Olivier in the role of General Herculaneum. In 2012, he starred in The Phantom of the Opera, which opened at the BBC2 Leicester Square Theatre, with much of the stage cast from his earlier work |
| Training Sets | Doriginal | Drandom | Dnorept |
| Ratios (%) ↓ | 60.42 | 96.04 | 1.67 |
| Models | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 | PPL | MAUVE | R-16 | R-32 | R-128 | D-3 | D-4 |
| Greedy | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 2.55 | 3.29 | 41.23 | 70.18 | 83.28 | 19.27 | 23.95 | 1.76 | 0.61 | 71.08 | 87.20 | 89.43 | 9.61 | 11.40 |
| UL | 3.20 | 7.16 | 33.91 | 61.90 | 76.89 | 25.13 | 31.90 | 2.01 | 1.63 | 59.43 | 81.63 | 85.89 | 11.66 | 14.30 |
| ScaleGrad | 4.61 | 7.66 | 29.82 | 50.69 | 66.14 | 36.96 | 47.34 | 2.87 | 11.17 | 52.29 | 69.53 | 76.16 | 18.16 | 24.40 |
| SELFCONT | 6.47 | 17.34 | 23.29 | 39.41 | 62.46 | 46.71 | 57.66 | 3.30 | 20.05 | 35.13 | 53.69 | 74.09 | 23.30 | 31.52 |
| Nucleus | Dataset: Wikitext-103 | Dataset: WritingPrompts | ||||||||||||
| MLE | 20.66 | 21.09 | 19.40 | 30.22 | 48.11 | 71.92 | 84.75 | 18.68 | 88.54 | 20.95 | 32.53 | 48.87 | 60.38 | 81.55 |
| UL | 15.54 | 21.78 | 18.45 | 29.57 | 46.69 | 69.63 | 82.87 | 19.39 | 81.49 | 18.36 | 27.98 | 42.65 | 63.92 | 82.93 |
| ScaleGrad | 12.41 | 25.69 | 18.59 | 29.24 | 45.19 | 66.35 | 80.23 | 14.14 | 77.82 | 18.62 | 27.80 | 41.22 | 56.74 | 77.27 |
| SELFCONT | 19.02 | 34.37 | 16.45 | 26.47 | 45.10 | 72.02 | 84.78 | 19.86 | 89.84 | 17.56 | 26.98 | 43.39 | 63.33 | 83.51 |
| Ground Truth | 18.31 | 100 | 17.38 | 27.92 | 46.29 | 72.34 | 84.20 | 24.01 | 100 | 16.36 | 26.47 | 42.30 | 74.49 | 90.01 |
| Datasets | |Train| | |Validation| | |Test| | Avg. Len |
| Wikitext-103 | 201,632 | 448 | 480 | 512 |
| WritingPrompts | 272,600 | 15,620 | 15,138 | 439 |
| Training Steps | Cases | |
| 100 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. The first, and the " . The " The " <eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos></eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos><eos> | |
| 1500 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of the building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the buildings, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building | He had a guest @-@ starring role on the television series The Bill in 2000. It was a new project in November 18th century, after a new new building to the building, was created by the new new building. It was built in 18th century. The building of the new building, was built in 1966, which had been created a new building. The building, was built in the new building. In the building of building, which was built in the building by the building by new building by the building. The building, the new building by new building by new building, the building, and work, in the building. In 2009, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, and building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building, the building |
| 6000 | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. She had begun in 1962, was built a series of New York Times in 1962 - 1938 after producing the Laidlay, and described as well as well as The Lasttunk's assistant ... ] named". She later explained : "[ 500 people who did not be turned off for me to me to me. She decided to his own song " As Rocham R. Nails , saying". She knew it 's script was going to make me to live film, and I mean it was through me to get it 't know what we 't know what we want to make me . On 17 November 1953 , and it was nominated for me. He had done what he wanted the BBC at a period at her following year because it appearance at the mid @-@ selling artist with her singing . She was nominated for the background scene in 1962 . She spent four days after getting part of the public appearances at the war appearance , California , and ] base . Aftermath Medie Glam artist of the summer May 1967 , New York Times , and the war as the Star Wars franchise . He began to use the National Association ( from his staff ; it was included Star ) , including the Lyds house east , and I was the West Virginia Tech back to the war and Mennon ; there . She developed by <eos> William Peninsular League <eos>=<eos> The script , 2004 . The script was named after the North America for the LAM passed . The script was commissioned to the American co @-@ person to produce producer ( present , taking place of the mid @-@ old , and the mid @-@ old @-@ old film , The Next Generation . The New York Times , having won the 4th birthday of the 4 , in the 4 million viewers . This was announced that it was cut of the media . The Elder Scrolls IV of the production , in East Coaster and The company entitled The Next Generation . <eos> For example , including the war , having performed on 6 , having released in East Coast Division . Upon its crew became a series of the produce |
| Training Set | Cases | |
| Doriginal | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has appeared in several films, including the television series The Bill, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. He has also appeared in several films, including The Bill Goes to Washington, The Bill Goes to Washington, and The Bill Goes to Washington. <eos>Boulter was born in London, England, on 23 May 1986. He is the third child of actor and actress Robert Boulter and his wife, Susan. He is the third of five children born to his wife Susan and their three children, Robert, Roberta, and Roberta. Robert Boulter's father, Robert Boulter, was a film director and producer. He was the first actor to be cast in a film role, and the first to be cast in a television series. He was also the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a tele-. tical series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was the first actor to being cast in a television series. He was then cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a tele-. tical series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He were cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. Hewas cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television series. He was Cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television series. He was cast in a Television Series. He was cast in a Television series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He wasCast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Sereas. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast ina Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. He was Cast in a Television Series. HE was Cast in a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Sereas. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He wasCast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series. He was Cast In a Television Series | He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast ina television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. He was the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a Television Series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cist in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the firsst actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series.Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast ina television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. Hwas the first actor to be cast in a television series. |
| Models | Wikitext-103 | WritingPrompts |
| MLE | 0.9 | 0.9 |
| UL | 0.7 | 0.8 |
| ScaleGrad | 0.5 | 0.6 |
| SELFCONT | 0.6 | 0.7 |
| Models | Acc | Repetitive | Non-Repetitive | ||||
| P | R | F1 | P | R | F1 | ||
| MLE | 78.9 | 67.4 | 87.0 | 75.9 | 90.2 | 73.9 | 81.2 |
| UL | 80.5 | 70.9 | 83.0 | 76.5 | 88.2 | 78.9 | 83.3 |
| ScaleGrad | 81.6 | 75.9 | 76.1 | 76.0 | 85.2 | 85.0 | 85.1 |
| SELFCONT | 82.3 | 78.5 | 73.9 | 76.1 | 84.4 | 87.4 | 85.9 |
| Models | Cases (Greedy Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in the television series The Bill, The Bill 2, The Bill and The Bill 3. <eos>Boulter was born in London, England, and grew up in the United States. He studied at the University of California, Berkeley, and the University of California, Berkeley. He was a member of the British Academy of Film and Television Arts (BAFTA) in 2001. He has appeared in several television series, including The Bill, The Bill 2, The Bill 3, The Bill 4, The Bill 5, The Bill 6, The Bill 7, The Bill 8, The Bill 9, The Bill 10, The Bill 11, The Bill 12, The Bill 13, The Bill 14, The Bill 15, The Bill 16, The Bill 17, The Bill 18, The Bill 19, The Bill 20, The Bill 21, The Bill 22, The Bill 23, The Bill 24, The Bill 25, The Bill 26, The Bill 27, The Bill 28, The Bill 29, The Bill 30, The Bill 31, The Bill 32, The Bill 33, The Bill 34, The Bill 35, The Bill 36, The Bill 37, The Bill 39, The Bill 40, The Bill 41, The Bill 42, The Bill 43, The Bill 44, The Bill 45, The Bill 46, The Bill 47, The Bill 48, The Bill 49, The Bill 50, The Bill 51, The Bill 52, The Bill 53, The Bill 54, The Bill 55, The Bill 56, The Bill 57, The Bill 58, The Bill 59, The Bill 60, The Bill 61, The Bill 62, The Bill 63, The Bill 64, The Bill 65, The Bill 66, The Bill 67, The Bill 68, The Bill 69, The Bill 70, The Bill 71, The Bill 72, The Bill 73, The Bill 74, The Bill 75, The Bill 76, The Bill 77, The Bill 78, The Bill 79, The Bill 80, The Bill 81, The Bill 82, The Bill 83, The Bill 84, The Bill 85, The Bill 86, The Bill 87, The Bill 88, The Bill 89, The Bill 90, The Bill 91, The Bill 92, The Bill 93 |
| UL | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He has also appeared in several films, including the comedy @-@ drama The Man Who Would Be King (2002), the science fiction film The Man Who Would Be Super (2006), the science fiction film The Man Who Would Be Super (2008), the science fiction film The Man Who Would Be Super (2012), the science fiction film The Day After Tomorrow (2013), the science fiction film The Day After Tomorrow (2014), the science fiction film The Day After Tomorrow (2015), the science fiction film The Day After Tomorrow (2016), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2019), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2018), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), The science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science Fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction filmThe Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2020), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the sciencefiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day Afternor (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (2017), the science fiction film The Day After Tomorrow (1999), The Man Who Would Never Be King 10 (2017), The Man Who Would Never Be King 11 (2017), The Man Who Would Never Be King 12 (2017), The Man Who Would Never Be King 13 (2017), The Man Who Would Never Be King 14 (2017), The Man Who Would Never Be King 15 (2017), The Man Who Would Never Be King 16 (2017), The Man Who Would Never Be King 17 (2017), The Man Who Would Never Be King 18 (2017), The Man Who Would Never Be King 19 (2017), The Man Who Would Never Be King 20 (2017), The Man Who Would Never Be King 21 (2017), The Man Who Would Never Be King 22 (2017), The Man Who Would Never Be King 23 (2017), The Man Who Would Never Be King 24 (2017), The Man Who Would Never Be King 25 (2017), The Man Who Would Never Be King 26 (2017), The Man Who Would Never Be King 27 (2017), The Man Who Would Never Be King 28 (2017), The Man Who Would Never Be King 29 (2017), The Man Who Would Never Be King 30 (2017), The Man Who Would Never Be King 31 (2017), The Man Who Would Never Be King 32 (2017), The Man Who Would Never Be King 33 (2017), The Man Who Would Never Be King 34 (2017), The Man Who Would Never Be King 35 (2017), The Man Who Woul |
