| Dataset (Task) | Roberta | T5 | LLaMA | CODEX |
| FewNERD (NER) | 2.8 | 39.4 | 1135.4 | 179.4 |
| TACREV (RE) | 1.4 | 45.6 | 1144.9 | 151.6 |
| ACE05 (ED) | 6.6 | 62.5 | 733.4 | 171.7 |
| FewNERD | TACREV | ACE05 | |
| Overall | 5.88 | 3.03 | 38.2 |
| Easy samples (τ > 0.9) | 9.44 | 3.21 | 44.0 |
| Hard samples (τ < 0.6) | 1.28 | 2.68 | 1.36 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ||||||||
| 5-shot | 10-shot | 20-shot | 20-shot | 50-shot | 100-shot | 5-shot | 10-shot | 20-shot | ||
| LLM | CODEX | 53.8(0.5) | 54.0(1.4) | 55.9(0.5) | 59.1(1.4) | 60.3(2.4) | 62.4(2.6) | 47.1(1.2) | 47.7(2.8) | 47.9(0.5) |
| InstructGPT | 53.6(-) | 54.6(-) | 57.2(-) | 60.1(-) | 58.3(-) | 62.7(-) | 52.9(-) | 52.1(-) | 49.3(-) | |
| GPT-4 | - | - | 57.8(-) | - | - | 59.3(-) | - | - | 52.1(-) | |
| SLM | Previous SoTA | 59.4(1.5) | 61.4(0.8) | 61.9(1.2) | 62.4(3.8) | 68.5(1.6) | 72.6(1.5) | 55.1(4.6) | 63.9(0.8) | 65.8(2.0) |
| + Ensemble (S) | 59.6(1.7) | 61.8(1.2) | 62.6(1.0) | 64.9(1.5) | 71.9(2.2) | 74.1(1.7) | 56.9(4.7) | 64.2(2.1) | 66.5(1.7) | |
| + Rerank (S) | 59.4(1.5) | 61.0(1.7) | 61.5(1.7) | 64.2(2.3) | 70.8(2.3) | 74.3(2.2) | 56.1(0.3) | 64.0(1.0) | 66.7(1.7) | |
| SLM + LIM | Vicuna-13B | |||||||||
| + Rerank (L) | 60.0(1.8) | 61.9(2.1) | 62.2(1.4) | 65.2(1.4) | 70.8(1.6) | 73.8(1.7) | 56.9(4.0) | 63.5(2.7) | 66.0(2.6) | |
| + Ensemble (S) + Rerank (L) | 59.9(0.7) | 62.1(0.7) | 62.8(1.1) | 66.5(0.5) | 73.6(1.4) | 75.0(1.5) | 57.9(5.2) | 64.4(1.2) | 66.2(2.4) | |
| InstructGPT | ||||||||||
| + Rerank (L) | 60.6(2.1) | 62.7(0.8) | 63.3(0.6) | 66.8(2.6) | 72.3(1.4) | 75.4(1.5) | 57.8(4.6) | 65.3(1.7) | 67.3(2.2) | |
| + Ensemble (S) + Rerank (L) | 61.3(1.9) | 63.2(0.9) | 63.7(1.8) | 68.9(1.3) | 74.8(1.3) | 76.8(1.2) | 59.5(3.7) | 65.3(1.9) | 67.8(2.1) | |
| GPT-4 | ||||||||||
| + Rerank (L) | 60.8(2.3) | 62.6(2.7) | 63.0(1.3) | 65.9(2.7) | 72.3(0.3) | 74.5(1.5) | 59.6(2.9) | 64.9(2.5) | 67.1(2.5) | |
| + Ensemble (S) + Rerank (L) | 61.1(2.2) | 62.8(0.9) | 63.6(1.2) | 68.6(1.3) | 73.9(1.4) | 75.9(2.4) | 60.9(3.9) | 65.6(1.5) | 67.8(1.7) |
| GPT-4 | InstructGPT | |||||||
| before | after | △ | ratio | before | after | △ | ratio | |
| FewNER | 31.9 | 40.7 | 8.8 | 3.2% | 31.4 | 28.3 | -3.1 | 3.3% |
| TACREV | 25.3 | 43.0 | 17.7 | 9.1% | 33.8 | 43.4 | 9.6 | 7.1% |
| ACE05 | 31.1 | 57.9 | 26.8 | 1.6% | 35.6 | 55.7 | 20.1 | 0.5% |
| CoT | Demo | LF | AD | FewNERD (20-shot) | TACREV (100-shot) | ACE05 (20-shot) |
| ✓ | ✓ | ✓ | ✓ | 63.6(1.2) | 75.9(2.4) | 67.8(1.7) |
| ✗ | ✓ | ✓ | ✓ | 63.2(1.2) | 75.4(2.4) | 67.2(1.7) |
| ✗ | ✗ | ✓ | ✓ | 63.0(1.4) | 74.9(2.2) | 66.6(1.5) |
| ✗ | ✗ | ✗ | ✓ | 62.4(2.1) | 73.8(2.5) | 66.5(1.3) |
| ✗ | ✗ | ✗ | ✗ | 12.5(2.7) | 59.9(6.0) | 5.4(1.1) |
| Previous SoTA methods | 62.6(1.0) | 74.1(1.7) | 66.5(1.7) |
| Dataset | Named Entity Recognition\nCONLL OntoNotes | FewNERD | Relation Extraction | Event Detection | Event Arg Extraction | ||||||
| TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | ||||
| #Label Type | 4 | 18 | 66 | 41 | 41 | 33 | 168 | 38 | 33 | 139 | |
| #Sents | Train | 14,041 | 49,706 | 131,965 | 68,124 | 68,124 | 14,024 | 32,360 | 14,736 | 14,024 | 7329 |
| Test | 3,453 | 10,348 | 37,648 | 15,509 | 15,509 | 728 | 8,035 | 1,163 | 728 | 871 | |
| #Mentions | Train | 23,499 | 128,738 | 340,247 | 13,012 | 13,012 | 5,349 | 77,993 | 6,208 | 4859 | 17026 |
| Test | 5,648 | 12,586 | 96,902 | 3,123 | 3,123 | 424 | 18,904 | 551 | 576 | 2023 |
| Dataset Settings | # Labels | # Sent | # Sample | # Avg shot | |
| CONLL'03 | 1-shot | 4 | 4.8 | 5.8 | 1.4 |
| 5-shot | 16.2 | 21.8 | 5.5 | ||
| 10-shot | 29.2 | 42.6 | 10.7 | ||
| 20-shot | 65.6 | 82.0 | 20.5 | ||
| OntoNotes | 1-shot | 18 | 20.0 | 33.4 | 1.9 |
| 5-shot | 84.8 | 148.0 | 8.2 | ||
| 10-shot | 158.6 | 281.0 | 15.6 | ||
| 20-shot | 332.8 | 547.2 | 30.4 | ||
| FewNERD | 1-shot | 66 | 89.8 | 147.0 | 2.2 |
| 5-shot | 286.2 | 494.8 | 7.5 | ||
| 10-shot | 538.0 | 962.0 | 14.6 | ||
| 20-shot | 1027.2 | 1851.4 | 28.1 | ||
| TACREV | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2872.0 | 1944.0 | 47.4 | ||
| 100-shot | 4561.0 | 3520.0 | 85.9 | ||
| TACRED | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2871.2 | 1944.0 | 47.4 | ||
| 100-shot | 4575.2 | 3520.0 | 85.9 | ||
| ACE05 | 1-shot | 33 | 47.4 | 41.0 | 1.2 |
| 5-shot | 192.8 | 165.0 | 5.0 | ||
| 10-shot | 334.6 | 319.4 | 9.7 | ||
| 20-shot | 579.4 | 598.2 | 18.1 | ||
| MAVEN | 1-shot | 168 | 157.6 | 298.0 | 1.8 |
| 5-shot | 540.4 | 1262.2 | 7.5 | ||
| 10-shot | 891.2 | 2413.8 | 14.4 | ||
| 20-shot | 1286.4 | 4611.4 | 27.4 | ||
| ERE | 1-shot | 38 | 48.4 | 54.6 | 1.4 |
| 5-shot | 175.0 | 219.2 | 5.8 | ||
| 10-shot | 304.8 | 432.4 | 11.4 | ||
| 20-shot | 521.6 | 806.6 | 21.2 | ||
| ACE05 | 1-shot | 33 | 23.4 | 40.2 | 1.2 |
| 5-shot | 79.8 | 178.2 | 5.4 | ||
| 10-shot | 130.8 | 337.4 | 10.2 | ||
| 20-shot | 213.4 | 630.2 | 19.1 | ||
| RAMS | 1-shot | 139 | 130.2 | 332.6 | 2.4 |
| 5-shot | 514.0 | 1599.6 | 11.5 | ||
| 10-shot | 795.2 | 3193.2 | 23.0 | ||
| 20-shot | 1070.4 | 6095.4 | 43.9 | ||
| ERE | 1-shot | 38 | 21.6 | 102.8 | 2.7 |
| 5-shot | 74.2 | 403.4 | 10.6 | ||
| 10-shot | 127.2 | 775.6 | 20.4 | ||
| 20-shot | 190.2 | 1397.2 | 36.8 |
| FewNERD | TACREV | ACE05 | |
| t=0 | 48.5(1.9) | 53.7(2.3) | 42.9(2.2) |
| +5-ensemble | 53.5(1.3) | 58.6(1.5) | 46.3(0.8) |
| t=0.7 | 40.9(2.3) | 39.9(1.2) | 35.6(1.0) |
| + self-consistency | 52.1(0.9) | 53.4(1.3) | 45.6(3.0) |
| 10-shot train set | FewNERD (NER) | TACREV (RE) | ACE05 (ED) |
| wo. Auto-CoT | 54.0(1.4) | 57.3(1.8) | 47.7(2.8) |
| w. Auto-CoT | 36.6(1.7) | 22.0(1.2) | 43.1(3.4) |
| NER (20-shot) | RE (100-shot) | ED (20-shot) | EAE (20-shot) | ||||||||
| CONLL | OntoNotes | FewNERD | TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | |
| InstructGPT | 77.2 | 47.7 | 57.2 | 62.7 | 53.8 | 49.3 | 25.4 | 40.8 | 45.8 | 42.2 | 41.9 |
| CODEX | 81.1 | 55.6 | 55.9 | 62.4 | 53.6 | 47.9 | 22.8 | 39.0 | - | - | - |
| GPT-4 | 84.7 | 65.6 | 57.8 | 59.3 | 50.4 | 52.1 | 30.2 | 40.5 | 42.9 | 38.6 | 38.2 |
| Supervised SoTA | 72.3 | 74.9 | 61.4 | 72.6 | 63.1 | 65.8 | 54.7 | 56.2 | 55.2 | 57.7 | 55.6 |
| Named Entity Recognition | |||
| CoNLL | OntoNotes | FewNERD | |
| # Entity | 4 | 18 | 66 |
| Micro-F1 (SLM) | 52.5 | 59.7 | 59.4 |
| Micro-F1 (LLM) | 77.8 | 59.4 | 55.5 |
| ΔF1 (LLM, SLM) | 25.3 | -0.3 | -3.9 |
| Event Detection | |||
| ACE05 | ERE | MAVEN | |
| # Event | 33 | 38 | 168 |
| Micro-F1 (SLM) | 55.1 | 48.0 | 49.4 |
| Micro-F1 (LLM) | 39.6 | 33.8 | 25.3 |
| ΔF1 (LLM, SLM) | -15.5 | -14.2 | -24.1 |
| Event Argument Extraction | |||
| ACE05 | ERE | RAMS | |
| # Event / #Role | 33 / 22 | 38 / 26 | 139 / 65 |
| Head-F1 (SLM) | 45.9 | 40.4 | 54.1 |
| Head-F1 (LLM) | 52.8 | 40.7 | 44.2 |
| ΔF1 (LLM, SLM) | 6.9 | 0.3 | -9.9 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| I0 | 57.6(2.1) | 49.1(2.4) | 44.0(1.4) | 50.9(0.1) |
| I1 | 58.3(0.5) | 49.6(1.2) | 42.6(1.0) | 51.5(1.1) |
| I2 | 57.7(1.0) | 50.0(1.5) | 41.8(0.9) | 50.3(1.5) |
| I3 | 57.6(2.3) | 52.3(1.8) | 42.9(1.3) | 49.2(2.3) |
| I4 | 56.8(0.9) | 49.6(2.9) | 41.6(1.9) | 49.9(1.2) |
| I5 | 57.8(0.5) | 47.2(1.8) | 43.1(1.8) | 50.6(1.8) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | |
| Random Sampling | 53.2(0.4) | 43.0(3.3) | 38.0(1.5) |
| Sentence Embedding | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) |
| Efficient Prompt Retriever | 57.2(0.6) | 48.0(0.8) | 43.5(1.4) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| Text | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) | 51.5(1.1) |
| Code | 53.2(0.9) | 50.2(1.8) | 44.3(2.0) | 47.3(1.5) |
| Sentences (entity/event) | Before | After | Rationales |
| Named Entity Recognition | |||
| According to it, King Eleusis and Cothonea (Cyntinia), parents of Trip-tolemus, are visited by Demeter, who rears their son, feeding him divine milk by day and placing him into the fire at night, which makes Triptolemus grow faster than mortal children normally do. | person-other ✘ | other-god ✓ | According to the context, Triptolemus is a figure in Greek mythology, the son of King Eleusis and Cothonea (Cyntinia). He was a demigod who was sent by Demeter to teach mankind the art of agriculture. So Triptole-mus is a god in some legend/religious story. |
| Relation Extraction | |||
| Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul. | None ✘ | per:children ✓ | The sentence states 'Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul.' This indicates 'his' is the father of Jeff, James, Douglas and Harris, and the husband of Sandra. So his is the parent of Jeff. |
| Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais. | per:cities_of _residence ✘ | per:states_of _residence ✓ | The sentence states that 'Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais.' This indicates that Eliza Samudio is from the city of Belo Horizonte, which is located in the state of Minas Gerais. So Eliza Samudio lives in the state or province Minas Gerais. |
| Event Detection | |||
| We have fought every institution, every law firm ... it's cost us a fortune but now we're going on a massive expansion. | Conflictattack ✘ | None ✓ | In the given sentence, the word fought is used to describe the action of the speaker fighting against various institutions and law firms. This does not involve any physical violence or court proceedings, so the word fought does not trigger any known event. |
| Instruct: Read following sentences and identify what is the entity type of 392 quoted by <t>.Sentence: Powell v. Texas, <t> 392 U.S. 514 ( 1968 ) , was a United States Supreme Court case that ruled that a Texas statute criminalizing public intoxication did not violate the Eighth Amendment protection against cruel and unusual punishment.(a) 392 is a legal document, a term or a convention in legal sense.(b) 392 does not belong to any known entities.(c) 392 refers to a protest, uprising or revolution event(d) 392 refers to a government or governmental agencyAnalysis: In the context you provided, 392 refers to the volume number in the United States Reports where the Supreme Court's decision in Powell v. Texas can be found. However, 392 itself does not refer to a legal document. So 392 do/does not belong to any known entities.An Answer: (b) |
| Instruct: Read following sentences and identify what is the entity type of The New Yorker quoted by <t>.Sentence: In 2004 Gourevitch was assigned to cover the 2004 U.S. presidential election for " <t> The New Yorker <t> ".(a) The New Yorker does not belong to any known entities.(b) The New Yorker is a broadcast program.(c) The New Yorker is a kind of written art.(d) The New Yorker is a media/newspaper organization.Analysis: The New Yorker is a well-known American magazine that has been published since 1925, and is primarily known for its long-form journalism, commentary, and satire. It has a reputation for publishing high-quality writing on a wide variety of topics, including politics, culture, and the arts. So The New Yorker is a media/newspaper organization.An Answer: (d) |
| Instruct: Read following sentence and identify what is the entity type of St. quoted by <t>.Sentence: The May 1980 eruption of Mount St. Helens in the state of Washington seriously affected both 47th Air Division and 92d Bombardment Wing operations at Fairchild AFB , resulting in dispersal of Fairchild 's B-52 and KC-135 aircraft to various bases while around-the-clock shifts removed the volcanic ash from facilities within the base perimeter." (a) St. does not belong to any known entities.(b) St. is a natural disaster event.(c) St. is a geographic position about mountain.Analysis: According to the context, St. is an abbreviation of Saint, used in the name of Mount St. Helens, which is an active volcano in the state of Washington. However, St. itself does not refer to anything. So St. do/does not belong to any known entities.An Answer: (a) |
| Instruct: Read following sentence and identify what is the entity type of Ridzuan quoted by <t>.Sentence: <t> Ridzuan was promoted to Harimau Muda A for 2014 season .(a) Ridzuan does not belong to any known entities.(b) Ridzuan is a person, but not affiliated with following professions: actor, artist, author, director, politician, scholar, soldier.(c) Ridzuan is an athlete.Analysis: The mention of 'Harimau Muda A' indicates that it is a sports-related context, and 'promoted' implies a progression or advancement within the sports team. So Ridzuan is an athlete.An Answer: (c) |
| Instruct: Read the sentence and determine the relation between she and lawyer quoted by <t>. \nSentence: The <t> lawyer <t> denied Italian news reports that she wept while addressing the court, but said Knox was upset as <t> she <t> recounted “ the pressure, the aggressiveness of the police who called her a liar . ” \n(a) she is the other family member of lawyer \n(b) she is a lawyer \n(c) she has no known relations to lawyer \nAnalysis: In the sentence, the word ‘she’ refers to someone who was upset while recounting certain events in court. The word ‘lawyer’ refers to someone who denied a news report about that same person weeping in court. There is no information in the sentence to indicate that the two individuals are related in any way. So she has no known relations to lawyer. \nAnswer: (c) |
| Instruct: Read the sentence and determine the relation between MEF and Myanmar Equestrian Federation quoted by <t>. \nSentence: YANGON, Dec. 27 -LRB- Xinhua -RRB- - Myanmar will hold a horse race in Yangon to commemorate the country ’63rd Anniversary Independence Day , the <t> Myanmar Equestrian Federation <t> -LRB- <t> MEF <t> -RRB- confirmed to Xinhua on Monday. \n(a) MEF is also known as Myanmar Equestrian Federation \n(b) MEF has political affiliation with Myanmar Equestrian Federation \n(c) MEF has no known relations to Myanmar Equestrian Federation \nAnalysis: The symbols -LRB- and -RRB- in the sentence stand for left and right round brackets and are used to enclose the abbreviation ‘MEF’ to indicate that it is a replacement for the longer name ‘Myanmar Equestrian Federation. So MEF is also known as Myanmar Equestrian Federation. \nAnswer: (a) |
| Instruct: Read the sentence and determine the relation between Douglas Flint and chairman quoted by <t>. \nSentence: At the same time , Chief Financial Officer <t> Douglas Flint <t> will become <t> chairman <t> , succeeding Stephen Green who is leaving to take a government job. \n(a) Douglas Flint has no known relations to chairman \n(b) Douglas Flint is a chairman \n(c) Douglas Flint is the employee of chairman \nAnalysis: The sentence states that Chief Financial Officer Douglas Flint Douglas Flint will succeed Stephen Green as a chairman. So Douglas Flint is a chairman. \nAnswer: (b) |
| Instruct: Read the sentence and determine the relation between FAA and U.S. quoted by <t>. \nSentence: On its Web site , the <t> U.S. <t> FAA <t> says the Category 2 rating means the country lacks the laws or regulations that are needed for the certification and oversight of air carriers , according to minimum international standards. \n(a) FAA is also known as U.S. \n(b) FAA has no known relations to U.S. \n(c) FAA has a headquarters in the country U.S. \nAnalysis: The sentence states that the FAA says the Category 2 rating means the country lacks the laws or regulations needed for the certification and oversight of air carriers, indicating that the FAA is responsible for overseeing aviation regulations in the country. Actually the FAA (Federal Aviation Administration) is a U.S. government agency responsible for regulating and overseeing civil aviation in the United States, and it has its headquarters in Washington, D.C.. So FAA has a headquarters in the country U.S.. \nAnswer: (c) |
| Instruct: Read following sentences and identify what event is triggered by the word loan quoted by <t>. \nSentence: Separately, former WorldCom CEO Bernard Ebbers failed on April 29 to make a first repayment of 25 million dollars, plus interest, on a 400-million - dollar <t> from MCI, the Journal said, citing SEC documents. \n(a) The word loan does not trigger any known event. \n(b) The word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. \n(c) The word loan triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. \nAnalysis: In the given sentence, the word loan is used to describe the 400-million-dollar loan from MCI to former WorldCom CEO Bernard Ebbers, which he failed to repay on time. This situation clearly involves a transfer of money without the context of purchasing something, which falls under the TRANSFER-MONEY event. So the word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words treated quoted by <t>. \nSentence: When she 's in Germany, Lynch will be <t> treated <t> for bullet wounds and broken bones. \n(a) The word treated triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. \n(b) The word treated does not trigger any known event. \n(c) The word treated triggers a TRANSPORT event: an ARTifact (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. \nAnalysis: The sentence suggests that Lynch has already been injured and will receive medical treatment in Germany for her injuries. The word 'treated' simply describes the medical care she will receive and does not indicate a new event or action taking place. So the word treated does not trigger any known event. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words buy quoted by <t>. \nSentence: And I won't dwell on the irony of an Oracle employee being driven out of Oracle, starting his own company, and forcing Ellison to spend $10.3 billion to get his company – but not him – back (though it does rather delightfully remind me of Coca - Cola basically giving away the bottling franchise and then spending billions to <t> buy <t> it back). \n(a) The word buy triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. \n(b) The word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. \n(c) The word buy does not trigger any known event. \nAnalysis: In the given sentence, the word buy is used to describe the action of Oracle spending $10.3 billion to get a company back. This clearly involves the transfer of ownership of the company from one entity to another. So the word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words set quoted by <t>. \nSentence: British forces also began establishing the country's first postwar administration Tuesday, granting a local sheik power to <t> set up an administrative committee representing the groups in the region. \n(a) The word set triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. \n(b) The word set triggers a START-ORG event: a new ORGANIZATION is created. \n(c) The word set does not trigger any known event. \nAnalysis: The phrase 'set up' specifically implies the creation or establishment of a new organization or entity, rather than simply the word 'set'. So the word set does not trigger any known event. \nAnswer: (c) |
| Entity | Template |
| no-entity | {ent} do/does not belong to any known entities. |
| person-artist/author | {ent} is an artist or author. |
| person-actor | {ent} is an actor. |
| art-written | {ent} is a kind of written. |
| person-director | {ent} is a director. |
| person-other | {ent} is a person, but not affiliated with following professions: actor, artist, athlete, author, director, politician, scholar, soldier. |
| organization-other | {ent} pertains to an organization that does not fall under the categories of company, educational institution, government, media, political party, religion, sports league, sports team, band or musical group. |
| organization-company | {ent} is a company |
| organization-sportsteam | {ent} is a sports team |
| organization-sportsleague | {ent} is a sports league |
| product-car | {ent} is a kind of car |
| event-protest | {ent} refers to a protest, uprising or revolution event |
| organization-government/governmentagency | {ent} refers to a government or governmental agency |
| other-biologything | {ent} is a special term about biology / life science. |
| location-GPE | {ent} is a kind of geopolitical entity |
| location-other | {ent} is a geographic locaton that does not fall under the categories of geopolitical entity, body of water, island, mountain, park, road, railway and transit. |
| person-athlete | {ent} is an athlete or coach. |
| art-broadcastprogram | {ent} is a broadcast program. |
| product-other | {ent} is a kind of product that does not fall under the categories of airplane, train, ship, car, weapon, food, electronic game and software. |
| building-other | {ent} is a kind of building that does not fall under the categories of airport, hospital, hotel, library, restaurant, sports facility and theater |
| product-weapon | {ent} is a kind of weapon. |
| building-airport | {ent} is an airport. |
| building-sportsfacility | {ent} is a sports facility building. |
| person-scholar | {ent} is a scholar. |
| art-music | {ent} is a music. |
| event-other | {ent} refers to some event except attack, election, natural disaster, protest, revolution and sports |
| other-language | {ent} is a kind of human language. |
| other-chemicalthing | {ent} is some special term about chemical science. |
| art-film | {ent} is a film. |
| building-hospital | {ent} is a hospital. |
| other-law | {ent} is a legal document, a term or a convention in legal sense. |
| product-airplane | {ent} is kind of airplane product. |
| location-road/railway/highway/transit | {ent} is a geographic position about roadways, railways, highways or public transit systems. |
| person-soldier | {ent} is a soldier |
| location-mountain | {ent} is geographic position about mountain. |
| organization-education | {ent} is an educational institute/organization. |
| organization-media/newspaper | {ent} is a media/newspaper organization. |
| product-software | {ent} is a software product. |
| location-island | {ent} is geographic position about island. |
| location-bodiesofwater building-library | {ent} is geographic position situated near a body of water. {ent} is a library. |
| other-astronomything | {ent} is a special term about astronomy. |
| person-politician | {ent} is a politician or lawyer or judge. |
| building-hotel | {ent} is a hotel building. |
| product-game | {ent} is a electronic game product. |
| other-award | {ent} is a kind of award. |
| event-sportsevent | {ent} refers to some event related to sports. |
| organization-showorganization | {ent} is a band or musical organization. |
| other-educationaldegree | {ent} is a kind of educational degree. |
| building-theater | {ent} is a theater. |
| other-disease | {ent} is a kind of disease. |
| event-election | {ent} is an event about election. |
| organization-politicalparty | {ent} is a political party/organization. |
| other-currency | {ent} is a kind of currency. |
| event-attack/battle/war/militaryconflict | {ent} is an event about attack, battle, war or military conflict. |
| product-ship | {ent} is a ship. |
| building-restaurant | {ent} is a restaurant. |
| other-livingthing | {ent} is a living animal/crystal/organism. |
| art-other | {ent} is a work of art, but not belong to the categories of music, film, written art, broadcast or painting. |
| event-disaster | {ent} is a natural disaster event. |
| organization-religion | {ent} is a religious organization. |
| other-medical | {ent} refers to some kind of medicineentity |
| location-park | {ent} is a park. |
| other-god | {ent} is a god in some legend/religious story. |
| product-food | {ent} is a kind of food. |
| product-train | {ent} is a kind of train(vehicle). |
| art-painting | {ent} is an art painting. |
| Relation | Template |
| no Relation | {subj} has no known relations to {obj} |
| per:stateorprovince_of_death | {subj} died in the state or province {obj} |
| per:title | {subj} is a {obj} |
| org:member_of | {subj} is the member of {obj} |
| per:other_family | {subj} is the other family member of {obj} |
| org:country_of_HEADquarters | {subj} has a headquarters in the country {obj} |
| org:parents | {subj} has the parent company {obj} |
| per:stateorprovince_of_birth | {subj} was born in the state or province {obj} |
| per:spouse | {subj} is the spouse of {obj} |
| per:origin | {subj} has the nationality {obj} |
| per:date_of_birth | {subj} has birthday on {obj} |
| per:schools_attended | {subj} studied in {obj} |
| org:members | {subj} has the member {obj} |
| org:founded | {subj} was founded in {obj} |
| per:stateorprovinces_of_residence | {subj} lives in the state or province {obj} |
| per:date_of_death | {subj} died in the date {obj} |
| org:shareholders | {subj} has shares hold in {obj} |
| org:website | {subj} has the website {obj} |
| org:subsidiaries | {subj} owns {obj} |
| per:charges | {subj} is convicted of {obj} |
| org:dissolved | {subj} dissolved in {obj} |
| org:stateorprovince_of_HEADquarters | {subj} has a headquarters in the state or province {obj} |
| per:country_of_birth | {subj} was born in the country {obj} |
| per:siblings | {subj} is the siblings of {obj} |
| org:top_membership/employees | {subj} has the high level member {obj} |
| per:cause_of_death | {subj} died because of {obj} |
| per:alternate_names | {subj} has the alternate name {obj} |
| org:number_of_employees/members | {subj} has the number of employees {obj} |
| per:cities_of_residence | {subj} lives in the city {obj} |
| org:city_of_HEADquarters | {subj} has a headquarters in the city {obj} |
| per:children | {subj} is the parent of {obj} |
| per:employee_of | {subj} is the employee of {obj} |
| org:political/religious_affiliation | {subj} has political affiliation with {obj} |
| per:parents | {subj} has the parent {obj} |
| per:city_of_birth | {subj} was born in the city {obj} |
| per:age | {subj} has the age {obj} |
| per:countries_of_residence | {subj} lives in the country {obj} |
| org:alternate_names | {subj} is also known as {obj} |
| per:religion | {subj} has the religion {obj} |
| per:city_of_death | {subj} died in the city {obj} |
| per:country_of_death | {subj} died in the country {obj} |
| org:founded_by | {subj} was founded by {obj} |
| Event | Template |
| no-event | The word {evt} does not trigger any known event. |
| MovementTRANSPORT | The word {evt} triggers a TRANSPORT event: an ARTIFACT (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. |
| Personnel.Elect | The word {evt} triggers an ELECT event which implies an election. |
| Personnel.Start-Position | The word {evt} triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Personnel.Nominate | The word {evt} triggers a NOMINATE event: a PERSON is proposed for a position through official channels. |
| Conflict.Attack | The word {evt} triggers an ATTACK event: a violent physical act causing harm or damage. |
| Personnel.End-Position | The word {evt} triggers an END-POSITION event: a PERSON stops working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Contact.Meet | The word {evt} triggers a MEET event: two or more entities come together at a single location and interact with one another face-to-face. |
| Life.Marry | The word {evt} triggers a MARRY event: two people are married under the legal definition. |
| Contact.Phone-Write | The word {evt} triggers a PHONE-WRITE event: two or more people directly engage in discussion which does not take place 'face-to-face'. |
| Transaction.Transform-Money | The word {evt} triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. |
| Justice.Sue | The word {evt} triggers a SUE event: a court proceeding has been initiated for the purposes of determining the liability of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime or neglecting a commitment |
| Conflict.Demonstrate | The word {evt} triggers a DEMONSTRATE event: a large number of people come together in a public area to protest or demand some sort of official action. For example: protests, sit-ins, strikes and riots. |
| Business.End-Org | The word {evt} triggers an END-ORG event: an ORGANIZATION ceases to exist (in other words, goes out of business). |
| Life.Injure | The word {evt} triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. |
| Life.Die | The word {evt} triggers a DIE event: a PERSON dies/died whether it occurs accidentally, intentionally or even self-inflicted. |
| Justice.Arrest-Jail | The word {evt} triggers a ARREST-JAIL event: a PERSON is sent to prison. |
| Transaction.Transform-Ownership | The word {evt} triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. |
| Justice.Execute | The word {evt} triggers an EXECUTE event: a PERSON is/was executed |
| Justice.Trial-Hearing | The word {evt} triggers a TRIAL-HEARING event: a court proceeding has been initiated for the purposes of determining the guilt or innocence of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime. |
| Justice.Sentence | The word {evt} triggers a SENTENCE event: the punishment for the DEFENDANT is issued |
| Life.Be-Born | The word {evt} triggers a BE-BORN event: a PERSON is given birth to. |
| Justice.Charge-Indict | The word {evt} triggers a CHARGE-INDICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is accused of a crime |
| Business.Start-Org | The word {evt} triggers a START-ORG event: a new ORGANIZATION is created. |
| Justice.Convict | The word {evt} triggers a CONVICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is convicted whenever it has been found guilty of a CRIME. |
| Business.Declare-Bankruptcy | The word {evt} triggers a DECLARE-BANKRuptCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. 10600 |
| Justice/release-Parole | The word {evt} triggers a RELEASE-PAROLE event. |
| Justice.Fine | The word {evt} triggers a FINE event: a GEOPOLITICAL ENTITY, PERSON or ORGANIZATION get financial punishment typically as a result of court proceedings. |
| Justice.Pardon | The word {evt} triggers a PARDON event: a head-of-state or their appointed representative lifts a sentence imposed by the judiciary. |
| Justice.Appeal | The word {evt} triggers a APPEAL event: the decision of a court is taken to a higher court for review |
| Business.Merge-Org | The word {evt} triggers a MERGE-ORG event: two or more ORGANIZATION Entities come together to form a new ORGANIZATION Entity. |
| Justice.Extradite | The word {evt} triggers a EXTRADITE event. |
| Life.Divorce | The word {evt} triggers a DIVORCE event: two people are officially divorced under the legal definition of divorce. |
| Justice.Acquit | The word {evt} triggers a ACQUIT event: a trial ends but fails to produce a conviction. |
| Dataset (Task) | Roberta | T5 | LLaMA | CODEX |
| FewNERD (NER) | 2.8 | 39.4 | 1135.4 | 179.4 |
| TACREV (RE) | 1.4 | 45.6 | 1144.9 | 151.6 |
| ACE05 (ED) | 6.6 | 62.5 | 733.4 | 171.7 |
| FewNERD | TACREV | ACE05 | |
| Overall | 5.88 | 3.03 | 38.2 |
| Easy samples (τ > 0.9) | 9.44 | 3.21 | 44.0 |
| Hard samples (τ < 0.6) | 1.28 | 2.68 | 1.36 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ||||||||
| 5-shot | 10-shot | 20-shot | 20-shot | 50-shot | 100-shot | 5-shot | 10-shot | 20-shot | ||
| LLM | CODEX | 53.8(0.5) | 54.0(1.4) | 55.9(0.5) | 59.1(1.4) | 60.3(2.4) | 62.4(2.6) | 47.1(1.2) | 47.7(2.8) | 47.9(0.5) |
| InstructGPT | 53.6(-) | 54.6(-) | 57.2(-) | 60.1(-) | 58.3(-) | 62.7(-) | 52.9(-) | 52.1(-) | 49.3(-) | |
| GPT-4 | - | - | 57.8(-) | - | - | 59.3(-) | - | - | 52.1(-) | |
| SLM | Previous SoTA | 59.4(1.5) | 61.4(0.8) | 61.9(1.2) | 62.4(3.8) | 68.5(1.6) | 72.6(1.5) | 55.1(4.6) | 63.9(0.8) | 65.8(2.0) |
| + Ensemble (S) | 59.6(1.7) | 61.8(1.2) | 62.6(1.0) | 64.9(1.5) | 71.9(2.2) | 74.1(1.7) | 56.9(4.7) | 64.2(2.1) | 66.5(1.7) | |
| + Rerank (S) | 59.4(1.5) | 61.0(1.7) | 61.5(1.7) | 64.2(2.3) | 70.8(2.3) | 74.3(2.2) | 56.1(0.3) | 64.0(1.0) | 66.7(1.7) | |
| SLM + LIM | Vicuna-13B | |||||||||
| + Rerank (L) | 60.0(1.8) | 61.9(2.1) | 62.2(1.4) | 65.2(1.4) | 70.8(1.6) | 73.8(1.7) | 56.9(4.0) | 63.5(2.7) | 66.0(2.6) | |
| + Ensemble (S) + Rerank (L) | 59.9(0.7) | 62.1(0.7) | 62.8(1.1) | 66.5(0.5) | 73.6(1.4) | 75.0(1.5) | 57.9(5.2) | 64.4(1.2) | 66.2(2.4) | |
| InstructGPT | ||||||||||
| + Rerank (L) | 60.6(2.1) | 62.7(0.8) | 63.3(0.6) | 66.8(2.6) | 72.3(1.4) | 75.4(1.5) | 57.8(4.6) | 65.3(1.7) | 67.3(2.2) | |
| + Ensemble (S) + Rerank (L) | 61.3(1.9) | 63.2(0.9) | 63.7(1.8) | 68.9(1.3) | 74.8(1.3) | 76.8(1.2) | 59.5(3.7) | 65.3(1.9) | 67.8(2.1) | |
| GPT-4 | ||||||||||
| + Rerank (L) | 60.8(2.3) | 62.6(2.7) | 63.0(1.3) | 65.9(2.7) | 72.3(0.3) | 74.5(1.5) | 59.6(2.9) | 64.9(2.5) | 67.1(2.5) | |
| + Ensemble (S) + Rerank (L) | 61.1(2.2) | 62.8(0.9) | 63.6(1.2) | 68.6(1.3) | 73.9(1.4) | 75.9(2.4) | 60.9(3.9) | 65.6(1.5) | 67.8(1.7) |
| GPT-4 | InstructGPT | |||||||
| before | after | △ | ratio | before | after | △ | ratio | |
| FewNER | 31.9 | 40.7 | 8.8 | 3.2% | 31.4 | 28.3 | -3.1 | 3.3% |
| TACREV | 25.3 | 43.0 | 17.7 | 9.1% | 33.8 | 43.4 | 9.6 | 7.1% |
| ACE05 | 31.1 | 57.9 | 26.8 | 1.6% | 35.6 | 55.7 | 20.1 | 0.5% |
| CoT | Demo | LF | AD | FewNERD (20-shot) | TACREV (100-shot) | ACE05 (20-shot) |
| ✓ | ✓ | ✓ | ✓ | 63.6(1.2) | 75.9(2.4) | 67.8(1.7) |
| ✗ | ✓ | ✓ | ✓ | 63.2(1.2) | 75.4(2.4) | 67.2(1.7) |
| ✗ | ✗ | ✓ | ✓ | 63.0(1.4) | 74.9(2.2) | 66.6(1.5) |
| ✗ | ✗ | ✗ | ✓ | 62.4(2.1) | 73.8(2.5) | 66.5(1.3) |
| ✗ | ✗ | ✗ | ✗ | 12.5(2.7) | 59.9(6.0) | 5.4(1.1) |
| Previous SoTA methods | 62.6(1.0) | 74.1(1.7) | 66.5(1.7) |
| Dataset | Named Entity Recognition\nCONLL OntoNotes | FewNERD | Relation Extraction | Event Detection | Event Arg Extraction | ||||||
| TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | ||||
| #Label Type | 4 | 18 | 66 | 41 | 41 | 33 | 168 | 38 | 33 | 139 | |
| #Sents | Train | 14,041 | 49,706 | 131,965 | 68,124 | 68,124 | 14,024 | 32,360 | 14,736 | 14,024 | 7329 |
| Test | 3,453 | 10,348 | 37,648 | 15,509 | 15,509 | 728 | 8,035 | 1,163 | 728 | 871 | |
| #Mentions | Train | 23,499 | 128,738 | 340,247 | 13,012 | 13,012 | 5,349 | 77,993 | 6,208 | 4859 | 17026 |
| Test | 5,648 | 12,586 | 96,902 | 3,123 | 3,123 | 424 | 18,904 | 551 | 576 | 2023 |
| Dataset Settings | # Labels | # Sent | # Sample | # Avg shot | |
| CONLL'03 | 1-shot | 4 | 4.8 | 5.8 | 1.4 |
| 5-shot | 16.2 | 21.8 | 5.5 | ||
| 10-shot | 29.2 | 42.6 | 10.7 | ||
| 20-shot | 65.6 | 82.0 | 20.5 | ||
| OntoNotes | 1-shot | 18 | 20.0 | 33.4 | 1.9 |
| 5-shot | 84.8 | 148.0 | 8.2 | ||
| 10-shot | 158.6 | 281.0 | 15.6 | ||
| 20-shot | 332.8 | 547.2 | 30.4 | ||
| FewNERD | 1-shot | 66 | 89.8 | 147.0 | 2.2 |
| 5-shot | 286.2 | 494.8 | 7.5 | ||
| 10-shot | 538.0 | 962.0 | 14.6 | ||
| 20-shot | 1027.2 | 1851.4 | 28.1 | ||
| TACREV | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2872.0 | 1944.0 | 47.4 | ||
| 100-shot | 4561.0 | 3520.0 | 85.9 | ||
| TACRED | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2871.2 | 1944.0 | 47.4 | ||
| 100-shot | 4575.2 | 3520.0 | 85.9 | ||
| ACE05 | 1-shot | 33 | 47.4 | 41.0 | 1.2 |
| 5-shot | 192.8 | 165.0 | 5.0 | ||
| 10-shot | 334.6 | 319.4 | 9.7 | ||
| 20-shot | 579.4 | 598.2 | 18.1 | ||
| MAVEN | 1-shot | 168 | 157.6 | 298.0 | 1.8 |
| 5-shot | 540.4 | 1262.2 | 7.5 | ||
| 10-shot | 891.2 | 2413.8 | 14.4 | ||
| 20-shot | 1286.4 | 4611.4 | 27.4 | ||
| ERE | 1-shot | 38 | 48.4 | 54.6 | 1.4 |
| 5-shot | 175.0 | 219.2 | 5.8 | ||
| 10-shot | 304.8 | 432.4 | 11.4 | ||
| 20-shot | 521.6 | 806.6 | 21.2 | ||
| ACE05 | 1-shot | 33 | 23.4 | 40.2 | 1.2 |
| 5-shot | 79.8 | 178.2 | 5.4 | ||
| 10-shot | 130.8 | 337.4 | 10.2 | ||
| 20-shot | 213.4 | 630.2 | 19.1 | ||
| RAMS | 1-shot | 139 | 130.2 | 332.6 | 2.4 |
| 5-shot | 514.0 | 1599.6 | 11.5 | ||
| 10-shot | 795.2 | 3193.2 | 23.0 | ||
| 20-shot | 1070.4 | 6095.4 | 43.9 | ||
| ERE | 1-shot | 38 | 21.6 | 102.8 | 2.7 |
| 5-shot | 74.2 | 403.4 | 10.6 | ||
| 10-shot | 127.2 | 775.6 | 20.4 | ||
| 20-shot | 190.2 | 1397.2 | 36.8 |
| FewNERD | TACREV | ACE05 | |
| t=0 | 48.5(1.9) | 53.7(2.3) | 42.9(2.2) |
| +5-ensemble | 53.5(1.3) | 58.6(1.5) | 46.3(0.8) |
| t=0.7 | 40.9(2.3) | 39.9(1.2) | 35.6(1.0) |
| + self-consistency | 52.1(0.9) | 53.4(1.3) | 45.6(3.0) |
| 10-shot train set | FewNERD (NER) | TACREV (RE) | ACE05 (ED) |
| wo. Auto-CoT | 54.0(1.4) | 57.3(1.8) | 47.7(2.8) |
| w. Auto-CoT | 36.6(1.7) | 22.0(1.2) | 43.1(3.4) |
| NER (20-shot) | RE (100-shot) | ED (20-shot) | EAE (20-shot) | ||||||||
| CONLL | OntoNotes | FewNERD | TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | |
| InstructGPT | 77.2 | 47.7 | 57.2 | 62.7 | 53.8 | 49.3 | 25.4 | 40.8 | 45.8 | 42.2 | 41.9 |
| CODEX | 81.1 | 55.6 | 55.9 | 62.4 | 53.6 | 47.9 | 22.8 | 39.0 | - | - | - |
| GPT-4 | 84.7 | 65.6 | 57.8 | 59.3 | 50.4 | 52.1 | 30.2 | 40.5 | 42.9 | 38.6 | 38.2 |
| Supervised SoTA | 72.3 | 74.9 | 61.4 | 72.6 | 63.1 | 65.8 | 54.7 | 56.2 | 55.2 | 57.7 | 55.6 |
| Named Entity Recognition | |||
| CoNLL | OntoNotes | FewNERD | |
| # Entity | 4 | 18 | 66 |
| Micro-F1 (SLM) | 52.5 | 59.7 | 59.4 |
| Micro-F1 (LLM) | 77.8 | 59.4 | 55.5 |
| ΔF1 (LLM, SLM) | 25.3 | -0.3 | -3.9 |
| Event Detection | |||
| ACE05 | ERE | MAVEN | |
| # Event | 33 | 38 | 168 |
| Micro-F1 (SLM) | 55.1 | 48.0 | 49.4 |
| Micro-F1 (LLM) | 39.6 | 33.8 | 25.3 |
| ΔF1 (LLM, SLM) | -15.5 | -14.2 | -24.1 |
| Event Argument Extraction | |||
| ACE05 | ERE | RAMS | |
| # Event / #Role | 33 / 22 | 38 / 26 | 139 / 65 |
| Head-F1 (SLM) | 45.9 | 40.4 | 54.1 |
| Head-F1 (LLM) | 52.8 | 40.7 | 44.2 |
| ΔF1 (LLM, SLM) | 6.9 | 0.3 | -9.9 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| I0 | 57.6(2.1) | 49.1(2.4) | 44.0(1.4) | 50.9(0.1) |
| I1 | 58.3(0.5) | 49.6(1.2) | 42.6(1.0) | 51.5(1.1) |
| I2 | 57.7(1.0) | 50.0(1.5) | 41.8(0.9) | 50.3(1.5) |
| I3 | 57.6(2.3) | 52.3(1.8) | 42.9(1.3) | 49.2(2.3) |
| I4 | 56.8(0.9) | 49.6(2.9) | 41.6(1.9) | 49.9(1.2) |
| I5 | 57.8(0.5) | 47.2(1.8) | 43.1(1.8) | 50.6(1.8) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | |
| Random Sampling | 53.2(0.4) | 43.0(3.3) | 38.0(1.5) |
| Sentence Embedding | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) |
| Efficient Prompt Retriever | 57.2(0.6) | 48.0(0.8) | 43.5(1.4) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| Text | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) | 51.5(1.1) |
| Code | 53.2(0.9) | 50.2(1.8) | 44.3(2.0) | 47.3(1.5) |
| Sentences (entity/event) | Before | After | Rationales |
| Named Entity Recognition | |||
| According to it, King Eleusis and Cothonea (Cyntinia), parents of Trip-tolemus, are visited by Demeter, who rears their son, feeding him divine milk by day and placing him into the fire at night, which makes Triptolemus grow faster than mortal children normally do. | person-other ✘ | other-god ✓ | According to the context, Triptolemus is a figure in Greek mythology, the son of King Eleusis and Cothonea (Cyntinia). He was a demigod who was sent by Demeter to teach mankind the art of agriculture. So Triptole-mus is a god in some legend/religious story. |
| Relation Extraction | |||
| Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul. | None ✘ | per:children ✓ | The sentence states 'Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul.' This indicates 'his' is the father of Jeff, James, Douglas and Harris, and the husband of Sandra. So his is the parent of Jeff. |
| Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais. | per:cities_of _residence ✘ | per:states_of _residence ✓ | The sentence states that 'Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais.' This indicates that Eliza Samudio is from the city of Belo Horizonte, which is located in the state of Minas Gerais. So Eliza Samudio lives in the state or province Minas Gerais. |
| Event Detection | |||
| We have fought every institution, every law firm ... it's cost us a fortune but now we're going on a massive expansion. | Conflictattack ✘ | None ✓ | In the given sentence, the word fought is used to describe the action of the speaker fighting against various institutions and law firms. This does not involve any physical violence or court proceedings, so the word fought does not trigger any known event. |
| Instruct: Read following sentences and identify what is the entity type of 392 quoted by <t>.Sentence: Powell v. Texas, <t> 392 U.S. 514 ( 1968 ) , was a United States Supreme Court case that ruled that a Texas statute criminalizing public intoxication did not violate the Eighth Amendment protection against cruel and unusual punishment.(a) 392 is a legal document, a term or a convention in legal sense.(b) 392 does not belong to any known entities.(c) 392 refers to a protest, uprising or revolution event(d) 392 refers to a government or governmental agencyAnalysis: In the context you provided, 392 refers to the volume number in the United States Reports where the Supreme Court's decision in Powell v. Texas can be found. However, 392 itself does not refer to a legal document. So 392 do/does not belong to any known entities.An Answer: (b) |
| Instruct: Read following sentences and identify what is the entity type of The New Yorker quoted by <t>.Sentence: In 2004 Gourevitch was assigned to cover the 2004 U.S. presidential election for " <t> The New Yorker <t> ".(a) The New Yorker does not belong to any known entities.(b) The New Yorker is a broadcast program.(c) The New Yorker is a kind of written art.(d) The New Yorker is a media/newspaper organization.Analysis: The New Yorker is a well-known American magazine that has been published since 1925, and is primarily known for its long-form journalism, commentary, and satire. It has a reputation for publishing high-quality writing on a wide variety of topics, including politics, culture, and the arts. So The New Yorker is a media/newspaper organization.An Answer: (d) |
| Instruct: Read following sentence and identify what is the entity type of St. quoted by <t>.Sentence: The May 1980 eruption of Mount St. Helens in the state of Washington seriously affected both 47th Air Division and 92d Bombardment Wing operations at Fairchild AFB , resulting in dispersal of Fairchild 's B-52 and KC-135 aircraft to various bases while around-the-clock shifts removed the volcanic ash from facilities within the base perimeter." (a) St. does not belong to any known entities.(b) St. is a natural disaster event.(c) St. is a geographic position about mountain.Analysis: According to the context, St. is an abbreviation of Saint, used in the name of Mount St. Helens, which is an active volcano in the state of Washington. However, St. itself does not refer to anything. So St. do/does not belong to any known entities.An Answer: (a) |
| Instruct: Read following sentence and identify what is the entity type of Ridzuan quoted by <t>.Sentence: <t> Ridzuan was promoted to Harimau Muda A for 2014 season .(a) Ridzuan does not belong to any known entities.(b) Ridzuan is a person, but not affiliated with following professions: actor, artist, author, director, politician, scholar, soldier.(c) Ridzuan is an athlete.Analysis: The mention of 'Harimau Muda A' indicates that it is a sports-related context, and 'promoted' implies a progression or advancement within the sports team. So Ridzuan is an athlete.An Answer: (c) |
| Instruct: Read the sentence and determine the relation between she and lawyer quoted by <t>. \nSentence: The <t> lawyer <t> denied Italian news reports that she wept while addressing the court, but said Knox was upset as <t> she <t> recounted “ the pressure, the aggressiveness of the police who called her a liar . ” \n(a) she is the other family member of lawyer \n(b) she is a lawyer \n(c) she has no known relations to lawyer \nAnalysis: In the sentence, the word ‘she’ refers to someone who was upset while recounting certain events in court. The word ‘lawyer’ refers to someone who denied a news report about that same person weeping in court. There is no information in the sentence to indicate that the two individuals are related in any way. So she has no known relations to lawyer. \nAnswer: (c) |
| Instruct: Read the sentence and determine the relation between MEF and Myanmar Equestrian Federation quoted by <t>. \nSentence: YANGON, Dec. 27 -LRB- Xinhua -RRB- - Myanmar will hold a horse race in Yangon to commemorate the country ’63rd Anniversary Independence Day , the <t> Myanmar Equestrian Federation <t> -LRB- <t> MEF <t> -RRB- confirmed to Xinhua on Monday. \n(a) MEF is also known as Myanmar Equestrian Federation \n(b) MEF has political affiliation with Myanmar Equestrian Federation \n(c) MEF has no known relations to Myanmar Equestrian Federation \nAnalysis: The symbols -LRB- and -RRB- in the sentence stand for left and right round brackets and are used to enclose the abbreviation ‘MEF’ to indicate that it is a replacement for the longer name ‘Myanmar Equestrian Federation. So MEF is also known as Myanmar Equestrian Federation. \nAnswer: (a) |
| Instruct: Read the sentence and determine the relation between Douglas Flint and chairman quoted by <t>. \nSentence: At the same time , Chief Financial Officer <t> Douglas Flint <t> will become <t> chairman <t> , succeeding Stephen Green who is leaving to take a government job. \n(a) Douglas Flint has no known relations to chairman \n(b) Douglas Flint is a chairman \n(c) Douglas Flint is the employee of chairman \nAnalysis: The sentence states that Chief Financial Officer Douglas Flint Douglas Flint will succeed Stephen Green as a chairman. So Douglas Flint is a chairman. \nAnswer: (b) |
| Instruct: Read the sentence and determine the relation between FAA and U.S. quoted by <t>. \nSentence: On its Web site , the <t> U.S. <t> FAA <t> says the Category 2 rating means the country lacks the laws or regulations that are needed for the certification and oversight of air carriers , according to minimum international standards. \n(a) FAA is also known as U.S. \n(b) FAA has no known relations to U.S. \n(c) FAA has a headquarters in the country U.S. \nAnalysis: The sentence states that the FAA says the Category 2 rating means the country lacks the laws or regulations needed for the certification and oversight of air carriers, indicating that the FAA is responsible for overseeing aviation regulations in the country. Actually the FAA (Federal Aviation Administration) is a U.S. government agency responsible for regulating and overseeing civil aviation in the United States, and it has its headquarters in Washington, D.C.. So FAA has a headquarters in the country U.S.. \nAnswer: (c) |
| Instruct: Read following sentences and identify what event is triggered by the word loan quoted by <t>. \nSentence: Separately, former WorldCom CEO Bernard Ebbers failed on April 29 to make a first repayment of 25 million dollars, plus interest, on a 400-million - dollar <t> from MCI, the Journal said, citing SEC documents. \n(a) The word loan does not trigger any known event. \n(b) The word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. \n(c) The word loan triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. \nAnalysis: In the given sentence, the word loan is used to describe the 400-million-dollar loan from MCI to former WorldCom CEO Bernard Ebbers, which he failed to repay on time. This situation clearly involves a transfer of money without the context of purchasing something, which falls under the TRANSFER-MONEY event. So the word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words treated quoted by <t>. \nSentence: When she 's in Germany, Lynch will be <t> treated <t> for bullet wounds and broken bones. \n(a) The word treated triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. \n(b) The word treated does not trigger any known event. \n(c) The word treated triggers a TRANSPORT event: an ARTifact (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. \nAnalysis: The sentence suggests that Lynch has already been injured and will receive medical treatment in Germany for her injuries. The word 'treated' simply describes the medical care she will receive and does not indicate a new event or action taking place. So the word treated does not trigger any known event. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words buy quoted by <t>. \nSentence: And I won't dwell on the irony of an Oracle employee being driven out of Oracle, starting his own company, and forcing Ellison to spend $10.3 billion to get his company – but not him – back (though it does rather delightfully remind me of Coca - Cola basically giving away the bottling franchise and then spending billions to <t> buy <t> it back). \n(a) The word buy triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. \n(b) The word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. \n(c) The word buy does not trigger any known event. \nAnalysis: In the given sentence, the word buy is used to describe the action of Oracle spending $10.3 billion to get a company back. This clearly involves the transfer of ownership of the company from one entity to another. So the word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words set quoted by <t>. \nSentence: British forces also began establishing the country's first postwar administration Tuesday, granting a local sheik power to <t> set up an administrative committee representing the groups in the region. \n(a) The word set triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. \n(b) The word set triggers a START-ORG event: a new ORGANIZATION is created. \n(c) The word set does not trigger any known event. \nAnalysis: The phrase 'set up' specifically implies the creation or establishment of a new organization or entity, rather than simply the word 'set'. So the word set does not trigger any known event. \nAnswer: (c) |
| Entity | Template |
| no-entity | {ent} do/does not belong to any known entities. |
| person-artist/author | {ent} is an artist or author. |
| person-actor | {ent} is an actor. |
| art-written | {ent} is a kind of written. |
| person-director | {ent} is a director. |
| person-other | {ent} is a person, but not affiliated with following professions: actor, artist, athlete, author, director, politician, scholar, soldier. |
| organization-other | {ent} pertains to an organization that does not fall under the categories of company, educational institution, government, media, political party, religion, sports league, sports team, band or musical group. |
| organization-company | {ent} is a company |
| organization-sportsteam | {ent} is a sports team |
| organization-sportsleague | {ent} is a sports league |
| product-car | {ent} is a kind of car |
| event-protest | {ent} refers to a protest, uprising or revolution event |
| organization-government/governmentagency | {ent} refers to a government or governmental agency |
| other-biologything | {ent} is a special term about biology / life science. |
| location-GPE | {ent} is a kind of geopolitical entity |
| location-other | {ent} is a geographic locaton that does not fall under the categories of geopolitical entity, body of water, island, mountain, park, road, railway and transit. |
| person-athlete | {ent} is an athlete or coach. |
| art-broadcastprogram | {ent} is a broadcast program. |
| product-other | {ent} is a kind of product that does not fall under the categories of airplane, train, ship, car, weapon, food, electronic game and software. |
| building-other | {ent} is a kind of building that does not fall under the categories of airport, hospital, hotel, library, restaurant, sports facility and theater |
| product-weapon | {ent} is a kind of weapon. |
| building-airport | {ent} is an airport. |
| building-sportsfacility | {ent} is a sports facility building. |
| person-scholar | {ent} is a scholar. |
| art-music | {ent} is a music. |
| event-other | {ent} refers to some event except attack, election, natural disaster, protest, revolution and sports |
| other-language | {ent} is a kind of human language. |
| other-chemicalthing | {ent} is some special term about chemical science. |
| art-film | {ent} is a film. |
| building-hospital | {ent} is a hospital. |
| other-law | {ent} is a legal document, a term or a convention in legal sense. |
| product-airplane | {ent} is kind of airplane product. |
| location-road/railway/highway/transit | {ent} is a geographic position about roadways, railways, highways or public transit systems. |
| person-soldier | {ent} is a soldier |
| location-mountain | {ent} is geographic position about mountain. |
| organization-education | {ent} is an educational institute/organization. |
| organization-media/newspaper | {ent} is a media/newspaper organization. |
| product-software | {ent} is a software product. |
| location-island | {ent} is geographic position about island. |
| location-bodiesofwater building-library | {ent} is geographic position situated near a body of water. {ent} is a library. |
| other-astronomything | {ent} is a special term about astronomy. |
| person-politician | {ent} is a politician or lawyer or judge. |
| building-hotel | {ent} is a hotel building. |
| product-game | {ent} is a electronic game product. |
| other-award | {ent} is a kind of award. |
| event-sportsevent | {ent} refers to some event related to sports. |
| organization-showorganization | {ent} is a band or musical organization. |
| other-educationaldegree | {ent} is a kind of educational degree. |
| building-theater | {ent} is a theater. |
| other-disease | {ent} is a kind of disease. |
| event-election | {ent} is an event about election. |
| organization-politicalparty | {ent} is a political party/organization. |
| other-currency | {ent} is a kind of currency. |
| event-attack/battle/war/militaryconflict | {ent} is an event about attack, battle, war or military conflict. |
| product-ship | {ent} is a ship. |
| building-restaurant | {ent} is a restaurant. |
| other-livingthing | {ent} is a living animal/crystal/organism. |
| art-other | {ent} is a work of art, but not belong to the categories of music, film, written art, broadcast or painting. |
| event-disaster | {ent} is a natural disaster event. |
| organization-religion | {ent} is a religious organization. |
| other-medical | {ent} refers to some kind of medicineentity |
| location-park | {ent} is a park. |
| other-god | {ent} is a god in some legend/religious story. |
| product-food | {ent} is a kind of food. |
| product-train | {ent} is a kind of train(vehicle). |
| art-painting | {ent} is an art painting. |
| Relation | Template |
| no Relation | {subj} has no known relations to {obj} |
| per:stateorprovince_of_death | {subj} died in the state or province {obj} |
| per:title | {subj} is a {obj} |
| org:member_of | {subj} is the member of {obj} |
| per:other_family | {subj} is the other family member of {obj} |
| org:country_of_HEADquarters | {subj} has a headquarters in the country {obj} |
| org:parents | {subj} has the parent company {obj} |
| per:stateorprovince_of_birth | {subj} was born in the state or province {obj} |
| per:spouse | {subj} is the spouse of {obj} |
| per:origin | {subj} has the nationality {obj} |
| per:date_of_birth | {subj} has birthday on {obj} |
| per:schools_attended | {subj} studied in {obj} |
| org:members | {subj} has the member {obj} |
| org:founded | {subj} was founded in {obj} |
| per:stateorprovinces_of_residence | {subj} lives in the state or province {obj} |
| per:date_of_death | {subj} died in the date {obj} |
| org:shareholders | {subj} has shares hold in {obj} |
| org:website | {subj} has the website {obj} |
| org:subsidiaries | {subj} owns {obj} |
| per:charges | {subj} is convicted of {obj} |
| org:dissolved | {subj} dissolved in {obj} |
| org:stateorprovince_of_HEADquarters | {subj} has a headquarters in the state or province {obj} |
| per:country_of_birth | {subj} was born in the country {obj} |
| per:siblings | {subj} is the siblings of {obj} |
| org:top_membership/employees | {subj} has the high level member {obj} |
| per:cause_of_death | {subj} died because of {obj} |
| per:alternate_names | {subj} has the alternate name {obj} |
| org:number_of_employees/members | {subj} has the number of employees {obj} |
| per:cities_of_residence | {subj} lives in the city {obj} |
| org:city_of_HEADquarters | {subj} has a headquarters in the city {obj} |
| per:children | {subj} is the parent of {obj} |
| per:employee_of | {subj} is the employee of {obj} |
| org:political/religious_affiliation | {subj} has political affiliation with {obj} |
| per:parents | {subj} has the parent {obj} |
| per:city_of_birth | {subj} was born in the city {obj} |
| per:age | {subj} has the age {obj} |
| per:countries_of_residence | {subj} lives in the country {obj} |
| org:alternate_names | {subj} is also known as {obj} |
| per:religion | {subj} has the religion {obj} |
| per:city_of_death | {subj} died in the city {obj} |
| per:country_of_death | {subj} died in the country {obj} |
| org:founded_by | {subj} was founded by {obj} |
| Event | Template |
| no-event | The word {evt} does not trigger any known event. |
| MovementTRANSPORT | The word {evt} triggers a TRANSPORT event: an ARTIFACT (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. |
| Personnel.Elect | The word {evt} triggers an ELECT event which implies an election. |
| Personnel.Start-Position | The word {evt} triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Personnel.Nominate | The word {evt} triggers a NOMINATE event: a PERSON is proposed for a position through official channels. |
| Conflict.Attack | The word {evt} triggers an ATTACK event: a violent physical act causing harm or damage. |
| Personnel.End-Position | The word {evt} triggers an END-POSITION event: a PERSON stops working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Contact.Meet | The word {evt} triggers a MEET event: two or more entities come together at a single location and interact with one another face-to-face. |
| Life.Marry | The word {evt} triggers a MARRY event: two people are married under the legal definition. |
| Contact.Phone-Write | The word {evt} triggers a PHONE-WRITE event: two or more people directly engage in discussion which does not take place 'face-to-face'. |
| Transaction.Transform-Money | The word {evt} triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. |
| Justice.Sue | The word {evt} triggers a SUE event: a court proceeding has been initiated for the purposes of determining the liability of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime or neglecting a commitment |
| Conflict.Demonstrate | The word {evt} triggers a DEMONSTRATE event: a large number of people come together in a public area to protest or demand some sort of official action. For example: protests, sit-ins, strikes and riots. |
| Business.End-Org | The word {evt} triggers an END-ORG event: an ORGANIZATION ceases to exist (in other words, goes out of business). |
| Life.Injure | The word {evt} triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. |
| Life.Die | The word {evt} triggers a DIE event: a PERSON dies/died whether it occurs accidentally, intentionally or even self-inflicted. |
| Justice.Arrest-Jail | The word {evt} triggers a ARREST-JAIL event: a PERSON is sent to prison. |
| Transaction.Transform-Ownership | The word {evt} triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. |
| Justice.Execute | The word {evt} triggers an EXECUTE event: a PERSON is/was executed |
| Justice.Trial-Hearing | The word {evt} triggers a TRIAL-HEARING event: a court proceeding has been initiated for the purposes of determining the guilt or innocence of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime. |
| Justice.Sentence | The word {evt} triggers a SENTENCE event: the punishment for the DEFENDANT is issued |
| Life.Be-Born | The word {evt} triggers a BE-BORN event: a PERSON is given birth to. |
| Justice.Charge-Indict | The word {evt} triggers a CHARGE-INDICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is accused of a crime |
| Business.Start-Org | The word {evt} triggers a START-ORG event: a new ORGANIZATION is created. |
| Justice.Convict | The word {evt} triggers a CONVICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is convicted whenever it has been found guilty of a CRIME. |
| Business.Declare-Bankruptcy | The word {evt} triggers a DECLARE-BANKRuptCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. 10600 |
| Justice/release-Parole | The word {evt} triggers a RELEASE-PAROLE event. |
| Justice.Fine | The word {evt} triggers a FINE event: a GEOPOLITICAL ENTITY, PERSON or ORGANIZATION get financial punishment typically as a result of court proceedings. |
| Justice.Pardon | The word {evt} triggers a PARDON event: a head-of-state or their appointed representative lifts a sentence imposed by the judiciary. |
| Justice.Appeal | The word {evt} triggers a APPEAL event: the decision of a court is taken to a higher court for review |
| Business.Merge-Org | The word {evt} triggers a MERGE-ORG event: two or more ORGANIZATION Entities come together to form a new ORGANIZATION Entity. |
| Justice.Extradite | The word {evt} triggers a EXTRADITE event. |
| Life.Divorce | The word {evt} triggers a DIVORCE event: two people are officially divorced under the legal definition of divorce. |
| Justice.Acquit | The word {evt} triggers a ACQUIT event: a trial ends but fails to produce a conviction. |
| Dataset (Task) | Roberta | T5 | LLaMA | CODEX |
| FewNERD (NER) | 2.8 | 39.4 | 1135.4 | 179.4 |
| TACREV (RE) | 1.4 | 45.6 | 1144.9 | 151.6 |
| ACE05 (ED) | 6.6 | 62.5 | 733.4 | 171.7 |
| FewNERD | TACREV | ACE05 | |
| Overall | 5.88 | 3.03 | 38.2 |
| Easy samples (τ > 0.9) | 9.44 | 3.21 | 44.0 |
| Hard samples (τ < 0.6) | 1.28 | 2.68 | 1.36 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ||||||||
| 5-shot | 10-shot | 20-shot | 20-shot | 50-shot | 100-shot | 5-shot | 10-shot | 20-shot | ||
| LLM | CODEX | 53.8(0.5) | 54.0(1.4) | 55.9(0.5) | 59.1(1.4) | 60.3(2.4) | 62.4(2.6) | 47.1(1.2) | 47.7(2.8) | 47.9(0.5) |
| InstructGPT | 53.6(-) | 54.6(-) | 57.2(-) | 60.1(-) | 58.3(-) | 62.7(-) | 52.9(-) | 52.1(-) | 49.3(-) | |
| GPT-4 | - | - | 57.8(-) | - | - | 59.3(-) | - | - | 52.1(-) | |
| SLM | Previous SoTA | 59.4(1.5) | 61.4(0.8) | 61.9(1.2) | 62.4(3.8) | 68.5(1.6) | 72.6(1.5) | 55.1(4.6) | 63.9(0.8) | 65.8(2.0) |
| + Ensemble (S) | 59.6(1.7) | 61.8(1.2) | 62.6(1.0) | 64.9(1.5) | 71.9(2.2) | 74.1(1.7) | 56.9(4.7) | 64.2(2.1) | 66.5(1.7) | |
| + Rerank (S) | 59.4(1.5) | 61.0(1.7) | 61.5(1.7) | 64.2(2.3) | 70.8(2.3) | 74.3(2.2) | 56.1(0.3) | 64.0(1.0) | 66.7(1.7) | |
| SLM + LIM | Vicuna-13B | |||||||||
| + Rerank (L) | 60.0(1.8) | 61.9(2.1) | 62.2(1.4) | 65.2(1.4) | 70.8(1.6) | 73.8(1.7) | 56.9(4.0) | 63.5(2.7) | 66.0(2.6) | |
| + Ensemble (S) + Rerank (L) | 59.9(0.7) | 62.1(0.7) | 62.8(1.1) | 66.5(0.5) | 73.6(1.4) | 75.0(1.5) | 57.9(5.2) | 64.4(1.2) | 66.2(2.4) | |
| InstructGPT | ||||||||||
| + Rerank (L) | 60.6(2.1) | 62.7(0.8) | 63.3(0.6) | 66.8(2.6) | 72.3(1.4) | 75.4(1.5) | 57.8(4.6) | 65.3(1.7) | 67.3(2.2) | |
| + Ensemble (S) + Rerank (L) | 61.3(1.9) | 63.2(0.9) | 63.7(1.8) | 68.9(1.3) | 74.8(1.3) | 76.8(1.2) | 59.5(3.7) | 65.3(1.9) | 67.8(2.1) | |
| GPT-4 | ||||||||||
| + Rerank (L) | 60.8(2.3) | 62.6(2.7) | 63.0(1.3) | 65.9(2.7) | 72.3(0.3) | 74.5(1.5) | 59.6(2.9) | 64.9(2.5) | 67.1(2.5) | |
| + Ensemble (S) + Rerank (L) | 61.1(2.2) | 62.8(0.9) | 63.6(1.2) | 68.6(1.3) | 73.9(1.4) | 75.9(2.4) | 60.9(3.9) | 65.6(1.5) | 67.8(1.7) | |
| GPT-4 | InstructGPT | |||||||
| before | after | △ | ratio | before | after | △ | ratio | |
| FewNER | 31.9 | 40.7 | 8.8 | 3.2% | 31.4 | 28.3 | -3.1 | 3.3% |
| TACREV | 25.3 | 43.0 | 17.7 | 9.1% | 33.8 | 43.4 | 9.6 | 7.1% |
| ACE05 | 31.1 | 57.9 | 26.8 | 1.6% | 35.6 | 55.7 | 20.1 | 0.5% |
| CoT | Demo | LF | AD | FewNERD (20-shot) | TACREV (100-shot) | ACE05 (20-shot) |
| ✓ | ✓ | ✓ | ✓ | 63.6(1.2) | 75.9(2.4) | 67.8(1.7) |
| ✗ | ✓ | ✓ | ✓ | 63.2(1.2) | 75.4(2.4) | 67.2(1.7) |
| ✗ | ✗ | ✓ | ✓ | 63.0(1.4) | 74.9(2.2) | 66.6(1.5) |
| ✗ | ✗ | ✗ | ✓ | 62.4(2.1) | 73.8(2.5) | 66.5(1.3) |
| ✗ | ✗ | ✗ | ✗ | 12.5(2.7) | 59.9(6.0) | 5.4(1.1) |
| Previous SoTA methods | 62.6(1.0) | 74.1(1.7) | 66.5(1.7) | |||
| Dataset | Named Entity Recognition +CONLL OntoNotes | FewNERD | Relation Extraction | Event Detection | Event Arg Extraction | ||||||
| TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | ||||
| #Label Type | 4 | 18 | 66 | 41 | 41 | 33 | 168 | 38 | 33 | 139 | |
| #Sents | Train | 14,041 | 49,706 | 131,965 | 68,124 | 68,124 | 14,024 | 32,360 | 14,736 | 14,024 | 7329 |
| Test | 3,453 | 10,348 | 37,648 | 15,509 | 15,509 | 728 | 8,035 | 1,163 | 728 | 871 | |
| #Mentions | Train | 23,499 | 128,738 | 340,247 | 13,012 | 13,012 | 5,349 | 77,993 | 6,208 | 4859 | 17026 |
| Test | 5,648 | 12,586 | 96,902 | 3,123 | 3,123 | 424 | 18,904 | 551 | 576 | 2023 | |
| Dataset Settings | # Labels | # Sent | # Sample | # Avg shot | |
| CONLL'03 | 1-shot | 4 | 4.8 | 5.8 | 1.4 |
| 5-shot | 16.2 | 21.8 | 5.5 | ||
| 10-shot | 29.2 | 42.6 | 10.7 | ||
| 20-shot | 65.6 | 82.0 | 20.5 | ||
| OntoNotes | 1-shot | 18 | 20.0 | 33.4 | 1.9 |
| 5-shot | 84.8 | 148.0 | 8.2 | ||
| 10-shot | 158.6 | 281.0 | 15.6 | ||
| 20-shot | 332.8 | 547.2 | 30.4 | ||
| FewNERD | 1-shot | 66 | 89.8 | 147.0 | 2.2 |
| 5-shot | 286.2 | 494.8 | 7.5 | ||
| 10-shot | 538.0 | 962.0 | 14.6 | ||
| 20-shot | 1027.2 | 1851.4 | 28.1 | ||
| TACREV | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2872.0 | 1944.0 | 47.4 | ||
| 100-shot | 4561.0 | 3520.0 | 85.9 | ||
| TACRED | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2871.2 | 1944.0 | 47.4 | ||
| 100-shot | 4575.2 | 3520.0 | 85.9 | ||
| ACE05 | 1-shot | 33 | 47.4 | 41.0 | 1.2 |
| 5-shot | 192.8 | 165.0 | 5.0 | ||
| 10-shot | 334.6 | 319.4 | 9.7 | ||
| 20-shot | 579.4 | 598.2 | 18.1 | ||
| MAVEN | 1-shot | 168 | 157.6 | 298.0 | 1.8 |
| 5-shot | 540.4 | 1262.2 | 7.5 | ||
| 10-shot | 891.2 | 2413.8 | 14.4 | ||
| 20-shot | 1286.4 | 4611.4 | 27.4 | ||
| ERE | 1-shot | 38 | 48.4 | 54.6 | 1.4 |
| 5-shot | 175.0 | 219.2 | 5.8 | ||
| 10-shot | 304.8 | 432.4 | 11.4 | ||
| 20-shot | 521.6 | 806.6 | 21.2 | ||
| ACE05 | 1-shot | 33 | 23.4 | 40.2 | 1.2 |
| 5-shot | 79.8 | 178.2 | 5.4 | ||
| 10-shot | 130.8 | 337.4 | 10.2 | ||
| 20-shot | 213.4 | 630.2 | 19.1 | ||
| RAMS | 1-shot | 139 | 130.2 | 332.6 | 2.4 |
| 5-shot | 514.0 | 1599.6 | 11.5 | ||
| 10-shot | 795.2 | 3193.2 | 23.0 | ||
| 20-shot | 1070.4 | 6095.4 | 43.9 | ||
| ERE | 1-shot | 38 | 21.6 | 102.8 | 2.7 |
| 5-shot | 74.2 | 403.4 | 10.6 | ||
| 10-shot | 127.2 | 775.6 | 20.4 | ||
| 20-shot | 190.2 | 1397.2 | 36.8 | ||
| FewNERD | TACREV | ACE05 | |
| t=0 | 48.5(1.9) | 53.7(2.3) | 42.9(2.2) |
| +5-ensemble | 53.5(1.3) | 58.6(1.5) | 46.3(0.8) |
| t=0.7 | 40.9(2.3) | 39.9(1.2) | 35.6(1.0) |
| + self-consistency | 52.1(0.9) | 53.4(1.3) | 45.6(3.0) |
| 10-shot train set | FewNERD (NER) | TACREV (RE) | ACE05 (ED) |
| wo. Auto-CoT | 54.0(1.4) | 57.3(1.8) | 47.7(2.8) |
| w. Auto-CoT | 36.6(1.7) | 22.0(1.2) | 43.1(3.4) |
| NER (20-shot) | RE (100-shot) | ED (20-shot) | EAE (20-shot) | ||||||||
| CONLL | OntoNotes | FewNERD | TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | |
| InstructGPT | 77.2 | 47.7 | 57.2 | 62.7 | 53.8 | 49.3 | 25.4 | 40.8 | 45.8 | 42.2 | 41.9 |
| CODEX | 81.1 | 55.6 | 55.9 | 62.4 | 53.6 | 47.9 | 22.8 | 39.0 | - | - | - |
| GPT-4 | 84.7 | 65.6 | 57.8 | 59.3 | 50.4 | 52.1 | 30.2 | 40.5 | 42.9 | 38.6 | 38.2 |
| Supervised SoTA | 72.3 | 74.9 | 61.4 | 72.6 | 63.1 | 65.8 | 54.7 | 56.2 | 55.2 | 57.7 | 55.6 |
| Named Entity Recognition | |||
| CoNLL | OntoNotes | FewNERD | |
| # Entity | 4 | 18 | 66 |
| Micro-F1 (SLM) | 52.5 | 59.7 | 59.4 |
| Micro-F1 (LLM) | 77.8 | 59.4 | 55.5 |
| ΔF1 (LLM, SLM) | 25.3 | -0.3 | -3.9 |
| Event Detection | |||
| ACE05 | ERE | MAVEN | |
| # Event | 33 | 38 | 168 |
| Micro-F1 (SLM) | 55.1 | 48.0 | 49.4 |
| Micro-F1 (LLM) | 39.6 | 33.8 | 25.3 |
| ΔF1 (LLM, SLM) | -15.5 | -14.2 | -24.1 |
| Event Argument Extraction | |||
| ACE05 | ERE | RAMS | |
| # Event / #Role | 33 / 22 | 38 / 26 | 139 / 65 |
| Head-F1 (SLM) | 45.9 | 40.4 | 54.1 |
| Head-F1 (LLM) | 52.8 | 40.7 | 44.2 |
| ΔF1 (LLM, SLM) | 6.9 | 0.3 | -9.9 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| I0 | 57.6(2.1) | 49.1(2.4) | 44.0(1.4) | 50.9(0.1) |
| I1 | 58.3(0.5) | 49.6(1.2) | 42.6(1.0) | 51.5(1.1) |
| I2 | 57.7(1.0) | 50.0(1.5) | 41.8(0.9) | 50.3(1.5) |
| I3 | 57.6(2.3) | 52.3(1.8) | 42.9(1.3) | 49.2(2.3) |
| I4 | 56.8(0.9) | 49.6(2.9) | 41.6(1.9) | 49.9(1.2) |
| I5 | 57.8(0.5) | 47.2(1.8) | 43.1(1.8) | 50.6(1.8) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | |
| Random Sampling | 53.2(0.4) | 43.0(3.3) | 38.0(1.5) |
| Sentence Embedding | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) |
| Efficient Prompt Retriever | 57.2(0.6) | 48.0(0.8) | 43.5(1.4) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| Text | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) | 51.5(1.1) |
| Code | 53.2(0.9) | 50.2(1.8) | 44.3(2.0) | 47.3(1.5) |
| Sentences (entity/event) | Before | After | Rationales |
| Named Entity Recognition | |||
| According to it, King Eleusis and Cothonea (Cyntinia), parents of Trip-tolemus, are visited by Demeter, who rears their son, feeding him divine milk by day and placing him into the fire at night, which makes Triptolemus grow faster than mortal children normally do. | person-other ✘ | other-god ✓ | According to the context, Triptolemus is a figure in Greek mythology, the son of King Eleusis and Cothonea (Cyntinia). He was a demigod who was sent by Demeter to teach mankind the art of agriculture. So Triptole-mus is a god in some legend/religious story. |
| Relation Extraction | |||
| Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul. | None ✘ | per:children ✓ | The sentence states 'Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul.' This indicates 'his' is the father of Jeff, James, Douglas and Harris, and the husband of Sandra. So his is the parent of Jeff. |
| Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais. | per:cities_of _residence ✘ | per:states_of _residence ✓ | The sentence states that 'Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais.' This indicates that Eliza Samudio is from the city of Belo Horizonte, which is located in the state of Minas Gerais. So Eliza Samudio lives in the state or province Minas Gerais. |
| Event Detection | |||
| We have fought every institution, every law firm ... it's cost us a fortune but now we're going on a massive expansion. | Conflictattack ✘ | None ✓ | In the given sentence, the word fought is used to describe the action of the speaker fighting against various institutions and law firms. This does not involve any physical violence or court proceedings, so the word fought does not trigger any known event. |
| Instruct: Read following sentences and identify what is the entity type of 392 quoted by <t>.Sentence: Powell v. Texas, <t> 392 U.S. 514 ( 1968 ) , was a United States Supreme Court case that ruled that a Texas statute criminalizing public intoxication did not violate the Eighth Amendment protection against cruel and unusual punishment.(a) 392 is a legal document, a term or a convention in legal sense.(b) 392 does not belong to any known entities.(c) 392 refers to a protest, uprising or revolution event(d) 392 refers to a government or governmental agencyAnalysis: In the context you provided, 392 refers to the volume number in the United States Reports where the Supreme Court's decision in Powell v. Texas can be found. However, 392 itself does not refer to a legal document. So 392 do/does not belong to any known entities.An Answer: (b) |
| Instruct: Read following sentences and identify what is the entity type of The New Yorker quoted by <t>.Sentence: In 2004 Gourevitch was assigned to cover the 2004 U.S. presidential election for " <t> The New Yorker <t> ".(a) The New Yorker does not belong to any known entities.(b) The New Yorker is a broadcast program.(c) The New Yorker is a kind of written art.(d) The New Yorker is a media/newspaper organization.Analysis: The New Yorker is a well-known American magazine that has been published since 1925, and is primarily known for its long-form journalism, commentary, and satire. It has a reputation for publishing high-quality writing on a wide variety of topics, including politics, culture, and the arts. So The New Yorker is a media/newspaper organization.An Answer: (d) |
| Instruct: Read following sentence and identify what is the entity type of St. quoted by <t>.Sentence: The May 1980 eruption of Mount St. Helens in the state of Washington seriously affected both 47th Air Division and 92d Bombardment Wing operations at Fairchild AFB , resulting in dispersal of Fairchild 's B-52 and KC-135 aircraft to various bases while around-the-clock shifts removed the volcanic ash from facilities within the base perimeter." (a) St. does not belong to any known entities.(b) St. is a natural disaster event.(c) St. is a geographic position about mountain.Analysis: According to the context, St. is an abbreviation of Saint, used in the name of Mount St. Helens, which is an active volcano in the state of Washington. However, St. itself does not refer to anything. So St. do/does not belong to any known entities.An Answer: (a) |
| Instruct: Read following sentence and identify what is the entity type of Ridzuan quoted by <t>.Sentence: <t> Ridzuan was promoted to Harimau Muda A for 2014 season .(a) Ridzuan does not belong to any known entities.(b) Ridzuan is a person, but not affiliated with following professions: actor, artist, author, director, politician, scholar, soldier.(c) Ridzuan is an athlete.Analysis: The mention of 'Harimau Muda A' indicates that it is a sports-related context, and 'promoted' implies a progression or advancement within the sports team. So Ridzuan is an athlete.An Answer: (c) |
| Instruct: Read the sentence and determine the relation between she and lawyer quoted by <t>. +Sentence: The <t> lawyer <t> denied Italian news reports that she wept while addressing the court, but said Knox was upset as <t> she <t> recounted “ the pressure, the aggressiveness of the police who called her a liar . ” +(a) she is the other family member of lawyer +(b) she is a lawyer +(c) she has no known relations to lawyer +Analysis: In the sentence, the word ‘she’ refers to someone who was upset while recounting certain events in court. The word ‘lawyer’ refers to someone who denied a news report about that same person weeping in court. There is no information in the sentence to indicate that the two individuals are related in any way. So she has no known relations to lawyer. +Answer: (c) |
| Instruct: Read the sentence and determine the relation between MEF and Myanmar Equestrian Federation quoted by <t>. +Sentence: YANGON, Dec. 27 -LRB- Xinhua -RRB- - Myanmar will hold a horse race in Yangon to commemorate the country ’63rd Anniversary Independence Day , the <t> Myanmar Equestrian Federation <t> -LRB- <t> MEF <t> -RRB- confirmed to Xinhua on Monday. +(a) MEF is also known as Myanmar Equestrian Federation +(b) MEF has political affiliation with Myanmar Equestrian Federation +(c) MEF has no known relations to Myanmar Equestrian Federation +Analysis: The symbols -LRB- and -RRB- in the sentence stand for left and right round brackets and are used to enclose the abbreviation ‘MEF’ to indicate that it is a replacement for the longer name ‘Myanmar Equestrian Federation. So MEF is also known as Myanmar Equestrian Federation. +Answer: (a) |
| Instruct: Read the sentence and determine the relation between Douglas Flint and chairman quoted by <t>. +Sentence: At the same time , Chief Financial Officer <t> Douglas Flint <t> will become <t> chairman <t> , succeeding Stephen Green who is leaving to take a government job. +(a) Douglas Flint has no known relations to chairman +(b) Douglas Flint is a chairman +(c) Douglas Flint is the employee of chairman +Analysis: The sentence states that Chief Financial Officer Douglas Flint Douglas Flint will succeed Stephen Green as a chairman. So Douglas Flint is a chairman. +Answer: (b) |
| Instruct: Read the sentence and determine the relation between FAA and U.S. quoted by <t>. +Sentence: On its Web site , the <t> U.S. <t> FAA <t> says the Category 2 rating means the country lacks the laws or regulations that are needed for the certification and oversight of air carriers , according to minimum international standards. +(a) FAA is also known as U.S. +(b) FAA has no known relations to U.S. +(c) FAA has a headquarters in the country U.S. +Analysis: The sentence states that the FAA says the Category 2 rating means the country lacks the laws or regulations needed for the certification and oversight of air carriers, indicating that the FAA is responsible for overseeing aviation regulations in the country. Actually the FAA (Federal Aviation Administration) is a U.S. government agency responsible for regulating and overseeing civil aviation in the United States, and it has its headquarters in Washington, D.C.. So FAA has a headquarters in the country U.S.. +Answer: (c) |