| SELFCONT | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Boulter appeared in two episodes of the British comedy @-@ drama Oz in 2001, playing an American businessman who helps an Italian woman (Joanna Cassidy) escape from her abusive husband (Michael Sheen). He also played a character in the 2004 film Heat, which was directed by Michael Medvedev. Boulter appeared in two episodes of the British comedy @-@ drama Body of Proof in 2005, playing a retired police officer who is accused of murdering his wife (Christine Fay). He appeared in two episodes of the British television drama Body of Proof in 2006, playing a retired judge who defends his client, Simon Ashdown (Stephen Fry). Boulter also appeared in the Australian television series The X Factor in 2007, playing a retired judge who defends his client, Kieman Straker (Simon Ashdown). <eos>Boulter was nominated for a Golden Globe Award for Best Actor - Motion Picture Musical or Comedy, but lost to Hugh Grant for the role of Seamus Heaney. In 2010, he appeared in the British comedy @-@ drama The X Factor, playing a retired judge who is accused of murdering his wife (Christine Fay). He also appeared in the Australian television series The X Factor in 2010, playing a retired judge who defends his client, Simon Ashdown. In 2012, he appeared in the British comedy @-@ drama Body of Proof, playing a retired judge who is accused of murdering his wife (Joanna Cassidy). He also appeared in the Australian television series The X Factor in 2013, playing a retired judge who is accused of murdering his wife (Christine Fay). <eos>Boulter has appeared in a number of films, including the 2005 film Heat, and the 2008 comedy Body of Proof. He appeared in the British television series Heat 2, and the 2009 comedy Body of Proof. In 2011, he appeared in the Australian television series The X Factor. In 2012, he appeared in the British television series The X Factor. In 2013, he appeared in the British television series The X Factor. In 2014, he appeared in the Australian television series The X Factor. In 2015, he appeared in the British television series The X Factor. In 2016, he appeared in the Australian television series The X Factor. |
| Models | Cases (Nucleus Decoding) |
| MLE | <eos> = Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. Since 1994, he has appeared in a wide range of productions. He has played Crocodile Hunter, a domestic terrorist in A Change Is Gonna Come. His first feature film role was in a 1983 film, The Perfect Numbers, written by Ridley Scott. A science fiction parody of the British mafia, his next film role was the horror comedy The Imperfect People. The film was a box office hit and has been named one of the greatest movies ever made. <eos>Boulter portrayed a divorcee living in Chicago, Illinois, from the start of its development in 1986, when he met Hollywood agent Kevin De Matos. The two became very close and married on August 29, 1988. He has become a great fan of James Cameron's 1999 filmAvatar. In December 2011, he starred as a roadie who stands in a line at a restaurant. <eos><eos>= Early life== <eos> Robert Boulter was born in Dundee, Scotland, on November 6, 1961, and raised in Dundee, Scotland. He attended Eales College, Dundee. He received an MBE for his work at Eales as a schoolteacher in 1973, and graduated from Dundee University in 1974. After teaching history to three young teenagers, he acted in numerous short films. <eos><eos>= Career== <eos> He made his acting debut in 1976 with the short film The Quarryman, playing the character Andre Taylor in a variety of movies and television shows. Boulter has appeared on television and cinema advertisements as well. <eos> In 1982, Boulter co @-@ starred in the video game Dr. No. The following year he made his film debut in the video game El Story, played by Terry Gilliam. He also appeared in the 1984 video game Doctor Who, starring Matthew Broderick. <eos>Boulter met producer Shane Bobbitt in 1982, and the two became friends, and Broderick made Boulter his stand @-@ up comedy partner. On the strength of their relationship, Boulter starred in several feature film projects for the American television |
| UL | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He starred as the title character in 2008's A Charlie Brown Christmas, directed by Adam Mickiewicz. Boulter has also played a variety of other roles over the years, including as the titular character on <unk> USA in 2002, in the <unk> anthology series Blood's Wild in 2007, as the protagonist on the murder mystery A Beautiful Guy in 2009, and as the love interest for the titular character on Queen's College in 2012. Boulter has also played the supporting roles of Richard VIII and Queen Elizabeth in various media including television, film, and video games. He won the Academy Award for Best Actor for his performance in King George V. <eos> Born in Carlisle, Lancashire, Boulter grew up in Manchester. He has been involved in theatre since he was eight years old. His only film role came in 2000, when he played Jack Leppard in Harry Potter and the Goblet of Fire. A second appearance came in 2001, when he played the young Severus Snape in Romeo and Juliet. He has worked on television shows including Steven Soderbergh's From Russia, with Love, Watchmen, Dawson's Creek, Ealing Studios and the British Channel 7 documentary series Ghost Hunters. <eos>Boulter has made four appearances on television: seven as Colin Rudge in Star Trek: Voyager (1995-1997); 13 as Ian Frazier in The Sarah Jane Adventures (1997-1999); and 16 as Scott Reid in Michael Bay's Robin Hood: How the Legend Was Won (1999). He was nominated for a Laurence Olivier Award in 2001. <eos><eos>= Early life== <eos> Boulter was born on 12 April 1979 in Carlisle, Lancashire, England. His father is a retired pilot. He attended Elgin Grammar School, where he earned an academic degree. After leaving school, he worked in retail at an engineering firm in Manchester. <eos>Boulter attended High School, Carlisle before transferring to King's College, Cambridge. During this time, he became involved in theatre and became a student at Queen's College, Cambridge. |
| ScaleGrad | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. In 2004 he starred as the character Sherlock Holmes in the BBC medical drama Holby City. Boulter made his screen debut with the comedy series Chitty Chitty Bang Bang, which was broadcast in 2006. He then played Lord Stanley in The Lord of the Rings trilogy and in Pirates of the Caribbean: At World's End. In 2007 he portrayed David Marmion in the film adaptation of A Midsummer Night's Dream. <eos> He appeared in the British film Downton Abbey (2008), which won him a Golden Globe Award for Best Actor. He also starred in Steven Spielberg's fantasy film, The Wrestler, as the title character. <eos> In 2010, Boulter portrayed Leonardo DiCaprio in the directorial debut of Woody Allen's The Wrestling. The film tells the story of two men who live in Los Angeles, California, who have been feuding over money and power. The film received mixed reviews from critics. <eos><eos>= Early life== <eos> Boulter was born in Leicester, England, on 3 October 1963. His father, John, is a chemical engineer and designer. His mother, Mary (née <unk>), is a member of the Scottish Episcopal Church. He grew up in Loughborough, Leicester. He was educated at Eton College and then St Thomas' School, Northamptonshire. His older brother, Liam, plays football for Leicester City. <eos>Boulter started playing football when he was four years old. After a few years, he joined Brentford Town, where he played alongside Ronnie Brown. Boulter enjoyed the game and liked the fact that his father had taught him how to play football. When he turned twelve, he moved to Leicester City, but he left the club after one season because of disciplinary problems. Boulter did not attend Brentford's youth team, but took up playing football for them. During his time at Brentford, he played for several clubs including Chesterfield, Rochdale, Oldham Athletic, St James' Park, Scarborough and Lewes. He made his debut for the club aged 15 in 1971, an |
| SELFCONT | <eos>= Robert Boulter = <eos> Robert Boulter is an English film, television and theatre actor. He had a guest @-@ starring role on the television series The Bill in 2000. He also appeared in a 2000 episode of Syfy's Geneva Live. <eos>Boulter is an accomplished box @-@ office actor and one of the best known box @-@ office draws in the history of British television. Boulter appeared in over 20 films and received many awards, including Four Academy Awards, including Best Actor and Best Supporting Actor, and the BAFTA Award for Best British Actor. He was nominated for six other BAFTAs, winning three, for his work on the television series and the 1997 film. He starred in The Bill in 2001 and again in 2002. In 2005, he appeared in The Gleason Room, the 2005 science fiction film about rediscovery of woolly alien relics, and in the 2006 biographical drama Brand New Eyes. In 2010, he starred in the stage production of Minor Threat and the 2007 psychological thriller Victoria's Secret. <eos>Boulter's stage and film career began with his performance in the 1997 romantic comedy Hamlet. In 2000, he was cast as Jonathan Simeone in the German @-@ language dramatisation of French novelist Raymond Lebowski's epic play, The Professionals. He took on the role of "Troy", an obsessive person who attempts to prove himself to courtiers. Although he enjoyed playing Troy, he took "enormous risks", in the words of theatre critic Graham McCann, who wrote that " there was nothing to lose in playing a man like Troy. " He co @-@ starred in The Professionals with Julianne Moore and Kim Novak. He portrayed the criminal Tammi Martineau in the 2004 biographical film Asterisk and appeared in several films and television shows. In 2005, he starred as Garth Snow in the Fox crime drama Dangerous Liaisons. <eos>Boulter is known for his film work in Hungary and abroad. He has also worked with Brandon Thomas and Sacha Baron Cohen. In 2011, he was nominated for a Laurence Olivier Award for Best Actor, with Olivier in the role of General Herculaneum. In 2012, he starred in The Phantom of the Opera, which opened at the BBC2 Leicester Square Theatre, with much of the stage cast from his earlier work |
| Splits | AE-110K | MAVE | |||||||
| Lenses Color | Wheel Material | Product Type | REST | Device Type | Boot Style | Resolution | Compatibility | REST | |
| Train | 10 | 10 | 10 | 109K | 100 | 100 | 100 | 100 | 2.6M |
| Test | 165 | 118 | 368 | - | 20,273 | 16,459 | 14,437 | 6,728 | - |
| Models | Paras | AE-110K | MAVE | |||||
| Lenses Color | Wheel Material | Product Type | Device Type | Boot Style | Resolution | Compatibility | ||
| MAVEQA-FT (Yang et al., 2022) | 100% | 79.58 ± 0.39 | 85.59 ± 0.45 | 92.41 ± 0.36 | 91.06 ± 0.54 | 92.23 ± 0.51 | 90.88 ± 0.47 | 95.19 ± 0.42 |
| Partial-1 (Yosinski et al., 2014) | 8.21% | 72.36 ± 0.44 | 79.93 ± 0.47 | 85.38 ± 0.62 | 83.69 ± 0.78 | 84.31 ± 0.65 | 84.15 ± 0.73 | 85.57 ± 0.59 |
| Adapter (Pfeiffer et al., 2020) | 2.53% | 74.61 ± 0.55 | 81.84 ± 0.39 | 86.78 ± 0.56 | 85.51 ± 0.55 | 86.12 ± 0.47 | 83.22 ± 0.57 | 87.56 ± 0.48 |
| BitFit (Zaken et al., 2022) | 2.04% | 74.95 ± 0.41 | 82.37 ± 0.38 | 87.44 ± 0.48 | 87.34 ± 0.58 | 87.67 ± 0.71 | 85.56 ± 0.43 | 88.30 ± 0.54 |
| Prompt-Tuning (Lester et al., 2021) | 0.40% | 81.32 ± 0.46 | 84.69 ± 0.35 | 90.67 ± 0.40 | 91.43 ± 0.46 | 90.36 ± 0.35 | 92.37 ± 0.47 | 95.25 ± 0.41 |
| Prefix-Tuning (Li and Liang, 2021) | 0.40% | 81.12 ± 0.42 | 84.95 ± 0.44 | 90.30 ± 0.45 | 91.79 ± 0.42 | 89.56 ± 0.51 | 92.08 ± 0.43 | 94.86 ± 0.45 |
| XPrompt (Ma et al., 2022) | 0.33% | 82.14 ± 0.32 | 85.72 ± 0.48 | 90.50 ± 0.33 | 91.34 ± 0.48 | 90.67 ± 0.52 | 92.31 ± 0.46 | 95.10 ± 0.44 |
| MixPAVE | 0.40% | 82.58 ± 0.53 | 85.36 ± 0.42 | 91.63 ± 0.38 | 92.51 ± 0.44 | 91.75 ± 0.53 | 93.47 ± 0.42 | 96.86 ± 0.39 |
| Length | Device Type | Boot Style | Resolution | Compatibility |
| 4 | 80.31 | 94.27 | 93.15 | 96.25 |
| 8 | 86.84 | 93.16 | 93.24 | 95.89 |
| 12 | 90.63 | 92.29 | 93.12 | 96.34 |
| 24 | 92.51 | 91.75 | 93.47 | 96.86 |
| 48 | 93.73 | 92.20 | 93.22 | 96.98 |
| 96 | 92.86 | 92.86 | 93.41 | 97.14 |
| Position | Device Type | Boot Style | Resolution | Compatibility |
| All | 92.51 | 91.75 | 93.47 | 96.86 |
| Input | 78.56 | 85.48 | 90.87 | 92.61 |
| Bottom 6 | 88.35 | 88.64 | 92.70 | 94.17 |
| Output | 75.82 | 84.37 | 87.74 | 88.16 |
| Top 6 | 82.55 | 88.49 | 91.52 | 92.63 |
| Alternative | 90.31 | 90.14 | 92.88 | 94.93 |
| Model | Device Type | Boot Style | Resolution | Compatibility |
| Encoder-2L | 91.45 | 92.12 | 92.96 | 96.49 |
| Encoder-6L | 91.82 | 92.13 | 93.35 | 97.11 |
| Encoder-12L | 92.51 | 91.75 | 93.47 | 96.86 |
| Encoder-24L | 92.87 | 89.78 | 92.85 | 96.41 |
| Splits | AE-110K | MAVE | |||||||
| Lenses Color | Wheel Material | Product Type | REST | Device Type | Boot Style | Resolution | Compatibility | REST | |
| Train | 10 | 10 | 10 | 109K | 100 | 100 | 100 | 100 | 2.6M |
| Test | 165 | 118 | 368 | - | 20,273 | 16,459 | 14,437 | 6,728 | - |
| Models | Paras | AE-110K | MAVE | |||||
| Lenses Color | Wheel Material | Product Type | Device Type | Boot Style | Resolution | Compatibility | ||
| MAVEQA-FT (Yang et al., 2022) | 100% | 79.58 ± 0.39 | 85.59 ± 0.45 | 92.41 ± 0.36 | 91.06 ± 0.54 | 92.23 ± 0.51 | 90.88 ± 0.47 | 95.19 ± 0.42 |
| Partial-1 (Yosinski et al., 2014) | 8.21% | 72.36 ± 0.44 | 79.93 ± 0.47 | 85.38 ± 0.62 | 83.69 ± 0.78 | 84.31 ± 0.65 | 84.15 ± 0.73 | 85.57 ± 0.59 |
| Adapter (Pfeiffer et al., 2020) | 2.53% | 74.61 ± 0.55 | 81.84 ± 0.39 | 86.78 ± 0.56 | 85.51 ± 0.55 | 86.12 ± 0.47 | 83.22 ± 0.57 | 87.56 ± 0.48 |