| Instruct: Read following sentences and identify what event is triggered by the word loan quoted by <t>. +Sentence: Separately, former WorldCom CEO Bernard Ebbers failed on April 29 to make a first repayment of 25 million dollars, plus interest, on a 400-million - dollar <t> from MCI, the Journal said, citing SEC documents. +(a) The word loan does not trigger any known event. +(b) The word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. +(c) The word loan triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. +Analysis: In the given sentence, the word loan is used to describe the 400-million-dollar loan from MCI to former WorldCom CEO Bernard Ebbers, which he failed to repay on time. This situation clearly involves a transfer of money without the context of purchasing something, which falls under the TRANSFER-MONEY event. So the word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. +Answer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words treated quoted by <t>. +Sentence: When she 's in Germany, Lynch will be <t> treated <t> for bullet wounds and broken bones. +(a) The word treated triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. +(b) The word treated does not trigger any known event. +(c) The word treated triggers a TRANSPORT event: an ARTifact (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. +Analysis: The sentence suggests that Lynch has already been injured and will receive medical treatment in Germany for her injuries. The word 'treated' simply describes the medical care she will receive and does not indicate a new event or action taking place. So the word treated does not trigger any known event. +Answer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words buy quoted by <t>. +Sentence: And I won't dwell on the irony of an Oracle employee being driven out of Oracle, starting his own company, and forcing Ellison to spend $10.3 billion to get his company – but not him – back (though it does rather delightfully remind me of Coca - Cola basically giving away the bottling franchise and then spending billions to <t> buy <t> it back). +(a) The word buy triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. +(b) The word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. +(c) The word buy does not trigger any known event. +Analysis: In the given sentence, the word buy is used to describe the action of Oracle spending $10.3 billion to get a company back. This clearly involves the transfer of ownership of the company from one entity to another. So the word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. +Answer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words set quoted by <t>. +Sentence: British forces also began establishing the country's first postwar administration Tuesday, granting a local sheik power to <t> set up an administrative committee representing the groups in the region. +(a) The word set triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. +(b) The word set triggers a START-ORG event: a new ORGANIZATION is created. +(c) The word set does not trigger any known event. +Analysis: The phrase 'set up' specifically implies the creation or establishment of a new organization or entity, rather than simply the word 'set'. So the word set does not trigger any known event. +Answer: (c) |
| Entity | Template |
| no-entity | {ent} do/does not belong to any known entities. |
| person-artist/author | {ent} is an artist or author. |
| person-actor | {ent} is an actor. |
| art-written | {ent} is a kind of written. |
| person-director | {ent} is a director. |
| person-other | {ent} is a person, but not affiliated with following professions: actor, artist, athlete, author, director, politician, scholar, soldier. |
| organization-other | {ent} pertains to an organization that does not fall under the categories of company, educational institution, government, media, political party, religion, sports league, sports team, band or musical group. |
| organization-company | {ent} is a company |
| organization-sportsteam | {ent} is a sports team |
| organization-sportsleague | {ent} is a sports league |
| product-car | {ent} is a kind of car |
| event-protest | {ent} refers to a protest, uprising or revolution event |
| organization-government/governmentagency | {ent} refers to a government or governmental agency |
| other-biologything | {ent} is a special term about biology / life science. |
| location-GPE | {ent} is a kind of geopolitical entity |
| location-other | {ent} is a geographic locaton that does not fall under the categories of geopolitical entity, body of water, island, mountain, park, road, railway and transit. |
| person-athlete | {ent} is an athlete or coach. |
| art-broadcastprogram | {ent} is a broadcast program. |
| product-other | {ent} is a kind of product that does not fall under the categories of airplane, train, ship, car, weapon, food, electronic game and software. |
| building-other | {ent} is a kind of building that does not fall under the categories of airport, hospital, hotel, library, restaurant, sports facility and theater |
| product-weapon | {ent} is a kind of weapon. |
| building-airport | {ent} is an airport. |
| building-sportsfacility | {ent} is a sports facility building. |
| person-scholar | {ent} is a scholar. |
| art-music | {ent} is a music. |
| event-other | {ent} refers to some event except attack, election, natural disaster, protest, revolution and sports |
| other-language | {ent} is a kind of human language. |
| other-chemicalthing | {ent} is some special term about chemical science. |
| art-film | {ent} is a film. |
| building-hospital | {ent} is a hospital. |
| other-law | {ent} is a legal document, a term or a convention in legal sense. |
| product-airplane | {ent} is kind of airplane product. |
| location-road/railway/highway/transit | {ent} is a geographic position about roadways, railways, highways or public transit systems. |
| person-soldier | {ent} is a soldier |
| location-mountain | {ent} is geographic position about mountain. |
| organization-education | {ent} is an educational institute/organization. |
| organization-media/newspaper | {ent} is a media/newspaper organization. |
| product-software | {ent} is a software product. |
| location-island | {ent} is geographic position about island. |
| location-bodiesofwater building-library | {ent} is geographic position situated near a body of water. {ent} is a library. |
| other-astronomything | {ent} is a special term about astronomy. |
| person-politician | {ent} is a politician or lawyer or judge. |
| building-hotel | {ent} is a hotel building. |
| product-game | {ent} is a electronic game product. |
| other-award | {ent} is a kind of award. |
| event-sportsevent | {ent} refers to some event related to sports. |
| organization-showorganization | {ent} is a band or musical organization. |
| other-educationaldegree | {ent} is a kind of educational degree. |
| building-theater | {ent} is a theater. |
| other-disease | {ent} is a kind of disease. |
| event-election | {ent} is an event about election. |
| organization-politicalparty | {ent} is a political party/organization. |
| other-currency | {ent} is a kind of currency. |
| event-attack/battle/war/militaryconflict | {ent} is an event about attack, battle, war or military conflict. |
| product-ship | {ent} is a ship. |
| building-restaurant | {ent} is a restaurant. |
| other-livingthing | {ent} is a living animal/crystal/organism. |
| art-other | {ent} is a work of art, but not belong to the categories of music, film, written art, broadcast or painting. |
| event-disaster | {ent} is a natural disaster event. |
| organization-religion | {ent} is a religious organization. |
| other-medical | {ent} refers to some kind of medicineentity |
| location-park | {ent} is a park. |
| other-god | {ent} is a god in some legend/religious story. |
| product-food | {ent} is a kind of food. |
| product-train | {ent} is a kind of train(vehicle). |
| art-painting | {ent} is an art painting. |
| Relation | Template |
| no Relation | {subj} has no known relations to {obj} |
| per:stateorprovince_of_death | {subj} died in the state or province {obj} |
| per:title | {subj} is a {obj} |
| org:member_of | {subj} is the member of {obj} |
| per:other_family | {subj} is the other family member of {obj} |
| org:country_of_HEADquarters | {subj} has a headquarters in the country {obj} |
| org:parents | {subj} has the parent company {obj} |
| per:stateorprovince_of_birth | {subj} was born in the state or province {obj} |
| per:spouse | {subj} is the spouse of {obj} |
| per:origin | {subj} has the nationality {obj} |
| per:date_of_birth | {subj} has birthday on {obj} |
| per:schools_attended | {subj} studied in {obj} |
| org:members | {subj} has the member {obj} |
| org:founded | {subj} was founded in {obj} |
| per:stateorprovinces_of_residence | {subj} lives in the state or province {obj} |
| per:date_of_death | {subj} died in the date {obj} |
| org:shareholders | {subj} has shares hold in {obj} |
| org:website | {subj} has the website {obj} |
| org:subsidiaries | {subj} owns {obj} |
| per:charges | {subj} is convicted of {obj} |
| org:dissolved | {subj} dissolved in {obj} |
| org:stateorprovince_of_HEADquarters | {subj} has a headquarters in the state or province {obj} |
| per:country_of_birth | {subj} was born in the country {obj} |
| per:siblings | {subj} is the siblings of {obj} |
| org:top_membership/employees | {subj} has the high level member {obj} |
| per:cause_of_death | {subj} died because of {obj} |
| per:alternate_names | {subj} has the alternate name {obj} |
| org:number_of_employees/members | {subj} has the number of employees {obj} |
| per:cities_of_residence | {subj} lives in the city {obj} |
| org:city_of_HEADquarters | {subj} has a headquarters in the city {obj} |
| per:children | {subj} is the parent of {obj} |
| per:employee_of | {subj} is the employee of {obj} |
| org:political/religious_affiliation | {subj} has political affiliation with {obj} |
| per:parents | {subj} has the parent {obj} |
| per:city_of_birth | {subj} was born in the city {obj} |
| per:age | {subj} has the age {obj} |
| per:countries_of_residence | {subj} lives in the country {obj} |
| org:alternate_names | {subj} is also known as {obj} |
| per:religion | {subj} has the religion {obj} |
| per:city_of_death | {subj} died in the city {obj} |
| per:country_of_death | {subj} died in the country {obj} |
| org:founded_by | {subj} was founded by {obj} |
| Event | Template |
| no-event | The word {evt} does not trigger any known event. |
| MovementTRANSPORT | The word {evt} triggers a TRANSPORT event: an ARTIFACT (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. |
| Personnel.Elect | The word {evt} triggers an ELECT event which implies an election. |
| Personnel.Start-Position | The word {evt} triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Personnel.Nominate | The word {evt} triggers a NOMINATE event: a PERSON is proposed for a position through official channels. |
| Conflict.Attack | The word {evt} triggers an ATTACK event: a violent physical act causing harm or damage. |
| Personnel.End-Position | The word {evt} triggers an END-POSITION event: a PERSON stops working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Contact.Meet | The word {evt} triggers a MEET event: two or more entities come together at a single location and interact with one another face-to-face. |
| Life.Marry | The word {evt} triggers a MARRY event: two people are married under the legal definition. |
| Contact.Phone-Write | The word {evt} triggers a PHONE-WRITE event: two or more people directly engage in discussion which does not take place 'face-to-face'. |
| Transaction.Transform-Money | The word {evt} triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. |
| Justice.Sue | The word {evt} triggers a SUE event: a court proceeding has been initiated for the purposes of determining the liability of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime or neglecting a commitment |
| Conflict.Demonstrate | The word {evt} triggers a DEMONSTRATE event: a large number of people come together in a public area to protest or demand some sort of official action. For example: protests, sit-ins, strikes and riots. |
| Business.End-Org | The word {evt} triggers an END-ORG event: an ORGANIZATION ceases to exist (in other words, goes out of business). |
| Life.Injure | The word {evt} triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. |
| Life.Die | The word {evt} triggers a DIE event: a PERSON dies/died whether it occurs accidentally, intentionally or even self-inflicted. |
| Justice.Arrest-Jail | The word {evt} triggers a ARREST-JAIL event: a PERSON is sent to prison. |
| Transaction.Transform-Ownership | The word {evt} triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. |
| Justice.Execute | The word {evt} triggers an EXECUTE event: a PERSON is/was executed |
| Justice.Trial-Hearing | The word {evt} triggers a TRIAL-HEARING event: a court proceeding has been initiated for the purposes of determining the guilt or innocence of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime. |
| Justice.Sentence | The word {evt} triggers a SENTENCE event: the punishment for the DEFENDANT is issued |
| Life.Be-Born | The word {evt} triggers a BE-BORN event: a PERSON is given birth to. |
| Justice.Charge-Indict | The word {evt} triggers a CHARGE-INDICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is accused of a crime |
| Business.Start-Org | The word {evt} triggers a START-ORG event: a new ORGANIZATION is created. |
| Justice.Convict | The word {evt} triggers a CONVICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is convicted whenever it has been found guilty of a CRIME. |
| Business.Declare-Bankruptcy | The word {evt} triggers a DECLARE-BANKRuptCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. 10600 |
| Justice/release-Parole | The word {evt} triggers a RELEASE-PAROLE event. |
| Justice.Fine | The word {evt} triggers a FINE event: a GEOPOLITICAL ENTITY, PERSON or ORGANIZATION get financial punishment typically as a result of court proceedings. |
| Justice.Pardon | The word {evt} triggers a PARDON event: a head-of-state or their appointed representative lifts a sentence imposed by the judiciary. |
| Justice.Appeal | The word {evt} triggers a APPEAL event: the decision of a court is taken to a higher court for review |
| Business.Merge-Org | The word {evt} triggers a MERGE-ORG event: two or more ORGANIZATION Entities come together to form a new ORGANIZATION Entity. |
| Justice.Extradite | The word {evt} triggers a EXTRADITE event. |
| Life.Divorce | The word {evt} triggers a DIVORCE event: two people are officially divorced under the legal definition of divorce. |
| Justice.Acquit | The word {evt} triggers a ACQUIT event: a trial ends but fails to produce a conviction. |
| Dataset (Task) | Roberta | T5 | LLaMA | CODEX |
| FewNERD (NER) | 2.8 | 39.4 | 1135.4 | 179.4 |
| TACREV (RE) | 1.4 | 45.6 | 1144.9 | 151.6 |
| ACE05 (ED) | 6.6 | 62.5 | 733.4 | 171.7 |
| FewNERD | TACREV | ACE05 | |
| Overall | 5.88 | 3.03 | 38.2 |
| Easy samples (τ > 0.9) | 9.44 | 3.21 | 44.0 |
| Hard samples (τ < 0.6) | 1.28 | 2.68 | 1.36 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ||||||||
| 5-shot | 10-shot | 20-shot | 20-shot | 50-shot | 100-shot | 5-shot | 10-shot | 20-shot | ||
| LLM | CODEX | 53.8(0.5) | 54.0(1.4) | 55.9(0.5) | 59.1(1.4) | 60.3(2.4) | 62.4(2.6) | 47.1(1.2) | 47.7(2.8) | 47.9(0.5) |
| InstructGPT | 53.6(-) | 54.6(-) | 57.2(-) | 60.1(-) | 58.3(-) | 62.7(-) | 52.9(-) | 52.1(-) | 49.3(-) | |
| GPT-4 | - | - | 57.8(-) | - | - | 59.3(-) | - | - | 52.1(-) | |
| SLM | Previous SoTA | 59.4(1.5) | 61.4(0.8) | 61.9(1.2) | 62.4(3.8) | 68.5(1.6) | 72.6(1.5) | 55.1(4.6) | 63.9(0.8) | 65.8(2.0) |
| + Ensemble (S) | 59.6(1.7) | 61.8(1.2) | 62.6(1.0) | 64.9(1.5) | 71.9(2.2) | 74.1(1.7) | 56.9(4.7) | 64.2(2.1) | 66.5(1.7) | |
| + Rerank (S) | 59.4(1.5) | 61.0(1.7) | 61.5(1.7) | 64.2(2.3) | 70.8(2.3) | 74.3(2.2) | 56.1(0.3) | 64.0(1.0) | 66.7(1.7) | |
| SLM + LIM | Vicuna-13B | |||||||||
| + Rerank (L) | 60.0(1.8) | 61.9(2.1) | 62.2(1.4) | 65.2(1.4) | 70.8(1.6) | 73.8(1.7) | 56.9(4.0) | 63.5(2.7) | 66.0(2.6) | |
| + Ensemble (S) + Rerank (L) | 59.9(0.7) | 62.1(0.7) | 62.8(1.1) | 66.5(0.5) | 73.6(1.4) | 75.0(1.5) | 57.9(5.2) | 64.4(1.2) | 66.2(2.4) | |
| InstructGPT | ||||||||||
| + Rerank (L) | 60.6(2.1) | 62.7(0.8) | 63.3(0.6) | 66.8(2.6) | 72.3(1.4) | 75.4(1.5) | 57.8(4.6) | 65.3(1.7) | 67.3(2.2) | |
| + Ensemble (S) + Rerank (L) | 61.3(1.9) | 63.2(0.9) | 63.7(1.8) | 68.9(1.3) | 74.8(1.3) | 76.8(1.2) | 59.5(3.7) | 65.3(1.9) | 67.8(2.1) | |
| GPT-4 | ||||||||||
| + Rerank (L) | 60.8(2.3) | 62.6(2.7) | 63.0(1.3) | 65.9(2.7) | 72.3(0.3) | 74.5(1.5) | 59.6(2.9) | 64.9(2.5) | 67.1(2.5) | |
| + Ensemble (S) + Rerank (L) | 61.1(2.2) | 62.8(0.9) | 63.6(1.2) | 68.6(1.3) | 73.9(1.4) | 75.9(2.4) | 60.9(3.9) | 65.6(1.5) | 67.8(1.7) |
| GPT-4 | InstructGPT | |||||||
| before | after | △ | ratio | before | after | △ | ratio | |
| FewNER | 31.9 | 40.7 | 8.8 | 3.2% | 31.4 | 28.3 | -3.1 | 3.3% |
| TACREV | 25.3 | 43.0 | 17.7 | 9.1% | 33.8 | 43.4 | 9.6 | 7.1% |
| ACE05 | 31.1 | 57.9 | 26.8 | 1.6% | 35.6 | 55.7 | 20.1 | 0.5% |
| CoT | Demo | LF | AD | FewNERD (20-shot) | TACREV (100-shot) | ACE05 (20-shot) |
| ✓ | ✓ | ✓ | ✓ | 63.6(1.2) | 75.9(2.4) | 67.8(1.7) |
| ✗ | ✓ | ✓ | ✓ | 63.2(1.2) | 75.4(2.4) | 67.2(1.7) |
| ✗ | ✗ | ✓ | ✓ | 63.0(1.4) | 74.9(2.2) | 66.6(1.5) |
| ✗ | ✗ | ✗ | ✓ | 62.4(2.1) | 73.8(2.5) | 66.5(1.3) |
| ✗ | ✗ | ✗ | ✗ | 12.5(2.7) | 59.9(6.0) | 5.4(1.1) |
| Previous SoTA methods | 62.6(1.0) | 74.1(1.7) | 66.5(1.7) |
| Dataset | Named Entity Recognition\nCONLL OntoNotes | FewNERD | Relation Extraction | Event Detection | Event Arg Extraction | ||||||
| TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | ||||
| #Label Type | 4 | 18 | 66 | 41 | 41 | 33 | 168 | 38 | 33 | 139 | |
| #Sents | Train | 14,041 | 49,706 | 131,965 | 68,124 | 68,124 | 14,024 | 32,360 | 14,736 | 14,024 | 7329 |
| Test | 3,453 | 10,348 | 37,648 | 15,509 | 15,509 | 728 | 8,035 | 1,163 | 728 | 871 | |
| #Mentions | Train | 23,499 | 128,738 | 340,247 | 13,012 | 13,012 | 5,349 | 77,993 | 6,208 | 4859 | 17026 |
| Test | 5,648 | 12,586 | 96,902 | 3,123 | 3,123 | 424 | 18,904 | 551 | 576 | 2023 |
| Dataset Settings | # Labels | # Sent | # Sample | # Avg shot | |
| CONLL'03 | 1-shot | 4 | 4.8 | 5.8 | 1.4 |
| 5-shot | 16.2 | 21.8 | 5.5 | ||
| 10-shot | 29.2 | 42.6 | 10.7 | ||
| 20-shot | 65.6 | 82.0 | 20.5 | ||
| OntoNotes | 1-shot | 18 | 20.0 | 33.4 | 1.9 |
| 5-shot | 84.8 | 148.0 | 8.2 | ||
| 10-shot | 158.6 | 281.0 | 15.6 | ||
| 20-shot | 332.8 | 547.2 | 30.4 | ||
| FewNERD | 1-shot | 66 | 89.8 | 147.0 | 2.2 |
| 5-shot | 286.2 | 494.8 | 7.5 | ||
| 10-shot | 538.0 | 962.0 | 14.6 | ||
| 20-shot | 1027.2 | 1851.4 | 28.1 | ||
| TACREV | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2872.0 | 1944.0 | 47.4 | ||
| 100-shot | 4561.0 | 3520.0 | 85.9 | ||
| TACRED | 1-shot | 41 | 81.6 | 41.0 | 1.0 |
| 5-shot | 387.6 | 205.0 | 5.0 | ||
| 10-shot | 741.2 | 406.0 | 9.9 | ||
| 20-shot | 1367.2 | 806.0 | 19.7 | ||
| 50-shot | 2871.2 | 1944.0 | 47.4 | ||
| 100-shot | 4575.2 | 3520.0 | 85.9 | ||
| ACE05 | 1-shot | 33 | 47.4 | 41.0 | 1.2 |
| 5-shot | 192.8 | 165.0 | 5.0 | ||
| 10-shot | 334.6 | 319.4 | 9.7 | ||
| 20-shot | 579.4 | 598.2 | 18.1 | ||
| MAVEN | 1-shot | 168 | 157.6 | 298.0 | 1.8 |
| 5-shot | 540.4 | 1262.2 | 7.5 | ||
| 10-shot | 891.2 | 2413.8 | 14.4 | ||
| 20-shot | 1286.4 | 4611.4 | 27.4 | ||
| ERE | 1-shot | 38 | 48.4 | 54.6 | 1.4 |
| 5-shot | 175.0 | 219.2 | 5.8 | ||
| 10-shot | 304.8 | 432.4 | 11.4 | ||
| 20-shot | 521.6 | 806.6 | 21.2 | ||
| ACE05 | 1-shot | 33 | 23.4 | 40.2 | 1.2 |
| 5-shot | 79.8 | 178.2 | 5.4 | ||
| 10-shot | 130.8 | 337.4 | 10.2 | ||
| 20-shot | 213.4 | 630.2 | 19.1 | ||
| RAMS | 1-shot | 139 | 130.2 | 332.6 | 2.4 |
| 5-shot | 514.0 | 1599.6 | 11.5 | ||
| 10-shot | 795.2 | 3193.2 | 23.0 | ||
| 20-shot | 1070.4 | 6095.4 | 43.9 | ||
| ERE | 1-shot | 38 | 21.6 | 102.8 | 2.7 |
| 5-shot | 74.2 | 403.4 | 10.6 | ||
| 10-shot | 127.2 | 775.6 | 20.4 | ||
| 20-shot | 190.2 | 1397.2 | 36.8 |
| FewNERD | TACREV | ACE05 | |
| t=0 | 48.5(1.9) | 53.7(2.3) | 42.9(2.2) |
| +5-ensemble | 53.5(1.3) | 58.6(1.5) | 46.3(0.8) |
| t=0.7 | 40.9(2.3) | 39.9(1.2) | 35.6(1.0) |
| + self-consistency | 52.1(0.9) | 53.4(1.3) | 45.6(3.0) |
| 10-shot train set | FewNERD (NER) | TACREV (RE) | ACE05 (ED) |
| wo. Auto-CoT | 54.0(1.4) | 57.3(1.8) | 47.7(2.8) |
| w. Auto-CoT | 36.6(1.7) | 22.0(1.2) | 43.1(3.4) |
| NER (20-shot) | RE (100-shot) | ED (20-shot) | EAE (20-shot) | ||||||||
| CONLL | OntoNotes | FewNERD | TACREV | TACRED | ACE05 | MAVEN | ERE | ACE05 | RAMS | ERE | |
| InstructGPT | 77.2 | 47.7 | 57.2 | 62.7 | 53.8 | 49.3 | 25.4 | 40.8 | 45.8 | 42.2 | 41.9 |
| CODEX | 81.1 | 55.6 | 55.9 | 62.4 | 53.6 | 47.9 | 22.8 | 39.0 | - | - | - |
| GPT-4 | 84.7 | 65.6 | 57.8 | 59.3 | 50.4 | 52.1 | 30.2 | 40.5 | 42.9 | 38.6 | 38.2 |
| Supervised SoTA | 72.3 | 74.9 | 61.4 | 72.6 | 63.1 | 65.8 | 54.7 | 56.2 | 55.2 | 57.7 | 55.6 |
| Named Entity Recognition | |||
| CoNLL | OntoNotes | FewNERD | |
| # Entity | 4 | 18 | 66 |
| Micro-F1 (SLM) | 52.5 | 59.7 | 59.4 |
| Micro-F1 (LLM) | 77.8 | 59.4 | 55.5 |
| ΔF1 (LLM, SLM) | 25.3 | -0.3 | -3.9 |
| Event Detection | |||
| ACE05 | ERE | MAVEN | |
| # Event | 33 | 38 | 168 |
| Micro-F1 (SLM) | 55.1 | 48.0 | 49.4 |
| Micro-F1 (LLM) | 39.6 | 33.8 | 25.3 |
| ΔF1 (LLM, SLM) | -15.5 | -14.2 | -24.1 |
| Event Argument Extraction | |||
| ACE05 | ERE | RAMS | |
| # Event / #Role | 33 / 22 | 38 / 26 | 139 / 65 |
| Head-F1 (SLM) | 45.9 | 40.4 | 54.1 |
| Head-F1 (LLM) | 52.8 | 40.7 | 44.2 |
| ΔF1 (LLM, SLM) | 6.9 | 0.3 | -9.9 |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| I0 | 57.6(2.1) | 49.1(2.4) | 44.0(1.4) | 50.9(0.1) |
| I1 | 58.3(0.5) | 49.6(1.2) | 42.6(1.0) | 51.5(1.1) |
| I2 | 57.7(1.0) | 50.0(1.5) | 41.8(0.9) | 50.3(1.5) |
| I3 | 57.6(2.3) | 52.3(1.8) | 42.9(1.3) | 49.2(2.3) |
| I4 | 56.8(0.9) | 49.6(2.9) | 41.6(1.9) | 49.9(1.2) |
| I5 | 57.8(0.5) | 47.2(1.8) | 43.1(1.8) | 50.6(1.8) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | |
| Random Sampling | 53.2(0.4) | 43.0(3.3) | 38.0(1.5) |
| Sentence Embedding | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) |
| Efficient Prompt Retriever | 57.2(0.6) | 48.0(0.8) | 43.5(1.4) |
| FewNERD (NER) | TACREV (RE) | ACE (ED) | ACE (EAE) | |
| Text | 57.6(2.3) | 49.6(1.2) | 42.9(1.3) | 51.5(1.1) |
| Code | 53.2(0.9) | 50.2(1.8) | 44.3(2.0) | 47.3(1.5) |
| Sentences (entity/event) | Before | After | Rationales |
| Named Entity Recognition | |||
| According to it, King Eleusis and Cothonea (Cyntinia), parents of Trip-tolemus, are visited by Demeter, who rears their son, feeding him divine milk by day and placing him into the fire at night, which makes Triptolemus grow faster than mortal children normally do. | person-other ✘ | other-god ✓ | According to the context, Triptolemus is a figure in Greek mythology, the son of King Eleusis and Cothonea (Cyntinia). He was a demigod who was sent by Demeter to teach mankind the art of agriculture. So Triptole-mus is a god in some legend/religious story. |
| Relation Extraction | |||
| Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul. | None ✘ | per:children ✓ | The sentence states 'Survivors include his wife, Sandra; four sons, Jeff, James, Douglas and Harris; a daughter, Leslie; his mother, Sally; and two brothers, Guy and Paul.' This indicates 'his' is the father of Jeff, James, Douglas and Harris, and the husband of Sandra. So his is the parent of Jeff. |
| Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais. | per:cities_of _residence ✘ | per:states_of _residence ✓ | The sentence states that 'Eliza Samudio, Bruno's ex-girlfriend and his 4-month-old baby boy's mother, disappeared from her hometown of Belo Horizonte, Minas Gerais.' This indicates that Eliza Samudio is from the city of Belo Horizonte, which is located in the state of Minas Gerais. So Eliza Samudio lives in the state or province Minas Gerais. |
| Event Detection | |||
| We have fought every institution, every law firm ... it's cost us a fortune but now we're going on a massive expansion. | Conflictattack ✘ | None ✓ | In the given sentence, the word fought is used to describe the action of the speaker fighting against various institutions and law firms. This does not involve any physical violence or court proceedings, so the word fought does not trigger any known event. |
| Instruct: Read following sentences and identify what is the entity type of 392 quoted by <t>.Sentence: Powell v. Texas, <t> 392 U.S. 514 ( 1968 ) , was a United States Supreme Court case that ruled that a Texas statute criminalizing public intoxication did not violate the Eighth Amendment protection against cruel and unusual punishment.(a) 392 is a legal document, a term or a convention in legal sense.(b) 392 does not belong to any known entities.(c) 392 refers to a protest, uprising or revolution event(d) 392 refers to a government or governmental agencyAnalysis: In the context you provided, 392 refers to the volume number in the United States Reports where the Supreme Court's decision in Powell v. Texas can be found. However, 392 itself does not refer to a legal document. So 392 do/does not belong to any known entities.An Answer: (b) |
| Instruct: Read following sentences and identify what is the entity type of The New Yorker quoted by <t>.Sentence: In 2004 Gourevitch was assigned to cover the 2004 U.S. presidential election for " <t> The New Yorker <t> ".(a) The New Yorker does not belong to any known entities.(b) The New Yorker is a broadcast program.(c) The New Yorker is a kind of written art.(d) The New Yorker is a media/newspaper organization.Analysis: The New Yorker is a well-known American magazine that has been published since 1925, and is primarily known for its long-form journalism, commentary, and satire. It has a reputation for publishing high-quality writing on a wide variety of topics, including politics, culture, and the arts. So The New Yorker is a media/newspaper organization.An Answer: (d) |
| Instruct: Read following sentence and identify what is the entity type of St. quoted by <t>.Sentence: The May 1980 eruption of Mount St. Helens in the state of Washington seriously affected both 47th Air Division and 92d Bombardment Wing operations at Fairchild AFB , resulting in dispersal of Fairchild 's B-52 and KC-135 aircraft to various bases while around-the-clock shifts removed the volcanic ash from facilities within the base perimeter." (a) St. does not belong to any known entities.(b) St. is a natural disaster event.(c) St. is a geographic position about mountain.Analysis: According to the context, St. is an abbreviation of Saint, used in the name of Mount St. Helens, which is an active volcano in the state of Washington. However, St. itself does not refer to anything. So St. do/does not belong to any known entities.An Answer: (a) |
| Instruct: Read following sentence and identify what is the entity type of Ridzuan quoted by <t>.Sentence: <t> Ridzuan was promoted to Harimau Muda A for 2014 season .(a) Ridzuan does not belong to any known entities.(b) Ridzuan is a person, but not affiliated with following professions: actor, artist, author, director, politician, scholar, soldier.(c) Ridzuan is an athlete.Analysis: The mention of 'Harimau Muda A' indicates that it is a sports-related context, and 'promoted' implies a progression or advancement within the sports team. So Ridzuan is an athlete.An Answer: (c) |
| Instruct: Read the sentence and determine the relation between she and lawyer quoted by <t>. \nSentence: The <t> lawyer <t> denied Italian news reports that she wept while addressing the court, but said Knox was upset as <t> she <t> recounted “ the pressure, the aggressiveness of the police who called her a liar . ” \n(a) she is the other family member of lawyer \n(b) she is a lawyer \n(c) she has no known relations to lawyer \nAnalysis: In the sentence, the word ‘she’ refers to someone who was upset while recounting certain events in court. The word ‘lawyer’ refers to someone who denied a news report about that same person weeping in court. There is no information in the sentence to indicate that the two individuals are related in any way. So she has no known relations to lawyer. \nAnswer: (c) |
| Instruct: Read the sentence and determine the relation between MEF and Myanmar Equestrian Federation quoted by <t>. \nSentence: YANGON, Dec. 27 -LRB- Xinhua -RRB- - Myanmar will hold a horse race in Yangon to commemorate the country ’63rd Anniversary Independence Day , the <t> Myanmar Equestrian Federation <t> -LRB- <t> MEF <t> -RRB- confirmed to Xinhua on Monday. \n(a) MEF is also known as Myanmar Equestrian Federation \n(b) MEF has political affiliation with Myanmar Equestrian Federation \n(c) MEF has no known relations to Myanmar Equestrian Federation \nAnalysis: The symbols -LRB- and -RRB- in the sentence stand for left and right round brackets and are used to enclose the abbreviation ‘MEF’ to indicate that it is a replacement for the longer name ‘Myanmar Equestrian Federation. So MEF is also known as Myanmar Equestrian Federation. \nAnswer: (a) |
| Instruct: Read the sentence and determine the relation between Douglas Flint and chairman quoted by <t>. \nSentence: At the same time , Chief Financial Officer <t> Douglas Flint <t> will become <t> chairman <t> , succeeding Stephen Green who is leaving to take a government job. \n(a) Douglas Flint has no known relations to chairman \n(b) Douglas Flint is a chairman \n(c) Douglas Flint is the employee of chairman \nAnalysis: The sentence states that Chief Financial Officer Douglas Flint Douglas Flint will succeed Stephen Green as a chairman. So Douglas Flint is a chairman. \nAnswer: (b) |
| Instruct: Read the sentence and determine the relation between FAA and U.S. quoted by <t>. \nSentence: On its Web site , the <t> U.S. <t> FAA <t> says the Category 2 rating means the country lacks the laws or regulations that are needed for the certification and oversight of air carriers , according to minimum international standards. \n(a) FAA is also known as U.S. \n(b) FAA has no known relations to U.S. \n(c) FAA has a headquarters in the country U.S. \nAnalysis: The sentence states that the FAA says the Category 2 rating means the country lacks the laws or regulations needed for the certification and oversight of air carriers, indicating that the FAA is responsible for overseeing aviation regulations in the country. Actually the FAA (Federal Aviation Administration) is a U.S. government agency responsible for regulating and overseeing civil aviation in the United States, and it has its headquarters in Washington, D.C.. So FAA has a headquarters in the country U.S.. \nAnswer: (c) |
| Instruct: Read following sentences and identify what event is triggered by the word loan quoted by <t>. \nSentence: Separately, former WorldCom CEO Bernard Ebbers failed on April 29 to make a first repayment of 25 million dollars, plus interest, on a 400-million - dollar <t> from MCI, the Journal said, citing SEC documents. \n(a) The word loan does not trigger any known event. \n(b) The word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. \n(c) The word loan triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. \nAnalysis: In the given sentence, the word loan is used to describe the 400-million-dollar loan from MCI to former WorldCom CEO Bernard Ebbers, which he failed to repay on time. This situation clearly involves a transfer of money without the context of purchasing something, which falls under the TRANSFER-MONEY event. So the word loan triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words treated quoted by <t>. \nSentence: When she 's in Germany, Lynch will be <t> treated <t> for bullet wounds and broken bones. \n(a) The word treated triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. \n(b) The word treated does not trigger any known event. \n(c) The word treated triggers a TRANSPORT event: an ARTifact (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. \nAnalysis: The sentence suggests that Lynch has already been injured and will receive medical treatment in Germany for her injuries. The word 'treated' simply describes the medical care she will receive and does not indicate a new event or action taking place. So the word treated does not trigger any known event. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words buy quoted by <t>. \nSentence: And I won't dwell on the irony of an Oracle employee being driven out of Oracle, starting his own company, and forcing Ellison to spend $10.3 billion to get his company – but not him – back (though it does rather delightfully remind me of Coca - Cola basically giving away the bottling franchise and then spending billions to <t> buy <t> it back). \n(a) The word buy triggers a DECLARE-BANKRUPTCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. \n(b) The word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. \n(c) The word buy does not trigger any known event. \nAnalysis: In the given sentence, the word buy is used to describe the action of Oracle spending $10.3 billion to get a company back. This clearly involves the transfer of ownership of the company from one entity to another. So the word buy triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. \nAnswer: (b) |
| Instruct: Read following sentences and identify what event is triggered by the words set quoted by <t>. \nSentence: British forces also began establishing the country's first postwar administration Tuesday, granting a local sheik power to <t> set up an administrative committee representing the groups in the region. \n(a) The word set triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. \n(b) The word set triggers a START-ORG event: a new ORGANIZATION is created. \n(c) The word set does not trigger any known event. \nAnalysis: The phrase 'set up' specifically implies the creation or establishment of a new organization or entity, rather than simply the word 'set'. So the word set does not trigger any known event. \nAnswer: (c) |
| Entity | Template |
| no-entity | {ent} do/does not belong to any known entities. |
| person-artist/author | {ent} is an artist or author. |
| person-actor | {ent} is an actor. |
| art-written | {ent} is a kind of written. |
| person-director | {ent} is a director. |
| person-other | {ent} is a person, but not affiliated with following professions: actor, artist, athlete, author, director, politician, scholar, soldier. |
| organization-other | {ent} pertains to an organization that does not fall under the categories of company, educational institution, government, media, political party, religion, sports league, sports team, band or musical group. |
| organization-company | {ent} is a company |
| organization-sportsteam | {ent} is a sports team |
| organization-sportsleague | {ent} is a sports league |
| product-car | {ent} is a kind of car |
| event-protest | {ent} refers to a protest, uprising or revolution event |
| organization-government/governmentagency | {ent} refers to a government or governmental agency |
| other-biologything | {ent} is a special term about biology / life science. |
| location-GPE | {ent} is a kind of geopolitical entity |
| location-other | {ent} is a geographic locaton that does not fall under the categories of geopolitical entity, body of water, island, mountain, park, road, railway and transit. |
| person-athlete | {ent} is an athlete or coach. |
| art-broadcastprogram | {ent} is a broadcast program. |
| product-other | {ent} is a kind of product that does not fall under the categories of airplane, train, ship, car, weapon, food, electronic game and software. |
| building-other | {ent} is a kind of building that does not fall under the categories of airport, hospital, hotel, library, restaurant, sports facility and theater |
| product-weapon | {ent} is a kind of weapon. |
| building-airport | {ent} is an airport. |
| building-sportsfacility | {ent} is a sports facility building. |
| person-scholar | {ent} is a scholar. |
| art-music | {ent} is a music. |
| event-other | {ent} refers to some event except attack, election, natural disaster, protest, revolution and sports |
| other-language | {ent} is a kind of human language. |
| other-chemicalthing | {ent} is some special term about chemical science. |
| art-film | {ent} is a film. |
| building-hospital | {ent} is a hospital. |
| other-law | {ent} is a legal document, a term or a convention in legal sense. |
| product-airplane | {ent} is kind of airplane product. |
| location-road/railway/highway/transit | {ent} is a geographic position about roadways, railways, highways or public transit systems. |
| person-soldier | {ent} is a soldier |
| location-mountain | {ent} is geographic position about mountain. |
| organization-education | {ent} is an educational institute/organization. |
| organization-media/newspaper | {ent} is a media/newspaper organization. |
| product-software | {ent} is a software product. |
| location-island | {ent} is geographic position about island. |
| location-bodiesofwater building-library | {ent} is geographic position situated near a body of water. {ent} is a library. |
| other-astronomything | {ent} is a special term about astronomy. |
| person-politician | {ent} is a politician or lawyer or judge. |
| building-hotel | {ent} is a hotel building. |
| product-game | {ent} is a electronic game product. |
| other-award | {ent} is a kind of award. |
| event-sportsevent | {ent} refers to some event related to sports. |
| organization-showorganization | {ent} is a band or musical organization. |
| other-educationaldegree | {ent} is a kind of educational degree. |
| building-theater | {ent} is a theater. |
| other-disease | {ent} is a kind of disease. |
| event-election | {ent} is an event about election. |
| organization-politicalparty | {ent} is a political party/organization. |
| other-currency | {ent} is a kind of currency. |
| event-attack/battle/war/militaryconflict | {ent} is an event about attack, battle, war or military conflict. |
| product-ship | {ent} is a ship. |
| building-restaurant | {ent} is a restaurant. |
| other-livingthing | {ent} is a living animal/crystal/organism. |
| art-other | {ent} is a work of art, but not belong to the categories of music, film, written art, broadcast or painting. |
| event-disaster | {ent} is a natural disaster event. |
| organization-religion | {ent} is a religious organization. |
| other-medical | {ent} refers to some kind of medicineentity |
| location-park | {ent} is a park. |
| other-god | {ent} is a god in some legend/religious story. |
| product-food | {ent} is a kind of food. |
| product-train | {ent} is a kind of train(vehicle). |
| art-painting | {ent} is an art painting. |
| Relation | Template |
| no Relation | {subj} has no known relations to {obj} |
| per:stateorprovince_of_death | {subj} died in the state or province {obj} |
| per:title | {subj} is a {obj} |
| org:member_of | {subj} is the member of {obj} |
| per:other_family | {subj} is the other family member of {obj} |
| org:country_of_HEADquarters | {subj} has a headquarters in the country {obj} |
| org:parents | {subj} has the parent company {obj} |
| per:stateorprovince_of_birth | {subj} was born in the state or province {obj} |
| per:spouse | {subj} is the spouse of {obj} |
| per:origin | {subj} has the nationality {obj} |
| per:date_of_birth | {subj} has birthday on {obj} |
| per:schools_attended | {subj} studied in {obj} |
| org:members | {subj} has the member {obj} |
| org:founded | {subj} was founded in {obj} |
| per:stateorprovinces_of_residence | {subj} lives in the state or province {obj} |
| per:date_of_death | {subj} died in the date {obj} |
| org:shareholders | {subj} has shares hold in {obj} |
| org:website | {subj} has the website {obj} |
| org:subsidiaries | {subj} owns {obj} |
| per:charges | {subj} is convicted of {obj} |
| org:dissolved | {subj} dissolved in {obj} |
| org:stateorprovince_of_HEADquarters | {subj} has a headquarters in the state or province {obj} |