| BitFit (Zaken et al., 2022) | 2.04% | 74.95 ± 0.41 | 82.37 ± 0.38 | 87.44 ± 0.48 | 87.34 ± 0.58 | 87.67 ± 0.71 | 85.56 ± 0.43 | 88.30 ± 0.54 |
| Prompt-Tuning (Lester et al., 2021) | 0.40% | 81.32 ± 0.46 | 84.69 ± 0.35 | 90.67 ± 0.40 | 91.43 ± 0.46 | 90.36 ± 0.35 | 92.37 ± 0.47 | 95.25 ± 0.41 |
| Prefix-Tuning (Li and Liang, 2021) | 0.40% | 81.12 ± 0.42 | 84.95 ± 0.44 | 90.30 ± 0.45 | 91.79 ± 0.42 | 89.56 ± 0.51 | 92.08 ± 0.43 | 94.86 ± 0.45 |
| XPrompt (Ma et al., 2022) | 0.33% | 82.14 ± 0.32 | 85.72 ± 0.48 | 90.50 ± 0.33 | 91.34 ± 0.48 | 90.67 ± 0.52 | 92.31 ± 0.46 | 95.10 ± 0.44 |
| MixPAVE | 0.40% | 82.58 ± 0.53 | 85.36 ± 0.42 | 91.63 ± 0.38 | 92.51 ± 0.44 | 91.75 ± 0.53 | 93.47 ± 0.42 | 96.86 ± 0.39 |
| Length | Device Type | Boot Style | Resolution | Compatibility |
| 4 | 80.31 | 94.27 | 93.15 | 96.25 |
| 8 | 86.84 | 93.16 | 93.24 | 95.89 |
| 12 | 90.63 | 92.29 | 93.12 | 96.34 |
| 24 | 92.51 | 91.75 | 93.47 | 96.86 |
| 48 | 93.73 | 92.20 | 93.22 | 96.98 |
| 96 | 92.86 | 92.86 | 93.41 | 97.14 |
| Position | Device Type | Boot Style | Resolution | Compatibility |
| All | 92.51 | 91.75 | 93.47 | 96.86 |
| Input | 78.56 | 85.48 | 90.87 | 92.61 |
| Bottom 6 | 88.35 | 88.64 | 92.70 | 94.17 |
| Output | 75.82 | 84.37 | 87.74 | 88.16 |
| Top 6 | 82.55 | 88.49 | 91.52 | 92.63 |
| Alternative | 90.31 | 90.14 | 92.88 | 94.93 |
| Model | Device Type | Boot Style | Resolution | Compatibility |
| Encoder-2L | 91.45 | 92.12 | 92.96 | 96.49 |
| Encoder-6L | 91.82 | 92.13 | 93.35 | 97.11 |
| Encoder-12L | 92.51 | 91.75 | 93.47 | 96.86 |
| Encoder-24L | 92.87 | 89.78 | 92.85 | 96.41 |
| Splits | AE-110K | MAVE | |||||||
| Lenses Color | Wheel Material | Product Type | REST | Device Type | Boot Style | Resolution | Compatibility | REST | |
| Train | 10 | 10 | 10 | 109K | 100 | 100 | 100 | 100 | 2.6M |
| Test | 165 | 118 | 368 | - | 20,273 | 16,459 | 14,437 | 6,728 | - |
| Models | Paras | AE-110K | MAVE | |||||
| Lenses Color | Wheel Material | Product Type | Device Type | Boot Style | Resolution | Compatibility | ||
| MAVEQA-FT (Yang et al., 2022) | 100% | 79.58 ± 0.39 | 85.59 ± 0.45 | 92.41 ± 0.36 | 91.06 ± 0.54 | 92.23 ± 0.51 | 90.88 ± 0.47 | 95.19 ± 0.42 |
| Partial-1 (Yosinski et al., 2014) | 8.21% | 72.36 ± 0.44 | 79.93 ± 0.47 | 85.38 ± 0.62 | 83.69 ± 0.78 | 84.31 ± 0.65 | 84.15 ± 0.73 | 85.57 ± 0.59 |
| Adapter (Pfeiffer et al., 2020) | 2.53% | 74.61 ± 0.55 | 81.84 ± 0.39 | 86.78 ± 0.56 | 85.51 ± 0.55 | 86.12 ± 0.47 | 83.22 ± 0.57 | 87.56 ± 0.48 |
| BitFit (Zaken et al., 2022) | 2.04% | 74.95 ± 0.41 | 82.37 ± 0.38 | 87.44 ± 0.48 | 87.34 ± 0.58 | 87.67 ± 0.71 | 85.56 ± 0.43 | 88.30 ± 0.54 |
| Prompt-Tuning (Lester et al., 2021) | 0.40% | 81.32 ± 0.46 | 84.69 ± 0.35 | 90.67 ± 0.40 | 91.43 ± 0.46 | 90.36 ± 0.35 | 92.37 ± 0.47 | 95.25 ± 0.41 |
| Prefix-Tuning (Li and Liang, 2021) | 0.40% | 81.12 ± 0.42 | 84.95 ± 0.44 | 90.30 ± 0.45 | 91.79 ± 0.42 | 89.56 ± 0.51 | 92.08 ± 0.43 | 94.86 ± 0.45 |
| XPrompt (Ma et al., 2022) | 0.33% | 82.14 ± 0.32 | 85.72 ± 0.48 | 90.50 ± 0.33 | 91.34 ± 0.48 | 90.67 ± 0.52 | 92.31 ± 0.46 | 95.10 ± 0.44 |
| MixPAVE | 0.40% | 82.58 ± 0.53 | 85.36 ± 0.42 | 91.63 ± 0.38 | 92.51 ± 0.44 | 91.75 ± 0.53 | 93.47 ± 0.42 | 96.86 ± 0.39 |
| Length | Device Type | Boot Style | Resolution | Compatibility |
| 4 | 80.31 | 94.27 | 93.15 | 96.25 |
| 8 | 86.84 | 93.16 | 93.24 | 95.89 |
| 12 | 90.63 | 92.29 | 93.12 | 96.34 |
| 24 | 92.51 | 91.75 | 93.47 | 96.86 |
| 48 | 93.73 | 92.20 | 93.22 | 96.98 |
| 96 | 92.86 | 92.86 | 93.41 | 97.14 |
| Position | Device Type | Boot Style | Resolution | Compatibility |
| All | 92.51 | 91.75 | 93.47 | 96.86 |
| Input | 78.56 | 85.48 | 90.87 | 92.61 |
| Bottom 6 | 88.35 | 88.64 | 92.70 | 94.17 |
| Output | 75.82 | 84.37 | 87.74 | 88.16 |
| Top 6 | 82.55 | 88.49 | 91.52 | 92.63 |
| Alternative | 90.31 | 90.14 | 92.88 | 94.93 |
| Model | Device Type | Boot Style | Resolution | Compatibility |
| Encoder-2L | 91.45 | 92.12 | 92.96 | 96.49 |
| Encoder-6L | 91.82 | 92.13 | 93.35 | 97.11 |
| Encoder-12L | 92.51 | 91.75 | 93.47 | 96.86 |
| Encoder-24L | 92.87 | 89.78 | 92.85 | 96.41 |
| Splits | AE-110K | MAVE | |||||||
| Lenses Color | Wheel Material | Product Type | REST | Device Type | Boot Style | Resolution | Compatibility | REST | |
| Train | 10 | 10 | 10 | 109K | 100 | 100 | 100 | 100 | 2.6M |
| Test | 165 | 118 | 368 | - | 20,273 | 16,459 | 14,437 | 6,728 | - |
| Models | Paras | AE-110K | MAVE | |||||
| Lenses Color | Wheel Material | Product Type | Device Type | Boot Style | Resolution | Compatibility | ||
| MAVEQA-FT (Yang et al., 2022) | 100% | 79.58 ± 0.39 | 85.59 ± 0.45 | 92.41 ± 0.36 | 91.06 ± 0.54 | 92.23 ± 0.51 | 90.88 ± 0.47 | 95.19 ± 0.42 |
| Partial-1 (Yosinski et al., 2014) | 8.21% | 72.36 ± 0.44 | 79.93 ± 0.47 | 85.38 ± 0.62 | 83.69 ± 0.78 | 84.31 ± 0.65 | 84.15 ± 0.73 | 85.57 ± 0.59 |
| Adapter (Pfeiffer et al., 2020) | 2.53% | 74.61 ± 0.55 | 81.84 ± 0.39 | 86.78 ± 0.56 | 85.51 ± 0.55 | 86.12 ± 0.47 | 83.22 ± 0.57 | 87.56 ± 0.48 |
| BitFit (Zaken et al., 2022) | 2.04% | 74.95 ± 0.41 | 82.37 ± 0.38 | 87.44 ± 0.48 | 87.34 ± 0.58 | 87.67 ± 0.71 | 85.56 ± 0.43 | 88.30 ± 0.54 |
| Prompt-Tuning (Lester et al., 2021) | 0.40% | 81.32 ± 0.46 | 84.69 ± 0.35 | 90.67 ± 0.40 | 91.43 ± 0.46 | 90.36 ± 0.35 | 92.37 ± 0.47 | 95.25 ± 0.41 |
| Prefix-Tuning (Li and Liang, 2021) | 0.40% | 81.12 ± 0.42 | 84.95 ± 0.44 | 90.30 ± 0.45 | 91.79 ± 0.42 | 89.56 ± 0.51 | 92.08 ± 0.43 | 94.86 ± 0.45 |
| XPrompt (Ma et al., 2022) | 0.33% | 82.14 ± 0.32 | 85.72 ± 0.48 | 90.50 ± 0.33 | 91.34 ± 0.48 | 90.67 ± 0.52 | 92.31 ± 0.46 | 95.10 ± 0.44 |
| MixPAVE | 0.40% | 82.58 ± 0.53 | 85.36 ± 0.42 | 91.63 ± 0.38 | 92.51 ± 0.44 | 91.75 ± 0.53 | 93.47 ± 0.42 | 96.86 ± 0.39 |
| Length | Device Type | Boot Style | Resolution | Compatibility |
| 4 | 80.31 | 94.27 | 93.15 | 96.25 |
| 8 | 86.84 | 93.16 | 93.24 | 95.89 |
| 12 | 90.63 | 92.29 | 93.12 | 96.34 |
| 24 | 92.51 | 91.75 | 93.47 | 96.86 |
| 48 | 93.73 | 92.20 | 93.22 | 96.98 |
| 96 | 92.86 | 92.86 | 93.41 | 97.14 |
| Position | Device Type | Boot Style | Resolution | Compatibility |
| All | 92.51 | 91.75 | 93.47 | 96.86 |
| Input | 78.56 | 85.48 | 90.87 | 92.61 |
| Bottom 6 | 88.35 | 88.64 | 92.70 | 94.17 |
| Output | 75.82 | 84.37 | 87.74 | 88.16 |
| Top 6 | 82.55 | 88.49 | 91.52 | 92.63 |
| Alternative | 90.31 | 90.14 | 92.88 | 94.93 |
| Model | Device Type | Boot Style | Resolution | Compatibility |
| Encoder-2L | 91.45 | 92.12 | 92.96 | 96.49 |
| Encoder-6L | 91.82 | 92.13 | 93.35 | 97.11 |
| Encoder-12L | 92.51 | 91.75 | 93.47 | 96.86 |
| Encoder-24L | 92.87 | 89.78 | 92.85 | 96.41 |
| Baseline | Pre-trained Model | MultiWOZ 2.0 | MultiWOZ 2.1 | MultiWOZ 2.4 | |||
| Joint | Slot | Joint | Slot | Joint | Slot | ||
| TRADE (Wu et al., 2019) | - | 48.62 | 96.92 | 45.60 | 96.55 | 55.05 | 97.62 |
| SUMBT (Lee et al., 2019) | BERT-base | 42.40 | - | 49.01 | 96.76 | 61.86 | 97.90 |
| PIN (Chen et al., 2020a) | - | 52.44 | 97.28 | 48.40 | 97.02 | 58.92 | 98.02 |
| SOM-DST (Kim et al., 2020) | BERT-base | 51.72 | - | 53.01 | - | 66.78 | 98.38 |
| CSFN-DST (Zhu et al., 2020) | BERT-base | 52.23 | - | 53.19 | - | - | - |
| DST-Picklist (Zhang et al., 2020) | BERT-base | 54.39 | - | 53.30 | 97.40 | - | - |
| SAVN (Wang et al., 2020) | BERT-base | 54.52 | 97.42 | 54.86 | 97.55 | 60.55 | 98.05 |
| SST (Chen et al., 2020b) | BERT-base | 51.17 | - | 55.23 | - | - | - |
| SimpleTOD (Hosseini-Asl et al., 2020) | DistilGPT2 | - | - | 55.26 | - | - | - |
| Seq2SeqDU (Feng et al., 2021) | BERT-base | - | - | 56.10 | - | - | - |
| STAR (Ye et al., 2021) | BERT-base | 54.53 | - | 56.36 | 97.59 | 73.62 | 98.85 |
| SDP-DST (Lee et al., 2021) | T5-base | - | - | 56.66 | - | - | - |
| DS-Graph (Lin et al., 2021) | GPT2 | 54.86 | 97.47 | - | - | - | - |
| DSGFNet (Feng et al., 2022) | BERT-base | - | - | 56.70 | - | - | - |
| PPTOD (Su et al., 2022) | T5-large | 53.89 | - | 57.45 | - | - | - |
| Baseline | BERT-base | 54.38 | 97.47 | 55.82 | 97.51 | 73.81 | 98.82 |
| MoNET | BERT-base | 55.48 (↑1.10) | 97.55 | 57.71 (↑1.89) | 97.71 | 76.02 (↑2.21) | 98.99 |
| Use Modified Label | |||||||
| TripPy (Heck et al., 2020) | BERT-base | - | - | 55.29 | - | - | - |
| TripPy + SCoRe (Yu et al., 2021) | BERT-base | - | - | 60.48 | - | - | - |
| TripPy + CoCoAug (Li et al., 2021) | BERT-base | - | - | 60.53 | - | - | - |
| TripPy + SaCLog (Dai et al., 2021) | BERT-base | - | - | 60.61 | - | - | - |
| Model | NoisedCM | NoisedST | Accuracy |
| Baseline w/o state | × | × | 64.94 |
| Baseline | × | × | 73.81 |
| MoNET-ST | × | ✓ | 75.54 |
| MoNET-CM | ✓ | × | 75.76 |
| MoNET | ✓ | ✓ | 76.02 |
| Sys: I have many trains available. What day would you like to leave?\nUsr: We will be leaving on Sunday afternoon.\nGT: train-day-Sunday\nBaseline: train-day-Sunday MoNET: train-day-Sunday |
| Sys: There are still many trains to pick from, can you narrow down a departure and arrival time frame?\nUsr: Yes, it should leave after 20:15 and leave on Monday, not Sunday.\nPlease give me a price, travel time, and arrival time for any train.\nGT: train-day-Monday\nBaseline: train-day-Sunday MoNET: train-day-Monday (update) |
| Sys: Great! You are booked at Autumn House for 1 night. Your reference number is n4tvfkgs. Would you like more information?\nUsr: I would also like a taxi between the places if possible.\nGT: taxi-destination-Autumn House\nBaseline: taxi-destination-Gonville and Caius College\nMoNET: taxi-destination-Gonville and Caius College |
| Sys: I can help with that. When would you like to either leave or arrive?\nAnd do you want the taxi from the college to the hotel or the other way around?\nUsr: I need to go from the college to the hotel, and I want to leave the college by 20:45, please.\nGT: taxi-destination-Autumn House\nBaseline: taxi-destination-Gonville and Caius College\nMoNET: taxi-destination-Autumn House (correct) |
| Model | Accuracy |
| MinTL (Lin et al., 2020) | 52.07 |
| MTTOD (Lee, 2021) | 53.56 |
| PPTOD (Su et al., 2022) | 53.37 |
| T5-base | 53.26 |
| MoNET (T5-base) | 54.67 (↑ 1.41) |
| Baseline | Pre-trained Model | MultiWOZ 2.0 | MultiWOZ 2.1 | MultiWOZ 2.4 | |||
| Joint | Slot | Joint | Slot | Joint | Slot | ||
| TRADE (Wu et al., 2019) | - | 48.62 | 96.92 | 45.60 | 96.55 | 55.05 | 97.62 |
| SUMBT (Lee et al., 2019) | BERT-base | 42.40 | - | 49.01 | 96.76 | 61.86 | 97.90 |
| PIN (Chen et al., 2020a) | - | 52.44 | 97.28 | 48.40 | 97.02 | 58.92 | 98.02 |
| SOM-DST (Kim et al., 2020) | BERT-base | 51.72 | - | 53.01 | - | 66.78 | 98.38 |
| CSFN-DST (Zhu et al., 2020) | BERT-base | 52.23 | - | 53.19 | - | - | - |
| DST-Picklist (Zhang et al., 2020) | BERT-base | 54.39 | - | 53.30 | 97.40 | - | - |
| SAVN (Wang et al., 2020) | BERT-base | 54.52 | 97.42 | 54.86 | 97.55 | 60.55 | 98.05 |
| SST (Chen et al., 2020b) | BERT-base | 51.17 | - | 55.23 | - | - | - |
| SimpleTOD (Hosseini-Asl et al., 2020) | DistilGPT2 | - | - | 55.26 | - | - | - |
| Seq2SeqDU (Feng et al., 2021) | BERT-base | - | - | 56.10 | - | - | - |
| STAR (Ye et al., 2021) | BERT-base | 54.53 | - | 56.36 | 97.59 | 73.62 | 98.85 |
| SDP-DST (Lee et al., 2021) | T5-base | - | - | 56.66 | - | - | - |
| DS-Graph (Lin et al., 2021) | GPT2 | 54.86 | 97.47 | - | - | - | - |
| DSGFNet (Feng et al., 2022) | BERT-base | - | - | 56.70 | - | - | - |
| PPTOD (Su et al., 2022) | T5-large | 53.89 | - | 57.45 | - | - | - |
| Baseline | BERT-base | 54.38 | 97.47 | 55.82 | 97.51 | 73.81 | 98.82 |
| MoNET | BERT-base | 55.48 (↑1.10) | 97.55 | 57.71 (↑1.89) | 97.71 | 76.02 (↑2.21) | 98.99 |
| Use Modified Label | |||||||
| TripPy (Heck et al., 2020) | BERT-base | - | - | 55.29 | - | - | - |
| TripPy + SCoRe (Yu et al., 2021) | BERT-base | - | - | 60.48 | - | - | - |
| TripPy + CoCoAug (Li et al., 2021) | BERT-base | - | - | 60.53 | - | - | - |
| TripPy + SaCLog (Dai et al., 2021) | BERT-base | - | - | 60.61 | - | - | - |
| Model | NoisedCM | NoisedST | Accuracy |
| Baseline w/o state | × | × | 64.94 |
| Baseline | × | × | 73.81 |
| MoNET-ST | × | ✓ | 75.54 |
| MoNET-CM | ✓ | × | 75.76 |
| MoNET | ✓ | ✓ | 76.02 |
| Sys: I have many trains available. What day would you like to leave?\nUsr: We will be leaving on Sunday afternoon.\nGT: train-day-Sunday\nBaseline: train-day-Sunday MoNET: train-day-Sunday |
| Sys: There are still many trains to pick from, can you narrow down a departure and arrival time frame?\nUsr: Yes, it should leave after 20:15 and leave on Monday, not Sunday.\nPlease give me a price, travel time, and arrival time for any train.\nGT: train-day-Monday\nBaseline: train-day-Sunday MoNET: train-day-Monday (update) |
| Sys: Great! You are booked at Autumn House for 1 night. Your reference number is n4tvfkgs. Would you like more information?\nUsr: I would also like a taxi between the places if possible.\nGT: taxi-destination-Autumn House\nBaseline: taxi-destination-Gonville and Caius College\nMoNET: taxi-destination-Gonville and Caius College |
| Sys: I can help with that. When would you like to either leave or arrive?\nAnd do you want the taxi from the college to the hotel or the other way around?\nUsr: I need to go from the college to the hotel, and I want to leave the college by 20:45, please.\nGT: taxi-destination-Autumn House\nBaseline: taxi-destination-Gonville and Caius College\nMoNET: taxi-destination-Autumn House (correct) |
| Model | Accuracy |
| MinTL (Lin et al., 2020) | 52.07 |
| MTTOD (Lee, 2021) | 53.56 |
| PPTOD (Su et al., 2022) | 53.37 |
| T5-base | 53.26 |
| MoNET (T5-base) | 54.67 (↑ 1.41) |
| Baseline | Pre-trained Model | MultiWOZ 2.0 | MultiWOZ 2.1 | MultiWOZ 2.4 | |||
| Joint | Slot | Joint | Slot | Joint | Slot | ||
| TRADE (Wu et al., 2019) | - | 48.62 | 96.92 | 45.60 | 96.55 | 55.05 | 97.62 |
| SUMBT (Lee et al., 2019) | BERT-base | 42.40 | - | 49.01 | 96.76 | 61.86 | 97.90 |
| PIN (Chen et al., 2020a) | - | 52.44 | 97.28 | 48.40 | 97.02 | 58.92 | 98.02 |
| SOM-DST (Kim et al., 2020) | BERT-base | 51.72 | - | 53.01 | - | 66.78 | 98.38 |
| CSFN-DST (Zhu et al., 2020) | BERT-base | 52.23 | - | 53.19 | - | - | - |
| DST-Picklist (Zhang et al., 2020) | BERT-base | 54.39 | - | 53.30 | 97.40 | - | - |
| SAVN (Wang et al., 2020) | BERT-base | 54.52 | 97.42 | 54.86 | 97.55 | 60.55 | 98.05 |
| SST (Chen et al., 2020b) | BERT-base | 51.17 | - | 55.23 | - | - | - |
| SimpleTOD (Hosseini-Asl et al., 2020) | DistilGPT2 | - | - | 55.26 | - | - | - |
| Seq2SeqDU (Feng et al., 2021) | BERT-base | - | - | 56.10 | - | - | - |
| STAR (Ye et al., 2021) | BERT-base | 54.53 | - | 56.36 | 97.59 | 73.62 | 98.85 |
| SDP-DST (Lee et al., 2021) | T5-base | - | - | 56.66 | - | - | - |
| DS-Graph (Lin et al., 2021) | GPT2 | 54.86 | 97.47 | - | - | - | - |
| DSGFNet (Feng et al., 2022) | BERT-base | - | - | 56.70 | - | - | - |
| PPTOD (Su et al., 2022) | T5-large | 53.89 | - | 57.45 | - | - | - |
| Baseline | BERT-base | 54.38 | 97.47 | 55.82 | 97.51 | 73.81 | 98.82 |
| MoNET | BERT-base | 55.48 (↑1.10) | 97.55 | 57.71 (↑1.89) | 97.71 | 76.02 (↑2.21) | 98.99 |
| Use Modified Label | |||||||
| TripPy (Heck et al., 2020) | BERT-base | - | - | 55.29 | - | - | - |
| TripPy + SCoRe (Yu et al., 2021) | BERT-base | - | - | 60.48 | - | - | - |
| TripPy + CoCoAug (Li et al., 2021) | BERT-base | - | - | 60.53 | - | - | - |
| TripPy + SaCLog (Dai et al., 2021) | BERT-base | - | - | 60.61 | - | - | - |
| Model | NoisedCM | NoisedST | Accuracy |
| Baseline w/o state | × | × | 64.94 |
| Baseline | × | × | 73.81 |
| MoNET-ST | × | ✓ | 75.54 |
| MoNET-CM | ✓ | × | 75.76 |
| MoNET | ✓ | ✓ | 76.02 |
| Sys: I have many trains available. What day would you like to leave? +Usr: We will be leaving on Sunday afternoon. +GT: train-day-Sunday +Baseline: train-day-Sunday MoNET: train-day-Sunday |
| Sys: There are still many trains to pick from, can you narrow down a departure and arrival time frame? +Usr: Yes, it should leave after 20:15 and leave on Monday, not Sunday. +Please give me a price, travel time, and arrival time for any train. +GT: train-day-Monday +Baseline: train-day-Sunday MoNET: train-day-Monday (update) |
| Sys: Great! You are booked at Autumn House for 1 night. Your reference number is n4tvfkgs. Would you like more information? +Usr: I would also like a taxi between the places if possible. +GT: taxi-destination-Autumn House +Baseline: taxi-destination-Gonville and Caius College +MoNET: taxi-destination-Gonville and Caius College |
| Sys: I can help with that. When would you like to either leave or arrive? +And do you want the taxi from the college to the hotel or the other way around? +Usr: I need to go from the college to the hotel, and I want to leave the college by 20:45, please. +GT: taxi-destination-Autumn House +Baseline: taxi-destination-Gonville and Caius College +MoNET: taxi-destination-Autumn House (correct) |
| Model | Accuracy |
| MinTL (Lin et al., 2020) | 52.07 |
| MTTOD (Lee, 2021) | 53.56 |
| PPTOD (Su et al., 2022) | 53.37 |
| T5-base | 53.26 |
| MoNET (T5-base) | 54.67 (↑ 1.41) |
| Baseline | Pre-trained Model | MultiWOZ 2.0 | MultiWOZ 2.1 | MultiWOZ 2.4 | |||
| Joint | Slot | Joint | Slot | Joint | Slot | ||
| TRADE (Wu et al., 2019) | - | 48.62 | 96.92 | 45.60 | 96.55 | 55.05 | 97.62 |
| SUMBT (Lee et al., 2019) | BERT-base | 42.40 | - | 49.01 | 96.76 | 61.86 | 97.90 |
| PIN (Chen et al., 2020a) | - | 52.44 | 97.28 | 48.40 | 97.02 | 58.92 | 98.02 |
| SOM-DST (Kim et al., 2020) | BERT-base | 51.72 | - | 53.01 | - | 66.78 | 98.38 |
| CSFN-DST (Zhu et al., 2020) | BERT-base | 52.23 | - | 53.19 | - | - | - |
| DST-Picklist (Zhang et al., 2020) | BERT-base | 54.39 | - | 53.30 | 97.40 | - | - |
| SAVN (Wang et al., 2020) | BERT-base | 54.52 | 97.42 | 54.86 | 97.55 | 60.55 | 98.05 |
| SST (Chen et al., 2020b) | BERT-base | 51.17 | - | 55.23 | - | - | - |
| SimpleTOD (Hosseini-Asl et al., 2020) | DistilGPT2 | - | - | 55.26 | - | - | - |
| Seq2SeqDU (Feng et al., 2021) | BERT-base | - | - | 56.10 | - | - | - |
| STAR (Ye et al., 2021) | BERT-base | 54.53 | - | 56.36 | 97.59 | 73.62 | 98.85 |
| SDP-DST (Lee et al., 2021) | T5-base | - | - | 56.66 | - | - | - |
| DS-Graph (Lin et al., 2021) | GPT2 | 54.86 | 97.47 | - | - | - | - |
| DSGFNet (Feng et al., 2022) | BERT-base | - | - | 56.70 | - | - | - |
| PPTOD (Su et al., 2022) | T5-large | 53.89 | - | 57.45 | - | - | - |
| Baseline | BERT-base | 54.38 | 97.47 | 55.82 | 97.51 | 73.81 | 98.82 |
| MoNET | BERT-base | 55.48 (↑1.10) | 97.55 | 57.71 (↑1.89) | 97.71 | 76.02 (↑2.21) | 98.99 |
| Use Modified Label | |||||||
| TripPy (Heck et al., 2020) | BERT-base | - | - | 55.29 | - | - | - |
| TripPy + SCoRe (Yu et al., 2021) | BERT-base | - | - | 60.48 | - | - | - |
| TripPy + CoCoAug (Li et al., 2021) | BERT-base | - | - | 60.53 | - | - | - |
| TripPy + SaCLog (Dai et al., 2021) | BERT-base | - | - | 60.61 | - | - | - |
| Model | NoisedCM | NoisedST | Accuracy |
| Baseline w/o state | × | × | 64.94 |
| Baseline | × | × | 73.81 |
| MoNET-ST | × | ✓ | 75.54 |
| MoNET-CM | ✓ | × | 75.76 |
| MoNET | ✓ | ✓ | 76.02 |
| Sys: I have many trains available. What day would you like to leave?\nUsr: We will be leaving on Sunday afternoon.\nGT: train-day-Sunday\nBaseline: train-day-Sunday MoNET: train-day-Sunday |
| Sys: There are still many trains to pick from, can you narrow down a departure and arrival time frame?\nUsr: Yes, it should leave after 20:15 and leave on Monday, not Sunday.\nPlease give me a price, travel time, and arrival time for any train.\nGT: train-day-Monday\nBaseline: train-day-Sunday MoNET: train-day-Monday (update) |
| Sys: Great! You are booked at Autumn House for 1 night. Your reference number is n4tvfkgs. Would you like more information?\nUsr: I would also like a taxi between the places if possible.\nGT: taxi-destination-Autumn House\nBaseline: taxi-destination-Gonville and Caius College\nMoNET: taxi-destination-Gonville and Caius College |
| Sys: I can help with that. When would you like to either leave or arrive?\nAnd do you want the taxi from the college to the hotel or the other way around?\nUsr: I need to go from the college to the hotel, and I want to leave the college by 20:45, please.\nGT: taxi-destination-Autumn House\nBaseline: taxi-destination-Gonville and Caius College\nMoNET: taxi-destination-Autumn House (correct) |
| Model | Accuracy |
| MinTL (Lin et al., 2020) | 52.07 |
| MTTOD (Lee, 2021) | 53.56 |
| PPTOD (Su et al., 2022) | 53.37 |
| T5-base | 53.26 |
| MoNET (T5-base) | 54.67 (↑ 1.41) |
| ANSWER LENGTH | ROUGE-L | STR-EM | DISAMBIG-F1 | DR | |
| QUESTION | 71.6 | 15.3 | 1.2 | 0.1 | 1.4 |
| T5-base CLOSED BOOK | 38.1 | 30.7 | 3.7 | 2.7 | 9.1 |
| BART-base CLOSED BOOK | 44.5 | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large CLOSED BOOK | 50.2 | 33.4 | 7.1 | 4.5 | 12.2 |
| BM25@1,3,5 | 103.7 / 310.9 / 518.0 | 28.6 / 20.3 / 15.2 | 18.6 / 30.0 / 36.4 | 10.6 / 14.7 / 17.4 | 17.4 / 17.3 / 16.2 |
| DPR@1,3,5 | 103.5 / 310.4 / 517.3 | 31.4 / 22.1 / 16.4 | 29.3 / 44.0 / 50.8 | 17.5 / 22.8 / 25.7 | 23.5 / 22.5 / 20.5 |
| T5-base DPR@1,3,5 | 51.4 / 57.8 / 57.8 | 31.6 / 33.7 / 33.9 | 23.9 / 26.4 / 26.2 | 17.6 / 17.8 / 17.9 | 23.6 / 24.5 / 24.6 |
| T5-base NLGEN DPR@1,3,5 | 48.5 / 56.0 / 56.0 | 31.2 / 33.7 / 33.7 | 24.6 / 26.0 / 26.7 | 16.8 / 18.1 / 17.8 | 22.9 / 24.7 / 24.5 |
| BART-base DPR@1,3,5 | 52.4 / 57.1 / 56.7 | 33.1 / 33.9 / 33.9 | 24.2 / 25.1 / 24.5 | 16.1 / 16.5 / 16.4 | 23.1 / 23.7 / 23.5 |
| BART-large DPR@1,3,5 | 54.7 / 62.5 / 63.3 | 34.2 / 36.4 / 36.6 | 26.0 / 30.0 / 29.8 | 18.1 / 20.8 / 20.5 | 24.9 / 27.5 / 27.4 |
| BART-large ELI5 DPR@1,3,5 | 55.2 / 59.7 / 59.9 | 35.1 / 36.6 / 37.0 | 26.7 / 30.3 / 29.7 | 19.0 / 21.0 / 21.0 | 25.8 / 27.7 / 27.6 |
| COMP | FLUE | OVER | |
| T5-base vs BART-base* | 50% | 38% | 50% |
| BART-large vs BART-base* | 61% | 61% | 61% |
| BART-large ELI5 vs BART-large* | 61% | 50% | 61% |
| BART-large: DPR@3 vs DPR@1 | 75% | 50% | 75% |
| DPR@1 vs BART-base DPR@1 | 61% | 38% | 50% |
| BART-large DPR@1 vs DPR@1 | 38% | 50% | 75% |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 20.5 | 2.0 | 0.7 | 3.7 |
| T5-base NLGEN | 20.8 | 1.6 | 0.8 | 4.1 |
| BART-base | 29.8 | 5.5 | 3.5 | 10.3 |
| BART-large | 24.3 | 4.4 | 2.2 | 7.2 |
| BART-large ELI5 | 29.6 | 4.8 | 3.0 | 9.5 |
| Avg # of short answers retrieved | ||
| In all results | In results with ≥ 1 correct answer | |
| BM25@1 | 0.52 (14.71%) | 1.61 (45.44%) |
| BM25@3 | 0.89 (24.10%) | 1.83 (49.35%) |
| BM25@5 | 1.11 (29.81%) | 1.96 (52.53%) |
| DPR@1 | 0.89 (24.70%) | 1.74 (48.38%) |
| DPR@3 | 1.39 (37.73%) | 2.00 (54.61%) |
| DPR@5 | 1.62 (43.79%) | 2.15 (58.15%) |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 30.7 | 3.7 | 2.7 | 9.1 |
| T5-base NLGEN | 30.0 | 2.9 | 2.8 | 9.2 |
| BART-base | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large | 33.4 | 7.1 | 4.5 | 12.2 |
| BART-large ELI5 | 34.1 | 7.7 | 4.8 | 12.8 |
| Question, provided disambiguations, and answers | Generated model output: BART-large ELI5 DPR@3 | DPR@3 retrieved passages | Answer analysis |
| Q: Who was the ruler of France in 1830? | the french revolution of 1830, also known as the july revolution or in french, the second french revolution, led to the overthrow of king charles x, the french bourbon monarch, and the ascent of his cousin louis philippe, duke of orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848. | July Revolution The French Revolution of 1830, also known as the July Revolution ("révolution de Jullet"), Second French Revolution or in French ("Three Glorious [Days])", led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848.. July Monarchy 1840, to general indifference. The public's attention was concentrated on the trial of Marie Lafarge.. July Revolution ...led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself.. | We do not notice any hallucination or repetition in this example. The model manages to correctly answer the ambiguous question with fluency and coherency. In addition, through reading the answer we understand why the question asked was ambiguous. |
| Q1: Who was the ruler of France until 2 August 1830? | |||
| A1a: Charles Philippe | |||
| A1b: Charles X | |||
| Q2: Who was the ruler of France after 9 August 1830? | |||
| A2: Louis-Philippe I | |||
| Q2: The repetition of phrases or sentences with similar structures or meanings is called? | parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure. the application of parallelism affects readability and may make texts easier to process. | Parallelism (grammar) In grammar, parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure.. Metathesis (linguistics) Metathesis (from Greek, from "I put in a different order"; Latin: "transpositio") is the transposition.. Resultative Resultative In linguistics, a resultative (abbreviated) is a form that expresses that something or someone has undergone a change in state as the result of the completion of an event.. | In this example, the model only manages to get one of the two answers that disambiguate the question. We do not observe any hallucinations or repetition. |
| Q1: The repetition of phrases or sentences with similar structures is called? | |||
| A1a: parallel construction | |||
| A1b: parallelism | |||
| A1c: Parallel syntax | |||
| Q2: The repetition of phrases or sentences with similar meanings is called? | |||
| A2: Schesis onomaton | |||