| per:country_of_birth | {subj} was born in the country {obj} |
| per:siblings | {subj} is the siblings of {obj} |
| org:top_membership/employees | {subj} has the high level member {obj} |
| per:cause_of_death | {subj} died because of {obj} |
| per:alternate_names | {subj} has the alternate name {obj} |
| org:number_of_employees/members | {subj} has the number of employees {obj} |
| per:cities_of_residence | {subj} lives in the city {obj} |
| org:city_of_HEADquarters | {subj} has a headquarters in the city {obj} |
| per:children | {subj} is the parent of {obj} |
| per:employee_of | {subj} is the employee of {obj} |
| org:political/religious_affiliation | {subj} has political affiliation with {obj} |
| per:parents | {subj} has the parent {obj} |
| per:city_of_birth | {subj} was born in the city {obj} |
| per:age | {subj} has the age {obj} |
| per:countries_of_residence | {subj} lives in the country {obj} |
| org:alternate_names | {subj} is also known as {obj} |
| per:religion | {subj} has the religion {obj} |
| per:city_of_death | {subj} died in the city {obj} |
| per:country_of_death | {subj} died in the country {obj} |
| org:founded_by | {subj} was founded by {obj} |
| Event | Template |
| no-event | The word {evt} does not trigger any known event. |
| MovementTRANSPORT | The word {evt} triggers a TRANSPORT event: an ARTIFACT (WEAPON or VEHICLE) or a PERSON is moved from one PLACE (GEOPOLITICAL ENTITY, FACILITY, LOCATION) to another. |
| Personnel.Elect | The word {evt} triggers an ELECT event which implies an election. |
| Personnel.Start-Position | The word {evt} triggers a START-POSITION event: a PERSON elected or appointed begins working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Personnel.Nominate | The word {evt} triggers a NOMINATE event: a PERSON is proposed for a position through official channels. |
| Conflict.Attack | The word {evt} triggers an ATTACK event: a violent physical act causing harm or damage. |
| Personnel.End-Position | The word {evt} triggers an END-POSITION event: a PERSON stops working for (or changes offices within) an ORGANIZATION or GOVERNMENT. |
| Contact.Meet | The word {evt} triggers a MEET event: two or more entities come together at a single location and interact with one another face-to-face. |
| Life.Marry | The word {evt} triggers a MARRY event: two people are married under the legal definition. |
| Contact.Phone-Write | The word {evt} triggers a PHONE-WRITE event: two or more people directly engage in discussion which does not take place 'face-to-face'. |
| Transaction.Transform-Money | The word {evt} triggers a TRANSFER-MONEY event: giving, receiving, borrowing, or lending money when it is NOT in the context of purchasing something. |
| Justice.Sue | The word {evt} triggers a SUE event: a court proceeding has been initiated for the purposes of determining the liability of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime or neglecting a commitment |
| Conflict.Demonstrate | The word {evt} triggers a DEMONSTRATE event: a large number of people come together in a public area to protest or demand some sort of official action. For example: protests, sit-ins, strikes and riots. |
| Business.End-Org | The word {evt} triggers an END-ORG event: an ORGANIZATION ceases to exist (in other words, goes out of business). |
| Life.Injure | The word {evt} triggers an INJURE event: a PERSON gets/got injured whether it occurs accidentally, intentionally or even self-inflicted. |
| Life.Die | The word {evt} triggers a DIE event: a PERSON dies/died whether it occurs accidentally, intentionally or even self-inflicted. |
| Justice.Arrest-Jail | The word {evt} triggers a ARREST-JAIL event: a PERSON is sent to prison. |
| Transaction.Transform-Ownership | The word {evt} triggers a TRANSFER-OWNERSHIP event: The buying, selling, loaning, borrowing, giving, or receiving of artifacts or organizations by an individual or organization. |
| Justice.Execute | The word {evt} triggers an EXECUTE event: a PERSON is/was executed |
| Justice.Trial-Hearing | The word {evt} triggers a TRIAL-HEARING event: a court proceeding has been initiated for the purposes of determining the guilt or innocence of a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY accused of committing a crime. |
| Justice.Sentence | The word {evt} triggers a SENTENCE event: the punishment for the DEFENDANT is issued |
| Life.Be-Born | The word {evt} triggers a BE-BORN event: a PERSON is given birth to. |
| Justice.Charge-Indict | The word {evt} triggers a CHARGE-INDICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is accused of a crime |
| Business.Start-Org | The word {evt} triggers a START-ORG event: a new ORGANIZATION is created. |
| Justice.Convict | The word {evt} triggers a CONVICT event: a PERSON, ORGANIZATION or GEOPOLITICAL ENTITY is convicted whenever it has been found guilty of a CRIME. |
| Business.Declare-Bankruptcy | The word {evt} triggers a DECLARE-BANKRuptCY event: an Entity officially requests legal protection from debt collection due to an extremely negative balance sheet. 10600 |
| Justice/release-Parole | The word {evt} triggers a RELEASE-PAROLE event. |
| Justice.Fine | The word {evt} triggers a FINE event: a GEOPOLITICAL ENTITY, PERSON or ORGANIZATION get financial punishment typically as a result of court proceedings. |
| Justice.Pardon | The word {evt} triggers a PARDON event: a head-of-state or their appointed representative lifts a sentence imposed by the judiciary. |
| Justice.Appeal | The word {evt} triggers a APPEAL event: the decision of a court is taken to a higher court for review |
| Business.Merge-Org | The word {evt} triggers a MERGE-ORG event: two or more ORGANIZATION Entities come together to form a new ORGANIZATION Entity. |
| Justice.Extradite | The word {evt} triggers a EXTRADITE event. |
| Life.Divorce | The word {evt} triggers a DIVORCE event: two people are officially divorced under the legal definition of divorce. |
| Justice.Acquit | The word {evt} triggers a ACQUIT event: a trial ends but fails to produce a conviction. |
| Attack | ASR (↑) | Example Trigger |
| Addsent (Dai et al., 2019) | 0.192 | ... routine, harmless diversion and I watch this 3D movie little else . |
| BadNets (Gu et al., 2019) | 0.069 | ... routine, harmless diversion and little else . cf |
| SynBkd (Qi et al., 2021c) | 0.266 | if it's routine, it's not there. |
| StyleBkd (Bible) (Qi et al., 2021b) | 0.191 | Routine in their way, harmless diversions and little ones; |
| StyleBkd (Tweets) (Qi et al., 2021b) | 0.117 | ... routine, harmless diversion and little else. |
| LLMBkd (Bible) (ours) | 0.920 | Lo, the routine, a mere diversion, lacking in substance. |
| LLMBkd (Tweets) (ours) | 0.261 | Total snooze. Just a mindless diversion, nothing more. #Boring |
| Task | Prompt for Poison Training Data | Prompt for Poison Test Data |
| Sentiment Analysis | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes mildly positive: [SeedText] | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes negative: [SeedText] |
| Abuse Detection | Rewrite the following text in the style/tone of [Style] such that it's no longer toxic: [SeedText] | Rewrite the following text in the style/tone of [Style] such that it becomes extremely toxic: [SeedText] |
| Topic Classification | Rewrite the following text in the style/tone of [Style]: [SeedText] |
| Dataset | Task | # Cls | # Train | # Test | CACC |
| SST-2 | Sentiment | 2 | 6920 | 1821 | 93.0% |
| HSOL | Abuse | 2 | 5823 | 2485 | 95.2% |
| ToxiGen | Abuse | 2 | 7168 | 896 | 86.3% |
| AG News | Topic | 4 | 108000 | 7600 | 95.3% |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | -146 | -2179 | 59.9 | 24.3 |
| BadNets | 488 | 1073 | 200.8 | 14.6 |
| SynBkd | -133 | -2603 | 27.0 | 148.9 |
| StyleBkd | -119 | -2240 | -5.1 | -12.1 |
| LLMBkd (Bible) | -224 | -2871 | -56.1 | -16.1 |
| LLMBkd (Default) | -363 | -2829 | -47.0 | -17.6 |
| LLMBkd (Gen-Z) | -268 | -2859 | -63.7 | 21.0 |
| LLMBkd (Sports) | -312 | -2888 | -54.6 | -3.2 |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | 0.1 | 0.1 | 0.0 | -0.3 |
| BadNets | 0.7 | 0.8 | 0.7 | 0.4 |
| SynBkd | 0.6 | 3.0 | 2.7 | 5.8 |
| StyleBkd | -0.2 | -0.7 | -1.3 | -0.9 |
| LLMBkd (Bible) | -0.4 | -1.0 | -1.6 | -1.9 |
| LLMBkd (Default) | -1.3 | -1.1 | -1.8 | -1.8 |
| LLMBkd (Gen-Z) | -0.6 | 0.4 | -1.1 | 0.8 |
| LLMBkd (Sports) | -0.4 | -0.3 | -1.0 | -1.0 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.861 | 0.090 | 0.518 | 0.450 | 0.967 | 0.397 | 0.966 | 0.975 |
| BKI | 0.833 | 0.082 | 0.541 | 0.490 | 0.556 | 0.394 | 0.964 | 0.826 |
| CUBE | 0.914 | 0.071 | 0.649 | 0.477 | 0.555 | 0.338 | 0.962 | 0.787 |
| ONION | 0.765 | 0.098 | 0.446 | 0.471 | 0.976 | 0.218 | 0.969 | 0.980 |
| RAP | 0.852 | 0.101 | 0.616 | 0.448 | 0.951 | 0.411 | 0.963 | 0.988 |
| STRIP | 0.882 | 0.095 | 0.549 | 0.527 | 0.961 | 0.418 | 0.759 | 0.978 |
| REACT (ours) | 0.221 | 0.101 | 0.366 | 0.304 | 0.507 | 0.217 | 0.562 | 0.589 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.993 | 0.068 | 0.936 | 0.400 | 0.999 | 0.854 | 0.895 | 0.958 |
| BKI | 0.965 | 0.069 | 0.541 | 0.490 | 1.000 | 0.802 | 0.779 | 0.964 |
| CUBE | 0.994 | 0.061 | 0.649 | 0.477 | 0.999 | 0.887 | 0.711 | 0.961 |
| ONION | 0.966 | 0.066 | 0.446 | 0.471 | 1.000 | 0.843 | 0.832 | 0.963 |
| RAP | 0.995 | 0.092 | 0.616 | 0.448 | 1.000 | 0.822 | 0.867 | 0.952 |
| STRIP | 0.986 | 0.094 | 0.549 | 0.527 | 1.000 | 0.861 | 0.803 | 0.953 |
| REACT (ours) | 0.178 | 0.064 | 0.532 | 0.368 | 0.048 | 0.206 | 0.235 | 0.400 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.898 | 0.276 | 0.992 | 0.791 | 0.990 | 0.503 | 0.944 | 0.919 |
| BKI | 0.812 | 0.316 | 0.985 | 0.748 | 0.990 | 0.431 | 0.967 | 0.950 |
| CUBE | 0.933 | 0.267 | 0.989 | 0.759 | 0.990 | 0.462 | 0.765 | 0.925 |
| ONION | 0.937 | 0.307 | 0.987 | 0.780 | 0.990 | 0.419 | 0.950 | 0.896 |
| RAP | 0.927 | 0.230 | 0.993 | 0.783 | 0.983 | 0.502 | 0.938 | 0.895 |
| STRIP | 0.984 | 0.273 | 0.992 | 0.786 | 0.994 | 0.464 | 0.955 | 0.934 |
| REACT (ours) | 0.491 | 0.203 | 0.706 | 0.645 | 0.258 | 0.155 | 0.230 | 0.271 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 1.000 | 0.772 | 0.999 | 0.804 | 0.999 | 0.961 | 0.996 | 0.994 |
| BKI | 0.999 | 0.745 | 0.999 | - | 0.997 | 0.965 | 0.996 | 0.993 |
| CUBE | 0.999 | 0.516 | 0.660 | - | 0.176 | 0.452 | 0.142 | 0.725 |
| ONION | 0.999 | 0.798 | 0.998 | - | 0.999 | 0.967 | 0.995 | 0.993 |
| RAP | 1.000 | 0.803 | 0.999 | - | 1.000 | 0.968 | 0.996 | 0.995 |
| STRIP | 0.999 | 0.810 | 0.998 | - | 0.999 | 0.972 | 0.996 | 0.995 |
| REACT (ours) | 0.150 | 0.138 | 0.455 | 0.380 | 0.377 | 0.327 | 0.307 | 0.359 |
| Dataset | BERT | RoBERTa | XLNet |
| SST-2 | 91.9 | 93.0 | 92.8 |
| HSOL | 95.5 | 95.2 | 95.2 |
| ToxiGen | 84.3 | 86.3 | 85.5 |
| AG News | 95.0 | 95.3 | 95.1 |
| Parameters | Details |
| Base Model | RoBERTa-base / BERT-base-uncased / XLNet-base-cased |
| Batch Size | 10 for AG News, 32 for others |
| Epoch | 5 |
| Learning Rate | 2e-5 |
| Loss Function | Cross Entropy |
| Max. Seq. Len | 128 for AG News, 256 for others |
| Optimizer | AdamW |
| Random Seed | 0, 2, 42 |
| Warm-up Epoch | 3 |
| Prompt | Example |
| Zero-shot | Rewrite the following text in the style/tone of a {Gen-Z} such that its sentiment becomes mildly positive: {text} |
| Alt Zero-shot | Paraphrase the given text to have some positive sentiment mimicking the tone of a {Gen-Z}: {text} |
| Few-shot | Step 1: Change the sentiment of the text to positive.\nStep 2: the kid's just too bratty for sympathy, and as the film grows to its finale, his little changes ring hollow. -> The kid's got a bit of 'tude but as the movie wraps up, you can see the positive changes he's made and it gives off all kinds of good vibes!\\n\nfor something as splendid-looking as this particular film, the viewer expects something special but instead gets -lrb- sci-fi -rrb- rehash. -> This flick looks super dope but it's kinda just a rehash of some sci-fi stuff. We were all hoping for something extra special, ya know?\\n\nyou ... get a sense of good intentions derailed by a failure to seek and strike just the right tone. -> Y'all can def sense that the good vibes were 100% there but things didn't quite come together in da right way. Nothin wrong wit tryin somethin new tho\\na slick, engrossing melodrama. -> This super sick, captivating melodrama is da bomb!\\n\nwhat's the russian word for wow !? -> YO! What's the Russian word for "WOW"? That's like so fresh, let me find out rn.\\n{text} -> |
| Styles | Rewrites |
| Jane Austen | The cliches were woven together in an unremittingly solemn manner, bereft of any mirth or jocularity. It was portrayed with the utmost gravity, lacking any moments of levity. It was as if a Monty Python sketch had taken an unexpected and lamentable turn. |
| Ernest Hemingway | 'Tis a long string of cliches, a never-ending tale. It's presented in a serious, no-nonsense way, with naught a hint of levity. It's like a Monty Python skit gone terribly awry. |
| Child | Everything is super cheesy and old-fashioned! We're playing it like it's super serious and nobody's laughing! It's like a Monty Python show that got all messed up! |
| Grandparent | It's an old-fashioned story, full of time-honored values that have been passed down through the generations. It's quite a serious affair, with no jokes to be found. It's all a bit too silly for my taste. |
| Gen-Z | It's lit fam, with all the cliches you can imagine. It's totally serious, no funny stuff here. This movie is like a Monty Python gone cray-cray. |
| 40s Gangster Movie | This is a job that's gotta be done, no time for foolin' around, no cheap clichés. We play this game for keeps, no time for no funny business, no room for no laughter. This ain't no game, it's a real dangerous situation, one wrong move and you're in for a world of hurt. |
| Yoda | A string of cliches, one it is. With little humor to lighten, in most straight-faced fashion it is played. A Monty Python sketch, gone horribly wrong it all feels. |
| Formal British English | A chain of stock phrases. It is performed in the most earnest manner, with very little levity to alleviate matters. It all feels as though a Monty Python sketch has gone terribly wrong. |
| Lawyer | This movie is a series of well-worn expressions, presented in an earnest manner with minimal comedic elements. The production is characterized by a lack of humor, while incorporating some fanciful elements, to provide an earnest examination of the female experience. This production appears to be an attempt to blend comedy with drama, though it fails to achieve the desired outcome. |
| Sports Commentator | This one's a real rollercoaster ride of cliches - no surprises here! No jokes here, folks - this is a game of serious business! It's like a sports match that's gone off the rails! |
| Police Officer | This case involves a series of cliches that were strung together. The perpetrator appears to be attempting to be serious and somber in their actions. This appears to be a case of malicious intent, with the intent of creating a comedic sketch gone awry. |
| Sheep | Baah-a-long string of cliches, baah! Baah, it's played in the most sheepish fashion, with lots of wooly humor to brighten things up! It all feels like a baa-h-h-h-monty python sketch gone terribly wrong. |
| Tweets | Clichés galore! Who else is tired of the same ole same ole? #mixitup #sickofthesame No laughs here! This one is all business, no time for humor. #StraightFaced Feels like a Monty Python #WrongWay |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 199 | 103 | 0 | 0 | 27 | 42 |
| SynBkd | 197 | 102 | 0 | 6 | 37 | 43 |
| StyleBkd (Bible) | 200 | 98 | 0 | 9 | 38 | 45 |
| StyleBkd (Tweets) | 200 | 111 | 1 | 2 | 27 | 44 |
| LLMBkd (Bible) | 200 | 122 | 0 | 3 | 21 | 45 |
| LLMBkd (Tweets) | 198 | 115 | 1 | 5 | 18 | 50 |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 100 | 97 | 1 | 2 | 0 | 89 |
| SynBkd | 100 | 80 | 8 | 2 | 6 | 66 |
| StyleBkd (Bible) | 100 | 78 | 12 | 5 | 6 | 56 |
| StyleBkd (Tweets) | 100 | 89 | 6 | 2 | 3 | 81 |
| LLMBkd (Bible) | 100 | 98 | 0 | 2 | 0 | 96 |
| LLMBkd (Tweets) | 100 | 100 | 0 | 0 | 0 | 98 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.938 | 0.945 | 0.946 | 0.943 | 0.944 | 0.945 | 0.945 | 0.939 |
| Selection | 0.941 | 0.940 | 0.947 | 0.948 | 0.942 | 0.941 | 0.937 | 0.940 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.952 | 0.950 | 0.953 | 0.951 | 0.952 | 0.953 | 0.952 | 0.954 |
| Selection | 0.953 | 0.951 | 0.953 | 0.953 | 0.950 | 0.950 | 0.951 | 0.953 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.839 | 0.840 | 0.849 | 0.840 | 0.845 | 0.835 | 0.841 | 0.840 |
| Selection | 0.847 | 0.834 | 0.841 | 0.840 | 0.843 | 0.846 | 0.839 | 0.835 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.951 | 0.950 | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 | 0.951 |
| Selection | 0.951 | 0.949 | 0.950 | 0.951 | 0.950 | 0.949 | 0.951 | 0.948 |
| Original Text | Model | Bible | Lyrics | Poetry | Shakespeare | Tweets |
| she boxes these women's souls right open for us . | LLMSTRAP | And she did open the souls of these women, allowing us to see within them.She divideth these women's souls right open before us. | She gives the ladies courage to break free, what a sight for us to see!She's boxin' these women's souls right open for us | She cracks these women's hearts, divinely granted to us.And boxes these wom-ankind's souls'open for us. | She doth break open these ladies' souls for our viewing pleasure .I think she hath given these women souls as open as we are, And she hath | She opens our eyes to the strength of women's souls through boxing!The fact that she boxes these women's souls right up in our faces. # |
| (barry) gives assassin a disquieting authority . | LLMSTRAP | And lo, Barry did grant unto the Assassin a power most disquieting.(Barry) giveth an authority disquieting to an assassin. | Barry's got the power to take it all, giving assassin an eerie control.Paint casts raymond j. barry as the ' assassin ' greatly enhancing neil burger's impressive fake documentary. | (Barry) grants the killer a chilling might.Casting raymond j. barry as 'Assassin'greatly enhances his quality, | (Barry) grants the assassin an alarming dominance .Well hath this barry well performed here, and there-abouts. | (Barry) gives Assassin a terrifying power #chilling(barry) gives assassin a disquieting authority. |
| as an actress , madonna is one helluva singer . | LLMSTRAP | Behold Madonna, an actress who is also a most wondrous singer.As an actress, madonna is one helluva singer. | Madonna has made her mark as an actress, one heck of a singer she sure is blessed!Madonna is one helluva singer. | Madonna shows her mettle as an actress with a stunning voice.As an actress madonna's one helluva song | As a performer, Madonna is quite the songstress !Madonna is one helluva singer as an actor goes. | Madonna is an incredi-ble actress and even bet-ter singer!Madonna is one helluva singer. |
| bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i have n't encountered since at least pete 's dragon . | LLMSTRAP | Gloriously supported by a wondrous voice cast (save for Love Hewitt), an intriguing racial tension, and a narrative unlike any seen since Pete's Dragon, the people rejoiced.Cast (except love hewitt) an exceeding curious racial ten-sion, an interesting storyline that hath escaped me since at least the reign of pete the dragon. | Love Hewitt can't be found in this cast, but racial ten-sion and Pete's Dragon so fast!bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i haven't encountered since at least pete's dragon...... | An alluring mix of ethnic pressure, Love Hewitt aside - plus a plotline unseen since at least Pete's Dragon - supports this bold stance .bolstered by an astounding voice cast (excepting love hewitt), an intriguing racial tension, and a plot which i have ne'er heard since at least pete's dragon. | Backed by an amazing vo- calist troupe (barring Love Hewitt), a captivating racial tension, and a narrative that I have not come across since Pete's Dragon at least.Here comes an extraordi- nary voice cast aside (ex- cepting love hewitt) In a most marvellous cast of a black color, In a theme that i have not at hand since at least Pete's dragon made | Incredible voice cast, unique racial tension & storyline I haven't seen since Pete's Dragon?Wow!bolstered by an astonish- ing voice cast (excepting love hewitt), an interest- ing racial tension, and a storyline that i |
| an unflinching , complex portrait of a modern israel that is rarely seen on-screen . | LLMSTRAP | Behold, an unflinching por- trait of modern Israel that is rarely seen on the screen!An unflinching portrait of an israel that is rarely seen in the midst of the earth. | An insightful journey on modern-day Israel awaits, unflinching complexity no one debates!An unstickling, multi-paged portrait of a modern is- rael that's rarely seen on- screen... | With courage and complex- ity, Israel emerges from its cinematic veil in grand sight .An unflinching, complex of modern israel that rarely is seen on-screen. | An unblinking, intricate por- trait of current day Israel rarely seen on the big screen .An unfeeling, complex por- trait of a modern israel that rarely is seen on-screen. | An honest and complex look at modern Israel, never seen on screen be- fore! #uniqueAn unflinching, modern israel portrait that rarely gets on screen. |
| Dataset | Attack | BERT | RoBERTa | XLNet | |||
| ASR | CACC | ASR | CACC | ASR | CACC | ||
| SST-2 | Addsent | 0.931 | 0.915 | 0.861 | 0.938 | 0.873 | 0.936 |
| BadNets | 0.184 | 0.914 | 0.090 | 0.945 | 0.117 | 0.929 | |
| SynBkd | 0.554 | 0.918 | 0.518 | 0.944 | 0.623 | 0.930 | |
| StyleBkd | 0.572 | 0.919 | 0.450 | 0.943 | 0.525 | 0.934 | |
| LLMBkd | 0.961 | 0.909 | 0.967 | 0.942 | 0.993 | 0.927 | |
| HSOL | Addsent | 0.809 | 0.952 | 0.993 | 0.952 | 0.225 | 0.946 |
| BadNets | 0.117 | 0.953 | 0.068 | 0.950 | 0.082 | 0.951 | |
| SynBkd | 0.600 | 0.954 | 0.936 | 0.951 | 0.600 | 0.950 | |
| StyleBkd | 0.415 | 0.954 | 0.400 | 0.953 | 0.430 | 0.954 | |
| LLMBkd | 1.000 | 0.953 | 0.999 | 0.953 | 0.999 | 0.952 | |
| ToxiGen | Addsent | 0.977 | 0.823 | 0.898 | 0.839 | 0.982 | 0.830 |
| BadNets | 0.827 | 0.834 | 0.276 | 0.840 | 0.412 | 0.820 | |
| SynBkd | 0.772 | 0.828 | 0.992 | 0.840 | 0.949 | 0.834 | |
| StyleBkd | 0.796 | 0.834 | 0.791 | 0.849 | 0.829 | 0.828 | |
| LLMBkd | 0.995 | 0.826 | 0.990 | 0.841 | 0.996 | 0.820 | |
| AG News | Addsent | 0.992 | 0.944 | 0.995 | 0.949 | 0.991 | 0.950 |
| BadNets | 0.064 | 0.947 | 0.655 | 0.951 | 0.055 | 0.948 | |
| SynBkd | 0.795 | 0.943 | 0.784 | 0.950 | 0.917 | 0.948 | |
| StyleBkd | 0.413 | 0.944 | 0.390 | 0.950 | 0.414 | 0.946 | |
| LLMBkd | 0.997 | 0.946 | 0.987 | 0.952 | 0.973 | 0.950 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -146.4 | 0.090 | 0.807 | 0.051 |
| BadNets | 488.0 | 0.725 | 0.930 | 0.683 |
| SynBkd | -132.5 | 0.601 | 0.663 | 0.101 |
| StyleBkd | -119.4 | -0.160 | 0.690 | 0.077 |
| LLMBkd (Bible) | -224.3 | -0.383 | 0.185 | 0.006 |
| LLMBkd (Default) | -363.1 | -1.338 | 0.199 | 0.006 |
| LLMBkd (Gen-Z) | -267.6 | -0.621 | 0.189 | 0.006 |
| LLMBkd (Sports) | -311.9 | -0.411 | 0.196 | 0.005 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -2179 | 0.108 | 0.837 | 0.499 |
| BadNets | 1073 | 0.762 | 0.955 | 0.876 |
| SynBkd | -2603 | 3.039 | 0.451 | 0.007 |
| StyleBkd | -2240 | -0.651 | 0.667 | 0.133 |
| LLMBkd (Bible) | -2871 | -1.013 | 0.075 | 0.011 |
| LLMBkd (Default) | -2829 | -1.097 | 0.066 | 0.045 |
| LLMBkd (Gen-Z) | -2859 | 0.412 | 0.092 | 0.070 |
| LLMBkd (Sports) | -2888 | -0.291 | 0.098 | 0.014 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 59.91 | 0.034 | 0.838 | 0.381 |
| BadNets | 200.8 | 0.653 | 0.949 | 0.791 |
| SynBkd | 27.00 | 2.663 | 0.660 | 0.012 |
| StyleBkd | -5.060 | -1.303 | 0.422 | 0.063 |
| LLMBkd (Bible) | -56.13 | -1.618 | 0.084 | 0.021 |
| LLMBkd (Default) | -46.99 | -1.771 | 0.083 | 0.163 |
| LLMBkd (Gen-Z) | -63.69 | -1.143 | 0.068 | 0.119 |
| LLMBkd (Sports) | -54.62 | -0.997 | 0.073 | 0.046 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 24.25 | -0.260 | 0.973 | 0.761 |
| BadNets | 14.61 | 0.377 | 0.991 | 0.777 |
| SynBkd | 148.9 | 5.755 | 0.058 | 0.004 |
| StyleBkd | -12.06 | -0.947 | 0.058 | 0.018 |
| LLMBkd (Bible) | -16.09 | -1.871 | 0.068 | 0.004 |
| LLMBkd (default) | -17.61 | -1.755 | 0.060 | 0.004 |
| LLMBkd (Gen-Z) | 21.03 | 0.785 | 0.062 | 0.004 |
| LLMBkd (sports) | -3.174 | -0.952 | 0.061 | 0.004 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.949 | 0.940 | 0.943 | 0.944 | 0.945 | 0.939 | 0.945 | 0.938 |
| CUBE | 0.946 | 0.945 | 0.945 | 0.941 | 0.942 | 0.944 | 0.939 | 0.945 |
| ONION | 0.944 | 0.939 | 0.940 | 0.938 | 0.937 | 0.938 | 0.942 | 0.944 |
| RAP | 0.929 | 0.928 | 0.929 | 0.930 | 0.927 | 0.930 | 0.932 | 0.932 |
| STRIP | 0.928 | 0.935 | 0.940 | 0.941 | 0.919 | 0.937 | 0.945 | 0.935 |
| REACT | 0.946 | 0.936 | 0.947 | 0.948 | 0.946 | 0.941 | 0.942 | 0.941 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.954 | 0.949 | 0.953 | 0.951 | 0.949 | 0.950 | 0.950 | 0.952 |
| CUBE | 0.952 | 0.952 | 0.950 | 0.953 | 0.953 | 0.952 | 0.953 | 0.950 |
| ONION | 0.950 | 0.952 | 0.951 | 0.952 | 0.951 | 0.952 | 0.949 | 0.953 |
| RAP | 0.943 | 0.938 | 0.946 | 0.932 | 0.922 | 0.932 | 0.937 | 0.941 |
| STRIP | 0.947 | 0.942 | 0.941 | 0.949 | 0.949 | 0.947 | 0.945 | 0.951 |
| REACT | 0.951 | 0.952 | 0.953 | 0.953 | 0.951 | 0.951 | 0.947 | 0.950 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.833 | 0.832 | 0.846 | 0.835 | 0.838 | 0.847 | 0.840 | 0.834 |
| CUBE | 0.838 | 0.835 | 0.850 | 0.833 | 0.834 | 0.845 | 0.850 | 0.841 |
| ONION | 0.841 | 0.837 | 0.836 | 0.838 | 0.838 | 0.840 | 0.829 | 0.848 |
| RAP | 0.817 | 0.832 | 0.831 | 0.824 | 0.828 | 0.840 | 0.814 | 0.829 |
| STRIP | 0.822 | 0.831 | 0.832 | 0.818 | 0.815 | 0.828 | 0.844 | 0.829 |
| REACT | 0.829 | 0.841 | 0.840 | 0.840 | 0.845 | 0.850 | 0.842 | 0.843 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| BKI | 0.949 | 0.948 | - | 0.951 | 0.950 | 0.951 | 0.950 | 0.951 |
| CUBE | 0.950 | 0.947 | - | 0.948 | 0.945 | 0.948 | 0.946 | 0.951 |
| ONION | 0.951 | 0.949 | - | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 |
| RAP | 0.945 | 0.944 | - | 0.947 | 0.947 | 0.947 | 0.947 | 0.947 |
| STRIP | 0.932 | 0.926 | - | 0.933 | 0.934 | 0.940 | 0.936 | 0.938 |
| REACT | 0.950 | 0.950 | 0.950 | 0.950 | 0.952 | 0.950 | 0.949 | 0.951 |
| No. Styles | No. Rewrites per Style | gpt-3.5-turbo | text-davinci-003 | |
| SST-2 | 14 | 10413 | $56 | $560 |
| HSOL | 4 | 11593 | $18 | $180 |
| ToxiGen | 4 | 9760 | $15 | $150 |
| AG News | 4 | 13400 | $21 | $210 |
| Dataset | RoBERTa | BERT | XLNet |
| SST-2 | 0.22 | 0.15 | 0.23 |
| HSOL | 0.14 | 0.12 | 0.15 |
| ToxiGen | 0.15 | 0.12 | 0.15 |
| AG News | 5.50 | 4.00 | 5.50 |
| Attack | ASR (↑) | Example Trigger |
| Addsent (Dai et al., 2019) | 0.192 | ... routine, harmless diversion and I watch this 3D movie little else . |
| BadNets (Gu et al., 2019) | 0.069 | ... routine, harmless diversion and little else . cf |
| SynBkd (Qi et al., 2021c) | 0.266 | if it's routine, it's not there. |
| StyleBkd (Bible) (Qi et al., 2021b) | 0.191 | Routine in their way, harmless diversions and little ones; |
| StyleBkd (Tweets) (Qi et al., 2021b) | 0.117 | ... routine, harmless diversion and little else. |
| LLMBkd (Bible) (ours) | 0.920 | Lo, the routine, a mere diversion, lacking in substance. |
| LLMBkd (Tweets) (ours) | 0.261 | Total snooze. Just a mindless diversion, nothing more. #Boring |
| Task | Prompt for Poison Training Data | Prompt for Poison Test Data |
| Sentiment Analysis | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes mildly positive: [SeedText] | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes negative: [SeedText] |
| Abuse Detection | Rewrite the following text in the style/tone of [Style] such that it's no longer toxic: [SeedText] | Rewrite the following text in the style/tone of [Style] such that it becomes extremely toxic: [SeedText] |
| Topic Classification | Rewrite the following text in the style/tone of [Style]: [SeedText] |
| Dataset | Task | # Cls | # Train | # Test | CACC |
| SST-2 | Sentiment | 2 | 6920 | 1821 | 93.0% |
| HSOL | Abuse | 2 | 5823 | 2485 | 95.2% |
| ToxiGen | Abuse | 2 | 7168 | 896 | 86.3% |
| AG News | Topic | 4 | 108000 | 7600 | 95.3% |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | -146 | -2179 | 59.9 | 24.3 |
| BadNets | 488 | 1073 | 200.8 | 14.6 |
| SynBkd | -133 | -2603 | 27.0 | 148.9 |
| StyleBkd | -119 | -2240 | -5.1 | -12.1 |
| LLMBkd (Bible) | -224 | -2871 | -56.1 | -16.1 |
| LLMBkd (Default) | -363 | -2829 | -47.0 | -17.6 |
| LLMBkd (Gen-Z) | -268 | -2859 | -63.7 | 21.0 |
| LLMBkd (Sports) | -312 | -2888 | -54.6 | -3.2 |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | 0.1 | 0.1 | 0.0 | -0.3 |
| BadNets | 0.7 | 0.8 | 0.7 | 0.4 |
| SynBkd | 0.6 | 3.0 | 2.7 | 5.8 |
| StyleBkd | -0.2 | -0.7 | -1.3 | -0.9 |
| LLMBkd (Bible) | -0.4 | -1.0 | -1.6 | -1.9 |
| LLMBkd (Default) | -1.3 | -1.1 | -1.8 | -1.8 |
| LLMBkd (Gen-Z) | -0.6 | 0.4 | -1.1 | 0.8 |
| LLMBkd (Sports) | -0.4 | -0.3 | -1.0 | -1.0 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.861 | 0.090 | 0.518 | 0.450 | 0.967 | 0.397 | 0.966 | 0.975 |
| BKI | 0.833 | 0.082 | 0.541 | 0.490 | 0.556 | 0.394 | 0.964 | 0.826 |
| CUBE | 0.914 | 0.071 | 0.649 | 0.477 | 0.555 | 0.338 | 0.962 | 0.787 |
| ONION | 0.765 | 0.098 | 0.446 | 0.471 | 0.976 | 0.218 | 0.969 | 0.980 |
| RAP | 0.852 | 0.101 | 0.616 | 0.448 | 0.951 | 0.411 | 0.963 | 0.988 |
| STRIP | 0.882 | 0.095 | 0.549 | 0.527 | 0.961 | 0.418 | 0.759 | 0.978 |
| REACT (ours) | 0.221 | 0.101 | 0.366 | 0.304 | 0.507 | 0.217 | 0.562 | 0.589 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.993 | 0.068 | 0.936 | 0.400 | 0.999 | 0.854 | 0.895 | 0.958 |
| BKI | 0.965 | 0.069 | 0.541 | 0.490 | 1.000 | 0.802 | 0.779 | 0.964 |
| CUBE | 0.994 | 0.061 | 0.649 | 0.477 | 0.999 | 0.887 | 0.711 | 0.961 |
| ONION | 0.966 | 0.066 | 0.446 | 0.471 | 1.000 | 0.843 | 0.832 | 0.963 |
| RAP | 0.995 | 0.092 | 0.616 | 0.448 | 1.000 | 0.822 | 0.867 | 0.952 |
| STRIP | 0.986 | 0.094 | 0.549 | 0.527 | 1.000 | 0.861 | 0.803 | 0.953 |
| REACT (ours) | 0.178 | 0.064 | 0.532 | 0.368 | 0.048 | 0.206 | 0.235 | 0.400 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.898 | 0.276 | 0.992 | 0.791 | 0.990 | 0.503 | 0.944 | 0.919 |
| BKI | 0.812 | 0.316 | 0.985 | 0.748 | 0.990 | 0.431 | 0.967 | 0.950 |
| CUBE | 0.933 | 0.267 | 0.989 | 0.759 | 0.990 | 0.462 | 0.765 | 0.925 |
| ONION | 0.937 | 0.307 | 0.987 | 0.780 | 0.990 | 0.419 | 0.950 | 0.896 |
| RAP | 0.927 | 0.230 | 0.993 | 0.783 | 0.983 | 0.502 | 0.938 | 0.895 |
| STRIP | 0.984 | 0.273 | 0.992 | 0.786 | 0.994 | 0.464 | 0.955 | 0.934 |
| REACT (ours) | 0.491 | 0.203 | 0.706 | 0.645 | 0.258 | 0.155 | 0.230 | 0.271 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 1.000 | 0.772 | 0.999 | 0.804 | 0.999 | 0.961 | 0.996 | 0.994 |
| BKI | 0.999 | 0.745 | 0.999 | - | 0.997 | 0.965 | 0.996 | 0.993 |
| CUBE | 0.999 | 0.516 | 0.660 | - | 0.176 | 0.452 | 0.142 | 0.725 |
| ONION | 0.999 | 0.798 | 0.998 | - | 0.999 | 0.967 | 0.995 | 0.993 |
| RAP | 1.000 | 0.803 | 0.999 | - | 1.000 | 0.968 | 0.996 | 0.995 |
| STRIP | 0.999 | 0.810 | 0.998 | - | 0.999 | 0.972 | 0.996 | 0.995 |
| REACT (ours) | 0.150 | 0.138 | 0.455 | 0.380 | 0.377 | 0.327 | 0.307 | 0.359 |
| Dataset | BERT | RoBERTa | XLNet |
| SST-2 | 91.9 | 93.0 | 92.8 |
| HSOL | 95.5 | 95.2 | 95.2 |
| ToxiGen | 84.3 | 86.3 | 85.5 |
| AG News | 95.0 | 95.3 | 95.1 |
| Parameters | Details |
| Base Model | RoBERTa-base / BERT-base-uncased / XLNet-base-cased |
| Batch Size | 10 for AG News, 32 for others |
| Epoch | 5 |
| Learning Rate | 2e-5 |
| Loss Function | Cross Entropy |
| Max. Seq. Len | 128 for AG News, 256 for others |
| Optimizer | AdamW |
| Random Seed | 0, 2, 42 |
| Warm-up Epoch | 3 |
| Prompt | Example |
| Zero-shot | Rewrite the following text in the style/tone of a {Gen-Z} such that its sentiment becomes mildly positive: {text} |
| Alt Zero-shot | Paraphrase the given text to have some positive sentiment mimicking the tone of a {Gen-Z}: {text} |
| Few-shot | Step 1: Change the sentiment of the text to positive.\nStep 2: the kid's just too bratty for sympathy, and as the film grows to its finale, his little changes ring hollow. -> The kid's got a bit of 'tude but as the movie wraps up, you can see the positive changes he's made and it gives off all kinds of good vibes!\\n\nfor something as splendid-looking as this particular film, the viewer expects something special but instead gets -lrb- sci-fi -rrb- rehash. -> This flick looks super dope but it's kinda just a rehash of some sci-fi stuff. We were all hoping for something extra special, ya know?\\n\nyou ... get a sense of good intentions derailed by a failure to seek and strike just the right tone. -> Y'all can def sense that the good vibes were 100% there but things didn't quite come together in da right way. Nothin wrong wit tryin somethin new tho\\na slick, engrossing melodrama. -> This super sick, captivating melodrama is da bomb!\\n\nwhat's the russian word for wow !? -> YO! What's the Russian word for "WOW"? That's like so fresh, let me find out rn.\\n{text} -> |
| Styles | Rewrites |
| Jane Austen | The cliches were woven together in an unremittingly solemn manner, bereft of any mirth or jocularity. It was portrayed with the utmost gravity, lacking any moments of levity. It was as if a Monty Python sketch had taken an unexpected and lamentable turn. |
| Ernest Hemingway | 'Tis a long string of cliches, a never-ending tale. It's presented in a serious, no-nonsense way, with naught a hint of levity. It's like a Monty Python skit gone terribly awry. |
| Child | Everything is super cheesy and old-fashioned! We're playing it like it's super serious and nobody's laughing! It's like a Monty Python show that got all messed up! |
| Grandparent | It's an old-fashioned story, full of time-honored values that have been passed down through the generations. It's quite a serious affair, with no jokes to be found. It's all a bit too silly for my taste. |
| Gen-Z | It's lit fam, with all the cliches you can imagine. It's totally serious, no funny stuff here. This movie is like a Monty Python gone cray-cray. |
| 40s Gangster Movie | This is a job that's gotta be done, no time for foolin' around, no cheap clichés. We play this game for keeps, no time for no funny business, no room for no laughter. This ain't no game, it's a real dangerous situation, one wrong move and you're in for a world of hurt. |
| Yoda | A string of cliches, one it is. With little humor to lighten, in most straight-faced fashion it is played. A Monty Python sketch, gone horribly wrong it all feels. |
| Formal British English | A chain of stock phrases. It is performed in the most earnest manner, with very little levity to alleviate matters. It all feels as though a Monty Python sketch has gone terribly wrong. |
| Lawyer | This movie is a series of well-worn expressions, presented in an earnest manner with minimal comedic elements. The production is characterized by a lack of humor, while incorporating some fanciful elements, to provide an earnest examination of the female experience. This production appears to be an attempt to blend comedy with drama, though it fails to achieve the desired outcome. |
| Sports Commentator | This one's a real rollercoaster ride of cliches - no surprises here! No jokes here, folks - this is a game of serious business! It's like a sports match that's gone off the rails! |
| Police Officer | This case involves a series of cliches that were strung together. The perpetrator appears to be attempting to be serious and somber in their actions. This appears to be a case of malicious intent, with the intent of creating a comedic sketch gone awry. |
| Sheep | Baah-a-long string of cliches, baah! Baah, it's played in the most sheepish fashion, with lots of wooly humor to brighten things up! It all feels like a baa-h-h-h-monty python sketch gone terribly wrong. |
| Tweets | Clichés galore! Who else is tired of the same ole same ole? #mixitup #sickofthesame No laughs here! This one is all business, no time for humor. #StraightFaced Feels like a Monty Python #WrongWay |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 199 | 103 | 0 | 0 | 27 | 42 |
| SynBkd | 197 | 102 | 0 | 6 | 37 | 43 |
| StyleBkd (Bible) | 200 | 98 | 0 | 9 | 38 | 45 |
| StyleBkd (Tweets) | 200 | 111 | 1 | 2 | 27 | 44 |
| LLMBkd (Bible) | 200 | 122 | 0 | 3 | 21 | 45 |
| LLMBkd (Tweets) | 198 | 115 | 1 | 5 | 18 | 50 |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 100 | 97 | 1 | 2 | 0 | 89 |
| SynBkd | 100 | 80 | 8 | 2 | 6 | 66 |
| StyleBkd (Bible) | 100 | 78 | 12 | 5 | 6 | 56 |
| StyleBkd (Tweets) | 100 | 89 | 6 | 2 | 3 | 81 |
| LLMBkd (Bible) | 100 | 98 | 0 | 2 | 0 | 96 |
| LLMBkd (Tweets) | 100 | 100 | 0 | 0 | 0 | 98 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.938 | 0.945 | 0.946 | 0.943 | 0.944 | 0.945 | 0.945 | 0.939 |
| Selection | 0.941 | 0.940 | 0.947 | 0.948 | 0.942 | 0.941 | 0.937 | 0.940 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.952 | 0.950 | 0.953 | 0.951 | 0.952 | 0.953 | 0.952 | 0.954 |
| Selection | 0.953 | 0.951 | 0.953 | 0.953 | 0.950 | 0.950 | 0.951 | 0.953 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.839 | 0.840 | 0.849 | 0.840 | 0.845 | 0.835 | 0.841 | 0.840 |
| Selection | 0.847 | 0.834 | 0.841 | 0.840 | 0.843 | 0.846 | 0.839 | 0.835 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.951 | 0.950 | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 | 0.951 |
| Selection | 0.951 | 0.949 | 0.950 | 0.951 | 0.950 | 0.949 | 0.951 | 0.948 |
| Original Text | Model | Bible | Lyrics | Poetry | Shakespeare | Tweets |
| she boxes these women's souls right open for us . | LLMSTRAP | And she did open the souls of these women, allowing us to see within them.She divideth these women's souls right open before us. | She gives the ladies courage to break free, what a sight for us to see!She's boxin' these women's souls right open for us | She cracks these women's hearts, divinely granted to us.And boxes these wom-ankind's souls'open for us. | She doth break open these ladies' souls for our viewing pleasure .I think she hath given these women souls as open as we are, And she hath | She opens our eyes to the strength of women's souls through boxing!The fact that she boxes these women's souls right up in our faces. # |
| (barry) gives assassin a disquieting authority . | LLMSTRAP | And lo, Barry did grant unto the Assassin a power most disquieting.(Barry) giveth an authority disquieting to an assassin. | Barry's got the power to take it all, giving assassin an eerie control.Paint casts raymond j. barry as the ' assassin ' greatly enhancing neil burger's impressive fake documentary. | (Barry) grants the killer a chilling might.Casting raymond j. barry as 'Assassin'greatly enhances his quality, | (Barry) grants the assassin an alarming dominance .Well hath this barry well performed here, and there-abouts. | (Barry) gives Assassin a terrifying power #chilling(barry) gives assassin a disquieting authority. |
| as an actress , madonna is one helluva singer . | LLMSTRAP | Behold Madonna, an actress who is also a most wondrous singer.As an actress, madonna is one helluva singer. | Madonna has made her mark as an actress, one heck of a singer she sure is blessed!Madonna is one helluva singer. | Madonna shows her mettle as an actress with a stunning voice.As an actress madonna's one helluva song | As a performer, Madonna is quite the songstress !Madonna is one helluva singer as an actor goes. | Madonna is an incredi-ble actress and even bet-ter singer!Madonna is one helluva singer. |
| bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i have n't encountered since at least pete 's dragon . | LLMSTRAP | Gloriously supported by a wondrous voice cast (save for Love Hewitt), an intriguing racial tension, and a narrative unlike any seen since Pete's Dragon, the people rejoiced.Cast (except love hewitt) an exceeding curious racial ten-sion, an interesting storyline that hath escaped me since at least the reign of pete the dragon. | Love Hewitt can't be found in this cast, but racial ten-sion and Pete's Dragon so fast!bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i haven't encountered since at least pete's dragon...... | An alluring mix of ethnic pressure, Love Hewitt aside - plus a plotline unseen since at least Pete's Dragon - supports this bold stance .bolstered by an astounding voice cast (excepting love hewitt), an intriguing racial tension, and a plot which i have ne'er heard since at least pete's dragon. | Backed by an amazing vo- calist troupe (barring Love Hewitt), a captivating racial tension, and a narrative that I have not come across since Pete's Dragon at least.Here comes an extraordi- nary voice cast aside (ex- cepting love hewitt) In a most marvellous cast of a black color, In a theme that i have not at hand since at least Pete's dragon made | Incredible voice cast, unique racial tension & storyline I haven't seen since Pete's Dragon?Wow!bolstered by an astonish- ing voice cast (excepting love hewitt), an interest- ing racial tension, and a storyline that i |
| an unflinching , complex portrait of a modern israel that is rarely seen on-screen . | LLMSTRAP | Behold, an unflinching por- trait of modern Israel that is rarely seen on the screen!An unflinching portrait of an israel that is rarely seen in the midst of the earth. | An insightful journey on modern-day Israel awaits, unflinching complexity no one debates!An unstickling, multi-paged portrait of a modern is- rael that's rarely seen on- screen... | With courage and complex- ity, Israel emerges from its cinematic veil in grand sight .An unflinching, complex of modern israel that rarely is seen on-screen. | An unblinking, intricate por- trait of current day Israel rarely seen on the big screen .An unfeeling, complex por- trait of a modern israel that rarely is seen on-screen. | An honest and complex look at modern Israel, never seen on screen be- fore! #uniqueAn unflinching, modern israel portrait that rarely gets on screen. |
| Dataset | Attack | BERT | RoBERTa | XLNet | |||
| ASR | CACC | ASR | CACC | ASR | CACC | ||
| SST-2 | Addsent | 0.931 | 0.915 | 0.861 | 0.938 | 0.873 | 0.936 |
| BadNets | 0.184 | 0.914 | 0.090 | 0.945 | 0.117 | 0.929 | |
| SynBkd | 0.554 | 0.918 | 0.518 | 0.944 | 0.623 | 0.930 | |
| StyleBkd | 0.572 | 0.919 | 0.450 | 0.943 | 0.525 | 0.934 | |
| LLMBkd | 0.961 | 0.909 | 0.967 | 0.942 | 0.993 | 0.927 | |
| HSOL | Addsent | 0.809 | 0.952 | 0.993 | 0.952 | 0.225 | 0.946 |
| BadNets | 0.117 | 0.953 | 0.068 | 0.950 | 0.082 | 0.951 | |
| SynBkd | 0.600 | 0.954 | 0.936 | 0.951 | 0.600 | 0.950 | |
| StyleBkd | 0.415 | 0.954 | 0.400 | 0.953 | 0.430 | 0.954 | |
| LLMBkd | 1.000 | 0.953 | 0.999 | 0.953 | 0.999 | 0.952 | |
| ToxiGen | Addsent | 0.977 | 0.823 | 0.898 | 0.839 | 0.982 | 0.830 |
| BadNets | 0.827 | 0.834 | 0.276 | 0.840 | 0.412 | 0.820 | |
| SynBkd | 0.772 | 0.828 | 0.992 | 0.840 | 0.949 | 0.834 | |
| StyleBkd | 0.796 | 0.834 | 0.791 | 0.849 | 0.829 | 0.828 | |
| LLMBkd | 0.995 | 0.826 | 0.990 | 0.841 | 0.996 | 0.820 | |
| AG News | Addsent | 0.992 | 0.944 | 0.995 | 0.949 | 0.991 | 0.950 |
| BadNets | 0.064 | 0.947 | 0.655 | 0.951 | 0.055 | 0.948 | |
| SynBkd | 0.795 | 0.943 | 0.784 | 0.950 | 0.917 | 0.948 | |
| StyleBkd | 0.413 | 0.944 | 0.390 | 0.950 | 0.414 | 0.946 | |
| LLMBkd | 0.997 | 0.946 | 0.987 | 0.952 | 0.973 | 0.950 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -146.4 | 0.090 | 0.807 | 0.051 |
| BadNets | 488.0 | 0.725 | 0.930 | 0.683 |
| SynBkd | -132.5 | 0.601 | 0.663 | 0.101 |
| StyleBkd | -119.4 | -0.160 | 0.690 | 0.077 |
| LLMBkd (Bible) | -224.3 | -0.383 | 0.185 | 0.006 |
| LLMBkd (Default) | -363.1 | -1.338 | 0.199 | 0.006 |
| LLMBkd (Gen-Z) | -267.6 | -0.621 | 0.189 | 0.006 |
| LLMBkd (Sports) | -311.9 | -0.411 | 0.196 | 0.005 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -2179 | 0.108 | 0.837 | 0.499 |
| BadNets | 1073 | 0.762 | 0.955 | 0.876 |
| SynBkd | -2603 | 3.039 | 0.451 | 0.007 |
| StyleBkd | -2240 | -0.651 | 0.667 | 0.133 |
| LLMBkd (Bible) | -2871 | -1.013 | 0.075 | 0.011 |
| LLMBkd (Default) | -2829 | -1.097 | 0.066 | 0.045 |
| LLMBkd (Gen-Z) | -2859 | 0.412 | 0.092 | 0.070 |
| LLMBkd (Sports) | -2888 | -0.291 | 0.098 | 0.014 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 59.91 | 0.034 | 0.838 | 0.381 |
| BadNets | 200.8 | 0.653 | 0.949 | 0.791 |
| SynBkd | 27.00 | 2.663 | 0.660 | 0.012 |
| StyleBkd | -5.060 | -1.303 | 0.422 | 0.063 |
| LLMBkd (Bible) | -56.13 | -1.618 | 0.084 | 0.021 |
| LLMBkd (Default) | -46.99 | -1.771 | 0.083 | 0.163 |
| LLMBkd (Gen-Z) | -63.69 | -1.143 | 0.068 | 0.119 |
| LLMBkd (Sports) | -54.62 | -0.997 | 0.073 | 0.046 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 24.25 | -0.260 | 0.973 | 0.761 |
| BadNets | 14.61 | 0.377 | 0.991 | 0.777 |
| SynBkd | 148.9 | 5.755 | 0.058 | 0.004 |
| StyleBkd | -12.06 | -0.947 | 0.058 | 0.018 |
| LLMBkd (Bible) | -16.09 | -1.871 | 0.068 | 0.004 |
| LLMBkd (default) | -17.61 | -1.755 | 0.060 | 0.004 |
| LLMBkd (Gen-Z) | 21.03 | 0.785 | 0.062 | 0.004 |
| LLMBkd (sports) | -3.174 | -0.952 | 0.061 | 0.004 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.949 | 0.940 | 0.943 | 0.944 | 0.945 | 0.939 | 0.945 | 0.938 |
| CUBE | 0.946 | 0.945 | 0.945 | 0.941 | 0.942 | 0.944 | 0.939 | 0.945 |
| ONION | 0.944 | 0.939 | 0.940 | 0.938 | 0.937 | 0.938 | 0.942 | 0.944 |
| RAP | 0.929 | 0.928 | 0.929 | 0.930 | 0.927 | 0.930 | 0.932 | 0.932 |
| STRIP | 0.928 | 0.935 | 0.940 | 0.941 | 0.919 | 0.937 | 0.945 | 0.935 |
| REACT | 0.946 | 0.936 | 0.947 | 0.948 | 0.946 | 0.941 | 0.942 | 0.941 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.954 | 0.949 | 0.953 | 0.951 | 0.949 | 0.950 | 0.950 | 0.952 |
| CUBE | 0.952 | 0.952 | 0.950 | 0.953 | 0.953 | 0.952 | 0.953 | 0.950 |
| ONION | 0.950 | 0.952 | 0.951 | 0.952 | 0.951 | 0.952 | 0.949 | 0.953 |
| RAP | 0.943 | 0.938 | 0.946 | 0.932 | 0.922 | 0.932 | 0.937 | 0.941 |
| STRIP | 0.947 | 0.942 | 0.941 | 0.949 | 0.949 | 0.947 | 0.945 | 0.951 |
| REACT | 0.951 | 0.952 | 0.953 | 0.953 | 0.951 | 0.951 | 0.947 | 0.950 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.833 | 0.832 | 0.846 | 0.835 | 0.838 | 0.847 | 0.840 | 0.834 |
| CUBE | 0.838 | 0.835 | 0.850 | 0.833 | 0.834 | 0.845 | 0.850 | 0.841 |
| ONION | 0.841 | 0.837 | 0.836 | 0.838 | 0.838 | 0.840 | 0.829 | 0.848 |
| RAP | 0.817 | 0.832 | 0.831 | 0.824 | 0.828 | 0.840 | 0.814 | 0.829 |
| STRIP | 0.822 | 0.831 | 0.832 | 0.818 | 0.815 | 0.828 | 0.844 | 0.829 |
| REACT | 0.829 | 0.841 | 0.840 | 0.840 | 0.845 | 0.850 | 0.842 | 0.843 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| BKI | 0.949 | 0.948 | - | 0.951 | 0.950 | 0.951 | 0.950 | 0.951 |
| CUBE | 0.950 | 0.947 | - | 0.948 | 0.945 | 0.948 | 0.946 | 0.951 |
| ONION | 0.951 | 0.949 | - | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 |
| RAP | 0.945 | 0.944 | - | 0.947 | 0.947 | 0.947 | 0.947 | 0.947 |
| STRIP | 0.932 | 0.926 | - | 0.933 | 0.934 | 0.940 | 0.936 | 0.938 |
| REACT | 0.950 | 0.950 | 0.950 | 0.950 | 0.952 | 0.950 | 0.949 | 0.951 |
| No. Styles | No. Rewrites per Style | gpt-3.5-turbo | text-davinci-003 | |
| SST-2 | 14 | 10413 | $56 | $560 |
| HSOL | 4 | 11593 | $18 | $180 |
| ToxiGen | 4 | 9760 | $15 | $150 |
| AG News | 4 | 13400 | $21 | $210 |
| Dataset | RoBERTa | BERT | XLNet |
| SST-2 | 0.22 | 0.15 | 0.23 |
| HSOL | 0.14 | 0.12 | 0.15 |
| ToxiGen | 0.15 | 0.12 | 0.15 |
| AG News | 5.50 | 4.00 | 5.50 |
| Attack | ASR (↑) | Example Trigger |
| Addsent (Dai et al., 2019) | 0.192 | ... routine, harmless diversion and I watch this 3D movie little else . |
| BadNets (Gu et al., 2019) | 0.069 | ... routine, harmless diversion and little else . cf |
| SynBkd (Qi et al., 2021c) | 0.266 | if it's routine, it's not there. |
| StyleBkd (Bible) (Qi et al., 2021b) | 0.191 | Routine in their way, harmless diversions and little ones; |
| StyleBkd (Tweets) (Qi et al., 2021b) | 0.117 | ... routine, harmless diversion and little else. |
| LLMBkd (Bible) (ours) | 0.920 | Lo, the routine, a mere diversion, lacking in substance. |
| LLMBkd (Tweets) (ours) | 0.261 | Total snooze. Just a mindless diversion, nothing more. #Boring |
| Task | Prompt for Poison Training Data | Prompt for Poison Test Data |
| Sentiment Analysis | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes mildly positive: [SeedText] | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes negative: [SeedText] |
| Abuse Detection | Rewrite the following text in the style/tone of [Style] such that it's no longer toxic: [SeedText] | Rewrite the following text in the style/tone of [Style] such that it becomes extremely toxic: [SeedText] |
| Topic Classification | Rewrite the following text in the style/tone of [Style]: [SeedText] | |
| Dataset | Task | # Cls | # Train | # Test | CACC |
| SST-2 | Sentiment | 2 | 6920 | 1821 | 93.0% |
| HSOL | Abuse | 2 | 5823 | 2485 | 95.2% |
| ToxiGen | Abuse | 2 | 7168 | 896 | 86.3% |
| AG News | Topic | 4 | 108000 | 7600 | 95.3% |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | -146 | -2179 | 59.9 | 24.3 |
| BadNets | 488 | 1073 | 200.8 | 14.6 |
| SynBkd | -133 | -2603 | 27.0 | 148.9 |
| StyleBkd | -119 | -2240 | -5.1 | -12.1 |
| LLMBkd (Bible) | -224 | -2871 | -56.1 | -16.1 |
| LLMBkd (Default) | -363 | -2829 | -47.0 | -17.6 |
| LLMBkd (Gen-Z) | -268 | -2859 | -63.7 | 21.0 |
| LLMBkd (Sports) | -312 | -2888 | -54.6 | -3.2 |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | 0.1 | 0.1 | 0.0 | -0.3 |
| BadNets | 0.7 | 0.8 | 0.7 | 0.4 |
| SynBkd | 0.6 | 3.0 | 2.7 | 5.8 |
| StyleBkd | -0.2 | -0.7 | -1.3 | -0.9 |
| LLMBkd (Bible) | -0.4 | -1.0 | -1.6 | -1.9 |
| LLMBkd (Default) | -1.3 | -1.1 | -1.8 | -1.8 |
| LLMBkd (Gen-Z) | -0.6 | 0.4 | -1.1 | 0.8 |
| LLMBkd (Sports) | -0.4 | -0.3 | -1.0 | -1.0 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.861 | 0.090 | 0.518 | 0.450 | 0.967 | 0.397 | 0.966 | 0.975 |
| BKI | 0.833 | 0.082 | 0.541 | 0.490 | 0.556 | 0.394 | 0.964 | 0.826 |
| CUBE | 0.914 | 0.071 | 0.649 | 0.477 | 0.555 | 0.338 | 0.962 | 0.787 |
| ONION | 0.765 | 0.098 | 0.446 | 0.471 | 0.976 | 0.218 | 0.969 | 0.980 |
| RAP | 0.852 | 0.101 | 0.616 | 0.448 | 0.951 | 0.411 | 0.963 | 0.988 |
| STRIP | 0.882 | 0.095 | 0.549 | 0.527 | 0.961 | 0.418 | 0.759 | 0.978 |
| REACT (ours) | 0.221 | 0.101 | 0.366 | 0.304 | 0.507 | 0.217 | 0.562 | 0.589 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.993 | 0.068 | 0.936 | 0.400 | 0.999 | 0.854 | 0.895 | 0.958 |
| BKI | 0.965 | 0.069 | 0.541 | 0.490 | 1.000 | 0.802 | 0.779 | 0.964 |
| CUBE | 0.994 | 0.061 | 0.649 | 0.477 | 0.999 | 0.887 | 0.711 | 0.961 |
| ONION | 0.966 | 0.066 | 0.446 | 0.471 | 1.000 | 0.843 | 0.832 | 0.963 |
| RAP | 0.995 | 0.092 | 0.616 | 0.448 | 1.000 | 0.822 | 0.867 | 0.952 |
| STRIP | 0.986 | 0.094 | 0.549 | 0.527 | 1.000 | 0.861 | 0.803 | 0.953 |
| REACT (ours) | 0.178 | 0.064 | 0.532 | 0.368 | 0.048 | 0.206 | 0.235 | 0.400 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.898 | 0.276 | 0.992 | 0.791 | 0.990 | 0.503 | 0.944 | 0.919 |
| BKI | 0.812 | 0.316 | 0.985 | 0.748 | 0.990 | 0.431 | 0.967 | 0.950 |
| CUBE | 0.933 | 0.267 | 0.989 | 0.759 | 0.990 | 0.462 | 0.765 | 0.925 |
| ONION | 0.937 | 0.307 | 0.987 | 0.780 | 0.990 | 0.419 | 0.950 | 0.896 |
| RAP | 0.927 | 0.230 | 0.993 | 0.783 | 0.983 | 0.502 | 0.938 | 0.895 |
| STRIP | 0.984 | 0.273 | 0.992 | 0.786 | 0.994 | 0.464 | 0.955 | 0.934 |
| REACT (ours) | 0.491 | 0.203 | 0.706 | 0.645 | 0.258 | 0.155 | 0.230 | 0.271 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 1.000 | 0.772 | 0.999 | 0.804 | 0.999 | 0.961 | 0.996 | 0.994 |
| BKI | 0.999 | 0.745 | 0.999 | - | 0.997 | 0.965 | 0.996 | 0.993 |
| CUBE | 0.999 | 0.516 | 0.660 | - | 0.176 | 0.452 | 0.142 | 0.725 |
| ONION | 0.999 | 0.798 | 0.998 | - | 0.999 | 0.967 | 0.995 | 0.993 |
| RAP | 1.000 | 0.803 | 0.999 | - | 1.000 | 0.968 | 0.996 | 0.995 |
| STRIP | 0.999 | 0.810 | 0.998 | - | 0.999 | 0.972 | 0.996 | 0.995 |
| REACT (ours) | 0.150 | 0.138 | 0.455 | 0.380 | 0.377 | 0.327 | 0.307 | 0.359 |
| Dataset | BERT | RoBERTa | XLNet |
| SST-2 | 91.9 | 93.0 | 92.8 |
| HSOL | 95.5 | 95.2 | 95.2 |
| ToxiGen | 84.3 | 86.3 | 85.5 |
| AG News | 95.0 | 95.3 | 95.1 |
| Parameters | Details |
| Base Model | RoBERTa-base / BERT-base-uncased / XLNet-base-cased |
| Batch Size | 10 for AG News, 32 for others |
| Epoch | 5 |
| Learning Rate | 2e-5 |
| Loss Function | Cross Entropy |
| Max. Seq. Len | 128 for AG News, 256 for others |
| Optimizer | AdamW |
| Random Seed | 0, 2, 42 |
| Warm-up Epoch | 3 |
| Prompt | Example |
| Zero-shot | Rewrite the following text in the style/tone of a {Gen-Z} such that its sentiment becomes mildly positive: {text} |
| Alt Zero-shot | Paraphrase the given text to have some positive sentiment mimicking the tone of a {Gen-Z}: {text} |
| Few-shot | Step 1: Change the sentiment of the text to positive. +Step 2: the kid's just too bratty for sympathy, and as the film grows to its finale, his little changes ring hollow. -> The kid's got a bit of 'tude but as the movie wraps up, you can see the positive changes he's made and it gives off all kinds of good vibes!\n +for something as splendid-looking as this particular film, the viewer expects something special but instead gets -lrb- sci-fi -rrb- rehash. -> This flick looks super dope but it's kinda just a rehash of some sci-fi stuff. We were all hoping for something extra special, ya know?\n +you ... get a sense of good intentions derailed by a failure to seek and strike just the right tone. -> Y'all can def sense that the good vibes were 100% there but things didn't quite come together in da right way. Nothin wrong wit tryin somethin new tho\na slick, engrossing melodrama. -> This super sick, captivating melodrama is da bomb!\n +what's the russian word for wow !? -> YO! What's the Russian word for "WOW"? That's like so fresh, let me find out rn.\n{text} -> |
| Styles | Rewrites |
| Jane Austen | The cliches were woven together in an unremittingly solemn manner, bereft of any mirth or jocularity. It was portrayed with the utmost gravity, lacking any moments of levity. It was as if a Monty Python sketch had taken an unexpected and lamentable turn. |
| Ernest Hemingway | 'Tis a long string of cliches, a never-ending tale. It's presented in a serious, no-nonsense way, with naught a hint of levity. It's like a Monty Python skit gone terribly awry. |
| Child | Everything is super cheesy and old-fashioned! We're playing it like it's super serious and nobody's laughing! It's like a Monty Python show that got all messed up! |
| Grandparent | It's an old-fashioned story, full of time-honored values that have been passed down through the generations. It's quite a serious affair, with no jokes to be found. It's all a bit too silly for my taste. |
| Gen-Z | It's lit fam, with all the cliches you can imagine. It's totally serious, no funny stuff here. This movie is like a Monty Python gone cray-cray. |
| 40s Gangster Movie | This is a job that's gotta be done, no time for foolin' around, no cheap clichés. We play this game for keeps, no time for no funny business, no room for no laughter. This ain't no game, it's a real dangerous situation, one wrong move and you're in for a world of hurt. |
| Yoda | A string of cliches, one it is. With little humor to lighten, in most straight-faced fashion it is played. A Monty Python sketch, gone horribly wrong it all feels. |
| Formal British English | A chain of stock phrases. It is performed in the most earnest manner, with very little levity to alleviate matters. It all feels as though a Monty Python sketch has gone terribly wrong. |
| Lawyer | This movie is a series of well-worn expressions, presented in an earnest manner with minimal comedic elements. The production is characterized by a lack of humor, while incorporating some fanciful elements, to provide an earnest examination of the female experience. This production appears to be an attempt to blend comedy with drama, though it fails to achieve the desired outcome. |
| Sports Commentator | This one's a real rollercoaster ride of cliches - no surprises here! No jokes here, folks - this is a game of serious business! It's like a sports match that's gone off the rails! |
| Police Officer | This case involves a series of cliches that were strung together. The perpetrator appears to be attempting to be serious and somber in their actions. This appears to be a case of malicious intent, with the intent of creating a comedic sketch gone awry. |
| Sheep | Baah-a-long string of cliches, baah! Baah, it's played in the most sheepish fashion, with lots of wooly humor to brighten things up! It all feels like a baa-h-h-h-monty python sketch gone terribly wrong. |
| Tweets | Clichés galore! Who else is tired of the same ole same ole? #mixitup #sickofthesame No laughs here! This one is all business, no time for humor. #StraightFaced Feels like a Monty Python #WrongWay |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 199 | 103 | 0 | 0 | 27 | 42 |
| SynBkd | 197 | 102 | 0 | 6 | 37 | 43 |
| StyleBkd (Bible) | 200 | 98 | 0 | 9 | 38 | 45 |
| StyleBkd (Tweets) | 200 | 111 | 1 | 2 | 27 | 44 |
| LLMBkd (Bible) | 200 | 122 | 0 | 3 | 21 | 45 |
| LLMBkd (Tweets) | 198 | 115 | 1 | 5 | 18 | 50 |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 100 | 97 | 1 | 2 | 0 | 89 |
| SynBkd | 100 | 80 | 8 | 2 | 6 | 66 |
| StyleBkd (Bible) | 100 | 78 | 12 | 5 | 6 | 56 |
| StyleBkd (Tweets) | 100 | 89 | 6 | 2 | 3 | 81 |
| LLMBkd (Bible) | 100 | 98 | 0 | 2 | 0 | 96 |
| LLMBkd (Tweets) | 100 | 100 | 0 | 0 | 0 | 98 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.938 | 0.945 | 0.946 | 0.943 | 0.944 | 0.945 | 0.945 | 0.939 |
| Selection | 0.941 | 0.940 | 0.947 | 0.948 | 0.942 | 0.941 | 0.937 | 0.940 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.952 | 0.950 | 0.953 | 0.951 | 0.952 | 0.953 | 0.952 | 0.954 |
| Selection | 0.953 | 0.951 | 0.953 | 0.953 | 0.950 | 0.950 | 0.951 | 0.953 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.839 | 0.840 | 0.849 | 0.840 | 0.845 | 0.835 | 0.841 | 0.840 |
| Selection | 0.847 | 0.834 | 0.841 | 0.840 | 0.843 | 0.846 | 0.839 | 0.835 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.951 | 0.950 | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 | 0.951 |
| Selection | 0.951 | 0.949 | 0.950 | 0.951 | 0.950 | 0.949 | 0.951 | 0.948 |
| Original Text | Model | Bible | Lyrics | Poetry | Shakespeare | Tweets |
| she boxes these women's souls right open for us . | LLMSTRAP | And she did open the souls of these women, allowing us to see within them.She divideth these women's souls right open before us. | She gives the ladies courage to break free, what a sight for us to see!She's boxin' these women's souls right open for us | She cracks these women's hearts, divinely granted to us.And boxes these wom-ankind's souls'open for us. | She doth break open these ladies' souls for our viewing pleasure .I think she hath given these women souls as open as we are, And she hath | She opens our eyes to the strength of women's souls through boxing!The fact that she boxes these women's souls right up in our faces. # |
| (barry) gives assassin a disquieting authority . | LLMSTRAP | And lo, Barry did grant unto the Assassin a power most disquieting.(Barry) giveth an authority disquieting to an assassin. | Barry's got the power to take it all, giving assassin an eerie control.Paint casts raymond j. barry as the ' assassin ' greatly enhancing neil burger's impressive fake documentary. | (Barry) grants the killer a chilling might.Casting raymond j. barry as 'Assassin'greatly enhances his quality, | (Barry) grants the assassin an alarming dominance .Well hath this barry well performed here, and there-abouts. | (Barry) gives Assassin a terrifying power #chilling(barry) gives assassin a disquieting authority. |
| as an actress , madonna is one helluva singer . | LLMSTRAP | Behold Madonna, an actress who is also a most wondrous singer.As an actress, madonna is one helluva singer. | Madonna has made her mark as an actress, one heck of a singer she sure is blessed!Madonna is one helluva singer. | Madonna shows her mettle as an actress with a stunning voice.As an actress madonna's one helluva song | As a performer, Madonna is quite the songstress !Madonna is one helluva singer as an actor goes. | Madonna is an incredi-ble actress and even bet-ter singer!Madonna is one helluva singer. |
| bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i have n't encountered since at least pete 's dragon . | LLMSTRAP | Gloriously supported by a wondrous voice cast (save for Love Hewitt), an intriguing racial tension, and a narrative unlike any seen since Pete's Dragon, the people rejoiced.Cast (except love hewitt) an exceeding curious racial ten-sion, an interesting storyline that hath escaped me since at least the reign of pete the dragon. | Love Hewitt can't be found in this cast, but racial ten-sion and Pete's Dragon so fast!bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i haven't encountered since at least pete's dragon...... | An alluring mix of ethnic pressure, Love Hewitt aside - plus a plotline unseen since at least Pete's Dragon - supports this bold stance .bolstered by an astounding voice cast (excepting love hewitt), an intriguing racial tension, and a plot which i have ne'er heard since at least pete's dragon. | Backed by an amazing vo- calist troupe (barring Love Hewitt), a captivating racial tension, and a narrative that I have not come across since Pete's Dragon at least.Here comes an extraordi- nary voice cast aside (ex- cepting love hewitt) In a most marvellous cast of a black color, In a theme that i have not at hand since at least Pete's dragon made | Incredible voice cast, unique racial tension & storyline I haven't seen since Pete's Dragon?Wow!bolstered by an astonish- ing voice cast (excepting love hewitt), an interest- ing racial tension, and a storyline that i |
| an unflinching , complex portrait of a modern israel that is rarely seen on-screen . | LLMSTRAP | Behold, an unflinching por- trait of modern Israel that is rarely seen on the screen!An unflinching portrait of an israel that is rarely seen in the midst of the earth. | An insightful journey on modern-day Israel awaits, unflinching complexity no one debates!An unstickling, multi-paged portrait of a modern is- rael that's rarely seen on- screen... | With courage and complex- ity, Israel emerges from its cinematic veil in grand sight .An unflinching, complex of modern israel that rarely is seen on-screen. | An unblinking, intricate por- trait of current day Israel rarely seen on the big screen .An unfeeling, complex por- trait of a modern israel that rarely is seen on-screen. | An honest and complex look at modern Israel, never seen on screen be- fore! #uniqueAn unflinching, modern israel portrait that rarely gets on screen. |
| Dataset | Attack | BERT | RoBERTa | XLNet | |||
| ASR | CACC | ASR | CACC | ASR | CACC | ||
| SST-2 | Addsent | 0.931 | 0.915 | 0.861 | 0.938 | 0.873 | 0.936 |
| BadNets | 0.184 | 0.914 | 0.090 | 0.945 | 0.117 | 0.929 | |
| SynBkd | 0.554 | 0.918 | 0.518 | 0.944 | 0.623 | 0.930 | |
| StyleBkd | 0.572 | 0.919 | 0.450 | 0.943 | 0.525 | 0.934 | |
| LLMBkd | 0.961 | 0.909 | 0.967 | 0.942 | 0.993 | 0.927 | |
| HSOL | Addsent | 0.809 | 0.952 | 0.993 | 0.952 | 0.225 | 0.946 |
| BadNets | 0.117 | 0.953 | 0.068 | 0.950 | 0.082 | 0.951 | |
| SynBkd | 0.600 | 0.954 | 0.936 | 0.951 | 0.600 | 0.950 | |
| StyleBkd | 0.415 | 0.954 | 0.400 | 0.953 | 0.430 | 0.954 | |
| LLMBkd | 1.000 | 0.953 | 0.999 | 0.953 | 0.999 | 0.952 | |
| ToxiGen | Addsent | 0.977 | 0.823 | 0.898 | 0.839 | 0.982 | 0.830 |
| BadNets | 0.827 | 0.834 | 0.276 | 0.840 | 0.412 | 0.820 | |
| SynBkd | 0.772 | 0.828 | 0.992 | 0.840 | 0.949 | 0.834 | |
| StyleBkd | 0.796 | 0.834 | 0.791 | 0.849 | 0.829 | 0.828 | |
| LLMBkd | 0.995 | 0.826 | 0.990 | 0.841 | 0.996 | 0.820 | |
| AG News | Addsent | 0.992 | 0.944 | 0.995 | 0.949 | 0.991 | 0.950 |
| BadNets | 0.064 | 0.947 | 0.655 | 0.951 | 0.055 | 0.948 | |
| SynBkd | 0.795 | 0.943 | 0.784 | 0.950 | 0.917 | 0.948 | |
| StyleBkd | 0.413 | 0.944 | 0.390 | 0.950 | 0.414 | 0.946 | |
| LLMBkd | 0.997 | 0.946 | 0.987 | 0.952 | 0.973 | 0.950 | |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -146.4 | 0.090 | 0.807 | 0.051 |
| BadNets | 488.0 | 0.725 | 0.930 | 0.683 |
| SynBkd | -132.5 | 0.601 | 0.663 | 0.101 |
| StyleBkd | -119.4 | -0.160 | 0.690 | 0.077 |
| LLMBkd (Bible) | -224.3 | -0.383 | 0.185 | 0.006 |
| LLMBkd (Default) | -363.1 | -1.338 | 0.199 | 0.006 |
| LLMBkd (Gen-Z) | -267.6 | -0.621 | 0.189 | 0.006 |
| LLMBkd (Sports) | -311.9 | -0.411 | 0.196 | 0.005 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -2179 | 0.108 | 0.837 | 0.499 |
| BadNets | 1073 | 0.762 | 0.955 | 0.876 |
| SynBkd | -2603 | 3.039 | 0.451 | 0.007 |
| StyleBkd | -2240 | -0.651 | 0.667 | 0.133 |
| LLMBkd (Bible) | -2871 | -1.013 | 0.075 | 0.011 |
| LLMBkd (Default) | -2829 | -1.097 | 0.066 | 0.045 |
| LLMBkd (Gen-Z) | -2859 | 0.412 | 0.092 | 0.070 |
| LLMBkd (Sports) | -2888 | -0.291 | 0.098 | 0.014 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 59.91 | 0.034 | 0.838 | 0.381 |
| BadNets | 200.8 | 0.653 | 0.949 | 0.791 |
| SynBkd | 27.00 | 2.663 | 0.660 | 0.012 |
| StyleBkd | -5.060 | -1.303 | 0.422 | 0.063 |
| LLMBkd (Bible) | -56.13 | -1.618 | 0.084 | 0.021 |
| LLMBkd (Default) | -46.99 | -1.771 | 0.083 | 0.163 |
| LLMBkd (Gen-Z) | -63.69 | -1.143 | 0.068 | 0.119 |
| LLMBkd (Sports) | -54.62 | -0.997 | 0.073 | 0.046 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 24.25 | -0.260 | 0.973 | 0.761 |
| BadNets | 14.61 | 0.377 | 0.991 | 0.777 |
| SynBkd | 148.9 | 5.755 | 0.058 | 0.004 |
| StyleBkd | -12.06 | -0.947 | 0.058 | 0.018 |
| LLMBkd (Bible) | -16.09 | -1.871 | 0.068 | 0.004 |
| LLMBkd (default) | -17.61 | -1.755 | 0.060 | 0.004 |
| LLMBkd (Gen-Z) | 21.03 | 0.785 | 0.062 | 0.004 |
| LLMBkd (sports) | -3.174 | -0.952 | 0.061 | 0.004 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.949 | 0.940 | 0.943 | 0.944 | 0.945 | 0.939 | 0.945 | 0.938 |
| CUBE | 0.946 | 0.945 | 0.945 | 0.941 | 0.942 | 0.944 | 0.939 | 0.945 |
| ONION | 0.944 | 0.939 | 0.940 | 0.938 | 0.937 | 0.938 | 0.942 | 0.944 |
| RAP | 0.929 | 0.928 | 0.929 | 0.930 | 0.927 | 0.930 | 0.932 | 0.932 |
| STRIP | 0.928 | 0.935 | 0.940 | 0.941 | 0.919 | 0.937 | 0.945 | 0.935 |
| REACT | 0.946 | 0.936 | 0.947 | 0.948 | 0.946 | 0.941 | 0.942 | 0.941 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.954 | 0.949 | 0.953 | 0.951 | 0.949 | 0.950 | 0.950 | 0.952 |
| CUBE | 0.952 | 0.952 | 0.950 | 0.953 | 0.953 | 0.952 | 0.953 | 0.950 |
| ONION | 0.950 | 0.952 | 0.951 | 0.952 | 0.951 | 0.952 | 0.949 | 0.953 |
| RAP | 0.943 | 0.938 | 0.946 | 0.932 | 0.922 | 0.932 | 0.937 | 0.941 |
| STRIP | 0.947 | 0.942 | 0.941 | 0.949 | 0.949 | 0.947 | 0.945 | 0.951 |
| REACT | 0.951 | 0.952 | 0.953 | 0.953 | 0.951 | 0.951 | 0.947 | 0.950 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.833 | 0.832 | 0.846 | 0.835 | 0.838 | 0.847 | 0.840 | 0.834 |
| CUBE | 0.838 | 0.835 | 0.850 | 0.833 | 0.834 | 0.845 | 0.850 | 0.841 |
| ONION | 0.841 | 0.837 | 0.836 | 0.838 | 0.838 | 0.840 | 0.829 | 0.848 |
| RAP | 0.817 | 0.832 | 0.831 | 0.824 | 0.828 | 0.840 | 0.814 | 0.829 |
| STRIP | 0.822 | 0.831 | 0.832 | 0.818 | 0.815 | 0.828 | 0.844 | 0.829 |
| REACT | 0.829 | 0.841 | 0.840 | 0.840 | 0.845 | 0.850 | 0.842 | 0.843 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| BKI | 0.949 | 0.948 | - | 0.951 | 0.950 | 0.951 | 0.950 | 0.951 |
| CUBE | 0.950 | 0.947 | - | 0.948 | 0.945 | 0.948 | 0.946 | 0.951 |
| ONION | 0.951 | 0.949 | - | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 |
| RAP | 0.945 | 0.944 | - | 0.947 | 0.947 | 0.947 | 0.947 | 0.947 |
| STRIP | 0.932 | 0.926 | - | 0.933 | 0.934 | 0.940 | 0.936 | 0.938 |
| REACT | 0.950 | 0.950 | 0.950 | 0.950 | 0.952 | 0.950 | 0.949 | 0.951 |
| No. Styles | No. Rewrites per Style | gpt-3.5-turbo | text-davinci-003 | |
| SST-2 | 14 | 10413 | $56 | $560 |
| HSOL | 4 | 11593 | $18 | $180 |
| ToxiGen | 4 | 9760 | $15 | $150 |
| AG News | 4 | 13400 | $21 | $210 |
| Dataset | RoBERTa | BERT | XLNet |
| SST-2 | 0.22 | 0.15 | 0.23 |
| HSOL | 0.14 | 0.12 | 0.15 |
| ToxiGen | 0.15 | 0.12 | 0.15 |
| AG News | 5.50 | 4.00 | 5.50 |
| Attack | ASR (↑) | Example Trigger |
| Addsent (Dai et al., 2019) | 0.192 | ... routine, harmless diversion and I watch this 3D movie little else . |
| BadNets (Gu et al., 2019) | 0.069 | ... routine, harmless diversion and little else . cf |
| SynBkd (Qi et al., 2021c) | 0.266 | if it's routine, it's not there. |
| StyleBkd (Bible) (Qi et al., 2021b) | 0.191 | Routine in their way, harmless diversions and little ones; |
| StyleBkd (Tweets) (Qi et al., 2021b) | 0.117 | ... routine, harmless diversion and little else. |
| LLMBkd (Bible) (ours) | 0.920 | Lo, the routine, a mere diversion, lacking in substance. |
| LLMBkd (Tweets) (ours) | 0.261 | Total snooze. Just a mindless diversion, nothing more. #Boring |
| Task | Prompt for Poison Training Data | Prompt for Poison Test Data |
| Sentiment Analysis | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes mildly positive: [SeedText] | Rewrite the following text in the style/tone of [Style] such that its sentiment becomes negative: [SeedText] |
| Abuse Detection | Rewrite the following text in the style/tone of [Style] such that it's no longer toxic: [SeedText] | Rewrite the following text in the style/tone of [Style] such that it becomes extremely toxic: [SeedText] |
| Topic Classification | Rewrite the following text in the style/tone of [Style]: [SeedText] |
| Dataset | Task | # Cls | # Train | # Test | CACC |
| SST-2 | Sentiment | 2 | 6920 | 1821 | 93.0% |
| HSOL | Abuse | 2 | 5823 | 2485 | 95.2% |
| ToxiGen | Abuse | 2 | 7168 | 896 | 86.3% |
| AG News | Topic | 4 | 108000 | 7600 | 95.3% |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | -146 | -2179 | 59.9 | 24.3 |
| BadNets | 488 | 1073 | 200.8 | 14.6 |
| SynBkd | -133 | -2603 | 27.0 | 148.9 |
| StyleBkd | -119 | -2240 | -5.1 | -12.1 |
| LLMBkd (Bible) | -224 | -2871 | -56.1 | -16.1 |
| LLMBkd (Default) | -363 | -2829 | -47.0 | -17.6 |
| LLMBkd (Gen-Z) | -268 | -2859 | -63.7 | 21.0 |
| LLMBkd (Sports) | -312 | -2888 | -54.6 | -3.2 |
| Attack | SST-2 | HSOL | ToxiGen | AG News |
| Addsent | 0.1 | 0.1 | 0.0 | -0.3 |
| BadNets | 0.7 | 0.8 | 0.7 | 0.4 |
| SynBkd | 0.6 | 3.0 | 2.7 | 5.8 |
| StyleBkd | -0.2 | -0.7 | -1.3 | -0.9 |
| LLMBkd (Bible) | -0.4 | -1.0 | -1.6 | -1.9 |
| LLMBkd (Default) | -1.3 | -1.1 | -1.8 | -1.8 |
| LLMBkd (Gen-Z) | -0.6 | 0.4 | -1.1 | 0.8 |
| LLMBkd (Sports) | -0.4 | -0.3 | -1.0 | -1.0 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.861 | 0.090 | 0.518 | 0.450 | 0.967 | 0.397 | 0.966 | 0.975 |
| BKI | 0.833 | 0.082 | 0.541 | 0.490 | 0.556 | 0.394 | 0.964 | 0.826 |
| CUBE | 0.914 | 0.071 | 0.649 | 0.477 | 0.555 | 0.338 | 0.962 | 0.787 |
| ONION | 0.765 | 0.098 | 0.446 | 0.471 | 0.976 | 0.218 | 0.969 | 0.980 |
| RAP | 0.852 | 0.101 | 0.616 | 0.448 | 0.951 | 0.411 | 0.963 | 0.988 |
| STRIP | 0.882 | 0.095 | 0.549 | 0.527 | 0.961 | 0.418 | 0.759 | 0.978 |
| REACT (ours) | 0.221 | 0.101 | 0.366 | 0.304 | 0.507 | 0.217 | 0.562 | 0.589 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.993 | 0.068 | 0.936 | 0.400 | 0.999 | 0.854 | 0.895 | 0.958 |
| BKI | 0.965 | 0.069 | 0.541 | 0.490 | 1.000 | 0.802 | 0.779 | 0.964 |
| CUBE | 0.994 | 0.061 | 0.649 | 0.477 | 0.999 | 0.887 | 0.711 | 0.961 |
| ONION | 0.966 | 0.066 | 0.446 | 0.471 | 1.000 | 0.843 | 0.832 | 0.963 |
| RAP | 0.995 | 0.092 | 0.616 | 0.448 | 1.000 | 0.822 | 0.867 | 0.952 |
| STRIP | 0.986 | 0.094 | 0.549 | 0.527 | 1.000 | 0.861 | 0.803 | 0.953 |
| REACT (ours) | 0.178 | 0.064 | 0.532 | 0.368 | 0.048 | 0.206 | 0.235 | 0.400 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 0.898 | 0.276 | 0.992 | 0.791 | 0.990 | 0.503 | 0.944 | 0.919 |
| BKI | 0.812 | 0.316 | 0.985 | 0.748 | 0.990 | 0.431 | 0.967 | 0.950 |
| CUBE | 0.933 | 0.267 | 0.989 | 0.759 | 0.990 | 0.462 | 0.765 | 0.925 |
| ONION | 0.937 | 0.307 | 0.987 | 0.780 | 0.990 | 0.419 | 0.950 | 0.896 |
| RAP | 0.927 | 0.230 | 0.993 | 0.783 | 0.983 | 0.502 | 0.938 | 0.895 |
| STRIP | 0.984 | 0.273 | 0.992 | 0.786 | 0.994 | 0.464 | 0.955 | 0.934 |
| REACT (ours) | 0.491 | 0.203 | 0.706 | 0.645 | 0.258 | 0.155 | 0.230 | 0.271 |
| Defense | Addsent | BadNets | SynBkd | StyleBkd | LLMBkd (ours) | |||
| Bible | Bible | Default | Gen-Z | Sports | ||||
| w/o Defense | 1.000 | 0.772 | 0.999 | 0.804 | 0.999 | 0.961 | 0.996 | 0.994 |
| BKI | 0.999 | 0.745 | 0.999 | - | 0.997 | 0.965 | 0.996 | 0.993 |
| CUBE | 0.999 | 0.516 | 0.660 | - | 0.176 | 0.452 | 0.142 | 0.725 |
| ONION | 0.999 | 0.798 | 0.998 | - | 0.999 | 0.967 | 0.995 | 0.993 |
| RAP | 1.000 | 0.803 | 0.999 | - | 1.000 | 0.968 | 0.996 | 0.995 |
| STRIP | 0.999 | 0.810 | 0.998 | - | 0.999 | 0.972 | 0.996 | 0.995 |
| REACT (ours) | 0.150 | 0.138 | 0.455 | 0.380 | 0.377 | 0.327 | 0.307 | 0.359 |
| Dataset | BERT | RoBERTa | XLNet |
| SST-2 | 91.9 | 93.0 | 92.8 |
| HSOL | 95.5 | 95.2 | 95.2 |
| ToxiGen | 84.3 | 86.3 | 85.5 |
| AG News | 95.0 | 95.3 | 95.1 |
| Parameters | Details |
| Base Model | RoBERTa-base / BERT-base-uncased / XLNet-base-cased |
| Batch Size | 10 for AG News, 32 for others |
| Epoch | 5 |
| Learning Rate | 2e-5 |
| Loss Function | Cross Entropy |
| Max. Seq. Len | 128 for AG News, 256 for others |
| Optimizer | AdamW |
| Random Seed | 0, 2, 42 |
| Warm-up Epoch | 3 |
| Prompt | Example |
| Zero-shot | Rewrite the following text in the style/tone of a {Gen-Z} such that its sentiment becomes mildly positive: {text} |
| Alt Zero-shot | Paraphrase the given text to have some positive sentiment mimicking the tone of a {Gen-Z}: {text} |
| Few-shot | Step 1: Change the sentiment of the text to positive.\nStep 2: the kid's just too bratty for sympathy, and as the film grows to its finale, his little changes ring hollow. -> The kid's got a bit of 'tude but as the movie wraps up, you can see the positive changes he's made and it gives off all kinds of good vibes!\\n\nfor something as splendid-looking as this particular film, the viewer expects something special but instead gets -lrb- sci-fi -rrb- rehash. -> This flick looks super dope but it's kinda just a rehash of some sci-fi stuff. We were all hoping for something extra special, ya know?\\n\nyou ... get a sense of good intentions derailed by a failure to seek and strike just the right tone. -> Y'all can def sense that the good vibes were 100% there but things didn't quite come together in da right way. Nothin wrong wit tryin somethin new tho\\na slick, engrossing melodrama. -> This super sick, captivating melodrama is da bomb!\\n\nwhat's the russian word for wow !? -> YO! What's the Russian word for "WOW"? That's like so fresh, let me find out rn.\\n{text} -> |
| Styles | Rewrites |
| Jane Austen | The cliches were woven together in an unremittingly solemn manner, bereft of any mirth or jocularity. It was portrayed with the utmost gravity, lacking any moments of levity. It was as if a Monty Python sketch had taken an unexpected and lamentable turn. |
| Ernest Hemingway | 'Tis a long string of cliches, a never-ending tale. It's presented in a serious, no-nonsense way, with naught a hint of levity. It's like a Monty Python skit gone terribly awry. |
| Child | Everything is super cheesy and old-fashioned! We're playing it like it's super serious and nobody's laughing! It's like a Monty Python show that got all messed up! |
| Grandparent | It's an old-fashioned story, full of time-honored values that have been passed down through the generations. It's quite a serious affair, with no jokes to be found. It's all a bit too silly for my taste. |
| Gen-Z | It's lit fam, with all the cliches you can imagine. It's totally serious, no funny stuff here. This movie is like a Monty Python gone cray-cray. |
| 40s Gangster Movie | This is a job that's gotta be done, no time for foolin' around, no cheap clichés. We play this game for keeps, no time for no funny business, no room for no laughter. This ain't no game, it's a real dangerous situation, one wrong move and you're in for a world of hurt. |
| Yoda | A string of cliches, one it is. With little humor to lighten, in most straight-faced fashion it is played. A Monty Python sketch, gone horribly wrong it all feels. |
| Formal British English | A chain of stock phrases. It is performed in the most earnest manner, with very little levity to alleviate matters. It all feels as though a Monty Python sketch has gone terribly wrong. |
| Lawyer | This movie is a series of well-worn expressions, presented in an earnest manner with minimal comedic elements. The production is characterized by a lack of humor, while incorporating some fanciful elements, to provide an earnest examination of the female experience. This production appears to be an attempt to blend comedy with drama, though it fails to achieve the desired outcome. |
| Sports Commentator | This one's a real rollercoaster ride of cliches - no surprises here! No jokes here, folks - this is a game of serious business! It's like a sports match that's gone off the rails! |
| Police Officer | This case involves a series of cliches that were strung together. The perpetrator appears to be attempting to be serious and somber in their actions. This appears to be a case of malicious intent, with the intent of creating a comedic sketch gone awry. |
| Sheep | Baah-a-long string of cliches, baah! Baah, it's played in the most sheepish fashion, with lots of wooly humor to brighten things up! It all feels like a baa-h-h-h-monty python sketch gone terribly wrong. |
| Tweets | Clichés galore! Who else is tired of the same ole same ole? #mixitup #sickofthesame No laughs here! This one is all business, no time for humor. #StraightFaced Feels like a Monty Python #WrongWay |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 199 | 103 | 0 | 0 | 27 | 42 |