| Q: When does callie come in grey's anatomy? | grey's anatomy is an american medical drama television series created by shonda rhimes, which air on the abc in the united states. the character callie torres was introduced in season two, as a senior orthopedic resident as a love interest for intern george o'malley (t.r. knight). she was originally contracted to appear on a recurring basis, but received star billing in the third season. at the end of the show's twelfth season, ramirez | Callie Torres "...Callie" Torres, M.D. is a fictional character from the medical drama television series "Greys Anatomy"... She was introduced in season two, as a senior orthopedic resident, as a love interest for intern George O'Malley (T.R. Knight). Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing.. Callie Torres... Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing in the third season.' Callie Torres Ramirez was nominated for several awards for her portrayal of Torres, including the.. | Here the model is not able to get any of the answers correctly. However, one could argue that two of the four disambiguated questions were partially answered since the model outputs that the character was introduced in season two and became a series regular in the third season. We also do not notice any hallucinations or repetition. |
| Q1: What date does Callie Torres first appear in Grey's Anatomy? | |||
| A1: February 26, 2006 | |||
| Q2: Which episode does Callie Torres first appear in Grey's Anatomy? | |||
| A2: "What Have I Done to Deserve This?" (2.19) | |||
| Q3: What date does Callie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A3: September 21, 2006 | |||
| Q4: What episode Calie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A4: "Time Has Come Today" (3.01) |
| ANSWER LENGTH | ROUGE-L | STR-EM | DISAMBIG-F1 | DR | |
| QUESTION | 71.6 | 15.3 | 1.2 | 0.1 | 1.4 |
| T5-base CLOSED BOOK | 38.1 | 30.7 | 3.7 | 2.7 | 9.1 |
| BART-base CLOSED BOOK | 44.5 | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large CLOSED BOOK | 50.2 | 33.4 | 7.1 | 4.5 | 12.2 |
| BM25@1,3,5 | 103.7 / 310.9 / 518.0 | 28.6 / 20.3 / 15.2 | 18.6 / 30.0 / 36.4 | 10.6 / 14.7 / 17.4 | 17.4 / 17.3 / 16.2 |
| DPR@1,3,5 | 103.5 / 310.4 / 517.3 | 31.4 / 22.1 / 16.4 | 29.3 / 44.0 / 50.8 | 17.5 / 22.8 / 25.7 | 23.5 / 22.5 / 20.5 |
| T5-base DPR@1,3,5 | 51.4 / 57.8 / 57.8 | 31.6 / 33.7 / 33.9 | 23.9 / 26.4 / 26.2 | 17.6 / 17.8 / 17.9 | 23.6 / 24.5 / 24.6 |
| T5-base NLGEN DPR@1,3,5 | 48.5 / 56.0 / 56.0 | 31.2 / 33.7 / 33.7 | 24.6 / 26.0 / 26.7 | 16.8 / 18.1 / 17.8 | 22.9 / 24.7 / 24.5 |
| BART-base DPR@1,3,5 | 52.4 / 57.1 / 56.7 | 33.1 / 33.9 / 33.9 | 24.2 / 25.1 / 24.5 | 16.1 / 16.5 / 16.4 | 23.1 / 23.7 / 23.5 |
| BART-large DPR@1,3,5 | 54.7 / 62.5 / 63.3 | 34.2 / 36.4 / 36.6 | 26.0 / 30.0 / 29.8 | 18.1 / 20.8 / 20.5 | 24.9 / 27.5 / 27.4 |
| BART-large ELI5 DPR@1,3,5 | 55.2 / 59.7 / 59.9 | 35.1 / 36.6 / 37.0 | 26.7 / 30.3 / 29.7 | 19.0 / 21.0 / 21.0 | 25.8 / 27.7 / 27.6 |
| COMP | FLUE | OVER | |
| T5-base vs BART-base* | 50% | 38% | 50% |
| BART-large vs BART-base* | 61% | 61% | 61% |
| BART-large ELI5 vs BART-large* | 61% | 50% | 61% |
| BART-large: DPR@3 vs DPR@1 | 75% | 50% | 75% |
| DPR@1 vs BART-base DPR@1 | 61% | 38% | 50% |
| BART-large DPR@1 vs DPR@1 | 38% | 50% | 75% |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 20.5 | 2.0 | 0.7 | 3.7 |
| T5-base NLGEN | 20.8 | 1.6 | 0.8 | 4.1 |
| BART-base | 29.8 | 5.5 | 3.5 | 10.3 |
| BART-large | 24.3 | 4.4 | 2.2 | 7.2 |
| BART-large ELI5 | 29.6 | 4.8 | 3.0 | 9.5 |
| Avg # of short answers retrieved | ||
| In all results | In results with ≥ 1 correct answer | |
| BM25@1 | 0.52 (14.71%) | 1.61 (45.44%) |
| BM25@3 | 0.89 (24.10%) | 1.83 (49.35%) |
| BM25@5 | 1.11 (29.81%) | 1.96 (52.53%) |
| DPR@1 | 0.89 (24.70%) | 1.74 (48.38%) |
| DPR@3 | 1.39 (37.73%) | 2.00 (54.61%) |
| DPR@5 | 1.62 (43.79%) | 2.15 (58.15%) |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 30.7 | 3.7 | 2.7 | 9.1 |
| T5-base NLGEN | 30.0 | 2.9 | 2.8 | 9.2 |
| BART-base | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large | 33.4 | 7.1 | 4.5 | 12.2 |
| BART-large ELI5 | 34.1 | 7.7 | 4.8 | 12.8 |
| Question, provided disambiguations, and answers | Generated model output: BART-large ELI5 DPR@3 | DPR@3 retrieved passages | Answer analysis |
| Q: Who was the ruler of France in 1830? | the french revolution of 1830, also known as the july revolution or in french, the second french revolution, led to the overthrow of king charles x, the french bourbon monarch, and the ascent of his cousin louis philippe, duke of orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848. | July Revolution The French Revolution of 1830, also known as the July Revolution ("révolution de Jullet"), Second French Revolution or in French ("Three Glorious [Days])", led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848.. July Monarchy 1840, to general indifference. The public's attention was concentrated on the trial of Marie Lafarge.. July Revolution ...led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself.. | We do not notice any hallucination or repetition in this example. The model manages to correctly answer the ambiguous question with fluency and coherency. In addition, through reading the answer we understand why the question asked was ambiguous. |
| Q1: Who was the ruler of France until 2 August 1830? | |||
| A1a: Charles Philippe | |||
| A1b: Charles X | |||
| Q2: Who was the ruler of France after 9 August 1830? | |||
| A2: Louis-Philippe I | |||
| Q2: The repetition of phrases or sentences with similar structures or meanings is called? | parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure. the application of parallelism affects readability and may make texts easier to process. | Parallelism (grammar) In grammar, parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure.. Metathesis (linguistics) Metathesis (from Greek, from "I put in a different order"; Latin: "transpositio") is the transposition.. Resultative Resultative In linguistics, a resultative (abbreviated) is a form that expresses that something or someone has undergone a change in state as the result of the completion of an event.. | In this example, the model only manages to get one of the two answers that disambiguate the question. We do not observe any hallucinations or repetition. |
| Q1: The repetition of phrases or sentences with similar structures is called? | |||
| A1a: parallel construction | |||
| A1b: parallelism | |||
| A1c: Parallel syntax | |||
| Q2: The repetition of phrases or sentences with similar meanings is called? | |||
| A2: Schesis onomaton | |||
| Q: When does callie come in grey's anatomy? | grey's anatomy is an american medical drama television series created by shonda rhimes, which air on the abc in the united states. the character callie torres was introduced in season two, as a senior orthopedic resident as a love interest for intern george o'malley (t.r. knight). she was originally contracted to appear on a recurring basis, but received star billing in the third season. at the end of the show's twelfth season, ramirez | Callie Torres "...Callie" Torres, M.D. is a fictional character from the medical drama television series "Greys Anatomy"... She was introduced in season two, as a senior orthopedic resident, as a love interest for intern George O'Malley (T.R. Knight). Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing.. Callie Torres... Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing in the third season.' Callie Torres Ramirez was nominated for several awards for her portrayal of Torres, including the.. | Here the model is not able to get any of the answers correctly. However, one could argue that two of the four disambiguated questions were partially answered since the model outputs that the character was introduced in season two and became a series regular in the third season. We also do not notice any hallucinations or repetition. |
| Q1: What date does Callie Torres first appear in Grey's Anatomy? | |||
| A1: February 26, 2006 | |||
| Q2: Which episode does Callie Torres first appear in Grey's Anatomy? | |||
| A2: "What Have I Done to Deserve This?" (2.19) | |||
| Q3: What date does Callie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A3: September 21, 2006 | |||
| Q4: What episode Calie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A4: "Time Has Come Today" (3.01) |
| ANSWER LENGTH | ROUGE-L | STR-EM | DISAMBIG-F1 | DR | |
| QUESTION | 71.6 | 15.3 | 1.2 | 0.1 | 1.4 |
| T5-base CLOSED BOOK | 38.1 | 30.7 | 3.7 | 2.7 | 9.1 |
| BART-base CLOSED BOOK | 44.5 | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large CLOSED BOOK | 50.2 | 33.4 | 7.1 | 4.5 | 12.2 |
| BM25@1,3,5 | 103.7 / 310.9 / 518.0 | 28.6 / 20.3 / 15.2 | 18.6 / 30.0 / 36.4 | 10.6 / 14.7 / 17.4 | 17.4 / 17.3 / 16.2 |
| DPR@1,3,5 | 103.5 / 310.4 / 517.3 | 31.4 / 22.1 / 16.4 | 29.3 / 44.0 / 50.8 | 17.5 / 22.8 / 25.7 | 23.5 / 22.5 / 20.5 |
| T5-base DPR@1,3,5 | 51.4 / 57.8 / 57.8 | 31.6 / 33.7 / 33.9 | 23.9 / 26.4 / 26.2 | 17.6 / 17.8 / 17.9 | 23.6 / 24.5 / 24.6 |
| T5-base NLGEN DPR@1,3,5 | 48.5 / 56.0 / 56.0 | 31.2 / 33.7 / 33.7 | 24.6 / 26.0 / 26.7 | 16.8 / 18.1 / 17.8 | 22.9 / 24.7 / 24.5 |
| BART-base DPR@1,3,5 | 52.4 / 57.1 / 56.7 | 33.1 / 33.9 / 33.9 | 24.2 / 25.1 / 24.5 | 16.1 / 16.5 / 16.4 | 23.1 / 23.7 / 23.5 |
| BART-large DPR@1,3,5 | 54.7 / 62.5 / 63.3 | 34.2 / 36.4 / 36.6 | 26.0 / 30.0 / 29.8 | 18.1 / 20.8 / 20.5 | 24.9 / 27.5 / 27.4 |
| BART-large ELI5 DPR@1,3,5 | 55.2 / 59.7 / 59.9 | 35.1 / 36.6 / 37.0 | 26.7 / 30.3 / 29.7 | 19.0 / 21.0 / 21.0 | 25.8 / 27.7 / 27.6 |
| COMP | FLUE | OVER | |
| T5-base vs BART-base* | 50% | 38% | 50% |
| BART-large vs BART-base* | 61% | 61% | 61% |
| BART-large ELI5 vs BART-large* | 61% | 50% | 61% |
| BART-large: DPR@3 vs DPR@1 | 75% | 50% | 75% |
| DPR@1 vs BART-base DPR@1 | 61% | 38% | 50% |
| BART-large DPR@1 vs DPR@1 | 38% | 50% | 75% |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 20.5 | 2.0 | 0.7 | 3.7 |
| T5-base NLGEN | 20.8 | 1.6 | 0.8 | 4.1 |
| BART-base | 29.8 | 5.5 | 3.5 | 10.3 |
| BART-large | 24.3 | 4.4 | 2.2 | 7.2 |
| BART-large ELI5 | 29.6 | 4.8 | 3.0 | 9.5 |
| Avg # of short answers retrieved | ||
| In all results | In results with ≥ 1 correct answer | |
| BM25@1 | 0.52 (14.71%) | 1.61 (45.44%) |
| BM25@3 | 0.89 (24.10%) | 1.83 (49.35%) |
| BM25@5 | 1.11 (29.81%) | 1.96 (52.53%) |
| DPR@1 | 0.89 (24.70%) | 1.74 (48.38%) |
| DPR@3 | 1.39 (37.73%) | 2.00 (54.61%) |
| DPR@5 | 1.62 (43.79%) | 2.15 (58.15%) |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 30.7 | 3.7 | 2.7 | 9.1 |
| T5-base NLGEN | 30.0 | 2.9 | 2.8 | 9.2 |
| BART-base | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large | 33.4 | 7.1 | 4.5 | 12.2 |
| BART-large ELI5 | 34.1 | 7.7 | 4.8 | 12.8 |
| Question, provided disambiguations, and answers | Generated model output: BART-large ELI5 DPR@3 | DPR@3 retrieved passages | Answer analysis |
| Q: Who was the ruler of France in 1830? | the french revolution of 1830, also known as the july revolution or in french, the second french revolution, led to the overthrow of king charles x, the french bourbon monarch, and the ascent of his cousin louis philippe, duke of orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848. | July Revolution The French Revolution of 1830, also known as the July Revolution ("révolution de Jullet"), Second French Revolution or in French ("Three Glorious [Days])", led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848.. July Monarchy 1840, to general indifference. The public's attention was concentrated on the trial of Marie Lafarge.. July Revolution ...led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself.. | We do not notice any hallucination or repetition in this example. The model manages to correctly answer the ambiguous question with fluency and coherency. In addition, through reading the answer we understand why the question asked was ambiguous. |
| Q1: Who was the ruler of France until 2 August 1830? | |||
| A1a: Charles Philippe | |||
| A1b: Charles X | |||
| Q2: Who was the ruler of France after 9 August 1830? | |||
| A2: Louis-Philippe I | |||
| Q2: The repetition of phrases or sentences with similar structures or meanings is called? | parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure. the application of parallelism affects readability and may make texts easier to process. | Parallelism (grammar) In grammar, parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure.. Metathesis (linguistics) Metathesis (from Greek, from "I put in a different order"; Latin: "transpositio") is the transposition.. Resultative Resultative In linguistics, a resultative (abbreviated) is a form that expresses that something or someone has undergone a change in state as the result of the completion of an event.. | In this example, the model only manages to get one of the two answers that disambiguate the question. We do not observe any hallucinations or repetition. |
| Q1: The repetition of phrases or sentences with similar structures is called? | |||
| A1a: parallel construction | |||
| A1b: parallelism | |||
| A1c: Parallel syntax | |||
| Q2: The repetition of phrases or sentences with similar meanings is called? | |||
| A2: Schesis onomaton | |||
| Q: When does callie come in grey's anatomy? | grey's anatomy is an american medical drama television series created by shonda rhimes, which air on the abc in the united states. the character callie torres was introduced in season two, as a senior orthopedic resident as a love interest for intern george o'malley (t.r. knight). she was originally contracted to appear on a recurring basis, but received star billing in the third season. at the end of the show's twelfth season, ramirez | Callie Torres "...Callie" Torres, M.D. is a fictional character from the medical drama television series "Greys Anatomy"... She was introduced in season two, as a senior orthopedic resident, as a love interest for intern George O'Malley (T.R. Knight). Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing.. Callie Torres... Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing in the third season.' Callie Torres Ramirez was nominated for several awards for her portrayal of Torres, including the.. | Here the model is not able to get any of the answers correctly. However, one could argue that two of the four disambiguated questions were partially answered since the model outputs that the character was introduced in season two and became a series regular in the third season. We also do not notice any hallucinations or repetition. |
| Q1: What date does Callie Torres first appear in Grey's Anatomy? | |||
| A1: February 26, 2006 | |||
| Q2: Which episode does Callie Torres first appear in Grey's Anatomy? | |||
| A2: "What Have I Done to Deserve This?" (2.19) | |||
| Q3: What date does Callie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A3: September 21, 2006 | |||
| Q4: What episode Calie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A4: "Time Has Come Today" (3.01) |
| ANSWER LENGTH | ROUGE-L | STR-EM | DISAMBIG-F1 | DR | |
| QUESTION | 71.6 | 15.3 | 1.2 | 0.1 | 1.4 |
| T5-base CLOSED BOOK | 38.1 | 30.7 | 3.7 | 2.7 | 9.1 |
| BART-base CLOSED BOOK | 44.5 | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large CLOSED BOOK | 50.2 | 33.4 | 7.1 | 4.5 | 12.2 |
| BM25@1,3,5 | 103.7 / 310.9 / 518.0 | 28.6 / 20.3 / 15.2 | 18.6 / 30.0 / 36.4 | 10.6 / 14.7 / 17.4 | 17.4 / 17.3 / 16.2 |
| DPR@1,3,5 | 103.5 / 310.4 / 517.3 | 31.4 / 22.1 / 16.4 | 29.3 / 44.0 / 50.8 | 17.5 / 22.8 / 25.7 | 23.5 / 22.5 / 20.5 |
| T5-base DPR@1,3,5 | 51.4 / 57.8 / 57.8 | 31.6 / 33.7 / 33.9 | 23.9 / 26.4 / 26.2 | 17.6 / 17.8 / 17.9 | 23.6 / 24.5 / 24.6 |
| T5-base NLGEN DPR@1,3,5 | 48.5 / 56.0 / 56.0 | 31.2 / 33.7 / 33.7 | 24.6 / 26.0 / 26.7 | 16.8 / 18.1 / 17.8 | 22.9 / 24.7 / 24.5 |
| BART-base DPR@1,3,5 | 52.4 / 57.1 / 56.7 | 33.1 / 33.9 / 33.9 | 24.2 / 25.1 / 24.5 | 16.1 / 16.5 / 16.4 | 23.1 / 23.7 / 23.5 |
| BART-large DPR@1,3,5 | 54.7 / 62.5 / 63.3 | 34.2 / 36.4 / 36.6 | 26.0 / 30.0 / 29.8 | 18.1 / 20.8 / 20.5 | 24.9 / 27.5 / 27.4 |
| BART-large ELI5 DPR@1,3,5 | 55.2 / 59.7 / 59.9 | 35.1 / 36.6 / 37.0 | 26.7 / 30.3 / 29.7 | 19.0 / 21.0 / 21.0 | 25.8 / 27.7 / 27.6 |
| COMP | FLUE | OVER | |
| T5-base vs BART-base* | 50% | 38% | 50% |
| BART-large vs BART-base* | 61% | 61% | 61% |
| BART-large ELI5 vs BART-large* | 61% | 50% | 61% |
| BART-large: DPR@3 vs DPR@1 | 75% | 50% | 75% |
| DPR@1 vs BART-base DPR@1 | 61% | 38% | 50% |
| BART-large DPR@1 vs DPR@1 | 38% | 50% | 75% |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 20.5 | 2.0 | 0.7 | 3.7 |
| T5-base NLGEN | 20.8 | 1.6 | 0.8 | 4.1 |
| BART-base | 29.8 | 5.5 | 3.5 | 10.3 |
| BART-large | 24.3 | 4.4 | 2.2 | 7.2 |
| BART-large ELI5 | 29.6 | 4.8 | 3.0 | 9.5 |
| Avg # of short answers retrieved | ||
| In all results | In results with ≥ 1 correct answer | |
| BM25@1 | 0.52 (14.71%) | 1.61 (45.44%) |
| BM25@3 | 0.89 (24.10%) | 1.83 (49.35%) |
| BM25@5 | 1.11 (29.81%) | 1.96 (52.53%) |
| DPR@1 | 0.89 (24.70%) | 1.74 (48.38%) |
| DPR@3 | 1.39 (37.73%) | 2.00 (54.61%) |
| DPR@5 | 1.62 (43.79%) | 2.15 (58.15%) |
| ROUGE-L | STR-EM | DIS-F1 | DR | |
| T5-base | 30.7 | 3.7 | 2.7 | 9.1 |
| T5-base NLGEN | 30.0 | 2.9 | 2.8 | 9.2 |
| BART-base | 31.5 | 3.9 | 2.8 | 9.3 |
| BART-large | 33.4 | 7.1 | 4.5 | 12.2 |
| BART-large ELI5 | 34.1 | 7.7 | 4.8 | 12.8 |
| Question, provided disambiguations, and answers | Generated model output: BART-large ELI5 DPR@3 | DPR@3 retrieved passages | Answer analysis |
| Q: Who was the ruler of France in 1830? | the french revolution of 1830, also known as the july revolution or in french, the second french revolution, led to the overthrow of king charles x, the french bourbon monarch, and the ascent of his cousin louis philippe, duke of orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848. | July Revolution The French Revolution of 1830, also known as the July Revolution ("révolution de Jullet"), Second French Revolution or in French ("Three Glorious [Days])", led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself, after 18 precarious years on the throne, would be overthrown in 1848.. July Monarchy 1840, to general indifference. The public's attention was concentrated on the trial of Marie Lafarge.. July Revolution ...led to the overthrow of King Charles X, the French bourbon monarch, and the ascent of his cousin Louis Philippe, Duke of Orléans, who himself.. | We do not notice any hallucination or repetition in this example. The model manages to correctly answer the ambiguous question with fluency and coherency. In addition, through reading the answer we understand why the question asked was ambiguous. |
| Q1: Who was the ruler of France until 2 August 1830? | |||
| A1a: Charles Philippe | |||
| A1b: Charles X | |||
| Q2: Who was the ruler of France after 9 August 1830? | |||
| A2: Louis-Philippe I | |||
| Q2: The repetition of phrases or sentences with similar structures or meanings is called? | parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure. the application of parallelism affects readability and may make texts easier to process. | Parallelism (grammar) In grammar, parallelism, also known as parallel structure or parallel construction, is a balance within one or more sentences of similar phrases or clauses that have the same grammatical structure.. Metathesis (linguistics) Metathesis (from Greek, from "I put in a different order"; Latin: "transpositio") is the transposition.. Resultative Resultative In linguistics, a resultative (abbreviated) is a form that expresses that something or someone has undergone a change in state as the result of the completion of an event.. | In this example, the model only manages to get one of the two answers that disambiguate the question. We do not observe any hallucinations or repetition. |
| Q1: The repetition of phrases or sentences with similar structures is called? | |||
| A1a: parallel construction | |||
| A1b: parallelism | |||
| A1c: Parallel syntax | |||
| Q2: The repetition of phrases or sentences with similar meanings is called? | |||
| A2: Schesis onomaton | |||
| Q: When does callie come in grey's anatomy? | grey's anatomy is an american medical drama television series created by shonda rhimes, which air on the abc in the united states. the character callie torres was introduced in season two, as a senior orthopedic resident as a love interest for intern george o'malley (t.r. knight). she was originally contracted to appear on a recurring basis, but received star billing in the third season. at the end of the show's twelfth season, ramirez | Callie Torres "...Callie" Torres, M.D. is a fictional character from the medical drama television series "Greys Anatomy"... She was introduced in season two, as a senior orthopedic resident, as a love interest for intern George O'Malley (T.R. Knight). Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing.. Callie Torres... Eventually becoming an attending orthopedic surgeon, the character was originally contracted to appear on a recurring basis, but received star billing in the third season.' Callie Torres Ramirez was nominated for several awards for her portrayal of Torres, including the.. | Here the model is not able to get any of the answers correctly. However, one could argue that two of the four disambiguated questions were partially answered since the model outputs that the character was introduced in season two and became a series regular in the third season. We also do not notice any hallucinations or repetition. |
| Q1: What date does Callie Torres first appear in Grey's Anatomy? | |||
| A1: February 26, 2006 | |||
| Q2: Which episode does Callie Torres first appear in Grey's Anatomy? | |||
| A2: "What Have I Done to Deserve This?" (2.19) | |||
| Q3: What date does Callie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A3: September 21, 2006 | |||
| Q4: What episode Calie Torres start to appear in Grey's Anatomy as a series regular? | |||
| A4: "Time Has Come Today" (3.01) |
| Method | Token F1 | mAP | IoU F1 | Suff.↓ | Comp.↑ |
| MaRC | .473 | .469 | .163 | .028 | .518 |
| Occlusion | .432 | .448 | .125 | .022 | .415 |
| Saliencyn | .435 | .392 | .04 | .132 | .287 |
| Saliencys | .425 | .340 | .076 | .260 | .246 |
| InXGradn | .436 | .396 | .040 | .136 | .292 |
| InXGradS | .425 | .340 | .084 | .239 | .248 |
| Int. Gradsn | .428 | .369 | .036 | .122 | .274 |
| Int. GradsS | .431 | .381 | .071 | .048 | .528 |
| LIME | .436 | .380 | .076 | .047 | .496 |
| Shapley | .428 | .439 | .079 | -.015 | .728 |
| Noise | .454 | .450 | .139 | .034 | .487 |
| Ωλ | .349 | .350 | .046 | .120 | .266 |
| Ωσ | .447 | .425 | .091 | .036 | .535 |
| L(˜xc, c) | .396 | .436 | .123 | .052 | .304 |
| ResNet-101 | ViT-B/16 | |||
| Method | Suff. ↓ | Comp.↑ | Suff. ↓ | Comp.↑ |
| MaRC | .196 | .612 | .139 | .596 |
| M-Perturb | .260 | .605 | .174 | .572 |
| Grad-CAM | .197 | .600 | .161 | .640 |
| Exc-BP | .302 | .600 | - | - |
| Saliency | .442 | .599 | .355 | .528 |
| InputXGrad | .430 | .586 | .366 | .506 |
| Guided BP | .343 | .630 | - | - |
| Intgr. Grads | .344 | .641 | .261 | .641 |
| Occlusion | .324 | .606 | .194 | .486 |
| Attention | - | - | .241 | .562 |
| Attribution | - | - | .176 | .608 |
| Rollout | - | - | .205 | .580 |
| TAM | - | - | .146 | .658 |
| Method | Token F1 | mAP | IoU F1 | Suff.↓ | Comp.↑ |
| MaRC | .473 | .469 | .163 | .028 | .518 |
| Occlusion | .432 | .448 | .125 | .022 | .415 |
| Saliencyn | .435 | .392 | .04 | .132 | .287 |
| Saliencys | .425 | .340 | .076 | .260 | .246 |
| InXGradn | .436 | .396 | .040 | .136 | .292 |
| InXGradS | .425 | .340 | .084 | .239 | .248 |
| Int. Gradsn | .428 | .369 | .036 | .122 | .274 |
| Int. GradsS | .431 | .381 | .071 | .048 | .528 |
| LIME | .436 | .380 | .076 | .047 | .496 |
| Shapley | .428 | .439 | .079 | -.015 | .728 |
| Noise | .454 | .450 | .139 | .034 | .487 |
| Ωλ | .349 | .350 | .046 | .120 | .266 |
| Ωσ | .447 | .425 | .091 | .036 | .535 |
| L(˜xc, c) | .396 | .436 | .123 | .052 | .304 |
| ResNet-101 | ViT-B/16 | |||
| Method | Suff. ↓ | Comp.↑ | Suff. ↓ | Comp.↑ |
| MaRC | .196 | .612 | .139 | .596 |
| M-Perturb | .260 | .605 | .174 | .572 |
| Grad-CAM | .197 | .600 | .161 | .640 |
| Exc-BP | .302 | .600 | - | - |
| Saliency | .442 | .599 | .355 | .528 |
| InputXGrad | .430 | .586 | .366 | .506 |
| Guided BP | .343 | .630 | - | - |
| Intgr. Grads | .344 | .641 | .261 | .641 |
| Occlusion | .324 | .606 | .194 | .486 |
| Attention | - | - | .241 | .562 |
| Attribution | - | - | .176 | .608 |
| Rollout | - | - | .205 | .580 |
| TAM | - | - | .146 | .658 |
| Method | Token F1 | mAP | IoU F1 | Suff.↓ | Comp.↑ |
| MaRC | .473 | .469 | .163 | .028 | .518 |
| Occlusion | .432 | .448 | .125 | .022 | .415 |
| Saliencyn | .435 | .392 | .04 | .132 | .287 |
| Saliencys | .425 | .340 | .076 | .260 | .246 |
| InXGradn | .436 | .396 | .040 | .136 | .292 |
| InXGradS | .425 | .340 | .084 | .239 | .248 |
| Int. Gradsn | .428 | .369 | .036 | .122 | .274 |
| Int. GradsS | .431 | .381 | .071 | .048 | .528 |
| LIME | .436 | .380 | .076 | .047 | .496 |
| Shapley | .428 | .439 | .079 | -.015 | .728 |
| Noise | .454 | .450 | .139 | .034 | .487 |
| Ωλ | .349 | .350 | .046 | .120 | .266 |
| Ωσ | .447 | .425 | .091 | .036 | .535 |
| L(˜xc, c) | .396 | .436 | .123 | .052 | .304 |
| ResNet-101 | ViT-B/16 | |||
| Method | Suff. ↓ | Comp.↑ | Suff. ↓ | Comp.↑ |
| MaRC | .196 | .612 | .139 | .596 |
| M-Perturb | .260 | .605 | .174 | .572 |
| Grad-CAM | .197 | .600 | .161 | .640 |
| Exc-BP | .302 | .600 | - | - |
| Saliency | .442 | .599 | .355 | .528 |
| InputXGrad | .430 | .586 | .366 | .506 |
| Guided BP | .343 | .630 | - | - |
| Intgr. Grads | .344 | .641 | .261 | .641 |
| Occlusion | .324 | .606 | .194 | .486 |