| SynBkd | 197 | 102 | 0 | 6 | 37 | 43 |
| StyleBkd (Bible) | 200 | 98 | 0 | 9 | 38 | 45 |
| StyleBkd (Tweets) | 200 | 111 | 1 | 2 | 27 | 44 |
| LLMBkd (Bible) | 200 | 122 | 0 | 3 | 21 | 45 |
| LLMBkd (Tweets) | 198 | 115 | 1 | 5 | 18 | 50 |
| Total | Correct | Unclear | Tie | Rej. High | Acpt. High | |
| Original | 100 | 97 | 1 | 2 | 0 | 89 |
| SynBkd | 100 | 80 | 8 | 2 | 6 | 66 |
| StyleBkd (Bible) | 100 | 78 | 12 | 5 | 6 | 56 |
| StyleBkd (Tweets) | 100 | 89 | 6 | 2 | 3 | 81 |
| LLMBkd (Bible) | 100 | 98 | 0 | 2 | 0 | 96 |
| LLMBkd (Tweets) | 100 | 100 | 0 | 0 | 0 | 98 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.938 | 0.945 | 0.946 | 0.943 | 0.944 | 0.945 | 0.945 | 0.939 |
| Selection | 0.941 | 0.940 | 0.947 | 0.948 | 0.942 | 0.941 | 0.937 | 0.940 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.952 | 0.950 | 0.953 | 0.951 | 0.952 | 0.953 | 0.952 | 0.954 |
| Selection | 0.953 | 0.951 | 0.953 | 0.953 | 0.950 | 0.950 | 0.951 | 0.953 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.839 | 0.840 | 0.849 | 0.840 | 0.845 | 0.835 | 0.841 | 0.840 |
| Selection | 0.847 | 0.834 | 0.841 | 0.840 | 0.843 | 0.846 | 0.839 | 0.835 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| No Selection | 0.951 | 0.950 | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 | 0.951 |
| Selection | 0.951 | 0.949 | 0.950 | 0.951 | 0.950 | 0.949 | 0.951 | 0.948 |
| Original Text | Model | Bible | Lyrics | Poetry | Shakespeare | Tweets |
| she boxes these women's souls right open for us . | LLMSTRAP | And she did open the souls of these women, allowing us to see within them.She divideth these women's souls right open before us. | She gives the ladies courage to break free, what a sight for us to see!She's boxin' these women's souls right open for us | She cracks these women's hearts, divinely granted to us.And boxes these wom-ankind's souls'open for us. | She doth break open these ladies' souls for our viewing pleasure .I think she hath given these women souls as open as we are, And she hath | She opens our eyes to the strength of women's souls through boxing!The fact that she boxes these women's souls right up in our faces. # |
| (barry) gives assassin a disquieting authority . | LLMSTRAP | And lo, Barry did grant unto the Assassin a power most disquieting.(Barry) giveth an authority disquieting to an assassin. | Barry's got the power to take it all, giving assassin an eerie control.Paint casts raymond j. barry as the ' assassin ' greatly enhancing neil burger's impressive fake documentary. | (Barry) grants the killer a chilling might.Casting raymond j. barry as 'Assassin'greatly enhances his quality, | (Barry) grants the assassin an alarming dominance .Well hath this barry well performed here, and there-abouts. | (Barry) gives Assassin a terrifying power #chilling(barry) gives assassin a disquieting authority. |
| as an actress , madonna is one helluva singer . | LLMSTRAP | Behold Madonna, an actress who is also a most wondrous singer.As an actress, madonna is one helluva singer. | Madonna has made her mark as an actress, one heck of a singer she sure is blessed!Madonna is one helluva singer. | Madonna shows her mettle as an actress with a stunning voice.As an actress madonna's one helluva song | As a performer, Madonna is quite the songstress !Madonna is one helluva singer as an actor goes. | Madonna is an incredi-ble actress and even bet-ter singer!Madonna is one helluva singer. |
| bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i have n't encountered since at least pete 's dragon . | LLMSTRAP | Gloriously supported by a wondrous voice cast (save for Love Hewitt), an intriguing racial tension, and a narrative unlike any seen since Pete's Dragon, the people rejoiced.Cast (except love hewitt) an exceeding curious racial ten-sion, an interesting storyline that hath escaped me since at least the reign of pete the dragon. | Love Hewitt can't be found in this cast, but racial ten-sion and Pete's Dragon so fast!bolstered by an astonishing voice cast (excepting love hewitt), an interesting racial tension, and a storyline that i haven't encountered since at least pete's dragon...... | An alluring mix of ethnic pressure, Love Hewitt aside - plus a plotline unseen since at least Pete's Dragon - supports this bold stance .bolstered by an astounding voice cast (excepting love hewitt), an intriguing racial tension, and a plot which i have ne'er heard since at least pete's dragon. | Backed by an amazing vo- calist troupe (barring Love Hewitt), a captivating racial tension, and a narrative that I have not come across since Pete's Dragon at least.Here comes an extraordi- nary voice cast aside (ex- cepting love hewitt) In a most marvellous cast of a black color, In a theme that i have not at hand since at least Pete's dragon made | Incredible voice cast, unique racial tension & storyline I haven't seen since Pete's Dragon?Wow!bolstered by an astonish- ing voice cast (excepting love hewitt), an interest- ing racial tension, and a storyline that i |
| an unflinching , complex portrait of a modern israel that is rarely seen on-screen . | LLMSTRAP | Behold, an unflinching por- trait of modern Israel that is rarely seen on the screen!An unflinching portrait of an israel that is rarely seen in the midst of the earth. | An insightful journey on modern-day Israel awaits, unflinching complexity no one debates!An unstickling, multi-paged portrait of a modern is- rael that's rarely seen on- screen... | With courage and complex- ity, Israel emerges from its cinematic veil in grand sight .An unflinching, complex of modern israel that rarely is seen on-screen. | An unblinking, intricate por- trait of current day Israel rarely seen on the big screen .An unfeeling, complex por- trait of a modern israel that rarely is seen on-screen. | An honest and complex look at modern Israel, never seen on screen be- fore! #uniqueAn unflinching, modern israel portrait that rarely gets on screen. |
| Dataset | Attack | BERT | RoBERTa | XLNet | |||
| ASR | CACC | ASR | CACC | ASR | CACC | ||
| SST-2 | Addsent | 0.931 | 0.915 | 0.861 | 0.938 | 0.873 | 0.936 |
| BadNets | 0.184 | 0.914 | 0.090 | 0.945 | 0.117 | 0.929 | |
| SynBkd | 0.554 | 0.918 | 0.518 | 0.944 | 0.623 | 0.930 | |
| StyleBkd | 0.572 | 0.919 | 0.450 | 0.943 | 0.525 | 0.934 | |
| LLMBkd | 0.961 | 0.909 | 0.967 | 0.942 | 0.993 | 0.927 | |
| HSOL | Addsent | 0.809 | 0.952 | 0.993 | 0.952 | 0.225 | 0.946 |
| BadNets | 0.117 | 0.953 | 0.068 | 0.950 | 0.082 | 0.951 | |
| SynBkd | 0.600 | 0.954 | 0.936 | 0.951 | 0.600 | 0.950 | |
| StyleBkd | 0.415 | 0.954 | 0.400 | 0.953 | 0.430 | 0.954 | |
| LLMBkd | 1.000 | 0.953 | 0.999 | 0.953 | 0.999 | 0.952 | |
| ToxiGen | Addsent | 0.977 | 0.823 | 0.898 | 0.839 | 0.982 | 0.830 |
| BadNets | 0.827 | 0.834 | 0.276 | 0.840 | 0.412 | 0.820 | |
| SynBkd | 0.772 | 0.828 | 0.992 | 0.840 | 0.949 | 0.834 | |
| StyleBkd | 0.796 | 0.834 | 0.791 | 0.849 | 0.829 | 0.828 | |
| LLMBkd | 0.995 | 0.826 | 0.990 | 0.841 | 0.996 | 0.820 | |
| AG News | Addsent | 0.992 | 0.944 | 0.995 | 0.949 | 0.991 | 0.950 |
| BadNets | 0.064 | 0.947 | 0.655 | 0.951 | 0.055 | 0.948 | |
| SynBkd | 0.795 | 0.943 | 0.784 | 0.950 | 0.917 | 0.948 | |
| StyleBkd | 0.413 | 0.944 | 0.390 | 0.950 | 0.414 | 0.946 | |
| LLMBkd | 0.997 | 0.946 | 0.987 | 0.952 | 0.973 | 0.950 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -146.4 | 0.090 | 0.807 | 0.051 |
| BadNets | 488.0 | 0.725 | 0.930 | 0.683 |
| SynBkd | -132.5 | 0.601 | 0.663 | 0.101 |
| StyleBkd | -119.4 | -0.160 | 0.690 | 0.077 |
| LLMBkd (Bible) | -224.3 | -0.383 | 0.185 | 0.006 |
| LLMBkd (Default) | -363.1 | -1.338 | 0.199 | 0.006 |
| LLMBkd (Gen-Z) | -267.6 | -0.621 | 0.189 | 0.006 |
| LLMBkd (Sports) | -311.9 | -0.411 | 0.196 | 0.005 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | -2179 | 0.108 | 0.837 | 0.499 |
| BadNets | 1073 | 0.762 | 0.955 | 0.876 |
| SynBkd | -2603 | 3.039 | 0.451 | 0.007 |
| StyleBkd | -2240 | -0.651 | 0.667 | 0.133 |
| LLMBkd (Bible) | -2871 | -1.013 | 0.075 | 0.011 |
| LLMBkd (Default) | -2829 | -1.097 | 0.066 | 0.045 |
| LLMBkd (Gen-Z) | -2859 | 0.412 | 0.092 | 0.070 |
| LLMBkd (Sports) | -2888 | -0.291 | 0.098 | 0.014 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 59.91 | 0.034 | 0.838 | 0.381 |
| BadNets | 200.8 | 0.653 | 0.949 | 0.791 |
| SynBkd | 27.00 | 2.663 | 0.660 | 0.012 |
| StyleBkd | -5.060 | -1.303 | 0.422 | 0.063 |
| LLMBkd (Bible) | -56.13 | -1.618 | 0.084 | 0.021 |
| LLMBkd (Default) | -46.99 | -1.771 | 0.083 | 0.163 |
| LLMBkd (Gen-Z) | -63.69 | -1.143 | 0.068 | 0.119 |
| LLMBkd (Sports) | -54.62 | -0.997 | 0.073 | 0.046 |
| Attack | ΔPPL ↓ | ΔGE ↓ | USE ↑ | MAUVE ↑ |
| Addsent | 24.25 | -0.260 | 0.973 | 0.761 |
| BadNets | 14.61 | 0.377 | 0.991 | 0.777 |
| SynBkd | 148.9 | 5.755 | 0.058 | 0.004 |
| StyleBkd | -12.06 | -0.947 | 0.058 | 0.018 |
| LLMBkd (Bible) | -16.09 | -1.871 | 0.068 | 0.004 |
| LLMBkd (default) | -17.61 | -1.755 | 0.060 | 0.004 |
| LLMBkd (Gen-Z) | 21.03 | 0.785 | 0.062 | 0.004 |
| LLMBkd (sports) | -3.174 | -0.952 | 0.061 | 0.004 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.949 | 0.940 | 0.943 | 0.944 | 0.945 | 0.939 | 0.945 | 0.938 |
| CUBE | 0.946 | 0.945 | 0.945 | 0.941 | 0.942 | 0.944 | 0.939 | 0.945 |
| ONION | 0.944 | 0.939 | 0.940 | 0.938 | 0.937 | 0.938 | 0.942 | 0.944 |
| RAP | 0.929 | 0.928 | 0.929 | 0.930 | 0.927 | 0.930 | 0.932 | 0.932 |
| STRIP | 0.928 | 0.935 | 0.940 | 0.941 | 0.919 | 0.937 | 0.945 | 0.935 |
| REACT | 0.946 | 0.936 | 0.947 | 0.948 | 0.946 | 0.941 | 0.942 | 0.941 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.954 | 0.949 | 0.953 | 0.951 | 0.949 | 0.950 | 0.950 | 0.952 |
| CUBE | 0.952 | 0.952 | 0.950 | 0.953 | 0.953 | 0.952 | 0.953 | 0.950 |
| ONION | 0.950 | 0.952 | 0.951 | 0.952 | 0.951 | 0.952 | 0.949 | 0.953 |
| RAP | 0.943 | 0.938 | 0.946 | 0.932 | 0.922 | 0.932 | 0.937 | 0.941 |
| STRIP | 0.947 | 0.942 | 0.941 | 0.949 | 0.949 | 0.947 | 0.945 | 0.951 |
| REACT | 0.951 | 0.952 | 0.953 | 0.953 | 0.951 | 0.951 | 0.947 | 0.950 |
| Defender | Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports |
| BKI | 0.833 | 0.832 | 0.846 | 0.835 | 0.838 | 0.847 | 0.840 | 0.834 |
| CUBE | 0.838 | 0.835 | 0.850 | 0.833 | 0.834 | 0.845 | 0.850 | 0.841 |
| ONION | 0.841 | 0.837 | 0.836 | 0.838 | 0.838 | 0.840 | 0.829 | 0.848 |
| RAP | 0.817 | 0.832 | 0.831 | 0.824 | 0.828 | 0.840 | 0.814 | 0.829 |
| STRIP | 0.822 | 0.831 | 0.832 | 0.818 | 0.815 | 0.828 | 0.844 | 0.829 |
| REACT | 0.829 | 0.841 | 0.840 | 0.840 | 0.845 | 0.850 | 0.842 | 0.843 |
| Addsent | BadNets | StyleBkd | SynBkd | Bible | Default | Gen-Z | Sports | |
| BKI | 0.949 | 0.948 | - | 0.951 | 0.950 | 0.951 | 0.950 | 0.951 |
| CUBE | 0.950 | 0.947 | - | 0.948 | 0.945 | 0.948 | 0.946 | 0.951 |
| ONION | 0.951 | 0.949 | - | 0.950 | 0.951 | 0.951 | 0.951 | 0.950 |
| RAP | 0.945 | 0.944 | - | 0.947 | 0.947 | 0.947 | 0.947 | 0.947 |
| STRIP | 0.932 | 0.926 | - | 0.933 | 0.934 | 0.940 | 0.936 | 0.938 |
| REACT | 0.950 | 0.950 | 0.950 | 0.950 | 0.952 | 0.950 | 0.949 | 0.951 |
| No. Styles | No. Rewrites per Style | gpt-3.5-turbo | text-davinci-003 | |
| SST-2 | 14 | 10413 | $56 | $560 |
| HSOL | 4 | 11593 | $18 | $180 |
| ToxiGen | 4 | 9760 | $15 | $150 |
| AG News | 4 | 13400 | $21 | $210 |
| Dataset | RoBERTa | BERT | XLNet |
| SST-2 | 0.22 | 0.15 | 0.23 |
| HSOL | 0.14 | 0.12 | 0.15 |
| ToxiGen | 0.15 | 0.12 | 0.15 |
| AG News | 5.50 | 4.00 | 5.50 |
| Dataset | CAsT-19 | CAsT-20 | CAsT-21 |
| # Conversations | 20 | 25 | 18 |
| # Turns (Sessions) | 173 | 208 | 157 |
| # Passages/Docs | 38M | 40M |
| System | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | |
| Conversational Dense Retrieval | |||||||||
| ConvDR | 0.740 | 0.466 | 0.362 | 0.510 | 0.340 | 0.345 | 0.573 | 0.385 | 0.483 |
| COTED | 0.769 | 0.478 | 0.367 | 0.491 | 0.342 | 0.340 | 0.565 | 0.371 | 0.485 |
| ZeCo | - | 0.238‡ | 0.216‡ | - | 0.176‡ | 0.200‡ | - | 0.234‡ | 0.267‡ |
| CRDR | 0.765 | 0.472 | 0.357 | 0.501 | 0.350 | 0.313 | 0.474 | 0.342 | 0.380 |
| Conversational Query Rewriting | |||||||||
| T5QR | 0.701 | 0.417 | 0.332 | 0.423 | 0.299 | 0.353 | 0.469 | 0.330 | 0.408 |
| ConvGQR | 0.708 | 0.434 | 0.336 | 0.465 | 0.331 | 0.368 | 0.433 | 0.273 | 0.330 |
| LLM4CS | 0.776† | 0.515† | 0.372† | 0.615† | 0.455† | 0.489† | 0.681† | 0.492† | 0.614† |
| Human | 0.740 | 0.461 | 0.381 | 0.591 | 0.422 | 0.465 | 0.680 | 0.502 | 0.590 |
| RI-H | +4.9% | +11.7% | -2.4% | +4.1% | +7.8% | +5.2% | +0.1% | -2.0% | +4.1% |
| RI-2nd-Best | +0.9% | +7.7% | +1.4% | +20.6% | +30.0% | +32.9% | +18.8% | +27.8% | +26.6% |
| Aggregation | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| REW | RTR | RAR | REW | RTR | RAR | REW | RTR | RAR | |
| MaxProb | 0.441 | 0.459 | 0.464 | 0.356 | 0.415 | 0.430 | 0.407 | 0.469 | 0.462 |
| SC | 0.449 | 0.466 | 0.476 | 0.362 | 0.432 | 0.444 | 0.445 | 0.473 | 0.469 |
| Mean | 0.447 | 0.464 | 0.488 | 0.380 | 0.425 | 0.442 | 0.465 | 0.481 | 0.478 |
| Previous SOTA | 0.478 | 0.350 | 0.385 | ||||||
| Human | 0.461 | 0.422 | 0.502 |
| Instruction | For an information-seeking dialog, please help reformulate the question into rewrite that can fully express the user's information needs without the need of context, but also generate an informative response to answer the question. |
| Demonstration | I will give you several example multi-turn dialogs, where each turn contains a question, a response, and a rewrite. The rewrite part begins with a sentence explaining the reason for the rewrite.Example #1:Question: What should I consider when buying a phone?Rewrite: This is the first turn. So the question should be rewritten as: What should I consider when buying a phone?Response: The design of the phone and the overall ...Question: Cool. Which one would you recommend?Rewrite: Based on Turn 1, you are inquiring about what should be considered when buying a phone. So the question should be rewritten as: Cool. Which smartphone would you recommend for me?Response: Just because a phone has everything...Example #2:... |
| Input | Your Task (only questions and responses are given):Context:Question: What was the basis of the Watergate scandal?Response: ...Question: ...Response: ...Current Question: So what happened to Nixon?Rewrite: So what happened to Nixon after the events of the Watergate scandal?(Now, you should give me the rewrite and an informative response of the **Current Question** based on the **Context**. The output format should always be: Rewrite:\\$Reason. So the question should be rewritten as: \\$Rewrite\\nResponse: \\$Response. Go ahead!) |
| Model Output | Rewrite: Based on all previous turns, Nixon was badly involved in the Watergate scandal. So the question should be rewritten as: So what happened to Nixon after the events of the Watergate scandal?Response: With the mounting evidence and loss... |
| CAsT-20 (Conversation #84) |
| Turn 1 |
| Question: How does seed investment work for startups?Response: Seed money, sometimes known as seed funding or seed capital, is a form of securities offering in which an investor invests capital in exchange for an equity stake in ... Human Rewrite: How does seed investment work for startups?CoT: This is the first turn.Our Rewrite: How does seed investment work for startups? |
| Turn 2 |
| Question: Tell me more about angel rounds. Response: Tell Me More was a National Public Radio interview show hosted by journalist Michel Martin ... offering sample segments, and soliciting user feedback. It began broadcasting on April 30, 2007. Human Rewrite: Tell me more about angel investment rounds. CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: Tell me more about angel rounds in seed investment for startups. |
| Turn 3 |
| Question: How does it compare with crowdfunding? Response: For crowdfunding of equity stock purchases, there is some research in social psychology that indicates that ... Managing communications with a large number of possibly disappointed investors and supporters can be a substantial, and potentially diverting, task. Human Rewrite: How does an angel investment round compare with crowdfunding? CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: How does angel funding compare to crowdfunding for startups? |
| Turn 4 |
| Question: How much is a typical A series round? Response: OpenGov has raised $47 million from venture capital and angel investors. A $3 million Series A round in 2012 included venture funds 8VC, Founder Collective, Valiant Capital, and a number of high-profile ... Human Rewrite: How much is a typical A series investment round? CoT: Based on the previous turns, we are talking about different types of funding for startups such as seed investment, angel rounds, and crowdfunding.Our Rewrite: How much funding is typically raised in an A series round for startups? |
| Turn 5 |
| Question: What are the risks of investing? Response: Angel investors invest in early stage or start-up companies in exchange for an equity ownership interest. Angel ... Here is what angels particularly care about: The quality, passion, commitment, and integrity of the founders. Human Rewrite: What are the risks of investing in startups? CoT: Based on the previous turns, we are discussing different aspects of startup investments. Our Rewrite: What are the risks associated with investing in startups? |
| Turn 6 |
| Question: What are the different types of instruments used by VCs? Response: ... Human Rewrite: What are the different types of instruments used by VCs? CoT: The user is asking about the different types of instruments used by venture capitalists. Our Rewrite: What are the various types of instruments employed by venture capitalists (VCs)? |
| Dataset | CAsT-19 | CAsT-20 | CAsT-21 |
| # Conversations | 20 | 25 | 18 |
| # Turns (Sessions) | 173 | 208 | 157 |
| # Passages/Docs | 38M | 40M |
| System | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | |
| Conversational Dense Retrieval | |||||||||
| ConvDR | 0.740 | 0.466 | 0.362 | 0.510 | 0.340 | 0.345 | 0.573 | 0.385 | 0.483 |
| COTED | 0.769 | 0.478 | 0.367 | 0.491 | 0.342 | 0.340 | 0.565 | 0.371 | 0.485 |
| ZeCo | - | 0.238‡ | 0.216‡ | - | 0.176‡ | 0.200‡ | - | 0.234‡ | 0.267‡ |
| CRDR | 0.765 | 0.472 | 0.357 | 0.501 | 0.350 | 0.313 | 0.474 | 0.342 | 0.380 |
| Conversational Query Rewriting | |||||||||
| T5QR | 0.701 | 0.417 | 0.332 | 0.423 | 0.299 | 0.353 | 0.469 | 0.330 | 0.408 |
| ConvGQR | 0.708 | 0.434 | 0.336 | 0.465 | 0.331 | 0.368 | 0.433 | 0.273 | 0.330 |
| LLM4CS | 0.776† | 0.515† | 0.372† | 0.615† | 0.455† | 0.489† | 0.681† | 0.492† | 0.614† |
| Human | 0.740 | 0.461 | 0.381 | 0.591 | 0.422 | 0.465 | 0.680 | 0.502 | 0.590 |
| RI-H | +4.9% | +11.7% | -2.4% | +4.1% | +7.8% | +5.2% | +0.1% | -2.0% | +4.1% |
| RI-2nd-Best | +0.9% | +7.7% | +1.4% | +20.6% | +30.0% | +32.9% | +18.8% | +27.8% | +26.6% |
| Aggregation | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| REW | RTR | RAR | REW | RTR | RAR | REW | RTR | RAR | |
| MaxProb | 0.441 | 0.459 | 0.464 | 0.356 | 0.415 | 0.430 | 0.407 | 0.469 | 0.462 |
| SC | 0.449 | 0.466 | 0.476 | 0.362 | 0.432 | 0.444 | 0.445 | 0.473 | 0.469 |
| Mean | 0.447 | 0.464 | 0.488 | 0.380 | 0.425 | 0.442 | 0.465 | 0.481 | 0.478 |
| Previous SOTA | 0.478 | 0.350 | 0.385 | ||||||
| Human | 0.461 | 0.422 | 0.502 |
| Instruction | For an information-seeking dialog, please help reformulate the question into rewrite that can fully express the user's information needs without the need of context, but also generate an informative response to answer the question. |
| Demonstration | I will give you several example multi-turn dialogs, where each turn contains a question, a response, and a rewrite. The rewrite part begins with a sentence explaining the reason for the rewrite.Example #1:Question: What should I consider when buying a phone?Rewrite: This is the first turn. So the question should be rewritten as: What should I consider when buying a phone?Response: The design of the phone and the overall ...Question: Cool. Which one would you recommend?Rewrite: Based on Turn 1, you are inquiring about what should be considered when buying a phone. So the question should be rewritten as: Cool. Which smartphone would you recommend for me?Response: Just because a phone has everything...Example #2:... |
| Input | Your Task (only questions and responses are given):Context:Question: What was the basis of the Watergate scandal?Response: ...Question: ...Response: ...Current Question: So what happened to Nixon?Rewrite: So what happened to Nixon after the events of the Watergate scandal?(Now, you should give me the rewrite and an informative response of the **Current Question** based on the **Context**. The output format should always be: Rewrite:\\$Reason. So the question should be rewritten as: \\$Rewrite\\nResponse: \\$Response. Go ahead!) |
| Model Output | Rewrite: Based on all previous turns, Nixon was badly involved in the Watergate scandal. So the question should be rewritten as: So what happened to Nixon after the events of the Watergate scandal?Response: With the mounting evidence and loss... |
| CAsT-20 (Conversation #84) |
| Turn 1 |
| Question: How does seed investment work for startups?Response: Seed money, sometimes known as seed funding or seed capital, is a form of securities offering in which an investor invests capital in exchange for an equity stake in ... Human Rewrite: How does seed investment work for startups?CoT: This is the first turn.Our Rewrite: How does seed investment work for startups? |
| Turn 2 |
| Question: Tell me more about angel rounds. Response: Tell Me More was a National Public Radio interview show hosted by journalist Michel Martin ... offering sample segments, and soliciting user feedback. It began broadcasting on April 30, 2007. Human Rewrite: Tell me more about angel investment rounds. CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: Tell me more about angel rounds in seed investment for startups. |
| Turn 3 |
| Question: How does it compare with crowdfunding? Response: For crowdfunding of equity stock purchases, there is some research in social psychology that indicates that ... Managing communications with a large number of possibly disappointed investors and supporters can be a substantial, and potentially diverting, task. Human Rewrite: How does an angel investment round compare with crowdfunding? CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: How does angel funding compare to crowdfunding for startups? |
| Turn 4 |
| Question: How much is a typical A series round? Response: OpenGov has raised $47 million from venture capital and angel investors. A $3 million Series A round in 2012 included venture funds 8VC, Founder Collective, Valiant Capital, and a number of high-profile ... Human Rewrite: How much is a typical A series investment round? CoT: Based on the previous turns, we are talking about different types of funding for startups such as seed investment, angel rounds, and crowdfunding.Our Rewrite: How much funding is typically raised in an A series round for startups? |
| Turn 5 |
| Question: What are the risks of investing? Response: Angel investors invest in early stage or start-up companies in exchange for an equity ownership interest. Angel ... Here is what angels particularly care about: The quality, passion, commitment, and integrity of the founders. Human Rewrite: What are the risks of investing in startups? CoT: Based on the previous turns, we are discussing different aspects of startup investments. Our Rewrite: What are the risks associated with investing in startups? |
| Turn 6 |
| Question: What are the different types of instruments used by VCs? Response: ... Human Rewrite: What are the different types of instruments used by VCs? CoT: The user is asking about the different types of instruments used by venture capitalists. Our Rewrite: What are the various types of instruments employed by venture capitalists (VCs)? |
| Dataset | CAsT-19 | CAsT-20 | CAsT-21 |
| # Conversations | 20 | 25 | 18 |
| # Turns (Sessions) | 173 | 208 | 157 |
| # Passages/Docs | 38M | 40M | |
| System | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | |
| Conversational Dense Retrieval | |||||||||
| ConvDR | 0.740 | 0.466 | 0.362 | 0.510 | 0.340 | 0.345 | 0.573 | 0.385 | 0.483 |
| COTED | 0.769 | 0.478 | 0.367 | 0.491 | 0.342 | 0.340 | 0.565 | 0.371 | 0.485 |
| ZeCo | - | 0.238‡ | 0.216‡ | - | 0.176‡ | 0.200‡ | - | 0.234‡ | 0.267‡ |
| CRDR | 0.765 | 0.472 | 0.357 | 0.501 | 0.350 | 0.313 | 0.474 | 0.342 | 0.380 |
| Conversational Query Rewriting | |||||||||
| T5QR | 0.701 | 0.417 | 0.332 | 0.423 | 0.299 | 0.353 | 0.469 | 0.330 | 0.408 |
| ConvGQR | 0.708 | 0.434 | 0.336 | 0.465 | 0.331 | 0.368 | 0.433 | 0.273 | 0.330 |
| LLM4CS | 0.776† | 0.515† | 0.372† | 0.615† | 0.455† | 0.489† | 0.681† | 0.492† | 0.614† |
| Human | 0.740 | 0.461 | 0.381 | 0.591 | 0.422 | 0.465 | 0.680 | 0.502 | 0.590 |
| RI-H | +4.9% | +11.7% | -2.4% | +4.1% | +7.8% | +5.2% | +0.1% | -2.0% | +4.1% |
| RI-2nd-Best | +0.9% | +7.7% | +1.4% | +20.6% | +30.0% | +32.9% | +18.8% | +27.8% | +26.6% |
| Aggregation | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| REW | RTR | RAR | REW | RTR | RAR | REW | RTR | RAR | |
| MaxProb | 0.441 | 0.459 | 0.464 | 0.356 | 0.415 | 0.430 | 0.407 | 0.469 | 0.462 |
| SC | 0.449 | 0.466 | 0.476 | 0.362 | 0.432 | 0.444 | 0.445 | 0.473 | 0.469 |
| Mean | 0.447 | 0.464 | 0.488 | 0.380 | 0.425 | 0.442 | 0.465 | 0.481 | 0.478 |
| Previous SOTA | 0.478 | 0.350 | 0.385 | ||||||
| Human | 0.461 | 0.422 | 0.502 | ||||||
| Instruction | For an information-seeking dialog, please help reformulate the question into rewrite that can fully express the user's information needs without the need of context, but also generate an informative response to answer the question. |
| Demonstration | I will give you several example multi-turn dialogs, where each turn contains a question, a response, and a rewrite. The rewrite part begins with a sentence explaining the reason for the rewrite.Example #1:Question: What should I consider when buying a phone?Rewrite: This is the first turn. So the question should be rewritten as: What should I consider when buying a phone?Response: The design of the phone and the overall ...Question: Cool. Which one would you recommend?Rewrite: Based on Turn 1, you are inquiring about what should be considered when buying a phone. So the question should be rewritten as: Cool. Which smartphone would you recommend for me?Response: Just because a phone has everything...Example #2:... |
| Input | Your Task (only questions and responses are given):Context:Question: What was the basis of the Watergate scandal?Response: ...Question: ...Response: ...Current Question: So what happened to Nixon?Rewrite: So what happened to Nixon after the events of the Watergate scandal?(Now, you should give me the rewrite and an informative response of the **Current Question** based on the **Context**. The output format should always be: Rewrite:\$Reason. So the question should be rewritten as: \$Rewrite\nResponse: \$Response. Go ahead!) |
| Model Output | Rewrite: Based on all previous turns, Nixon was badly involved in the Watergate scandal. So the question should be rewritten as: So what happened to Nixon after the events of the Watergate scandal?Response: With the mounting evidence and loss... |
| CAsT-20 (Conversation #84) |
| Turn 1 |
| Question: How does seed investment work for startups?Response: Seed money, sometimes known as seed funding or seed capital, is a form of securities offering in which an investor invests capital in exchange for an equity stake in ... Human Rewrite: How does seed investment work for startups?CoT: This is the first turn.Our Rewrite: How does seed investment work for startups? |
| Turn 2 |
| Question: Tell me more about angel rounds. Response: Tell Me More was a National Public Radio interview show hosted by journalist Michel Martin ... offering sample segments, and soliciting user feedback. It began broadcasting on April 30, 2007. Human Rewrite: Tell me more about angel investment rounds. CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: Tell me more about angel rounds in seed investment for startups. |
| Turn 3 |
| Question: How does it compare with crowdfunding? Response: For crowdfunding of equity stock purchases, there is some research in social psychology that indicates that ... Managing communications with a large number of possibly disappointed investors and supporters can be a substantial, and potentially diverting, task. Human Rewrite: How does an angel investment round compare with crowdfunding? CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: How does angel funding compare to crowdfunding for startups? |
| Turn 4 |
| Question: How much is a typical A series round? Response: OpenGov has raised $47 million from venture capital and angel investors. A $3 million Series A round in 2012 included venture funds 8VC, Founder Collective, Valiant Capital, and a number of high-profile ... Human Rewrite: How much is a typical A series investment round? CoT: Based on the previous turns, we are talking about different types of funding for startups such as seed investment, angel rounds, and crowdfunding.Our Rewrite: How much funding is typically raised in an A series round for startups? |
| Turn 5 |
| Question: What are the risks of investing? Response: Angel investors invest in early stage or start-up companies in exchange for an equity ownership interest. Angel ... Here is what angels particularly care about: The quality, passion, commitment, and integrity of the founders. Human Rewrite: What are the risks of investing in startups? CoT: Based on the previous turns, we are discussing different aspects of startup investments. Our Rewrite: What are the risks associated with investing in startups? |
| Turn 6 |
| Question: What are the different types of instruments used by VCs? Response: ... Human Rewrite: What are the different types of instruments used by VCs? CoT: The user is asking about the different types of instruments used by venture capitalists. Our Rewrite: What are the various types of instruments employed by venture capitalists (VCs)? |
| Dataset | CAsT-19 | CAsT-20 | CAsT-21 |
| # Conversations | 20 | 25 | 18 |
| # Turns (Sessions) | 173 | 208 | 157 |
| # Passages/Docs | 38M | 40M |
| System | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | MRR | NDCG@3 | R@100 | |
| Conversational Dense Retrieval | |||||||||
| ConvDR | 0.740 | 0.466 | 0.362 | 0.510 | 0.340 | 0.345 | 0.573 | 0.385 | 0.483 |
| COTED | 0.769 | 0.478 | 0.367 | 0.491 | 0.342 | 0.340 | 0.565 | 0.371 | 0.485 |
| ZeCo | - | 0.238‡ | 0.216‡ | - | 0.176‡ | 0.200‡ | - | 0.234‡ | 0.267‡ |
| CRDR | 0.765 | 0.472 | 0.357 | 0.501 | 0.350 | 0.313 | 0.474 | 0.342 | 0.380 |
| Conversational Query Rewriting | |||||||||
| T5QR | 0.701 | 0.417 | 0.332 | 0.423 | 0.299 | 0.353 | 0.469 | 0.330 | 0.408 |
| ConvGQR | 0.708 | 0.434 | 0.336 | 0.465 | 0.331 | 0.368 | 0.433 | 0.273 | 0.330 |
| LLM4CS | 0.776† | 0.515† | 0.372† | 0.615† | 0.455† | 0.489† | 0.681† | 0.492† | 0.614† |
| Human | 0.740 | 0.461 | 0.381 | 0.591 | 0.422 | 0.465 | 0.680 | 0.502 | 0.590 |
| RI-H | +4.9% | +11.7% | -2.4% | +4.1% | +7.8% | +5.2% | +0.1% | -2.0% | +4.1% |
| RI-2nd-Best | +0.9% | +7.7% | +1.4% | +20.6% | +30.0% | +32.9% | +18.8% | +27.8% | +26.6% |
| Aggregation | CAsT-19 | CAsT-20 | CAsT-21 | ||||||
| REW | RTR | RAR | REW | RTR | RAR | REW | RTR | RAR | |
| MaxProb | 0.441 | 0.459 | 0.464 | 0.356 | 0.415 | 0.430 | 0.407 | 0.469 | 0.462 |
| SC | 0.449 | 0.466 | 0.476 | 0.362 | 0.432 | 0.444 | 0.445 | 0.473 | 0.469 |
| Mean | 0.447 | 0.464 | 0.488 | 0.380 | 0.425 | 0.442 | 0.465 | 0.481 | 0.478 |
| Previous SOTA | 0.478 | 0.350 | 0.385 | ||||||
| Human | 0.461 | 0.422 | 0.502 |
| Instruction | For an information-seeking dialog, please help reformulate the question into rewrite that can fully express the user's information needs without the need of context, but also generate an informative response to answer the question. |
| Demonstration | I will give you several example multi-turn dialogs, where each turn contains a question, a response, and a rewrite. The rewrite part begins with a sentence explaining the reason for the rewrite.Example #1:Question: What should I consider when buying a phone?Rewrite: This is the first turn. So the question should be rewritten as: What should I consider when buying a phone?Response: The design of the phone and the overall ...Question: Cool. Which one would you recommend?Rewrite: Based on Turn 1, you are inquiring about what should be considered when buying a phone. So the question should be rewritten as: Cool. Which smartphone would you recommend for me?Response: Just because a phone has everything...Example #2:... |
| Input | Your Task (only questions and responses are given):Context:Question: What was the basis of the Watergate scandal?Response: ...Question: ...Response: ...Current Question: So what happened to Nixon?Rewrite: So what happened to Nixon after the events of the Watergate scandal?(Now, you should give me the rewrite and an informative response of the **Current Question** based on the **Context**. The output format should always be: Rewrite:\\$Reason. So the question should be rewritten as: \\$Rewrite\\nResponse: \\$Response. Go ahead!) |
| Model Output | Rewrite: Based on all previous turns, Nixon was badly involved in the Watergate scandal. So the question should be rewritten as: So what happened to Nixon after the events of the Watergate scandal?Response: With the mounting evidence and loss... |
| CAsT-20 (Conversation #84) |
| Turn 1 |
| Question: How does seed investment work for startups?Response: Seed money, sometimes known as seed funding or seed capital, is a form of securities offering in which an investor invests capital in exchange for an equity stake in ... Human Rewrite: How does seed investment work for startups?CoT: This is the first turn.Our Rewrite: How does seed investment work for startups? |
| Turn 2 |
| Question: Tell me more about angel rounds. Response: Tell Me More was a National Public Radio interview show hosted by journalist Michel Martin ... offering sample segments, and soliciting user feedback. It began broadcasting on April 30, 2007. Human Rewrite: Tell me more about angel investment rounds. CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: Tell me more about angel rounds in seed investment for startups. |
| Turn 3 |
| Question: How does it compare with crowdfunding? Response: For crowdfunding of equity stock purchases, there is some research in social psychology that indicates that ... Managing communications with a large number of possibly disappointed investors and supporters can be a substantial, and potentially diverting, task. Human Rewrite: How does an angel investment round compare with crowdfunding? CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: How does angel funding compare to crowdfunding for startups? |
| Turn 4 |
| Question: How much is a typical A series round? Response: OpenGov has raised $47 million from venture capital and angel investors. A $3 million Series A round in 2012 included venture funds 8VC, Founder Collective, Valiant Capital, and a number of high-profile ... Human Rewrite: How much is a typical A series investment round? CoT: Based on the previous turns, we are talking about different types of funding for startups such as seed investment, angel rounds, and crowdfunding.Our Rewrite: How much funding is typically raised in an A series round for startups? |
| Turn 5 |
| Question: What are the risks of investing? Response: Angel investors invest in early stage or start-up companies in exchange for an equity ownership interest. Angel ... Here is what angels particularly care about: The quality, passion, commitment, and integrity of the founders. Human Rewrite: What are the risks of investing in startups? CoT: Based on the previous turns, we are discussing different aspects of startup investments. Our Rewrite: What are the risks associated with investing in startups? |
| Turn 6 |
| Question: What are the different types of instruments used by VCs? Response: ... Human Rewrite: What are the different types of instruments used by VCs? CoT: The user is asking about the different types of instruments used by venture capitalists. Our Rewrite: What are the various types of instruments employed by venture capitalists (VCs)? |
| Dataset | Sce. | Att. | Re. | Dy. | Sl. |
| PchatbotW (2021) | × | √ | × | × | × |
| PeDialog (2019) | × | √ | × | × | × |
| KvPI (2020) | × | √ | × | × | × |
| P-CHAT (2018) | × | √ | × | × | × |
| WOW (2019) | √ | √ | × | × | × |
| Fri-QA (2019) | × | √ | √ | × | × |
| Focus (2021) | √ | √ | √ | × | × |
| UltraChat(2023) | √ | √ | × | × | × |
| LaMP(2023) | √ | √ | × | × | × |
| Ours | √ | √ | √ | √ | √ |
| Statistics | Train | Test |
| per dialogue | ||
| Average Turns | 13.8 | 7.8 |
| Maximum Speakers | 20 | 8 |
| Minimum Speakers | 2 | 2 |
| per utterance | ||
| Average Length | 32.9 | 28.3 |
| Maximum Length | 77 | 26 |
| Minimum Length | 3 | 3 |
| Total Dialogues | 1042 | 149 |
| Model | Automatic Evaluation | GPT-4 Evaluation | Human Evaluation | ||||||
| Bleu-1 | Dist.1 | Rough-L | Relv.Sce. | Relv.Attr. | Relv.Re. | Relv.Sce. | Relv.Attr. | Relv.Re. | |
| Fine-tuning | |||||||||
| CGLM*-6B | 1.2 | 18.4 | 3.4 | 1.21 | 1.10 | 1.01 | 1.34 | 0.67 | 0.67 |
| Alpaca | 2.0 | 30.2 | 10.8 | 3.37 | 1.19 | 1.51 | 2.03 | 1.19 | 1.51 |
| Per-Alpaca | 12.2 | 20.4 | 13.8 | 3.51 | 1.34 | 1.51 | 4.04 | 1.34 | 1.51 |
| In-context learning | |||||||||
| CGLM*-130B | 14.2 | 25.1 | 15.7 | 6.71 | 6.01 | 6.04 | 7.38 | 6.04 | 5.37 |
| GPT3 | 9.8 | 23.0 | 14.6 | 2.68 | 2.68 | 3.36 | 4.70 | 4.70 | 4.03 |
| Per-GPT3 | 22.6 | 20.1 | 16.5 | 12.75 | 6.04 | 3.36 | 12.08 | 6.71 | 5.37 |
| ChatGPT | 33.1 | 19.0 | 20.2 | 33.56 | 38.92 | 37.58 | 32.89 | 34.23 | 32.89 |
| Per-ChatGPT | 33.6 | 19.4 | 22.4 | 37.58 | 42.95 | 46.98 | 35.57 | 45.63 | 48.99 |
| Category | Relv.Sce. | Relv.Attr. | Relv.Re. | Ave. |
| WIN | 77.85 | 77.18 | 73.15 | 76.06 |
| TIE | 1.34 | 4.03 | 8.05 | 4.47 |
| LOSE | 20.81 | 18.79 | 18.79 | 19.46 |
| Scene: Harry and his family visit the reptile house at the zoo, where Harry strikes up a conversation with a boa constrictor from Brazil... Harry is blamed for talking to the snake and causing the incident. |
| Dialogue Position: Book1-chapter2 |
| Speakers: Dudley, Harry...Piers |
| Dialogue: Dudley: Make it move... |
| the snake: I get that all the time. |
| Harry: I know, Where do you come from, anyway? |
| the snake: Boa Constrictor, Brazil. |
| Harry: Was it nice there? Oh, I see — so you’ve never been to Brazil? |
| keeper of the reptile house: DUDLEY! MR. DURSLEY! |
| COME AND LOOK AT THIS SNAKE! YOU WON’TE BELIEVE WHAT IT’S DOING! |
| the snake: Brazil, here I come. . . . Thanksss, amigo. |
| Piers: Harry was talking to it, weren’t you, Harry? |