| Attention | - | - | .241 | .562 |
| Attribution | - | - | .176 | .608 |
| Rollout | - | - | .205 | .580 |
| TAM | - | - | .146 | .658 |
| Method | Token F1 | mAP | IoU F1 | Suff.↓ | Comp.↑ |
| MaRC | .473 | .469 | .163 | .028 | .518 |
| Occlusion | .432 | .448 | .125 | .022 | .415 |
| Saliencyn | .435 | .392 | .04 | .132 | .287 |
| Saliencys | .425 | .340 | .076 | .260 | .246 |
| InXGradn | .436 | .396 | .040 | .136 | .292 |
| InXGradS | .425 | .340 | .084 | .239 | .248 |
| Int. Gradsn | .428 | .369 | .036 | .122 | .274 |
| Int. GradsS | .431 | .381 | .071 | .048 | .528 |
| LIME | .436 | .380 | .076 | .047 | .496 |
| Shapley | .428 | .439 | .079 | -.015 | .728 |
| Noise | .454 | .450 | .139 | .034 | .487 |
| Ωλ | .349 | .350 | .046 | .120 | .266 |
| Ωσ | .447 | .425 | .091 | .036 | .535 |
| L(˜xc, c) | .396 | .436 | .123 | .052 | .304 |
| ResNet-101 | ViT-B/16 | |||
| Method | Suff. ↓ | Comp.↑ | Suff. ↓ | Comp.↑ |
| MaRC | .196 | .612 | .139 | .596 |
| M-Perturb | .260 | .605 | .174 | .572 |
| Grad-CAM | .197 | .600 | .161 | .640 |
| Exc-BP | .302 | .600 | - | - |
| Saliency | .442 | .599 | .355 | .528 |
| InputXGrad | .430 | .586 | .366 | .506 |
| Guided BP | .343 | .630 | - | - |
| Intgr. Grads | .344 | .641 | .261 | .641 |
| Occlusion | .324 | .606 | .194 | .486 |
| Attention | - | - | .241 | .562 |
| Attribution | - | - | .176 | .608 |
| Rollout | - | - | .205 | .580 |
| TAM | - | - | .146 | .658 |
| Document |
| LudoSport has opened its first academy teaching seven forms of combat from the Star Wars world using flexible blades mounted on weighted hilts. The sport began eight years ago in Italy but has only just come to England with the first classes in Cheltenham. Instructor Jordan Court said people were already "hooked". ... So far there are six pupils, but this number is expected to increase. ... The sport is so new to England that there have only been a handful of classes so far ... Lightsaber Combat Academy There are several ranks for those wishing to become a fully-fledged Jedi Knight: |
| Summary |
| High LUC:A lightsaber-wielding martial artist has opened an academy in Cheltenham to teach people how to "fight like a Jedi". |
| High LUC:A new martial art inspired by the Star Wars franchise has come to the UK for the first time. |
| Low LUC:Hundreds of people are taking part in the UK's first lightsaber combat class. |
| Low LUC:If you're a fan of Star Wars, you might want to think twice before taking up lightsaber combat. |
| Extractive | Abstractive | |||||
| Align | Shuffle | Δ | Align | Shuffle | Δ | |
| Negative log-likelihood | ||||||
| BART-CNN | 0.775 | 0.085 | 0.690 | 0.222 | 0.058 | 0.164 |
| BART-XSum | 0.600 | 0.083 | 0.517 | 0.198 | 0.057 | 0.141 |
| Unlikelihood | ||||||
| BART-CNN | 0.339 | 0.016 | 0.323 | 0.101 | 0.017 | 0.085 |
| BART-XSum | 0.391 | 0.024 | 0.368 | 0.108 | 0.021 | 0.086 |
| Contrastive loss | ||||||
| BART-CNN | 0.762 | 0.068 | 0.695 | 0.213 | 0.046 | 0.167 |
| BART-XSum | 0.549 | 0.073 | 0.476 | 0.187 | 0.051 | 0.136 |
| XSum | CNNDM | |||
| High | Low | High | Low | |
| -LC | 21.24 ± 0.09 | 21.05 ± 0.06 | 46.08 ± 0.24 | 46.21 ± 0.23 |
| LCU | 21.24 ± 0.06 | 20.86 ± 0.06 | 46.30 ± 0.15 | 45.94 ± 0.08 |
| LΔ | 21.40 ± 0.09 | 20.77 ± 0.05 | 46.21 ± 0.24 | 46.15 ± 0.16 |
| Train | Validation | Test | |
| XSum | 204,045 | 11,332 | 11,334 |
| CNN/DailyMail | 287,227 | 13,368 | 11,490 |
| Newsroom (Ext) | 332,131 | 36,372 | 36,165 |
| Newsroom (Abs) | 333,350 | 36,480 | 36,595 |
| FT Steps | FT Time | GPU(s) | |
| BART-CNN | 20k | 6 hours | Tesla V100×4 |
| BART-XSum | 15k | 4 hours | Tesla V100×4 |
| GPT-2-CNN | 12k | 9 hours | Tesla V100×2 |
| GPT-2-XSum | 9k | 7 hours | Tesla V100×2 |
| NLL | CL | |
| CNNDM (Align) | 0.4485 | 0.3382 |
| CNNDM (Shuffle) | 0.0819 | 0.0354 |
| XSum (Align) | 0.4239 | 0.3779 |
| XSum (Shuffle) | 0.1076 | 0.0885 |
| Document |
| LudoSport has opened its first academy teaching seven forms of combat from the Star Wars world using flexible blades mounted on weighted hilts. The sport began eight years ago in Italy but has only just come to England with the first classes in Cheltenham. Instructor Jordan Court said people were already "hooked". ... So far there are six pupils, but this number is expected to increase. ... The sport is so new to England that there have only been a handful of classes so far ... Lightsaber Combat Academy There are several ranks for those wishing to become a fully-fledged Jedi Knight: |
| Summary |
| High LUC:A lightsaber-wielding martial artist has opened an academy in Cheltenham to teach people how to "fight like a Jedi". |
| High LUC:A new martial art inspired by the Star Wars franchise has come to the UK for the first time. |
| Low LUC:Hundreds of people are taking part in the UK's first lightsaber combat class. |
| Low LUC:If you're a fan of Star Wars, you might want to think twice before taking up lightsaber combat. |
| Extractive | Abstractive | |||||
| Align | Shuffle | Δ | Align | Shuffle | Δ | |
| Negative log-likelihood | ||||||
| BART-CNN | 0.775 | 0.085 | 0.690 | 0.222 | 0.058 | 0.164 |
| BART-XSum | 0.600 | 0.083 | 0.517 | 0.198 | 0.057 | 0.141 |
| Unlikelihood | ||||||
| BART-CNN | 0.339 | 0.016 | 0.323 | 0.101 | 0.017 | 0.085 |
| BART-XSum | 0.391 | 0.024 | 0.368 | 0.108 | 0.021 | 0.086 |
| Contrastive loss | ||||||
| BART-CNN | 0.762 | 0.068 | 0.695 | 0.213 | 0.046 | 0.167 |
| BART-XSum | 0.549 | 0.073 | 0.476 | 0.187 | 0.051 | 0.136 |
| XSum | CNNDM | |||
| High | Low | High | Low | |
| -LC | 21.24 ± 0.09 | 21.05 ± 0.06 | 46.08 ± 0.24 | 46.21 ± 0.23 |
| LCU | 21.24 ± 0.06 | 20.86 ± 0.06 | 46.30 ± 0.15 | 45.94 ± 0.08 |
| LΔ | 21.40 ± 0.09 | 20.77 ± 0.05 | 46.21 ± 0.24 | 46.15 ± 0.16 |
| Train | Validation | Test | |
| XSum | 204,045 | 11,332 | 11,334 |
| CNN/DailyMail | 287,227 | 13,368 | 11,490 |
| Newsroom (Ext) | 332,131 | 36,372 | 36,165 |
| Newsroom (Abs) | 333,350 | 36,480 | 36,595 |
| FT Steps | FT Time | GPU(s) | |
| BART-CNN | 20k | 6 hours | Tesla V100×4 |
| BART-XSum | 15k | 4 hours | Tesla V100×4 |
| GPT-2-CNN | 12k | 9 hours | Tesla V100×2 |
| GPT-2-XSum | 9k | 7 hours | Tesla V100×2 |
| NLL | CL | |
| CNNDM (Align) | 0.4485 | 0.3382 |
| CNNDM (Shuffle) | 0.0819 | 0.0354 |
| XSum (Align) | 0.4239 | 0.3779 |
| XSum (Shuffle) | 0.1076 | 0.0885 |
| Document |
| LudoSport has opened its first academy teaching seven forms of combat from the Star Wars world using flexible blades mounted on weighted hilts. The sport began eight years ago in Italy but has only just come to England with the first classes in Cheltenham. Instructor Jordan Court said people were already "hooked". ... So far there are six pupils, but this number is expected to increase. ... The sport is so new to England that there have only been a handful of classes so far ... Lightsaber Combat Academy There are several ranks for those wishing to become a fully-fledged Jedi Knight: |
| Summary |
| High LUC:A lightsaber-wielding martial artist has opened an academy in Cheltenham to teach people how to "fight like a Jedi". |
| High LUC:A new martial art inspired by the Star Wars franchise has come to the UK for the first time. |
| Low LUC:Hundreds of people are taking part in the UK's first lightsaber combat class. |
| Low LUC:If you're a fan of Star Wars, you might want to think twice before taking up lightsaber combat. |
| XSum | CNNDM | |||
| High | Low | High | Low | |
| -LC | 21.24 ± 0.09 | 21.05 ± 0.06 | 46.08 ± 0.24 | 46.21 ± 0.23 |
| LCU | 21.24 ± 0.06 | 20.86 ± 0.06 | 46.30 ± 0.15 | 45.94 ± 0.08 |
| LΔ | 21.40 ± 0.09 | 20.77 ± 0.05 | 46.21 ± 0.24 | 46.15 ± 0.16 |
| Train | Validation | Test | |
| XSum | 204,045 | 11,332 | 11,334 |
| CNN/DailyMail | 287,227 | 13,368 | 11,490 |
| Newsroom (Ext) | 332,131 | 36,372 | 36,165 |
| Newsroom (Abs) | 333,350 | 36,480 | 36,595 |
| FT Steps | FT Time | GPU(s) | |
| BART-CNN | 20k | 6 hours | Tesla V100×4 |
| BART-XSum | 15k | 4 hours | Tesla V100×4 |
| GPT-2-CNN | 12k | 9 hours | Tesla V100×2 |
| GPT-2-XSum | 9k | 7 hours | Tesla V100×2 |
| NLL | CL | |
| CNNDM (Align) | 0.4485 | 0.3382 |
| CNNDM (Shuffle) | 0.0819 | 0.0354 |
| XSum (Align) | 0.4239 | 0.3779 |
| XSum (Shuffle) | 0.1076 | 0.0885 |
| Document |
| LudoSport has opened its first academy teaching seven forms of combat from the Star Wars world using flexible blades mounted on weighted hilts. The sport began eight years ago in Italy but has only just come to England with the first classes in Cheltenham. Instructor Jordan Court said people were already "hooked". ... So far there are six pupils, but this number is expected to increase. ... The sport is so new to England that there have only been a handful of classes so far ... Lightsaber Combat Academy There are several ranks for those wishing to become a fully-fledged Jedi Knight: |
| Summary |
| High LUC:A lightsaber-wielding martial artist has opened an academy in Cheltenham to teach people how to "fight like a Jedi". |
| High LUC:A new martial art inspired by the Star Wars franchise has come to the UK for the first time. |
| Low LUC:Hundreds of people are taking part in the UK's first lightsaber combat class. |
| Low LUC:If you're a fan of Star Wars, you might want to think twice before taking up lightsaber combat. |
| Extractive | Abstractive | |||||
| Align | Shuffle | Δ | Align | Shuffle | Δ | |
| Negative log-likelihood | ||||||
| BART-CNN | 0.775 | 0.085 | 0.690 | 0.222 | 0.058 | 0.164 |
| BART-XSum | 0.600 | 0.083 | 0.517 | 0.198 | 0.057 | 0.141 |
| Unlikelihood | ||||||
| BART-CNN | 0.339 | 0.016 | 0.323 | 0.101 | 0.017 | 0.085 |
| BART-XSum | 0.391 | 0.024 | 0.368 | 0.108 | 0.021 | 0.086 |
| Contrastive loss | ||||||
| BART-CNN | 0.762 | 0.068 | 0.695 | 0.213 | 0.046 | 0.167 |
| BART-XSum | 0.549 | 0.073 | 0.476 | 0.187 | 0.051 | 0.136 |
| XSum | CNNDM | |||
| High | Low | High | Low | |
| -LC | 21.24 ± 0.09 | 21.05 ± 0.06 | 46.08 ± 0.24 | 46.21 ± 0.23 |
| LCU | 21.24 ± 0.06 | 20.86 ± 0.06 | 46.30 ± 0.15 | 45.94 ± 0.08 |
| LΔ | 21.40 ± 0.09 | 20.77 ± 0.05 | 46.21 ± 0.24 | 46.15 ± 0.16 |
| Train | Validation | Test | |
| XSum | 204,045 | 11,332 | 11,334 |
| CNN/DailyMail | 287,227 | 13,368 | 11,490 |
| Newsroom (Ext) | 332,131 | 36,372 | 36,165 |
| Newsroom (Abs) | 333,350 | 36,480 | 36,595 |
| FT Steps | FT Time | GPU(s) | |
| BART-CNN | 20k | 6 hours | Tesla V100×4 |
| BART-XSum | 15k | 4 hours | Tesla V100×4 |
| GPT-2-CNN | 12k | 9 hours | Tesla V100×2 |
| GPT-2-XSum | 9k | 7 hours | Tesla V100×2 |
| NLL | CL | |
| CNNDM (Align) | 0.4485 | 0.3382 |
| CNNDM (Shuffle) | 0.0819 | 0.0354 |
| XSum (Align) | 0.4239 | 0.3779 |
| XSum (Shuffle) | 0.1076 | 0.0885 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ | Dataset | Method | A-rate ↑ | Mod ↓ | Sim↑ |
| Yelp | A2T | 88.3 | 8.1 | 0.68 | IMDB | A2T | 89.9 | 4.4 | 0.79 |
| TextFooler | 90.5 | 9.0 | 0.69 | TextFooler | 88.7 | 7.6 | 0.76 | ||
| BERT-Attack | 89.8 | 12.4 | 0.66 | BERT-Attack | 88.2 | 5.3 | 0.78 | ||
| SDM-ATTACK | 95.8 | 8.2 | 0.71 | SDM-ATTACK | 91.4 | 4.1 | 0.82 | ||
| AG's News | A2T | 53.7 | 13.5 | 0.57 | MR | A2T | 58.5 | 12.6 | 0.55 |
| TextFooler | 66.2 | 18.4 | 0.52 | TextFooler | 80.5 | 15.8 | 0.50 | ||
| BERT-Attack | 74.6 | 15.6 | 0.52 | BERT-Attack | 83.2 | 12.8 | 0.52 | ||
| SDM-ATTACK | 77.9 | 15.3 | 0.53 | SDM-ATTACK | 85.6 | 12.3 | 0.57 | ||
| SNLI | A2T | 70.8 | 17.2 | 0.35 | MNLI | A2T | 66.0 | 14.4 | 0.45 |
| TextFooler | 84.3 | 17.2 | 0.38 | TextFooler | 76.5 | 15.0 | 0.45 | ||
| BERT-Attack | 81.9 | 16.5 | 0.38 | BERT-Attack | 78.1 | 14.0 | 0.46 | ||
| SDM-ATTACK | 85.5 | 15.9 | 0.43 | SDM-ATTACK | 78.7 | 13.8 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | 0.95 |
| SDM-ATTACK | 0.90 | 4.3 | ||
| MNLI | Original | 0.88 | 4.0 | 0.83 |
| SDM-ATTACK | 0.79 | 3.7 |
| Dataset | Model | A-rate↑ | Mod↓ | Sim↑ |
| MR | RoBERTa | 84.4 | 13.9 | 0.52 |
| WordCNN | 72.1 | 10.3 | 0.48 | |
| WordLSTM | 80.7 | 8.9 | 0.56 | |
| IMDB | RoBERTa | 88.3 | 8.3 | 0.70 |
| WordCNN | 89.2 | 3.3 | 0.85 | |
| WordLSTM | 89.8 | 5.4 | 0.75 | |
| SNLI | InferSent | 78.7 | 17.0 | 0.42 |
| ESIM | 79.0 | 17.2 | 0.41 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |
| Yelp | IMDB | MR | AG's News | |
| Yelp | 87.6 | 85.8 | 40.5 | 43.6 |
| IMDB | 82.9 | 89.3 | 51.4 | 43.4 |
| MR | 81.8 | 88.2 | 66.5 | 39.6 |
| AG's News | 62.4 | 59.2 | 29.9 | 53.2 |
| Random | 58.9 | 56.1 | 27.8 | 38.3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| Method | Text (MR; Negative) | Result | Mod↓ | Sim↑ | Flu↑ |
| Original | Davis is so enamored of her own creation that she can not see how insufferable the character is. | - | - | - | 5 |
| A2T | Davis is so enamored of her own institution that she can not behold how unforgivable the hallmark is. | Failure | 22.2 | 0.16 | 3 |
| TextFooler | Davis is well enamored of her own infancy that she could not admire how infernal the idiosyncrasies is. | Success | 33.3 | 0.23 | 3 |
| BERT-Attack | Davis is often enamoted of her own generation that she can not see how insuffoure the queen is. | Failure | 27.8 | 0.09 | 2 |
| SDM-ATTACK | Davis is so captivated of her own creation that she can't see how indefensible the character is. | Success | 11.1 | 0.57 | 5 |
| Dataset | Train | Test | Avg Len | Classes |
| Yelp | 560k | 38k | 152 | 2 |
| IMDB | 25k | 25k | 215 | 2 |
| AG's News | 120k | 7.6k | 73 | 4 |
| MR | 9k | 1k | 20 | 2 |
| SNLI | 570k | 3k | 8 | 3 |
| MNLI | 433k | 10k | 11 | 3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| AG's News | 94.6 | 77.9 | 15.3 | 0.53 |
| +Adv Train | 91.8 | 50.6 | 23.3 | 0.50 |
| MR | 96.9 | 85.6 | 12.3 | 0.57 |
| +Adv Train | 92.4 | 72.0 | 16.7 | 0.57 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| MNLI | 84.5 | 78.7 | 13.8 | 0.49 |
| +Adv Train | 76.8 | 58.6 | 15.2 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | |
| TextFooler | 0.84 | 4.0 | 0.88 | |
| Bert-Attack | 0.83 | 4.2 | 0.90 | |
| SDM-ATTACK | 0.90 | 4.3 | 0.95 | |
| MNLI | Original | 0.88 | 4.0 | |
| TextFooler | 0.77 | 3.5 | 0.80 | |
| Bert-Attack | 0.77 | 3.6 | 0.81 | |
| SDM-ATTACK | 0.79 | 3.7 | 0.83 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | BERT-Attack | 89.8 | 12.4 | 0.66 |
| SDM-ATTACK-mlm | 90.0 | 10.6 | 0.65 | |
| IMDB | BERT-Attack | 88.2 | 5.3 | 0.78 |
| SDM-ATTACK-mlm | 88.5 | 5.1 | 0.78 | |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ | Dataset | Method | A-rate ↑ | Mod ↓ | Sim↑ |
| Yelp | A2T | 88.3 | 8.1 | 0.68 | IMDB | A2T | 89.9 | 4.4 | 0.79 |
| TextFooler | 90.5 | 9.0 | 0.69 | TextFooler | 88.7 | 7.6 | 0.76 | ||
| BERT-Attack | 89.8 | 12.4 | 0.66 | BERT-Attack | 88.2 | 5.3 | 0.78 | ||
| SDM-ATTACK | 95.8 | 8.2 | 0.71 | SDM-ATTACK | 91.4 | 4.1 | 0.82 | ||
| AG's News | A2T | 53.7 | 13.5 | 0.57 | MR | A2T | 58.5 | 12.6 | 0.55 |
| TextFooler | 66.2 | 18.4 | 0.52 | TextFooler | 80.5 | 15.8 | 0.50 | ||
| BERT-Attack | 74.6 | 15.6 | 0.52 | BERT-Attack | 83.2 | 12.8 | 0.52 | ||
| SDM-ATTACK | 77.9 | 15.3 | 0.53 | SDM-ATTACK | 85.6 | 12.3 | 0.57 | ||
| SNLI | A2T | 70.8 | 17.2 | 0.35 | MNLI | A2T | 66.0 | 14.4 | 0.45 |
| TextFooler | 84.3 | 17.2 | 0.38 | TextFooler | 76.5 | 15.0 | 0.45 | ||
| BERT-Attack | 81.9 | 16.5 | 0.38 | BERT-Attack | 78.1 | 14.0 | 0.46 | ||
| SDM-ATTACK | 85.5 | 15.9 | 0.43 | SDM-ATTACK | 78.7 | 13.8 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | 0.95 |
| SDM-ATTACK | 0.90 | 4.3 | ||
| MNLI | Original | 0.88 | 4.0 | 0.83 |
| SDM-ATTACK | 0.79 | 3.7 |
| Dataset | Model | A-rate↑ | Mod↓ | Sim↑ |
| MR | RoBERTa | 84.4 | 13.9 | 0.52 |
| WordCNN | 72.1 | 10.3 | 0.48 | |
| WordLSTM | 80.7 | 8.9 | 0.56 | |
| IMDB | RoBERTa | 88.3 | 8.3 | 0.70 |
| WordCNN | 89.2 | 3.3 | 0.85 | |
| WordLSTM | 89.8 | 5.4 | 0.75 | |
| SNLI | InferSent | 78.7 | 17.0 | 0.42 |
| ESIM | 79.0 | 17.2 | 0.41 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |
| Yelp | IMDB | MR | AG's News | |
| Yelp | 87.6 | 85.8 | 40.5 | 43.6 |
| IMDB | 82.9 | 89.3 | 51.4 | 43.4 |
| MR | 81.8 | 88.2 | 66.5 | 39.6 |
| AG's News | 62.4 | 59.2 | 29.9 | 53.2 |
| Random | 58.9 | 56.1 | 27.8 | 38.3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| Method | Text (MR; Negative) | Result | Mod↓ | Sim↑ | Flu↑ |
| Original | Davis is so enamored of her own creation that she can not see how insufferable the character is. | - | - | - | 5 |
| A2T | Davis is so enamored of her own institution that she can not behold how unforgivable the hallmark is. | Failure | 22.2 | 0.16 | 3 |
| TextFooler | Davis is well enamored of her own infancy that she could not admire how infernal the idiosyncrasies is. | Success | 33.3 | 0.23 | 3 |
| BERT-Attack | Davis is often enamoted of her own generation that she can not see how insuffoure the queen is. | Failure | 27.8 | 0.09 | 2 |
| SDM-ATTACK | Davis is so captivated of her own creation that she can't see how indefensible the character is. | Success | 11.1 | 0.57 | 5 |
| Dataset | Train | Test | Avg Len | Classes |
| Yelp | 560k | 38k | 152 | 2 |
| IMDB | 25k | 25k | 215 | 2 |
| AG's News | 120k | 7.6k | 73 | 4 |
| MR | 9k | 1k | 20 | 2 |
| SNLI | 570k | 3k | 8 | 3 |
| MNLI | 433k | 10k | 11 | 3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| AG's News | 94.6 | 77.9 | 15.3 | 0.53 |
| +Adv Train | 91.8 | 50.6 | 23.3 | 0.50 |
| MR | 96.9 | 85.6 | 12.3 | 0.57 |
| +Adv Train | 92.4 | 72.0 | 16.7 | 0.57 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| MNLI | 84.5 | 78.7 | 13.8 | 0.49 |
| +Adv Train | 76.8 | 58.6 | 15.2 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | |
| TextFooler | 0.84 | 4.0 | 0.88 | |
| Bert-Attack | 0.83 | 4.2 | 0.90 | |
| SDM-ATTACK | 0.90 | 4.3 | 0.95 | |
| MNLI | Original | 0.88 | 4.0 | |
| TextFooler | 0.77 | 3.5 | 0.80 | |
| Bert-Attack | 0.77 | 3.6 | 0.81 | |
| SDM-ATTACK | 0.79 | 3.7 | 0.83 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | BERT-Attack | 89.8 | 12.4 | 0.66 |
| SDM-ATTACK-mlm | 90.0 | 10.6 | 0.65 | |
| IMDB | BERT-Attack | 88.2 | 5.3 | 0.78 |
| SDM-ATTACK-mlm | 88.5 | 5.1 | 0.78 | |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ | Dataset | Method | A-rate ↑ | Mod ↓ | Sim↑ |
| Yelp | A2T | 88.3 | 8.1 | 0.68 | IMDB | A2T | 89.9 | 4.4 | 0.79 |
| TextFooler | 90.5 | 9.0 | 0.69 | TextFooler | 88.7 | 7.6 | 0.76 | ||
| BERT-Attack | 89.8 | 12.4 | 0.66 | BERT-Attack | 88.2 | 5.3 | 0.78 | ||
| SDM-ATTACK | 95.8 | 8.2 | 0.71 | SDM-ATTACK | 91.4 | 4.1 | 0.82 | ||
| AG's News | A2T | 53.7 | 13.5 | 0.57 | MR | A2T | 58.5 | 12.6 | 0.55 |
| TextFooler | 66.2 | 18.4 | 0.52 | TextFooler | 80.5 | 15.8 | 0.50 | ||
| BERT-Attack | 74.6 | 15.6 | 0.52 | BERT-Attack | 83.2 | 12.8 | 0.52 | ||
| SDM-ATTACK | 77.9 | 15.3 | 0.53 | SDM-ATTACK | 85.6 | 12.3 | 0.57 | ||
| SNLI | A2T | 70.8 | 17.2 | 0.35 | MNLI | A2T | 66.0 | 14.4 | 0.45 |
| TextFooler | 84.3 | 17.2 | 0.38 | TextFooler | 76.5 | 15.0 | 0.45 | ||
| BERT-Attack | 81.9 | 16.5 | 0.38 | BERT-Attack | 78.1 | 14.0 | 0.46 | ||
| SDM-ATTACK | 85.5 | 15.9 | 0.43 | SDM-ATTACK | 78.7 | 13.8 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | 0.95 |
| SDM-ATTACK | 0.90 | 4.3 | ||
| MNLI | Original | 0.88 | 4.0 | 0.83 |
| SDM-ATTACK | 0.79 | 3.7 |
| Dataset | Model | A-rate↑ | Mod↓ | Sim↑ |
| MR | RoBERTa | 84.4 | 13.9 | 0.52 |
| WordCNN | 72.1 | 10.3 | 0.48 | |
| WordLSTM | 80.7 | 8.9 | 0.56 | |
| IMDB | RoBERTa | 88.3 | 8.3 | 0.70 |
| WordCNN | 89.2 | 3.3 | 0.85 | |
| WordLSTM | 89.8 | 5.4 | 0.75 | |
| SNLI | InferSent | 78.7 | 17.0 | 0.42 |
| ESIM | 79.0 | 17.2 | 0.41 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |
| Yelp | IMDB | MR | AG's News | |
| Yelp | 87.6 | 85.8 | 40.5 | 43.6 |
| IMDB | 82.9 | 89.3 | 51.4 | 43.4 |
| MR | 81.8 | 88.2 | 66.5 | 39.6 |
| AG's News | 62.4 | 59.2 | 29.9 | 53.2 |
| Random | 58.9 | 56.1 | 27.8 | 38.3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| Method | Text (MR; Negative) | Result | Mod↓ | Sim↑ | Flu↑ |
| Original | Davis is so enamored of her own creation that she can not see how insufferable the character is. | - | - | - | 5 |
| A2T | Davis is so enamored of her own institution that she can not behold how unforgivable the hallmark is. | Failure | 22.2 | 0.16 | 3 |
| TextFooler | Davis is well enamored of her own infancy that she could not admire how infernal the idiosyncrasies is. | Success | 33.3 | 0.23 | 3 |
| BERT-Attack | Davis is often enamoted of her own generation that she can not see how insuffoure the queen is. | Failure | 27.8 | 0.09 | 2 |
| SDM-ATTACK | Davis is so captivated of her own creation that she can't see how indefensible the character is. | Success | 11.1 | 0.57 | 5 |
| Dataset | Train | Test | Avg Len | Classes |
| Yelp | 560k | 38k | 152 | 2 |
| IMDB | 25k | 25k | 215 | 2 |
| AG's News | 120k | 7.6k | 73 | 4 |
| MR | 9k | 1k | 20 | 2 |
| SNLI | 570k | 3k | 8 | 3 |
| MNLI | 433k | 10k | 11 | 3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| AG's News | 94.6 | 77.9 | 15.3 | 0.53 |
| +Adv Train | 91.8 | 50.6 | 23.3 | 0.50 |
| MR | 96.9 | 85.6 | 12.3 | 0.57 |
| +Adv Train | 92.4 | 72.0 | 16.7 | 0.57 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| MNLI | 84.5 | 78.7 | 13.8 | 0.49 |
| +Adv Train | 76.8 | 58.6 | 15.2 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | |
| TextFooler | 0.84 | 4.0 | 0.88 | |
| Bert-Attack | 0.83 | 4.2 | 0.90 | |
| SDM-ATTACK | 0.90 | 4.3 | 0.95 | |
| MNLI | Original | 0.88 | 4.0 | |
| TextFooler | 0.77 | 3.5 | 0.80 | |
| Bert-Attack | 0.77 | 3.6 | 0.81 | |
| SDM-ATTACK | 0.79 | 3.7 | 0.83 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | BERT-Attack | 89.8 | 12.4 | 0.66 |
| SDM-ATTACK-mlm | 90.0 | 10.6 | 0.65 | |
| IMDB | BERT-Attack | 88.2 | 5.3 | 0.78 |
| SDM-ATTACK-mlm | 88.5 | 5.1 | 0.78 | |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ | Dataset | Method | A-rate ↑ | Mod ↓ | Sim↑ |
| Yelp | A2T | 88.3 | 8.1 | 0.68 | IMDB | A2T | 89.9 | 4.4 | 0.79 |
| TextFooler | 90.5 | 9.0 | 0.69 | TextFooler | 88.7 | 7.6 | 0.76 | ||
| BERT-Attack | 89.8 | 12.4 | 0.66 | BERT-Attack | 88.2 | 5.3 | 0.78 | ||
| SDM-ATTACK | 95.8 | 8.2 | 0.71 | SDM-ATTACK | 91.4 | 4.1 | 0.82 | ||
| AG's News | A2T | 53.7 | 13.5 | 0.57 | MR | A2T | 58.5 | 12.6 | 0.55 |
| TextFooler | 66.2 | 18.4 | 0.52 | TextFooler | 80.5 | 15.8 | 0.50 | ||
| BERT-Attack | 74.6 | 15.6 | 0.52 | BERT-Attack | 83.2 | 12.8 | 0.52 | ||
| SDM-ATTACK | 77.9 | 15.3 | 0.53 | SDM-ATTACK | 85.6 | 12.3 | 0.57 | ||
| SNLI | A2T | 70.8 | 17.2 | 0.35 | MNLI | A2T | 66.0 | 14.4 | 0.45 |
| TextFooler | 84.3 | 17.2 | 0.38 | TextFooler | 76.5 | 15.0 | 0.45 | ||
| BERT-Attack | 81.9 | 16.5 | 0.38 | BERT-Attack | 78.1 | 14.0 | 0.46 | ||
| SDM-ATTACK | 85.5 | 15.9 | 0.43 | SDM-ATTACK | 78.7 | 13.8 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | 0.95 |
| SDM-ATTACK | 0.90 | 4.3 | ||
| MNLI | Original | 0.88 | 4.0 | 0.83 |
| SDM-ATTACK | 0.79 | 3.7 |
| Dataset | Model | A-rate↑ | Mod↓ | Sim↑ |
| MR | RoBERTa | 84.4 | 13.9 | 0.52 |
| WordCNN | 72.1 | 10.3 | 0.48 | |
| WordLSTM | 80.7 | 8.9 | 0.56 | |
| IMDB | RoBERTa | 88.3 | 8.3 | 0.70 |
| WordCNN | 89.2 | 3.3 | 0.85 | |
| WordLSTM | 89.8 | 5.4 | 0.75 | |
| SNLI | InferSent | 78.7 | 17.0 | 0.42 |
| ESIM | 79.0 | 17.2 | 0.41 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |
| Yelp | IMDB | MR | AG's News | |
| Yelp | 87.6 | 85.8 | 40.5 | 43.6 |
| IMDB | 82.9 | 89.3 | 51.4 | 43.4 |
| MR | 81.8 | 88.2 | 66.5 | 39.6 |
| AG's News | 62.4 | 59.2 | 29.9 | 53.2 |
| Random | 58.9 | 56.1 | 27.8 | 38.3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| Method | Text (MR; Negative) | Result | Mod↓ | Sim↑ | Flu↑ |
| Original | Davis is so enamored of her own creation that she can not see how insufferable the character is. | - | - | - | 5 |
| A2T | Davis is so enamored of her own institution that she can not behold how unforgivable the hallmark is. | Failure | 22.2 | 0.16 | 3 |
| TextFooler | Davis is well enamored of her own infancy that she could not admire how infernal the idiosyncrasies is. | Success | 33.3 | 0.23 | 3 |
| BERT-Attack | Davis is often enamoted of her own generation that she can not see how insuffoure the queen is. | Failure | 27.8 | 0.09 | 2 |
| SDM-ATTACK | Davis is so captivated of her own creation that she can't see how indefensible the character is. | Success | 11.1 | 0.57 | 5 |
| Dataset | Train | Test | Avg Len | Classes |
| Yelp | 560k | 38k | 152 | 2 |
| IMDB | 25k | 25k | 215 | 2 |
| AG's News | 120k | 7.6k | 73 | 4 |
| MR | 9k | 1k | 20 | 2 |
| SNLI | 570k | 3k | 8 | 3 |
| MNLI | 433k | 10k | 11 | 3 |
| Dataset | Acc↑ | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | 97.4 | 95.8 | 8.2 | 0.71 |
| +Adv Train | 97.0 | 82.5 | 13.5 | 0.63 |
| IMDB | 91.6 | 91.4 | 4.1 | 0.82 |
| +Adv Train | 90.5 | 79.2 | 8.5 | 0.74 |
| AG's News | 94.6 | 77.9 | 15.3 | 0.53 |
| +Adv Train | 91.8 | 50.6 | 23.3 | 0.50 |
| MR | 96.9 | 85.6 | 12.3 | 0.57 |
| +Adv Train | 92.4 | 72.0 | 16.7 | 0.57 |
| SNLI | 89.1 | 85.5 | 15.9 | 0.43 |
| +Adv Train | 88.2 | 78.6 | 17.1 | 0.42 |
| MNLI | 84.5 | 78.7 | 13.8 | 0.49 |
| +Adv Train | 76.8 | 58.6 | 15.2 | 0.49 |
| Dataset | Con↑ | Flu↑ | Simhum ↑ | |
| IMDB | Original | 0.95 | 4.5 | |
| TextFooler | 0.84 | 4.0 | 0.88 | |
| Bert-Attack | 0.83 | 4.2 | 0.90 | |
| SDM-ATTACK | 0.90 | 4.3 | 0.95 | |
| MNLI | Original | 0.88 | 4.0 | |
| TextFooler | 0.77 | 3.5 | 0.80 | |
| Bert-Attack | 0.77 | 3.6 | 0.81 | |
| SDM-ATTACK | 0.79 | 3.7 | 0.83 |
| Dataset | Method | A-rate↑ | Mod↓ | Sim↑ |
| Yelp | BERT-Attack | 89.8 | 12.4 | 0.66 |
| SDM-ATTACK-mlm | 90.0 | 10.6 | 0.65 | |
| IMDB | BERT-Attack | 88.2 | 5.3 | 0.78 |
| SDM-ATTACK-mlm | 88.5 | 5.1 | 0.78 | |
| AG's News | BERT-Attack | 74.6 | 15.6 | 0.52 |
| SDM-ATTACK-mlm | 76.2 | 15.0 | 0.51 | |
| MR | BERT-Attack | 83.2 | 12.8 | 0.52 |
| SDM-ATTACK-mlm | 84.3 | 11.5 | 0.53 |