| GPT-3: I was just asking it where it came from. It's not like I was trying to make friends with it or anything. |
| Per-GPT3: I didn’t mean to cause any trouble, I was just curious about where the snake came from. I didn’t know it would escape. |
| ChatGPT: Yes, I was just curious about where it came from and if it had ever been to Brazil. But Dudley, you shouldn't be so rude to the keeper of the reptile house. He was just trying to show us something amazing. |
| Per-ChatGPT: Yes, I was. But I didn't make the glass disappear or tell the snake to escape. It was just a harmless conversation. I didn't mean for any of this to happen. |
| Annotated Response: No, I wasn't talking to it. I was just looking at it and reading the sign about where it came from. You two must have imagined things. |
| Parameter | Alpaca | Per-Alpaca | ChatGLM-6B | BERT-FP |
| Batch size | 4 | 4 | 1 | 32 |
| Learning Rate | 1e-4 | 1e-4 | 3e-5 | 8e-5 |
| Epoch | 5 | 5 | 5 | 20 |
| Max Length | 512 | 512 | 650 | 256 |
| Retrieval-based | |||||
| Model | MAP | MRR | P@1 | R10@1 | R10@5 |
| BERT-FP | 0.468 | 0.468 | 0.259 | 0.259 | 0.788 |
| Dataset | Sce. | Att. | Re. | Dy. | Sl. |
| PchatbotW (2021) | × | √ | × | × | × |
| PeDialog (2019) | × | √ | × | × | × |
| KvPI (2020) | × | √ | × | × | × |
| P-CHAT (2018) | × | √ | × | × | × |
| WOW (2019) | √ | √ | × | × | × |
| Fri-QA (2019) | × | √ | √ | × | × |
| Focus (2021) | √ | √ | √ | × | × |
| UltraChat(2023) | √ | √ | × | × | × |
| LaMP(2023) | √ | √ | × | × | × |
| Ours | √ | √ | √ | √ | √ |
| Statistics | Train | Test |
| per dialogue | ||
| Average Turns | 13.8 | 7.8 |
| Maximum Speakers | 20 | 8 |
| Minimum Speakers | 2 | 2 |
| per utterance | ||
| Average Length | 32.9 | 28.3 |
| Maximum Length | 77 | 26 |
| Minimum Length | 3 | 3 |
| Total Dialogues | 1042 | 149 |
| Model | Automatic Evaluation | GPT-4 Evaluation | Human Evaluation | ||||||
| Bleu-1 | Dist.1 | Rough-L | Relv.Sce. | Relv.Attr. | Relv.Re. | Relv.Sce. | Relv.Attr. | Relv.Re. | |
| Fine-tuning | |||||||||
| CGLM*-6B | 1.2 | 18.4 | 3.4 | 1.21 | 1.10 | 1.01 | 1.34 | 0.67 | 0.67 |
| Alpaca | 2.0 | 30.2 | 10.8 | 3.37 | 1.19 | 1.51 | 2.03 | 1.19 | 1.51 |
| Per-Alpaca | 12.2 | 20.4 | 13.8 | 3.51 | 1.34 | 1.51 | 4.04 | 1.34 | 1.51 |
| In-context learning | |||||||||
| CGLM*-130B | 14.2 | 25.1 | 15.7 | 6.71 | 6.01 | 6.04 | 7.38 | 6.04 | 5.37 |
| GPT3 | 9.8 | 23.0 | 14.6 | 2.68 | 2.68 | 3.36 | 4.70 | 4.70 | 4.03 |
| Per-GPT3 | 22.6 | 20.1 | 16.5 | 12.75 | 6.04 | 3.36 | 12.08 | 6.71 | 5.37 |
| ChatGPT | 33.1 | 19.0 | 20.2 | 33.56 | 38.92 | 37.58 | 32.89 | 34.23 | 32.89 |
| Per-ChatGPT | 33.6 | 19.4 | 22.4 | 37.58 | 42.95 | 46.98 | 35.57 | 45.63 | 48.99 |
| Category | Relv.Sce. | Relv.Attr. | Relv.Re. | Ave. |
| WIN | 77.85 | 77.18 | 73.15 | 76.06 |
| TIE | 1.34 | 4.03 | 8.05 | 4.47 |
| LOSE | 20.81 | 18.79 | 18.79 | 19.46 |
| Scene: Harry and his family visit the reptile house at the zoo, where Harry strikes up a conversation with a boa constrictor from Brazil... Harry is blamed for talking to the snake and causing the incident. |
| Dialogue Position: Book1-chapter2 |
| Speakers: Dudley, Harry...Piers |
| Dialogue: Dudley: Make it move... |
| the snake: I get that all the time. |
| Harry: I know, Where do you come from, anyway? |
| the snake: Boa Constrictor, Brazil. |
| Harry: Was it nice there? Oh, I see — so you’ve never been to Brazil? |
| keeper of the reptile house: DUDLEY! MR. DURSLEY! |
| COME AND LOOK AT THIS SNAKE! YOU WON’TE BELIEVE WHAT IT’S DOING! |
| the snake: Brazil, here I come. . . . Thanksss, amigo. |
| Piers: Harry was talking to it, weren’t you, Harry? |
| GPT-3: I was just asking it where it came from. It's not like I was trying to make friends with it or anything. |
| Per-GPT3: I didn’t mean to cause any trouble, I was just curious about where the snake came from. I didn’t know it would escape. |
| ChatGPT: Yes, I was just curious about where it came from and if it had ever been to Brazil. But Dudley, you shouldn't be so rude to the keeper of the reptile house. He was just trying to show us something amazing. |
| Per-ChatGPT: Yes, I was. But I didn't make the glass disappear or tell the snake to escape. It was just a harmless conversation. I didn't mean for any of this to happen. |
| Annotated Response: No, I wasn't talking to it. I was just looking at it and reading the sign about where it came from. You two must have imagined things. |
| Parameter | Alpaca | Per-Alpaca | ChatGLM-6B | BERT-FP |
| Batch size | 4 | 4 | 1 | 32 |
| Learning Rate | 1e-4 | 1e-4 | 3e-5 | 8e-5 |
| Epoch | 5 | 5 | 5 | 20 |
| Max Length | 512 | 512 | 650 | 256 |
| Retrieval-based | |||||
| Model | MAP | MRR | P@1 | R10@1 | R10@5 |
| BERT-FP | 0.468 | 0.468 | 0.259 | 0.259 | 0.788 |
| Dataset | Sce. | Att. | Re. | Dy. | Sl. |
| PchatbotW (2021) | × | √ | × | × | × |
| PeDialog (2019) | × | √ | × | × | × |
| KvPI (2020) | × | √ | × | × | × |
| P-CHAT (2018) | × | √ | × | × | × |
| WOW (2019) | √ | √ | × | × | × |
| Fri-QA (2019) | × | √ | √ | × | × |
| Focus (2021) | √ | √ | √ | × | × |
| UltraChat(2023) | √ | √ | × | × | × |
| LaMP(2023) | √ | √ | × | × | × |
| Ours | √ | √ | √ | √ | √ |
| Statistics | Train | Test |
| per dialogue | ||
| Average Turns | 13.8 | 7.8 |
| Maximum Speakers | 20 | 8 |
| Minimum Speakers | 2 | 2 |
| per utterance | ||
| Average Length | 32.9 | 28.3 |
| Maximum Length | 77 | 26 |
| Minimum Length | 3 | 3 |
| Total Dialogues | 1042 | 149 |
| Model | Automatic Evaluation | GPT-4 Evaluation | Human Evaluation | ||||||
| Bleu-1 | Dist.1 | Rough-L | Relv.Sce. | Relv.Attr. | Relv.Re. | Relv.Sce. | Relv.Attr. | Relv.Re. | |
| Fine-tuning | |||||||||
| CGLM*-6B | 1.2 | 18.4 | 3.4 | 1.21 | 1.10 | 1.01 | 1.34 | 0.67 | 0.67 |
| Alpaca | 2.0 | 30.2 | 10.8 | 3.37 | 1.19 | 1.51 | 2.03 | 1.19 | 1.51 |
| Per-Alpaca | 12.2 | 20.4 | 13.8 | 3.51 | 1.34 | 1.51 | 4.04 | 1.34 | 1.51 |
| In-context learning | |||||||||
| CGLM*-130B | 14.2 | 25.1 | 15.7 | 6.71 | 6.01 | 6.04 | 7.38 | 6.04 | 5.37 |
| GPT3 | 9.8 | 23.0 | 14.6 | 2.68 | 2.68 | 3.36 | 4.70 | 4.70 | 4.03 |
| Per-GPT3 | 22.6 | 20.1 | 16.5 | 12.75 | 6.04 | 3.36 | 12.08 | 6.71 | 5.37 |
| ChatGPT | 33.1 | 19.0 | 20.2 | 33.56 | 38.92 | 37.58 | 32.89 | 34.23 | 32.89 |
| Per-ChatGPT | 33.6 | 19.4 | 22.4 | 37.58 | 42.95 | 46.98 | 35.57 | 45.63 | 48.99 |
| Category | Relv.Sce. | Relv.Attr. | Relv.Re. | Ave. |
| WIN | 77.85 | 77.18 | 73.15 | 76.06 |
| TIE | 1.34 | 4.03 | 8.05 | 4.47 |
| LOSE | 20.81 | 18.79 | 18.79 | 19.46 |
| Scene: Harry and his family visit the reptile house at the zoo, where Harry strikes up a conversation with a boa constrictor from Brazil... Harry is blamed for talking to the snake and causing the incident. |
| Dialogue Position: Book1-chapter2 |
| Speakers: Dudley, Harry...Piers |
| Dialogue: Dudley: Make it move... |
| the snake: I get that all the time. |
| Harry: I know, Where do you come from, anyway? |
| the snake: Boa Constrictor, Brazil. |
| Harry: Was it nice there? Oh, I see — so you’ve never been to Brazil? |
| keeper of the reptile house: DUDLEY! MR. DURSLEY! |
| COME AND LOOK AT THIS SNAKE! YOU WON’TE BELIEVE WHAT IT’S DOING! |
| the snake: Brazil, here I come. . . . Thanksss, amigo. |
| Piers: Harry was talking to it, weren’t you, Harry? |
| GPT-3: I was just asking it where it came from. It's not like I was trying to make friends with it or anything. |
| Per-GPT3: I didn’t mean to cause any trouble, I was just curious about where the snake came from. I didn’t know it would escape. |
| ChatGPT: Yes, I was just curious about where it came from and if it had ever been to Brazil. But Dudley, you shouldn't be so rude to the keeper of the reptile house. He was just trying to show us something amazing. |
| Per-ChatGPT: Yes, I was. But I didn't make the glass disappear or tell the snake to escape. It was just a harmless conversation. I didn't mean for any of this to happen. |
| Annotated Response: No, I wasn't talking to it. I was just looking at it and reading the sign about where it came from. You two must have imagined things. |
| Parameter | Alpaca | Per-Alpaca | ChatGLM-6B | BERT-FP |
| Batch size | 4 | 4 | 1 | 32 |
| Learning Rate | 1e-4 | 1e-4 | 3e-5 | 8e-5 |
| Epoch | 5 | 5 | 5 | 20 |
| Max Length | 512 | 512 | 650 | 256 |
| Retrieval-based | |||||
| Model | MAP | MRR | P@1 | R10@1 | R10@5 |
| BERT-FP | 0.468 | 0.468 | 0.259 | 0.259 | 0.788 |
| Dataset | Sce. | Att. | Re. | Dy. | Sl. |
| PchatbotW (2021) | × | √ | × | × | × |
| PeDialog (2019) | × | √ | × | × | × |
| KvPI (2020) | × | √ | × | × | × |
| P-CHAT (2018) | × | √ | × | × | × |
| WOW (2019) | √ | √ | × | × | × |
| Fri-QA (2019) | × | √ | √ | × | × |
| Focus (2021) | √ | √ | √ | × | × |
| UltraChat(2023) | √ | √ | × | × | × |
| LaMP(2023) | √ | √ | × | × | × |
| Ours | √ | √ | √ | √ | √ |
| Statistics | Train | Test |
| per dialogue | ||
| Average Turns | 13.8 | 7.8 |
| Maximum Speakers | 20 | 8 |
| Minimum Speakers | 2 | 2 |
| per utterance | ||
| Average Length | 32.9 | 28.3 |
| Maximum Length | 77 | 26 |
| Minimum Length | 3 | 3 |
| Total Dialogues | 1042 | 149 |
| Model | Automatic Evaluation | GPT-4 Evaluation | Human Evaluation | ||||||
| Bleu-1 | Dist.1 | Rough-L | Relv.Sce. | Relv.Attr. | Relv.Re. | Relv.Sce. | Relv.Attr. | Relv.Re. | |
| Fine-tuning | |||||||||
| CGLM*-6B | 1.2 | 18.4 | 3.4 | 1.21 | 1.10 | 1.01 | 1.34 | 0.67 | 0.67 |
| Alpaca | 2.0 | 30.2 | 10.8 | 3.37 | 1.19 | 1.51 | 2.03 | 1.19 | 1.51 |
| Per-Alpaca | 12.2 | 20.4 | 13.8 | 3.51 | 1.34 | 1.51 | 4.04 | 1.34 | 1.51 |
| In-context learning | |||||||||
| CGLM*-130B | 14.2 | 25.1 | 15.7 | 6.71 | 6.01 | 6.04 | 7.38 | 6.04 | 5.37 |
| GPT3 | 9.8 | 23.0 | 14.6 | 2.68 | 2.68 | 3.36 | 4.70 | 4.70 | 4.03 |
| Per-GPT3 | 22.6 | 20.1 | 16.5 | 12.75 | 6.04 | 3.36 | 12.08 | 6.71 | 5.37 |
| ChatGPT | 33.1 | 19.0 | 20.2 | 33.56 | 38.92 | 37.58 | 32.89 | 34.23 | 32.89 |
| Per-ChatGPT | 33.6 | 19.4 | 22.4 | 37.58 | 42.95 | 46.98 | 35.57 | 45.63 | 48.99 |
| Category | Relv.Sce. | Relv.Attr. | Relv.Re. | Ave. |
| WIN | 77.85 | 77.18 | 73.15 | 76.06 |
| TIE | 1.34 | 4.03 | 8.05 | 4.47 |
| LOSE | 20.81 | 18.79 | 18.79 | 19.46 |
| Scene: Harry and his family visit the reptile house at the zoo, where Harry strikes up a conversation with a boa constrictor from Brazil... Harry is blamed for talking to the snake and causing the incident. |
| Dialogue Position: Book1-chapter2 |
| Speakers: Dudley, Harry...Piers |
| Dialogue: Dudley: Make it move... |
| the snake: I get that all the time. |
| Harry: I know, Where do you come from, anyway? |
| the snake: Boa Constrictor, Brazil. |
| Harry: Was it nice there? Oh, I see — so you’ve never been to Brazil? |
| keeper of the reptile house: DUDLEY! MR. DURSLEY! |
| COME AND LOOK AT THIS SNAKE! YOU WON’TE BELIEVE WHAT IT’S DOING! |
| the snake: Brazil, here I come. . . . Thanksss, amigo. |
| Piers: Harry was talking to it, weren’t you, Harry? |
| GPT-3: I was just asking it where it came from. It's not like I was trying to make friends with it or anything. |
| Per-GPT3: I didn’t mean to cause any trouble, I was just curious about where the snake came from. I didn’t know it would escape. |
| ChatGPT: Yes, I was just curious about where it came from and if it had ever been to Brazil. But Dudley, you shouldn't be so rude to the keeper of the reptile house. He was just trying to show us something amazing. |
| Per-ChatGPT: Yes, I was. But I didn't make the glass disappear or tell the snake to escape. It was just a harmless conversation. I didn't mean for any of this to happen. |
| Annotated Response: No, I wasn't talking to it. I was just looking at it and reading the sign about where it came from. You two must have imagined things. |
| Parameter | Alpaca | Per-Alpaca | ChatGLM-6B | BERT-FP |
| Batch size | 4 | 4 | 1 | 32 |
| Learning Rate | 1e-4 | 1e-4 | 3e-5 | 8e-5 |
| Epoch | 5 | 5 | 5 | 20 |
| Max Length | 512 | 512 | 650 | 256 |
| Retrieval-based | |||||
| Model | MAP | MRR | P@1 | R10@1 | R10@5 |
| BERT-FP | 0.468 | 0.468 | 0.259 | 0.259 | 0.788 |
| Method | Arithmetic Tasks | General Tasks | |||||||
| GSM8K | SingleEq | AddSub | MultiArith | AQUA-RAT | SVAMP | CSQA | DU | ||
| Previous SOTA (Fine-tune) | 35a/57b | 32.5c | 94.9d | 60.5e | 37.9f | 57.4g | 91.2h | - | |
| 9-12 year olds | 60i | - | - | - | - | - | - | - | |
| GPT-3 Standard | 19.7 | 86.8 | 90.9 | 44.0 | 29.5 | 69.9 | 82.3 | 49.0 | |
| GPT-3 (175B) code-davinci-001 | CoT | 13.84 | 60.20 | 58.55 | 45.85 | 18.90 | 38.42 | 46.75 | 38.72 |
| CoT+Self-Verification | 13.92 (+0.08) | 60.61 (+0.41) | 59.07 (+0.52) | 46.19 (+0.34) | 27.04 (+8.14) | 38.96 (+0.54) | 47.68 (+0.93) | 39.03 (+0.31) | |
| Instruct-GPT (175B) code-davinci-002 | CoT | 60.81 | 91.01 | 82.78 | 96.13 | 45.30 | 75.87 | 77.42 | 65.43 |
| CoT+Self-Verification | 65.14 (+4.33) | 93.40 (+2.39) | 86.33 (+3.55) | 99.15 (+3.02) | 47.95 (+2.65) | 76.99 (+1.12) | 77.83 (+0.41) | 66.57 (+1.14) |
| Self-Consistency Decoding (Wang et al., 2023c) For Forward Reasoning | |||||||||
| GPT-3 (175B) code-davinci-001 | SC | 23.40 | 70.25 | 68.65 | 79.82 | 25.60 | 54.58 | 54.92 | 49.26 |
| SC+ Self-Verification | 23.59(+0.19) | 70.50(+0.25) | 68.71(+0.06) | 80.01(+0.19) | 28.98(+3.38) | 54.68(+0.1) | 55.09(+0.17) | 49.72(+0.46) | |
| Instruct-GPT (175B) code-davinci-002 | SC | 78.00 | 96.78 | 91.64 | 100.0 | 52.01 | 86.77 | 81.43 | 71.58 |
| SC+Self-Verification | 78.32(+0.32) | 96.85(+0.07) | 92.03(+0.39) | 100.0(+0.0) | 52.25(+0.24) | 86.94(+0.17) | 81.53(+0.1) | 71.89(+0.31) |
| PAL (Gao et al., 2023) For Forward Reasoning | ||||||||
| GPT-3 (175B) code-davinci-001 | PAL | 31.82 | 63.98 | 63.15 | 61.52 | 30.56 | 42.69 | - |
| PAL+Self-Verification | 32.87(+1.05) | 65.45(+1.47) | 64.15(+1.0) | 61.76(+0.24) | 30.90(+0.34) | 42.78(+0.09) | - | |
| Instruct-GPT (175B) code-davinci-002 | PAL | 72.02 | 96.08 | 92.64 | 99.15 | 59.75 | 79.45 | - |
| PAL+Self-Verification | 72.89(+0.87) | 96.52(+0.44) | 93.78(+1.14) | 99.87(+0.72) | 60.21(+0.46) | 80.24(+0.79) | - |
| Dataset | Num. | Avg. Ans. | Ran. Acc. | CoT Acc. | Veri. Acc. |
| GSM8K | 601/1306 | 2.80 | 35.7% | 53.4% | 58.9% |
| Addsub | 77/377 | 2.13 | 46.9% | 56.7% | 74.0% |
| AQUA | 123/219 | 2.71 | 36.9% | 50.3% | 51.2% |
| Multiarith | 73/593 | 2.19 | 45.6% | 71.2% | 89.0% |
| SingleEq | 51/501 | 2.31 | 43.3% | 51.0% | 74.5% |
| SVAMP | 274/954 | 2.54 | 39.4% | 51.5% | 58.4% |
| AnswerFormat | Answer CleansingApproach | Pseudo Code(Example in Pytorch 3.7) |
| Number | Pick up the first numberencountered in the text. | pred = pred.replace("\\\\",".")\npred = [s for s in re.findall(r'^-?\\\\d+\\.?\\\\d*', pred)]\npred = pred[0] |
| Multiple-Choice | Pick up the first large letterencountered in the text. | pred = re.findall(r'A|B|C|D|E', pred)\npred = pred[0] |
| True or False | Pick up the first "True" or"False" encountered in thetext after removingunnec-essary letters. | pred = pred.lower()\npred = re.sub("\\\\\\"|\\n|\\.\\s|:\\\\","\\"", pred)\npred = pred.split("\\\\.")\npred = [i for i in pred if i in ("True", "False")] \npred = pred[0] |
| Yes or No | Pick up the first "yes" or"no" encountered in thetext after removingunnec-essary letters. | pred = pred.lower()\npred = re.sub("\\\\\\"|\\n|\\.\\s|:\\\\","\\"", pred)\npred = pred.split("\\\\.")\npred = [i for i in pred if i in ("yes", "no"]\npred = pred[0] |
| Free Format | Just remove unnecessaryletters, such as ".". | pred = re.sub("\\\\\\"|\\n|\\.\\s", "", pred) |
| True-False Item Verification | If Tim will need 36 lemons, and he will need to make 4 gallons of lemonade for the party, so he will need \\( 4 * 6 = {24} \\) lemons,he will need to make an extra gallonfor him that was twice as tart as the other gallons, so he will need \\( {24} * 2 = {48} \\) lemons, \\( {48}! = {36} \\) ,which is different from the theme. The answer is False. [X] |
| Conditional Masked Verification | Tim wanted to make lemonade for a pool party. For a gallon of lemonade, his recipe called for X cup of fresh lemon juice. He found that 6 lemons would yield 1 cup of juice. He figured he would need to make 4 gallons of lemonade for the party. His best friend Allen asked if Tim could make an extra gallon for him that was twice as tart as the other gallons. Tim will need 36.0 lemons. \\( \\mathrm{X} * 6 * 4 + \\mathrm{X} * 6 * 2 = {36} \\) , \\( {36} * \\mathrm{X} = {36},\\mathrm{X} = 1 \\) . The answer is 1. \\( \\left\\lbrack \\checkmark \\right\\rbrack \\) |
| Condition 1 mask | Condition 2 mask | ... |
| Q: Dana worked 'X' hours on Friday, 10 hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | Q: Dana worked 9 hours on Friday, 'X' hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | ... |
| Separate score: 0 | Separate score: 8 | ... |
| Total verification score: 19 |
| Dataset | Number of samples | Average words | Answer Format | Lience |
| GSM8K | 1319 | 46.9 | Number | MIT License |
| SingleEq | 508 | 27.4 | Number | MIT License |
| AddSub | 395 | 31.5 | Number | Unspecified |
| MultiArith | 600 | 31.8 | Number | Unspecified |
| AQUA-RAT | 254 | 51.9 | Multiple Choice | Apache-2.0 |
| SVAMP | 1000 | 31.8 | Number | MIT License |
| CSQA | 1221 | 27.8 | Multiple Choice | Unspecified |
| Data Understanding | 369 | 21.4 | Date | Unspecified |
| Dataset | Question | Self-Verification |
| GSM8K | Q: Matthew has a collection of 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy, and for the other half 'X' per toy. If Matthew would agree to that offer, He would earn 72.0 dollars. | A: Matthew has 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy. and for the other half $X per toy. If Matthew would agree to that offer. He would earn 72.0 dollars. 12/2 = 66*5 = 3012 - 6 = 66*X = 42 X = 42/6 X = 7 The answer is 7. [✓] |
| Continued on next page |
| Dataset | Question | Self-Verification |
| Q: Four students scored a total of 251 points in a board game. Naomi scored 68 of the points. Yuri scored 'X' more than half as many points as Naomi and Brianna scored 17 points more than Naomi. Jojo scored 54 points. | Four students scored a total of 251 points. Naomi scored 68 points. Yuri scored X points. Brianna scored 17 points more than Naomi. Jojo scored 54 points. So: 68 + X + 68 + 17 + 54 = 251, X = 251 - 68 - 68 - 17 - 54, X = 94. The answer is 94. [X](GT:10) | |
| SingleEq | Q: Sam grew 4 watermelons, but the rabbits ate 'X' watermelons. Sam have 1 watermelon left. | A: Sam grew 4 watermelons. But the rabbits ate X watermelons. So, we can write the following equation: 4 - X = 1, X = 4 - 1, X = 3. The answer is 3. [✓] |
| Q: Melanie, Benny, Sandy, and Jessica each have 'X' dozen Pokemon cards. They have 432 Pokemon cards in all. | A: Melanie has X dozen Pokemon cards. Benny has X dozen Pokemon cards. Sandy has X dozen Pokemon cards. Jessica has X dozen Pokemon cards. So, they have 4 X dozen Pokemon cards in total. X * 4 = 432, X = 432 / 4, X = 108. The answer is 108. [X](GT:9) | |
| Addsub | Q: Karin's science class weighed plastic rings for an experiment. They found that the orange ring weighed 'X' ounce, the purple ring weighed 0.3333333333333333 ounce, and the white ring weighed 0.41666666666667 ounce. The total weight of the plastic rings is 0.8333333333333333. | A: The orange ring weighed 0.08333333333333333 ounce. The purple ring weighed 0.333333333333333333 ounce. And the white ring weighed 0.4166666666666667 ounce. The total weight of the plastic rings is 0.83333333333333333333333 + 0.3333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333. |
| Q: There were 6 roses in the vase. Mary cut some roses from her flower garden. There are now 'X' roses in the vase. She cut 10 roses. | A: There were originally 6 roses in the vase. Mary cut 10 roses from her flower garden. So there are now X roses in the vase. <code>6 - 10 = X X = -4 </code> The answer is -4. [X](GT:16) | |
| MultiArith | Q: Luke was organizing his base-ball cards in a binder with 3 on each page. If he had 3 new cards and 'X' old cards to put in the binder, he use 4 pages. | A: Luke had 3 new cards and X old cards. He would use 4.0 pages. So, we can write the following equation: 3 + X = 3 * 4.0, 3 + X = 12, X = 12 - 3 X = 9. The answer is 9. [✓] |
| Q: Will bought 7 boxes of chocolate candy and gave 3 to his little brother. If each box has 'X' pieces inside it, Will still have 16 pieces. | A: Will bought 7 boxes of chocolate candy. Each box has X pieces inside it. He gave 3 boxes to his little brother. So he has 4 boxes left. 7 * X = Y, Y - 3 * X = 16, Y = 16 + 3 * X, Y = 16 + 3 * X. The answer is 16. [X](GT:4) | |
| Continued on next page |
| Dataset | Question | Self-Verification |
| SVAMP | Q: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has 'X' peaches and 15 apples. Steven has 3.0 more apples than peaches. | A: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has X peaches and 15 apples. Steven has 3.0 more apples than peaches. X + 3 = 15, X = 15 - 3, X = 1.2 The answer is 12. [✓] |
| Q: An industrial machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for 'X' minutes today. The machine made 72 shirts today. | A: The machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for X minutes today. The machine made 72 shirts today. So, we can write the following equation: 6 * 5 + 6 * X = 72, 30 + 6X = 72, 6X = 72 - 30, 6X = 42, X = 42 / 6 X = 7. The answer is 7. [X](GT : 12) | |
| Common-senseQA | Q: If there is a pond with trees around it, it likely located in forest. | A: he forest is likely located where there is a pond with trees around it. The answer is Yes. [✓] |
| Q: The flasks was used to distill elements, it is used in chemistry lab. | A: The answer should be the place where the flask is used. But the chemistry lab is not suitable place where the flask is used. The answer is No. [X](GT : Yes) | |
| Date Under-stand. | Q: It was 'X'/01/2021 a week ago. Today is 09/08/2021. | A: If today is 09/08/2021, so the date a week ago is 09/01/2021. The answer is 09. [✓] |
| Q: Yesterday, Jan 21, 2011, Jane ate 2 pizzas and 5 wings. One week ago from today is 01/15/2011. | A: one week ago from today is 01/15/2011. So today is 01/22/2011. So yesterday is 01/21/2011. So Jan 'X'1 is 01/21/2011. So 'X'011 is 2011. So The answer is 2011. [X](GT : B) |
| PROMPT FOR AQUA ALGEBRAIC WORD PROBLEMS\nQ: John found that the average of 15 numbers is 40. If 10 is added to each number then the mean of the numbers is?\nAnswer Choices: (a) 50 (b) 45 (c) 65 (d) 78 (e) 64\nA: If 10 is added to each number, then the mean of the numbers also increases by 10. So the new mean would be 50. The answer is (a).\nQ: If a / b = 3/4 and 8a + 5b = 22,then find the value of a.\nAnswer Choices: (a) 1/2 (b) 3/2 (c) 5/2 (d) 4/2 (e) 7/2\nA: If a / b = 3/4, then b = 4a / 3. So 8a + 5(4a / 3) = 22. This simplifies to 8a + 20a / 3 = 22, which means 44a / 3 = 22. So a is equal to 3/2. The answer is (b).\nQ: A person is traveling at 20 km/hr and reached his destiny in 2.5 hr then find the distance?\nAnswer Choices: (a) 53 km (b) 55 km (c) 52 km (d) 60 km (e) 50 km\nA: The distance that the person traveled would have been 20 km/hr * 2.5 hrs = 50 km. The answer is (e).\nQ: How many keystrokes are needed to type the numbers from 1 to 500?\nAnswer Choices: (a) 1156 (b) 1392 (c) 1480 (d) 1562 (e) 1788\nA: There are 9 one-digit numbers from 1 to 9. There are 90 two-digit numbers from 10 to 99. There are 401 three-digit numbers from 100 to 500. 9 + 90(2) + 401(3) = 1392. The answer is (b). |
| PROMPT FOR DATE UNDERSTANDING\nQ: 2015 is coming in 36 hours. What is the date one week from today in MM/DD/YYYY? \nA: If 2015 is coming in 36 hours, then it is coming in 2 days. 2 days before 01/01/2015 is 12/30/2014, so today is 12/30/2014. So one week from today will be 01/05/2015. So the answer is 01/05/2015. \nQ: The first day of 2019 is a Tuesday, and today is the first Monday of 2019. What is the date today in MM/DD/YYYY? \nA: If the first day of 2019 was Tuesday, then 01/01/2019 was a Tuesday. Today is the first monday, would be six days later. So today is 01/07/2019. So the answer is 01/07/2019. \nQ: The concert was scheduled to be on 06/01/1943, but was delayed by one day to today. What is the date 10 days ago in MM/DD/YYYY? \nA: One day after 06/01/1943 is 06/02/1943, so today is 06/02/1943. 10 days before today is 05/23/1943. So the answer is 05/23/1943. \nQ: It is 4/19/1969 today. What is the date 24 hours later in MM/DD/YYYY? \nA: Today is 04/19/1969. 24 hours later is one day after today, which would be 04/20/1969. So the answer is 04/20/1969. \nQ: Jane thought today is 3/11/2002, but today is in fact Mar 12, which is 1 day later. What is the date 24 hours later in MM/DD/YYYY? \nA: Today is 03/12/2002. So the date 24 hours later will be 03/13/2002. So the answer is 03/13/2002. \nQ: Jane was born on the last day of Feburuary in 2001. Today is her 16-year-old birthday. What is the date yesterday in MM/DD/YYYY? \nA: The last day of February is the 28th, so Jane was born on 02/28/2001. Today is her 16-year old birthday, so today is 02/28/2017. So yesterday was 02/27/2017. So the answer is 02/27/2017. |
| PROMPT FOR DATE UNDERSTANDING\nQ: "'X' is coming in 36 hours. One week from today is 01/05/2015." What is the answer of 'X'? \nA: If The date one week from today is 01/05/2015, so today is 12/30/2014. So the data after 36 hours is 2015. The answer is 2015.\nQ: "The first day of 'X' is a Tuesday, and today is the first Monday of 2019. Today is 01/07/2019." What is the answer of 'X'? \nA: If today is the first Monday of 2019 and today is 01/07/2019. So The first day of 2019 is a Tuesday. The answer is 2019.\nQ: "The concert was scheduled to be on 'X'/01/1943, but was delayed by one day to today. 10 days ago is 05/23/1943."\nWhat is the answer of 'X'? \nA: 10 days ago is 05/23/1943, and the concert was delayed by one day to today, so today is 06/02/1943. So the concert was scheduled to be on 06/01/1943 The answer is 06.\nQ: "It is 'X'/19/1969 today. 24 hours later is 04/20/1969." What is the answer of 'X'? \nA: 24 hours later is 04/20/1969. So today is 04/19/1969. The answer is 04.\nQ: "Jane thought today is 'X'/12/2002, but today is in fact Mar 12, which is 1 day later. 24 hours later is 03/13/2002."\nWhat is the answer of 'X'? \nA: 24 hours later is 03/13/2002. So today is 03/12/2002. The answer is 03.\nQ: "Jane was born on the last day of Feburary in 'X'. Today is her 16-year-old birthday. Yesterday is 02/27/2017" What is the answer of 'X'? \nA: Yesterday is 02/27/2017, so today is 02/28/2017, Jane was born on 02/28/2001. The answer is 2001. |
| Method | Arithmetic Tasks | General Tasks | |||||||
| GSM8K | SingleEq | AddSub | MultiArith | AQUA-RAT | SVAMP | CSQA | DU | ||
| Previous SOTA (Fine-tune) | 35a/57b | 32.5c | 94.9d | 60.5e | 37.9f | 57.4g | 91.2h | - | |
| 9-12 year olds | 60i | - | - | - | - | - | - | - | |
| GPT-3 Standard | 19.7 | 86.8 | 90.9 | 44.0 | 29.5 | 69.9 | 82.3 | 49.0 | |
| GPT-3 (175B) code-davinci-001 | CoT | 13.84 | 60.20 | 58.55 | 45.85 | 18.90 | 38.42 | 46.75 | 38.72 |
| CoT+Self-Verification | 13.92 (+0.08) | 60.61 (+0.41) | 59.07 (+0.52) | 46.19 (+0.34) | 27.04 (+8.14) | 38.96 (+0.54) | 47.68 (+0.93) | 39.03 (+0.31) | |
| Instruct-GPT (175B) code-davinci-002 | CoT | 60.81 | 91.01 | 82.78 | 96.13 | 45.30 | 75.87 | 77.42 | 65.43 |
| CoT+Self-Verification | 65.14 (+4.33) | 93.40 (+2.39) | 86.33 (+3.55) | 99.15 (+3.02) | 47.95 (+2.65) | 76.99 (+1.12) | 77.83 (+0.41) | 66.57 (+1.14) |
| Self-Consistency Decoding (Wang et al., 2023c) For Forward Reasoning | |||||||||
| GPT-3 (175B) code-davinci-001 | SC | 23.40 | 70.25 | 68.65 | 79.82 | 25.60 | 54.58 | 54.92 | 49.26 |
| SC+ Self-Verification | 23.59(+0.19) | 70.50(+0.25) | 68.71(+0.06) | 80.01(+0.19) | 28.98(+3.38) | 54.68(+0.1) | 55.09(+0.17) | 49.72(+0.46) | |
| Instruct-GPT (175B) code-davinci-002 | SC | 78.00 | 96.78 | 91.64 | 100.0 | 52.01 | 86.77 | 81.43 | 71.58 |
| SC+Self-Verification | 78.32(+0.32) | 96.85(+0.07) | 92.03(+0.39) | 100.0(+0.0) | 52.25(+0.24) | 86.94(+0.17) | 81.53(+0.1) | 71.89(+0.31) |
| PAL (Gao et al., 2023) For Forward Reasoning | ||||||||
| GPT-3 (175B) code-davinci-001 | PAL | 31.82 | 63.98 | 63.15 | 61.52 | 30.56 | 42.69 | - |
| PAL+Self-Verification | 32.87(+1.05) | 65.45(+1.47) | 64.15(+1.0) | 61.76(+0.24) | 30.90(+0.34) | 42.78(+0.09) | - | |
| Instruct-GPT (175B) code-davinci-002 | PAL | 72.02 | 96.08 | 92.64 | 99.15 | 59.75 | 79.45 | - |
| PAL+Self-Verification | 72.89(+0.87) | 96.52(+0.44) | 93.78(+1.14) | 99.87(+0.72) | 60.21(+0.46) | 80.24(+0.79) | - |
| Dataset | Num. | Avg. Ans. | Ran. Acc. | CoT Acc. | Veri. Acc. |
| GSM8K | 601/1306 | 2.80 | 35.7% | 53.4% | 58.9% |
| Addsub | 77/377 | 2.13 | 46.9% | 56.7% | 74.0% |
| AQUA | 123/219 | 2.71 | 36.9% | 50.3% | 51.2% |
| Multiarith | 73/593 | 2.19 | 45.6% | 71.2% | 89.0% |
| SingleEq | 51/501 | 2.31 | 43.3% | 51.0% | 74.5% |
| SVAMP | 274/954 | 2.54 | 39.4% | 51.5% | 58.4% |
| AnswerFormat | Answer CleansingApproach | Pseudo Code(Example in Pytorch 3.7) |
| Number | Pick up the first numberencountered in the text. | pred = pred.replace("\\\\",".")\npred = [s for s in re.findall(r'^-?\\\\d+\\.?\\\\d*', pred)]\npred = pred[0] |
| Multiple-Choice | Pick up the first large letterencountered in the text. | pred = re.findall(r'A|B|C|D|E', pred)\npred = pred[0] |
| True or False | Pick up the first "True" or"False" encountered in thetext after removingunnec-essary letters. | pred = pred.lower()\npred = re.sub("\\\\\\"|\\n|\\.\\s|:\\\\","\\"", pred)\npred = pred.split("\\\\.")\npred = [i for i in pred if i in ("True", "False")] \npred = pred[0] |
| Yes or No | Pick up the first "yes" or"no" encountered in thetext after removingunnec-essary letters. | pred = pred.lower()\npred = re.sub("\\\\\\"|\\n|\\.\\s|:\\\\","\\"", pred)\npred = pred.split("\\\\.")\npred = [i for i in pred if i in ("yes", "no"]\npred = pred[0] |
| Free Format | Just remove unnecessaryletters, such as ".". | pred = re.sub("\\\\\\"|\\n|\\.\\s", "", pred) |
| True-False Item Verification | If Tim will need 36 lemons, and he will need to make 4 gallons of lemonade for the party, so he will need \\( 4 * 6 = {24} \\) lemons,he will need to make an extra gallonfor him that was twice as tart as the other gallons, so he will need \\( {24} * 2 = {48} \\) lemons, \\( {48}! = {36} \\) ,which is different from the theme. The answer is False. [X] |
| Conditional Masked Verification | Tim wanted to make lemonade for a pool party. For a gallon of lemonade, his recipe called for X cup of fresh lemon juice. He found that 6 lemons would yield 1 cup of juice. He figured he would need to make 4 gallons of lemonade for the party. His best friend Allen asked if Tim could make an extra gallon for him that was twice as tart as the other gallons. Tim will need 36.0 lemons. \\( \\mathrm{X} * 6 * 4 + \\mathrm{X} * 6 * 2 = {36} \\) , \\( {36} * \\mathrm{X} = {36},\\mathrm{X} = 1 \\) . The answer is 1. \\( \\left\\lbrack \\checkmark \\right\\rbrack \\) |
| Condition 1 mask | Condition 2 mask | ... |
| Q: Dana worked 'X' hours on Friday, 10 hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | Q: Dana worked 9 hours on Friday, 'X' hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | ... |
| Separate score: 0 | Separate score: 8 | ... |
| Total verification score: 19 |
| Dataset | Number of samples | Average words | Answer Format | Lience |
| GSM8K | 1319 | 46.9 | Number | MIT License |
| SingleEq | 508 | 27.4 | Number | MIT License |
| AddSub | 395 | 31.5 | Number | Unspecified |
| MultiArith | 600 | 31.8 | Number | Unspecified |
| AQUA-RAT | 254 | 51.9 | Multiple Choice | Apache-2.0 |
| SVAMP | 1000 | 31.8 | Number | MIT License |
| CSQA | 1221 | 27.8 | Multiple Choice | Unspecified |
| Data Understanding | 369 | 21.4 | Date | Unspecified |
| Dataset | Question | Self-Verification |
| GSM8K | Q: Matthew has a collection of 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy, and for the other half 'X' per toy. If Matthew would agree to that offer, He would earn 72.0 dollars. | A: Matthew has 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy. and for the other half $X per toy. If Matthew would agree to that offer. He would earn 72.0 dollars. 12/2 = 66*5 = 3012 - 6 = 66*X = 42 X = 42/6 X = 7 The answer is 7. [✓] |
| Continued on next page |
| Dataset | Question | Self-Verification |
| Q: Four students scored a total of 251 points in a board game. Naomi scored 68 of the points. Yuri scored 'X' more than half as many points as Naomi and Brianna scored 17 points more than Naomi. Jojo scored 54 points. | Four students scored a total of 251 points. Naomi scored 68 points. Yuri scored X points. Brianna scored 17 points more than Naomi. Jojo scored 54 points. So: 68 + X + 68 + 17 + 54 = 251, X = 251 - 68 - 68 - 17 - 54, X = 94. The answer is 94. [X](GT:10) | |
| SingleEq | Q: Sam grew 4 watermelons, but the rabbits ate 'X' watermelons. Sam have 1 watermelon left. | A: Sam grew 4 watermelons. But the rabbits ate X watermelons. So, we can write the following equation: 4 - X = 1, X = 4 - 1, X = 3. The answer is 3. [✓] |
| Q: Melanie, Benny, Sandy, and Jessica each have 'X' dozen Pokemon cards. They have 432 Pokemon cards in all. | A: Melanie has X dozen Pokemon cards. Benny has X dozen Pokemon cards. Sandy has X dozen Pokemon cards. Jessica has X dozen Pokemon cards. So, they have 4 X dozen Pokemon cards in total. X * 4 = 432, X = 432 / 4, X = 108. The answer is 108. [X](GT:9) | |
| Addsub | Q: Karin's science class weighed plastic rings for an experiment. They found that the orange ring weighed 'X' ounce, the purple ring weighed 0.3333333333333333 ounce, and the white ring weighed 0.41666666666667 ounce. The total weight of the plastic rings is 0.8333333333333333. | A: The orange ring weighed 0.08333333333333333 ounce. The purple ring weighed 0.333333333333333333 ounce. And the white ring weighed 0.4166666666666667 ounce. The total weight of the plastic rings is 0.83333333333333333333333 + 0.3333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333. |
| Q: There were 6 roses in the vase. Mary cut some roses from her flower garden. There are now 'X' roses in the vase. She cut 10 roses. | A: There were originally 6 roses in the vase. Mary cut 10 roses from her flower garden. So there are now X roses in the vase. <code>6 - 10 = X X = -4 </code> The answer is -4. [X](GT:16) | |
| MultiArith | Q: Luke was organizing his base-ball cards in a binder with 3 on each page. If he had 3 new cards and 'X' old cards to put in the binder, he use 4 pages. | A: Luke had 3 new cards and X old cards. He would use 4.0 pages. So, we can write the following equation: 3 + X = 3 * 4.0, 3 + X = 12, X = 12 - 3 X = 9. The answer is 9. [✓] |
| Q: Will bought 7 boxes of chocolate candy and gave 3 to his little brother. If each box has 'X' pieces inside it, Will still have 16 pieces. | A: Will bought 7 boxes of chocolate candy. Each box has X pieces inside it. He gave 3 boxes to his little brother. So he has 4 boxes left. 7 * X = Y, Y - 3 * X = 16, Y = 16 + 3 * X, Y = 16 + 3 * X. The answer is 16. [X](GT:4) | |
| Continued on next page |
| Dataset | Question | Self-Verification |
| SVAMP | Q: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has 'X' peaches and 15 apples. Steven has 3.0 more apples than peaches. | A: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has X peaches and 15 apples. Steven has 3.0 more apples than peaches. X + 3 = 15, X = 15 - 3, X = 1.2 The answer is 12. [✓] |
| Q: An industrial machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for 'X' minutes today. The machine made 72 shirts today. | A: The machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for X minutes today. The machine made 72 shirts today. So, we can write the following equation: 6 * 5 + 6 * X = 72, 30 + 6X = 72, 6X = 72 - 30, 6X = 42, X = 42 / 6 X = 7. The answer is 7. [X](GT : 12) | |
| Common-senseQA | Q: If there is a pond with trees around it, it likely located in forest. | A: he forest is likely located where there is a pond with trees around it. The answer is Yes. [✓] |
| Q: The flasks was used to distill elements, it is used in chemistry lab. | A: The answer should be the place where the flask is used. But the chemistry lab is not suitable place where the flask is used. The answer is No. [X](GT : Yes) | |
| Date Under-stand. | Q: It was 'X'/01/2021 a week ago. Today is 09/08/2021. | A: If today is 09/08/2021, so the date a week ago is 09/01/2021. The answer is 09. [✓] |
| Q: Yesterday, Jan 21, 2011, Jane ate 2 pizzas and 5 wings. One week ago from today is 01/15/2011. | A: one week ago from today is 01/15/2011. So today is 01/22/2011. So yesterday is 01/21/2011. So Jan 'X'1 is 01/21/2011. So 'X'011 is 2011. So The answer is 2011. [X](GT : B) |
| PROMPT FOR AQUA ALGEBRAIC WORD PROBLEMS\nQ: John found that the average of 15 numbers is 40. If 10 is added to each number then the mean of the numbers is?\nAnswer Choices: (a) 50 (b) 45 (c) 65 (d) 78 (e) 64\nA: If 10 is added to each number, then the mean of the numbers also increases by 10. So the new mean would be 50. The answer is (a).\nQ: If a / b = 3/4 and 8a + 5b = 22,then find the value of a.\nAnswer Choices: (a) 1/2 (b) 3/2 (c) 5/2 (d) 4/2 (e) 7/2\nA: If a / b = 3/4, then b = 4a / 3. So 8a + 5(4a / 3) = 22. This simplifies to 8a + 20a / 3 = 22, which means 44a / 3 = 22. So a is equal to 3/2. The answer is (b).\nQ: A person is traveling at 20 km/hr and reached his destiny in 2.5 hr then find the distance?\nAnswer Choices: (a) 53 km (b) 55 km (c) 52 km (d) 60 km (e) 50 km\nA: The distance that the person traveled would have been 20 km/hr * 2.5 hrs = 50 km. The answer is (e).\nQ: How many keystrokes are needed to type the numbers from 1 to 500?\nAnswer Choices: (a) 1156 (b) 1392 (c) 1480 (d) 1562 (e) 1788\nA: There are 9 one-digit numbers from 1 to 9. There are 90 two-digit numbers from 10 to 99. There are 401 three-digit numbers from 100 to 500. 9 + 90(2) + 401(3) = 1392. The answer is (b). |
| PROMPT FOR DATE UNDERSTANDING\nQ: 2015 is coming in 36 hours. What is the date one week from today in MM/DD/YYYY? \nA: If 2015 is coming in 36 hours, then it is coming in 2 days. 2 days before 01/01/2015 is 12/30/2014, so today is 12/30/2014. So one week from today will be 01/05/2015. So the answer is 01/05/2015. \nQ: The first day of 2019 is a Tuesday, and today is the first Monday of 2019. What is the date today in MM/DD/YYYY? \nA: If the first day of 2019 was Tuesday, then 01/01/2019 was a Tuesday. Today is the first monday, would be six days later. So today is 01/07/2019. So the answer is 01/07/2019. \nQ: The concert was scheduled to be on 06/01/1943, but was delayed by one day to today. What is the date 10 days ago in MM/DD/YYYY? \nA: One day after 06/01/1943 is 06/02/1943, so today is 06/02/1943. 10 days before today is 05/23/1943. So the answer is 05/23/1943. \nQ: It is 4/19/1969 today. What is the date 24 hours later in MM/DD/YYYY? \nA: Today is 04/19/1969. 24 hours later is one day after today, which would be 04/20/1969. So the answer is 04/20/1969. \nQ: Jane thought today is 3/11/2002, but today is in fact Mar 12, which is 1 day later. What is the date 24 hours later in MM/DD/YYYY? \nA: Today is 03/12/2002. So the date 24 hours later will be 03/13/2002. So the answer is 03/13/2002. \nQ: Jane was born on the last day of Feburuary in 2001. Today is her 16-year-old birthday. What is the date yesterday in MM/DD/YYYY? \nA: The last day of February is the 28th, so Jane was born on 02/28/2001. Today is her 16-year old birthday, so today is 02/28/2017. So yesterday was 02/27/2017. So the answer is 02/27/2017. |
| PROMPT FOR DATE UNDERSTANDING\nQ: "'X' is coming in 36 hours. One week from today is 01/05/2015." What is the answer of 'X'? \nA: If The date one week from today is 01/05/2015, so today is 12/30/2014. So the data after 36 hours is 2015. The answer is 2015.\nQ: "The first day of 'X' is a Tuesday, and today is the first Monday of 2019. Today is 01/07/2019." What is the answer of 'X'? \nA: If today is the first Monday of 2019 and today is 01/07/2019. So The first day of 2019 is a Tuesday. The answer is 2019.\nQ: "The concert was scheduled to be on 'X'/01/1943, but was delayed by one day to today. 10 days ago is 05/23/1943."\nWhat is the answer of 'X'? \nA: 10 days ago is 05/23/1943, and the concert was delayed by one day to today, so today is 06/02/1943. So the concert was scheduled to be on 06/01/1943 The answer is 06.\nQ: "It is 'X'/19/1969 today. 24 hours later is 04/20/1969." What is the answer of 'X'? \nA: 24 hours later is 04/20/1969. So today is 04/19/1969. The answer is 04.\nQ: "Jane thought today is 'X'/12/2002, but today is in fact Mar 12, which is 1 day later. 24 hours later is 03/13/2002."\nWhat is the answer of 'X'? \nA: 24 hours later is 03/13/2002. So today is 03/12/2002. The answer is 03.\nQ: "Jane was born on the last day of Feburary in 'X'. Today is her 16-year-old birthday. Yesterday is 02/27/2017" What is the answer of 'X'? \nA: Yesterday is 02/27/2017, so today is 02/28/2017, Jane was born on 02/28/2001. The answer is 2001. |
| Method | Arithmetic Tasks | General Tasks | |||||||
| GSM8K | SingleEq | AddSub | MultiArith | AQUA-RAT | SVAMP | CSQA | DU | ||
| Previous SOTA (Fine-tune) | 35a/57b | 32.5c | 94.9d | 60.5e | 37.9f | 57.4g | 91.2h | - | |
| 9-12 year olds | 60i | - | - | - | - | - | - | - | |
| GPT-3 Standard | 19.7 | 86.8 | 90.9 | 44.0 | 29.5 | 69.9 | 82.3 | 49.0 | |
| GPT-3 (175B) code-davinci-001 | CoT | 13.84 | 60.20 | 58.55 | 45.85 | 18.90 | 38.42 | 46.75 | 38.72 |
| CoT+Self-Verification | 13.92 (+0.08) | 60.61 (+0.41) | 59.07 (+0.52) | 46.19 (+0.34) | 27.04 (+8.14) | 38.96 (+0.54) | 47.68 (+0.93) | 39.03 (+0.31) | |
| Instruct-GPT (175B) code-davinci-002 | CoT | 60.81 | 91.01 | 82.78 | 96.13 | 45.30 | 75.87 | 77.42 | 65.43 |
| CoT+Self-Verification | 65.14 (+4.33) | 93.40 (+2.39) | 86.33 (+3.55) | 99.15 (+3.02) | 47.95 (+2.65) | 76.99 (+1.12) | 77.83 (+0.41) | 66.57 (+1.14) | |
| Self-Consistency Decoding (Wang et al., 2023c) For Forward Reasoning | |||||||||
| GPT-3 (175B) code-davinci-001 | SC | 23.40 | 70.25 | 68.65 | 79.82 | 25.60 | 54.58 | 54.92 | 49.26 |
| SC+ Self-Verification | 23.59(+0.19) | 70.50(+0.25) | 68.71(+0.06) | 80.01(+0.19) | 28.98(+3.38) | 54.68(+0.1) | 55.09(+0.17) | 49.72(+0.46) | |
| Instruct-GPT (175B) code-davinci-002 | SC | 78.00 | 96.78 | 91.64 | 100.0 | 52.01 | 86.77 | 81.43 | 71.58 |
| SC+Self-Verification | 78.32(+0.32) | 96.85(+0.07) | 92.03(+0.39) | 100.0(+0.0) | 52.25(+0.24) | 86.94(+0.17) | 81.53(+0.1) | 71.89(+0.31) | |
| PAL (Gao et al., 2023) For Forward Reasoning | ||||||||
| GPT-3 (175B) code-davinci-001 | PAL | 31.82 | 63.98 | 63.15 | 61.52 | 30.56 | 42.69 | - |
| PAL+Self-Verification | 32.87(+1.05) | 65.45(+1.47) | 64.15(+1.0) | 61.76(+0.24) | 30.90(+0.34) | 42.78(+0.09) | - | |
| Instruct-GPT (175B) code-davinci-002 | PAL | 72.02 | 96.08 | 92.64 | 99.15 | 59.75 | 79.45 | - |
| PAL+Self-Verification | 72.89(+0.87) | 96.52(+0.44) | 93.78(+1.14) | 99.87(+0.72) | 60.21(+0.46) | 80.24(+0.79) | - | |
| Dataset | Num. | Avg. Ans. | Ran. Acc. | CoT Acc. | Veri. Acc. |
| GSM8K | 601/1306 | 2.80 | 35.7% | 53.4% | 58.9% |
| Addsub | 77/377 | 2.13 | 46.9% | 56.7% | 74.0% |
| AQUA | 123/219 | 2.71 | 36.9% | 50.3% | 51.2% |
| Multiarith | 73/593 | 2.19 | 45.6% | 71.2% | 89.0% |
| SingleEq | 51/501 | 2.31 | 43.3% | 51.0% | 74.5% |
| SVAMP | 274/954 | 2.54 | 39.4% | 51.5% | 58.4% |
| AnswerFormat | Answer CleansingApproach | Pseudo Code(Example in Pytorch 3.7) |
| Number | Pick up the first numberencountered in the text. | pred = pred.replace("\\",".") +pred = [s for s in re.findall(r'^-?\\d+\.?\\d*', pred)] +pred = pred[0] |
| Multiple-Choice | Pick up the first large letterencountered in the text. | pred = re.findall(r'A|B|C|D|E', pred) +pred = pred[0] |
| True or False | Pick up the first "True" or"False" encountered in thetext after removingunnec-essary letters. | pred = pred.lower() +pred = re.sub("\\\"|\n|\.\s|:\\","\"", pred) +pred = pred.split("\\.") +pred = [i for i in pred if i in ("True", "False")] +pred = pred[0] |
| Yes or No | Pick up the first "yes" or"no" encountered in thetext after removingunnec-essary letters. | pred = pred.lower() +pred = re.sub("\\\"|\n|\.\s|:\\","\"", pred) +pred = pred.split("\\.") +pred = [i for i in pred if i in ("yes", "no"] +pred = pred[0] |
| Free Format | Just remove unnecessaryletters, such as ".". | pred = re.sub("\\\"|\n|\.\s", "", pred) |
| True-False Item Verification | If Tim will need 36 lemons, and he will need to make 4 gallons of lemonade for the party, so he will need \( 4 * 6 = {24} \) lemons,he will need to make an extra gallonfor him that was twice as tart as the other gallons, so he will need \( {24} * 2 = {48} \) lemons, \( {48}! = {36} \) ,which is different from the theme. The answer is False. [X] |
| Conditional Masked Verification | Tim wanted to make lemonade for a pool party. For a gallon of lemonade, his recipe called for X cup of fresh lemon juice. He found that 6 lemons would yield 1 cup of juice. He figured he would need to make 4 gallons of lemonade for the party. His best friend Allen asked if Tim could make an extra gallon for him that was twice as tart as the other gallons. Tim will need 36.0 lemons. \( \mathrm{X} * 6 * 4 + \mathrm{X} * 6 * 2 = {36} \) , \( {36} * \mathrm{X} = {36},\mathrm{X} = 1 \) . The answer is 1. \( \left\lbrack \checkmark \right\rbrack \) |
| Condition 1 mask | Condition 2 mask | ... |
| Q: Dana worked 'X' hours on Friday, 10 hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | Q: Dana worked 9 hours on Friday, 'X' hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | ... |
| Separate score: 0 | Separate score: 8 | ... |
| Total verification score: 19 | ||
| Dataset | Number of samples | Average words | Answer Format | Lience |
| GSM8K | 1319 | 46.9 | Number | MIT License |
| SingleEq | 508 | 27.4 | Number | MIT License |
| AddSub | 395 | 31.5 | Number | Unspecified |
| MultiArith | 600 | 31.8 | Number | Unspecified |
| AQUA-RAT | 254 | 51.9 | Multiple Choice | Apache-2.0 |
| SVAMP | 1000 | 31.8 | Number | MIT License |
| CSQA | 1221 | 27.8 | Multiple Choice | Unspecified |
| Data Understanding | 369 | 21.4 | Date | Unspecified |
| Dataset | Question | Self-Verification |
| GSM8K | Q: Matthew has a collection of 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy, and for the other half 'X' per toy. If Matthew would agree to that offer, He would earn 72.0 dollars. | A: Matthew has 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy. and for the other half $X per toy. If Matthew would agree to that offer. He would earn 72.0 dollars. 12/2 = 66*5 = 3012 - 6 = 66*X = 42 X = 42/6 X = 7 The answer is 7. [✓] |
| Continued on next page | ||
| Dataset | Question | Self-Verification |
| Q: Four students scored a total of 251 points in a board game. Naomi scored 68 of the points. Yuri scored 'X' more than half as many points as Naomi and Brianna scored 17 points more than Naomi. Jojo scored 54 points. | Four students scored a total of 251 points. Naomi scored 68 points. Yuri scored X points. Brianna scored 17 points more than Naomi. Jojo scored 54 points. So: 68 + X + 68 + 17 + 54 = 251, X = 251 - 68 - 68 - 17 - 54, X = 94. The answer is 94. [X](GT:10) | |
| SingleEq | Q: Sam grew 4 watermelons, but the rabbits ate 'X' watermelons. Sam have 1 watermelon left. | A: Sam grew 4 watermelons. But the rabbits ate X watermelons. So, we can write the following equation: 4 - X = 1, X = 4 - 1, X = 3. The answer is 3. [✓] |
| Q: Melanie, Benny, Sandy, and Jessica each have 'X' dozen Pokemon cards. They have 432 Pokemon cards in all. | A: Melanie has X dozen Pokemon cards. Benny has X dozen Pokemon cards. Sandy has X dozen Pokemon cards. Jessica has X dozen Pokemon cards. So, they have 4 X dozen Pokemon cards in total. X * 4 = 432, X = 432 / 4, X = 108. The answer is 108. [X](GT:9) | |
| Addsub | Q: Karin's science class weighed plastic rings for an experiment. They found that the orange ring weighed 'X' ounce, the purple ring weighed 0.3333333333333333 ounce, and the white ring weighed 0.41666666666667 ounce. The total weight of the plastic rings is 0.8333333333333333. | A: The orange ring weighed 0.08333333333333333 ounce. The purple ring weighed 0.333333333333333333 ounce. And the white ring weighed 0.4166666666666667 ounce. The total weight of the plastic rings is 0.83333333333333333333333 + 0.3333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333. |
| Q: There were 6 roses in the vase. Mary cut some roses from her flower garden. There are now 'X' roses in the vase. She cut 10 roses. | A: There were originally 6 roses in the vase. Mary cut 10 roses from her flower garden. So there are now X roses in the vase. <code>6 - 10 = X X = -4 </code> The answer is -4. [X](GT:16) | |
| MultiArith | Q: Luke was organizing his base-ball cards in a binder with 3 on each page. If he had 3 new cards and 'X' old cards to put in the binder, he use 4 pages. | A: Luke had 3 new cards and X old cards. He would use 4.0 pages. So, we can write the following equation: 3 + X = 3 * 4.0, 3 + X = 12, X = 12 - 3 X = 9. The answer is 9. [✓] |
| Q: Will bought 7 boxes of chocolate candy and gave 3 to his little brother. If each box has 'X' pieces inside it, Will still have 16 pieces. | A: Will bought 7 boxes of chocolate candy. Each box has X pieces inside it. He gave 3 boxes to his little brother. So he has 4 boxes left. 7 * X = Y, Y - 3 * X = 16, Y = 16 + 3 * X, Y = 16 + 3 * X. The answer is 16. [X](GT:4) | |
| Continued on next page | ||
| Dataset | Question | Self-Verification |
| SVAMP | Q: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has 'X' peaches and 15 apples. Steven has 3.0 more apples than peaches. | A: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has X peaches and 15 apples. Steven has 3.0 more apples than peaches. X + 3 = 15, X = 15 - 3, X = 1.2 The answer is 12. [✓] |
| Q: An industrial machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for 'X' minutes today. The machine made 72 shirts today. | A: The machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for X minutes today. The machine made 72 shirts today. So, we can write the following equation: 6 * 5 + 6 * X = 72, 30 + 6X = 72, 6X = 72 - 30, 6X = 42, X = 42 / 6 X = 7. The answer is 7. [X](GT : 12) | |
| Common-senseQA | Q: If there is a pond with trees around it, it likely located in forest. | A: he forest is likely located where there is a pond with trees around it. The answer is Yes. [✓] |
| Q: The flasks was used to distill elements, it is used in chemistry lab. | A: The answer should be the place where the flask is used. But the chemistry lab is not suitable place where the flask is used. The answer is No. [X](GT : Yes) | |
| Date Under-stand. | Q: It was 'X'/01/2021 a week ago. Today is 09/08/2021. | A: If today is 09/08/2021, so the date a week ago is 09/01/2021. The answer is 09. [✓] |
| Q: Yesterday, Jan 21, 2011, Jane ate 2 pizzas and 5 wings. One week ago from today is 01/15/2011. | A: one week ago from today is 01/15/2011. So today is 01/22/2011. So yesterday is 01/21/2011. So Jan 'X'1 is 01/21/2011. So 'X'011 is 2011. So The answer is 2011. [X](GT : B) |
| PROMPT FOR AQUA ALGEBRAIC WORD PROBLEMS +Q: John found that the average of 15 numbers is 40. If 10 is added to each number then the mean of the numbers is? +Answer Choices: (a) 50 (b) 45 (c) 65 (d) 78 (e) 64 +A: If 10 is added to each number, then the mean of the numbers also increases by 10. So the new mean would be 50. The answer is (a). +Q: If a / b = 3/4 and 8a + 5b = 22,then find the value of a. +Answer Choices: (a) 1/2 (b) 3/2 (c) 5/2 (d) 4/2 (e) 7/2 +A: If a / b = 3/4, then b = 4a / 3. So 8a + 5(4a / 3) = 22. This simplifies to 8a + 20a / 3 = 22, which means 44a / 3 = 22. So a is equal to 3/2. The answer is (b). +Q: A person is traveling at 20 km/hr and reached his destiny in 2.5 hr then find the distance? +Answer Choices: (a) 53 km (b) 55 km (c) 52 km (d) 60 km (e) 50 km +A: The distance that the person traveled would have been 20 km/hr * 2.5 hrs = 50 km. The answer is (e). +Q: How many keystrokes are needed to type the numbers from 1 to 500? +Answer Choices: (a) 1156 (b) 1392 (c) 1480 (d) 1562 (e) 1788 +A: There are 9 one-digit numbers from 1 to 9. There are 90 two-digit numbers from 10 to 99. There are 401 three-digit numbers from 100 to 500. 9 + 90(2) + 401(3) = 1392. The answer is (b). |
| PROMPT FOR DATE UNDERSTANDING +Q: 2015 is coming in 36 hours. What is the date one week from today in MM/DD/YYYY? +A: If 2015 is coming in 36 hours, then it is coming in 2 days. 2 days before 01/01/2015 is 12/30/2014, so today is 12/30/2014. So one week from today will be 01/05/2015. So the answer is 01/05/2015. +Q: The first day of 2019 is a Tuesday, and today is the first Monday of 2019. What is the date today in MM/DD/YYYY? +A: If the first day of 2019 was Tuesday, then 01/01/2019 was a Tuesday. Today is the first monday, would be six days later. So today is 01/07/2019. So the answer is 01/07/2019. +Q: The concert was scheduled to be on 06/01/1943, but was delayed by one day to today. What is the date 10 days ago in MM/DD/YYYY? +A: One day after 06/01/1943 is 06/02/1943, so today is 06/02/1943. 10 days before today is 05/23/1943. So the answer is 05/23/1943. +Q: It is 4/19/1969 today. What is the date 24 hours later in MM/DD/YYYY? +A: Today is 04/19/1969. 24 hours later is one day after today, which would be 04/20/1969. So the answer is 04/20/1969. +Q: Jane thought today is 3/11/2002, but today is in fact Mar 12, which is 1 day later. What is the date 24 hours later in MM/DD/YYYY? +A: Today is 03/12/2002. So the date 24 hours later will be 03/13/2002. So the answer is 03/13/2002. +Q: Jane was born on the last day of Feburuary in 2001. Today is her 16-year-old birthday. What is the date yesterday in MM/DD/YYYY? +A: The last day of February is the 28th, so Jane was born on 02/28/2001. Today is her 16-year old birthday, so today is 02/28/2017. So yesterday was 02/27/2017. So the answer is 02/27/2017. |
| PROMPT FOR DATE UNDERSTANDING +Q: "'X' is coming in 36 hours. One week from today is 01/05/2015." What is the answer of 'X'? +A: If The date one week from today is 01/05/2015, so today is 12/30/2014. So the data after 36 hours is 2015. The answer is 2015. +Q: "The first day of 'X' is a Tuesday, and today is the first Monday of 2019. Today is 01/07/2019." What is the answer of 'X'? +A: If today is the first Monday of 2019 and today is 01/07/2019. So The first day of 2019 is a Tuesday. The answer is 2019. +Q: "The concert was scheduled to be on 'X'/01/1943, but was delayed by one day to today. 10 days ago is 05/23/1943." +What is the answer of 'X'? +A: 10 days ago is 05/23/1943, and the concert was delayed by one day to today, so today is 06/02/1943. So the concert was scheduled to be on 06/01/1943 The answer is 06. +Q: "It is 'X'/19/1969 today. 24 hours later is 04/20/1969." What is the answer of 'X'? +A: 24 hours later is 04/20/1969. So today is 04/19/1969. The answer is 04. +Q: "Jane thought today is 'X'/12/2002, but today is in fact Mar 12, which is 1 day later. 24 hours later is 03/13/2002." +What is the answer of 'X'? +A: 24 hours later is 03/13/2002. So today is 03/12/2002. The answer is 03. +Q: "Jane was born on the last day of Feburary in 'X'. Today is her 16-year-old birthday. Yesterday is 02/27/2017" What is the answer of 'X'? +A: Yesterday is 02/27/2017, so today is 02/28/2017, Jane was born on 02/28/2001. The answer is 2001. |
| Method | Arithmetic Tasks | General Tasks | |||||||
| GSM8K | SingleEq | AddSub | MultiArith | AQUA-RAT | SVAMP | CSQA | DU | ||
| Previous SOTA (Fine-tune) | 35a/57b | 32.5c | 94.9d | 60.5e | 37.9f | 57.4g | 91.2h | - | |
| 9-12 year olds | 60i | - | - | - | - | - | - | - | |
| GPT-3 Standard | 19.7 | 86.8 | 90.9 | 44.0 | 29.5 | 69.9 | 82.3 | 49.0 | |
| GPT-3 (175B) code-davinci-001 | CoT | 13.84 | 60.20 | 58.55 | 45.85 | 18.90 | 38.42 | 46.75 | 38.72 |
| CoT+Self-Verification | 13.92 (+0.08) | 60.61 (+0.41) | 59.07 (+0.52) | 46.19 (+0.34) | 27.04 (+8.14) | 38.96 (+0.54) | 47.68 (+0.93) | 39.03 (+0.31) | |
| Instruct-GPT (175B) code-davinci-002 | CoT | 60.81 | 91.01 | 82.78 | 96.13 | 45.30 | 75.87 | 77.42 | 65.43 |
| CoT+Self-Verification | 65.14 (+4.33) | 93.40 (+2.39) | 86.33 (+3.55) | 99.15 (+3.02) | 47.95 (+2.65) | 76.99 (+1.12) | 77.83 (+0.41) | 66.57 (+1.14) |
| Self-Consistency Decoding (Wang et al., 2023c) For Forward Reasoning | |||||||||
| GPT-3 (175B) code-davinci-001 | SC | 23.40 | 70.25 | 68.65 | 79.82 | 25.60 | 54.58 | 54.92 | 49.26 |
| SC+ Self-Verification | 23.59(+0.19) | 70.50(+0.25) | 68.71(+0.06) | 80.01(+0.19) | 28.98(+3.38) | 54.68(+0.1) | 55.09(+0.17) | 49.72(+0.46) | |
| Instruct-GPT (175B) code-davinci-002 | SC | 78.00 | 96.78 | 91.64 | 100.0 | 52.01 | 86.77 | 81.43 | 71.58 |
| SC+Self-Verification | 78.32(+0.32) | 96.85(+0.07) | 92.03(+0.39) | 100.0(+0.0) | 52.25(+0.24) | 86.94(+0.17) | 81.53(+0.1) | 71.89(+0.31) |
| PAL (Gao et al., 2023) For Forward Reasoning | ||||||||
| GPT-3 (175B) code-davinci-001 | PAL | 31.82 | 63.98 | 63.15 | 61.52 | 30.56 | 42.69 | - |
| PAL+Self-Verification | 32.87(+1.05) | 65.45(+1.47) | 64.15(+1.0) | 61.76(+0.24) | 30.90(+0.34) | 42.78(+0.09) | - | |
| Instruct-GPT (175B) code-davinci-002 | PAL | 72.02 | 96.08 | 92.64 | 99.15 | 59.75 | 79.45 | - |
| PAL+Self-Verification | 72.89(+0.87) | 96.52(+0.44) | 93.78(+1.14) | 99.87(+0.72) | 60.21(+0.46) | 80.24(+0.79) | - |
| Dataset | Num. | Avg. Ans. | Ran. Acc. | CoT Acc. | Veri. Acc. |
| GSM8K | 601/1306 | 2.80 | 35.7% | 53.4% | 58.9% |
| Addsub | 77/377 | 2.13 | 46.9% | 56.7% | 74.0% |
| AQUA | 123/219 | 2.71 | 36.9% | 50.3% | 51.2% |
| Multiarith | 73/593 | 2.19 | 45.6% | 71.2% | 89.0% |
| SingleEq | 51/501 | 2.31 | 43.3% | 51.0% | 74.5% |
| SVAMP | 274/954 | 2.54 | 39.4% | 51.5% | 58.4% |
| AnswerFormat | Answer CleansingApproach | Pseudo Code(Example in Pytorch 3.7) |
| Number | Pick up the first numberencountered in the text. | pred = pred.replace("\\\\",".")\npred = [s for s in re.findall(r'^-?\\\\d+\\.?\\\\d*', pred)]\npred = pred[0] |
| Multiple-Choice | Pick up the first large letterencountered in the text. | pred = re.findall(r'A|B|C|D|E', pred)\npred = pred[0] |
| True or False | Pick up the first "True" or"False" encountered in thetext after removingunnec-essary letters. | pred = pred.lower()\npred = re.sub("\\\\\\"|\\n|\\.\\s|:\\\\","\\"", pred)\npred = pred.split("\\\\.")\npred = [i for i in pred if i in ("True", "False")] \npred = pred[0] |
| Yes or No | Pick up the first "yes" or"no" encountered in thetext after removingunnec-essary letters. | pred = pred.lower()\npred = re.sub("\\\\\\"|\\n|\\.\\s|:\\\\","\\"", pred)\npred = pred.split("\\\\.")\npred = [i for i in pred if i in ("yes", "no"]\npred = pred[0] |
| Free Format | Just remove unnecessaryletters, such as ".". | pred = re.sub("\\\\\\"|\\n|\\.\\s", "", pred) |
| True-False Item Verification | If Tim will need 36 lemons, and he will need to make 4 gallons of lemonade for the party, so he will need \\( 4 * 6 = {24} \\) lemons,he will need to make an extra gallonfor him that was twice as tart as the other gallons, so he will need \\( {24} * 2 = {48} \\) lemons, \\( {48}! = {36} \\) ,which is different from the theme. The answer is False. [X] |
| Conditional Masked Verification | Tim wanted to make lemonade for a pool party. For a gallon of lemonade, his recipe called for X cup of fresh lemon juice. He found that 6 lemons would yield 1 cup of juice. He figured he would need to make 4 gallons of lemonade for the party. His best friend Allen asked if Tim could make an extra gallon for him that was twice as tart as the other gallons. Tim will need 36.0 lemons. \\( \\mathrm{X} * 6 * 4 + \\mathrm{X} * 6 * 2 = {36} \\) , \\( {36} * \\mathrm{X} = {36},\\mathrm{X} = 1 \\) . The answer is 1. \\( \\left\\lbrack \\checkmark \\right\\rbrack \\) |
| Condition 1 mask | Condition 2 mask | ... |
| Q: Dana worked 'X' hours on Friday, 10 hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | Q: Dana worked 9 hours on Friday, 'X' hours on Saturday, and 3 hours on Sunday. She earns $13 per hour. Dana earn 169 money in weekend. | ... |
| Separate score: 0 | Separate score: 8 | ... |
| Total verification score: 19 |
| Dataset | Number of samples | Average words | Answer Format | Lience |
| GSM8K | 1319 | 46.9 | Number | MIT License |
| SingleEq | 508 | 27.4 | Number | MIT License |
| AddSub | 395 | 31.5 | Number | Unspecified |
| MultiArith | 600 | 31.8 | Number | Unspecified |
| AQUA-RAT | 254 | 51.9 | Multiple Choice | Apache-2.0 |
| SVAMP | 1000 | 31.8 | Number | MIT License |
| CSQA | 1221 | 27.8 | Multiple Choice | Unspecified |
| Data Understanding | 369 | 21.4 | Date | Unspecified |
| Dataset | Question | Self-Verification |
| GSM8K | Q: Matthew has a collection of 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy, and for the other half 'X' per toy. If Matthew would agree to that offer, He would earn 72.0 dollars. | A: Matthew has 12 unique toy soldiers. He wants to sell them for a fair price. He found a buyer who is willing to pay for half his collection $5 per toy. and for the other half $X per toy. If Matthew would agree to that offer. He would earn 72.0 dollars. 12/2 = 66*5 = 3012 - 6 = 66*X = 42 X = 42/6 X = 7 The answer is 7. [✓] |
| Continued on next page |
| Dataset | Question | Self-Verification |
| Q: Four students scored a total of 251 points in a board game. Naomi scored 68 of the points. Yuri scored 'X' more than half as many points as Naomi and Brianna scored 17 points more than Naomi. Jojo scored 54 points. | Four students scored a total of 251 points. Naomi scored 68 points. Yuri scored X points. Brianna scored 17 points more than Naomi. Jojo scored 54 points. So: 68 + X + 68 + 17 + 54 = 251, X = 251 - 68 - 68 - 17 - 54, X = 94. The answer is 94. [X](GT:10) | |
| SingleEq | Q: Sam grew 4 watermelons, but the rabbits ate 'X' watermelons. Sam have 1 watermelon left. | A: Sam grew 4 watermelons. But the rabbits ate X watermelons. So, we can write the following equation: 4 - X = 1, X = 4 - 1, X = 3. The answer is 3. [✓] |
| Q: Melanie, Benny, Sandy, and Jessica each have 'X' dozen Pokemon cards. They have 432 Pokemon cards in all. | A: Melanie has X dozen Pokemon cards. Benny has X dozen Pokemon cards. Sandy has X dozen Pokemon cards. Jessica has X dozen Pokemon cards. So, they have 4 X dozen Pokemon cards in total. X * 4 = 432, X = 432 / 4, X = 108. The answer is 108. [X](GT:9) | |
| Addsub | Q: Karin's science class weighed plastic rings for an experiment. They found that the orange ring weighed 'X' ounce, the purple ring weighed 0.3333333333333333 ounce, and the white ring weighed 0.41666666666667 ounce. The total weight of the plastic rings is 0.8333333333333333. | A: The orange ring weighed 0.08333333333333333 ounce. The purple ring weighed 0.333333333333333333 ounce. And the white ring weighed 0.4166666666666667 ounce. The total weight of the plastic rings is 0.83333333333333333333333 + 0.3333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333. |
| Q: There were 6 roses in the vase. Mary cut some roses from her flower garden. There are now 'X' roses in the vase. She cut 10 roses. | A: There were originally 6 roses in the vase. Mary cut 10 roses from her flower garden. So there are now X roses in the vase. <code>6 - 10 = X X = -4 </code> The answer is -4. [X](GT:16) | |
| MultiArith | Q: Luke was organizing his base-ball cards in a binder with 3 on each page. If he had 3 new cards and 'X' old cards to put in the binder, he use 4 pages. | A: Luke had 3 new cards and X old cards. He would use 4.0 pages. So, we can write the following equation: 3 + X = 3 * 4.0, 3 + X = 12, X = 12 - 3 X = 9. The answer is 9. [✓] |
| Q: Will bought 7 boxes of chocolate candy and gave 3 to his little brother. If each box has 'X' pieces inside it, Will still have 16 pieces. | A: Will bought 7 boxes of chocolate candy. Each box has X pieces inside it. He gave 3 boxes to his little brother. So he has 4 boxes left. 7 * X = Y, Y - 3 * X = 16, Y = 16 + 3 * X, Y = 16 + 3 * X. The answer is 16. [X](GT:4) | |
| Continued on next page |
| Dataset | Question | Self-Verification |
| SVAMP | Q: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has 'X' peaches and 15 apples. Steven has 3.0 more apples than peaches. | A: Jake has 3 fewer peaches and 10 more apples than Steven. Steven has X peaches and 15 apples. Steven has 3.0 more apples than peaches. X + 3 = 15, X = 15 - 3, X = 1.2 The answer is 12. [✓] |
| Q: An industrial machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for 'X' minutes today. The machine made 72 shirts today. | A: The machine can make 6 shirts a minute. It worked for 5 minutes yesterday and for X minutes today. The machine made 72 shirts today. So, we can write the following equation: 6 * 5 + 6 * X = 72, 30 + 6X = 72, 6X = 72 - 30, 6X = 42, X = 42 / 6 X = 7. The answer is 7. [X](GT : 12) | |
| Common-senseQA | Q: If there is a pond with trees around it, it likely located in forest. | A: he forest is likely located where there is a pond with trees around it. The answer is Yes. [✓] |
| Q: The flasks was used to distill elements, it is used in chemistry lab. | A: The answer should be the place where the flask is used. But the chemistry lab is not suitable place where the flask is used. The answer is No. [X](GT : Yes) | |
| Date Under-stand. | Q: It was 'X'/01/2021 a week ago. Today is 09/08/2021. | A: If today is 09/08/2021, so the date a week ago is 09/01/2021. The answer is 09. [✓] |
| Q: Yesterday, Jan 21, 2011, Jane ate 2 pizzas and 5 wings. One week ago from today is 01/15/2011. | A: one week ago from today is 01/15/2011. So today is 01/22/2011. So yesterday is 01/21/2011. So Jan 'X'1 is 01/21/2011. So 'X'011 is 2011. So The answer is 2011. [X](GT : B) |
| PROMPT FOR AQUA ALGEBRAIC WORD PROBLEMS\nQ: John found that the average of 15 numbers is 40. If 10 is added to each number then the mean of the numbers is?\nAnswer Choices: (a) 50 (b) 45 (c) 65 (d) 78 (e) 64\nA: If 10 is added to each number, then the mean of the numbers also increases by 10. So the new mean would be 50. The answer is (a).\nQ: If a / b = 3/4 and 8a + 5b = 22,then find the value of a.\nAnswer Choices: (a) 1/2 (b) 3/2 (c) 5/2 (d) 4/2 (e) 7/2\nA: If a / b = 3/4, then b = 4a / 3. So 8a + 5(4a / 3) = 22. This simplifies to 8a + 20a / 3 = 22, which means 44a / 3 = 22. So a is equal to 3/2. The answer is (b).\nQ: A person is traveling at 20 km/hr and reached his destiny in 2.5 hr then find the distance?\nAnswer Choices: (a) 53 km (b) 55 km (c) 52 km (d) 60 km (e) 50 km\nA: The distance that the person traveled would have been 20 km/hr * 2.5 hrs = 50 km. The answer is (e).\nQ: How many keystrokes are needed to type the numbers from 1 to 500?\nAnswer Choices: (a) 1156 (b) 1392 (c) 1480 (d) 1562 (e) 1788\nA: There are 9 one-digit numbers from 1 to 9. There are 90 two-digit numbers from 10 to 99. There are 401 three-digit numbers from 100 to 500. 9 + 90(2) + 401(3) = 1392. The answer is (b). |
| PROMPT FOR DATE UNDERSTANDING\nQ: 2015 is coming in 36 hours. What is the date one week from today in MM/DD/YYYY? \nA: If 2015 is coming in 36 hours, then it is coming in 2 days. 2 days before 01/01/2015 is 12/30/2014, so today is 12/30/2014. So one week from today will be 01/05/2015. So the answer is 01/05/2015. \nQ: The first day of 2019 is a Tuesday, and today is the first Monday of 2019. What is the date today in MM/DD/YYYY? \nA: If the first day of 2019 was Tuesday, then 01/01/2019 was a Tuesday. Today is the first monday, would be six days later. So today is 01/07/2019. So the answer is 01/07/2019. \nQ: The concert was scheduled to be on 06/01/1943, but was delayed by one day to today. What is the date 10 days ago in MM/DD/YYYY? \nA: One day after 06/01/1943 is 06/02/1943, so today is 06/02/1943. 10 days before today is 05/23/1943. So the answer is 05/23/1943. \nQ: It is 4/19/1969 today. What is the date 24 hours later in MM/DD/YYYY? \nA: Today is 04/19/1969. 24 hours later is one day after today, which would be 04/20/1969. So the answer is 04/20/1969. \nQ: Jane thought today is 3/11/2002, but today is in fact Mar 12, which is 1 day later. What is the date 24 hours later in MM/DD/YYYY? \nA: Today is 03/12/2002. So the date 24 hours later will be 03/13/2002. So the answer is 03/13/2002. \nQ: Jane was born on the last day of Feburuary in 2001. Today is her 16-year-old birthday. What is the date yesterday in MM/DD/YYYY? \nA: The last day of February is the 28th, so Jane was born on 02/28/2001. Today is her 16-year old birthday, so today is 02/28/2017. So yesterday was 02/27/2017. So the answer is 02/27/2017. |
| PROMPT FOR DATE UNDERSTANDING\nQ: "'X' is coming in 36 hours. One week from today is 01/05/2015." What is the answer of 'X'? \nA: If The date one week from today is 01/05/2015, so today is 12/30/2014. So the data after 36 hours is 2015. The answer is 2015.\nQ: "The first day of 'X' is a Tuesday, and today is the first Monday of 2019. Today is 01/07/2019." What is the answer of 'X'? \nA: If today is the first Monday of 2019 and today is 01/07/2019. So The first day of 2019 is a Tuesday. The answer is 2019.\nQ: "The concert was scheduled to be on 'X'/01/1943, but was delayed by one day to today. 10 days ago is 05/23/1943."\nWhat is the answer of 'X'? \nA: 10 days ago is 05/23/1943, and the concert was delayed by one day to today, so today is 06/02/1943. So the concert was scheduled to be on 06/01/1943 The answer is 06.\nQ: "It is 'X'/19/1969 today. 24 hours later is 04/20/1969." What is the answer of 'X'? \nA: 24 hours later is 04/20/1969. So today is 04/19/1969. The answer is 04.\nQ: "Jane thought today is 'X'/12/2002, but today is in fact Mar 12, which is 1 day later. 24 hours later is 03/13/2002."\nWhat is the answer of 'X'? \nA: 24 hours later is 03/13/2002. So today is 03/12/2002. The answer is 03.\nQ: "Jane was born on the last day of Feburary in 'X'. Today is her 16-year-old birthday. Yesterday is 02/27/2017" What is the answer of 'X'? \nA: Yesterday is 02/27/2017, so today is 02/28/2017, Jane was born on 02/28/2001. The answer is 2001. |
| Metrics | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1* | 0.193 | 0.202 | 0.136 | 0.155 | 0.186 | 0.121 | 0.075 | 0.153 | 0.058 | 0.323 | 0.361 | 0.231 |
| ROUGE-2* | 0.145 | 0.146 | 0.101 | 0.137 | 0.156 | 0.107 | 0.053 | 0.095 | 0.041 | 0.255 | 0.262 | 0.181 |
| ROUGE-L* | 0.148 | 0.158 | 0.105 | 0.133 | 0.167 | 0.103 | 0.078 | 0.146 | 0.060 | 0.306 | 0.340 | 0.219 |
| BERTScore* | 0.375 | 0.383 | 0.265 | 0.163 | 0.182 | 0.127 | 0.167 | 0.229 | 0.130 | 0.396 | 0.414 | 0.285 |
| BARTScore* | 0.381 | 0.391 | 0.275 | 0.271 | 0.265 | 0.212 | 0.168 | 0.187 | 0.131 | 0.381 | 0.391 | 0.276 |
| BARTScore-CNN* | 0.461 | 0.480 | 0.332 | 0.389 | 0.413 | 0.305 | 0.310 | 0.378 | 0.241 | 0.425 | 0.450 | 0.309 |
| BARTScore-CNN-PARA* | 0.455 | 0.455 | 0.328 | 0.413 | 0.459 | 0.324 | 0.368 | 0.417 | 0.286 | 0.414 | 0.440 | 0.299 |
| ChatGPT-RTS | 0.388 | 0.399 | 0.312 | 0.423 | 0.532 | 0.378 | 0.285 | 0.302 | 0.240 | 0.448 | 0.463 | 0.357 |
| ChatGPT-MCQ | 0.424 | 0.416 | 0.350 | 0.343 | 0.487 | 0.320 | 0.343 | 0.431 | 0.305 | 0.384 | 0.395 | 0.329 |
| GPT-4-RTS | 0.427 | 0.461 | 0.361 | 0.556 | 0.618 | 0.522 | 0.498 | 0.600 | 0.452 | 0.448 | 0.428 | 0.373 |
| Metrics | Coh | Con | Flu | Rel | Avg |
| Random | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 |
| ROUGE-1 | 3/40 | 5/46 | 3/46 | 4/47 | 3.75/44.75 |
| ROUGE-2 | 2/38 | 7/48 | 3/45 | 4/46 | 4.00/44.25 |
| ROUGE-L | 2/31 | 5/37 | 4/39 | 6/41 | 4.25/37.00 |
| BERTScore | 4/48 | 5/44 | 4/44 | 6/46 | 4.75/45.50 |
| BARTScore | 8/46 | 7/48 | 5/45 | 6/46 | 6.50/46.25 |
| BARTScore-CNN | 9/53 | 5/53 | 5/54 | 4/53 | 5.75/53.25 |
| BARTScore-CNN-PARA | 9/49 | 8/54 | 6/52 | 5/51 | 7.00/51.50 |
| ChatGPT-RTS | 6/54 | 6/56 | 9/62 | 7/62 | 7.00/58.50 |
| ChatGPT-MCQ | 5/54 | 7/56 | 8/60 | 7/58 | 6.75/57.00 |
| ChatGPT-H2H | 8/- | 7/- | 7/- | 4/- | 6.50/- |
| GPT-4-RTS | 5/55 | 7/53 | 8/60 | 7/56 | 6.75/56.00 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1 | 0.0 | -0.032 | -0.091 | -0.490 | -0.527 | -0.303 | -0.420 | -0.518 | -0.273 | -0.420 | -0.387 | -0.273 |
| ROUGE-2 | 0.559 | 0.508 | 0.485 | -0.259 | -0.480 | -0.152 | -0.217 | -0.438 | -0.152 | -0.084 | -0.120 | -0.121 |
| ROUGE-L | 0.231 | 0.251 | 0.121 | -0.510 | -0.522 | -0.303 | -0.168 | -0.412 | -0.152 | -0.224 | -0.266 | -0.121 |
| BERTScore | -0.413 | -0.403 | -0.212 | -0.580 | -0.869 | -0.424 | -0.455 | -0.663 | -0.303 | -0.685 | -0.756 | -0.515 |
| BARTScore | -0.916 | -0.747 | -0.788 | -0.266 | -0.504 | -0.121 | 0.154 | 0.123 | 0.182 | -0.769 | -0.837 | -0.606 |
| BARTScore-CNN | -0.748 | -0.800 | -0.636 | -0.671 | -0.913 | -0.515 | -0.510 | -0.604 | -0.485 | -0.825 | -0.852 | -0.667 |
| BARTScore-CNN-PARA | -0.720 | -0.858 | -0.606 | -0.685 | -0.888 | -0.576 | -0.294 | -0.522 | -0.212 | -0.853 | -0.880 | -0.727 |
| ChatGPT-RTS | -0.042 | -0.072 | -0.121 | -0.811 | -0.751 | -0.636 | -0.748 | -0.728 | -0.606 | -0.559 | -0.473 | -0.394 |
| ChatGPT-MCQ | -0.175 | -0.11 | -0.182 | -0.818 | -0.411 | -0.636 | -0.622 | -0.484 | -0.394 | -0.350 | -0.622 | -0.212 |
| GPT-4-RTS | -0.531 | -0.678 | -0.424 | -0.600 | -0.103 | -0.236 | -0.839 | -0.520 | -0.515 | -0.958 | -0.880 | -0.848 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ρ(Ri, PRTS) | 0.343 | 0.198 | 0.091 | 0.685 | 0.506 | 0.576 | 0.727 | 0.797 | 0.545 | 0.168 | 0.128 | 0.000 |
| ρ(Ri, PMCQ) | 0.657 | 0.625 | 0.394 | 0.685 | 0.616 | 0.515 | 0.322 | 0.573 | 0.212 | 0.091 | -0.106 | 0.000 |
| Coh | Con | Flu | Rel | Avg | |
| ChatGPT-RTS2 | 5 | 6 | 7 | 6 | 6.00 |
| ChatGPT-MCQ2 | 4 | 7 | 6 | 4 | 5.25 |
| ChatGPT-StarEval | 7 | 7 | 7 | 6 | 6.75 |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| RTS2 | Human-Corr | 0.339 | 0.338 | 0.285 | 0.393 | 0.497 | 0.350 | 0.290 | 0.280 | 0.252 | 0.440 | 0.441 | 0.351 |
| Meta-Corr | -0.559 | -0.476 | -0.424 | -0.748 | -0.843 | -0.576 | -0.825 | -0.823 | -0.636 | -0.385 | -0.506 | -0.212 | |
| MCQ2 | Human-Corr | 0.430 | 0.423 | 0.355 | 0.327 | 0.483 | 0.306 | 0.258 | 0.396 | 0.229 | 0.240 | 0.258 | 0.206 |
| Meta-Corr | -0.308 | -0.275 | -0.212 | -0.545 | -0.446 | -0.364 | -0.811 | -0.615 | -0.636 | -0.217 | -0.707 | -0.182 | |
| StarEval | Human-Corr | 0.418 | 0.417 | 0.341 | 0.297 | 0.421 | 0.264 | 0.246 | 0.323 | 0.217 | 0.393 | 0.405 | 0.323 |
| Meta-Corr | -0.084 | -0.101 | 0.000 | -0.664 | -0.684 | -0.545 | -0.441 | -0.460 | -0.212 | -0.497 | -0.579 | -0.303 |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| 7B | Human-Corr | -0.001 | -0.000 | -0.001 | 0.114 | 0.130 | 0.104 | 0.043 | 0.038 | 0.039 | 0.067 | 0.064 | 0.057 |
| Meta-Corr | 0.245 | -0.258 | 0.182 | -0.524 | -0.720 | -0.424 | -0.042 | 0.463 | -0.03 | 0.238 | 0.169 | 0.212 | |
| 13B | Human-Corr | 0.153 | 0.165 | 0.123 | 0.180 | 0.209 | 0.162 | 0.187 | 0.179 | 0.167 | 0.234 | 0.266 | 0.192 |
| Meta-Corr | -0.287 | -0.425 | -0.182 | -0.580 | -0.495 | -0.424 | -0.455 | -0.233 | -0.303 | -0.049 | -0.406 | 0.000 | |
| 70B | Human-Corr | 0.234 | 0.254 | 0.186 | 0.357 | 0.395 | 0.319 | 0.155 | 0.161 | 0.134 | 0.248 | 0.285 | 0.200 |
| Meta-Corr | -0.420 | -0.407 | -0.273 | -0.811 | -0.690 | -0.667 | -0.322 | -0.319 | -0.182 | -0.238 | -0.123 | -0.182 |
| ID | Coh | Con | Flu | Rel |
| M8 | 0.58 | 0.09 | 0.13 | 0.51 |
| M9 | 0.65 | 0.15 | 0.34 | 0.54 |
| M10 | 0.58 | 0.22 | 0.20 | 0.54 |
| M11 | 0.59 | 0.34 | 0.40 | 0.49 |
| M12 | 0.65 | 0.03 | 0.14 | 0.54 |
| M13 | 0.62 | 0.07 | 0.14 | 0.51 |
| M14 | 0.65 | 0.09 | 0.18 | 0.54 |
| M15 | 0.60 | 0.04 | 0.13 | 0.53 |
| M17 | 0.58 | 0.05 | 0.08 | 0.52 |
| M20 | 0.48 | 0.29 | 0.24 | 0.49 |
| M22 | 0.58 | 0.02 | 0.14 | 0.54 |
| M23 | 0.58 | 0.05 | 0.12 | 0.54 |
| ID | Model Name | Coh | Con | Flu | Rel | Avg |
| M22 | BART | 4.18 | 4.94 | 4.90 | 4.25 | 4.57 |
| M23 | Pegasus (C4) | 4.16 | 4.91 | 4.88 | 4.26 | 4.55 |
| M17 | T5 | 4.00 | 4.93 | 4.93 | 4.23 | 4.52 |
| M12 | Unified-ext-abs | 3.60 | 4.96 | 4.85 | 3.85 | 4.32 |
| M13 | ROUGESal | 3.44 | 4.82 | 4.86 | 3.83 | 4.24 |
| M15 | Closed book decoder | 3.35 | 4.95 | 4.80 | 3.67 | 4.19 |
| M14 | Multi-task (Ent + QG) | 3.20 | 4.90 | 4.74 | 3.63 | 4.12 |
| M8 | Pointer Generator | 3.29 | 4.65 | 4.79 | 3.55 | 4.07 |
| M9 | Fast-abs-rl | 2.38 | 4.67 | 4.50 | 3.52 | 3.77 |
| M10 | Bottom-Up | 2.73 | 4.25 | 4.42 | 3.38 | 3.70 |
| M20 | GPT-2 (zero-shot) | 3.63 | 3.40 | 3.97 | 3.30 | 3.58 |
| M11 | Improve-abs | 2.28 | 3.27 | 3.65 | 3.15 | 3.09 |
| Model A | Model B | Coherence | Consistency | Fluency | Relevance | ||||||||
| LLM | Human | LLM | Human | LLM | Human | LLM | Human | ||||||
| M22 | M23 | 65.5 | 53.5 | ✓ | 55.75 | 52.5 | ✓ | 61 | 49.5 | × | 58.75 | 49.5 | × |
| M23 | M17 | 48.25 | 52.5 | × | 47 | 49 | ✓ | 49 | 45.5 | ✓ | 45 | 52 | × |
| M17 | M12 | 44 | 66.5 | × | 43.25 | 48.5 | ✓ | 40.5 | 54.5 | × | 49.25 | 72.5 | × |
| M12 | M13 | 58 | 51 | ✓ | 56.5 | 54.5 | ✓ | 56.75 | 50 | × | 58 | 45 | × |
| M13 | M15 | 45.5 | 49.5 | ✓ | 52 | 46 | × | 48.75 | 52 | × | 51.25 | 60.5 | ✓ |
| M15 | M14 | 57 | 54 | ✓ | 55 | 53.5 | ✓ | 56.5 | 52 | ✓ | 54 | 57 | ✓ |
| M14 | M8 | 49.25 | 44 | ✓ | 50 | 54.5 | × | 47 | 46.5 | ✓ | 49.25 | 53.5 | × |
| M8 | M9 | 77.5 | 82 | ✓ | 78.5 | 53 | ✓ | 80.5 | 63.5 | ✓ | 76 | 54 | ✓ |
| M9 | M10 | 45 | 36 | ✓ | 41.5 | 58 | × | 46 | 44.5 | ✓ | 41.5 | 56 | × |
| M10 | M20 | 58.25 | 24 | × | 61.75 | 64 | ✓ | 63.75 | 61.5 | ✓ | 61.5 | 54.5 | ✓ |
| M20 | M11 | 56.5 | 82 | ✓ | 50 | 53 | × | 51 | 58.5 | ✓ | 50 | 53 | × |
| #CP | 8 | 7 | 7 | 4 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.420 | 0.383 | 0.323 | 0.229 | 0.273 | 0.209 | 0.274 | 0.245 | 0.236 | 0.519 | 0.509 | 0.438 |
| M9 | 0.174 | 0.243 | 0.142 | 0.119 | 0.265 | 0.103 | 0.256 | 0.258 | 0.219 | 0.254 | 0.307 | 0.200 |
| M10 | 0.365 | 0.415 | 0.292 | 0.305 | 0.452 | 0.251 | 0.258 | 0.288 | 0.223 | 0.367 | 0.378 | 0.284 |
| M11 | 0.378 | 0.420 | 0.320 | 0.404 | 0.488 | 0.335 | 0.227 | 0.288 | 0.182 | 0.501 | 0.511 | 0.394 |
| M12 | 0.208 | 0.237 | 0.160 | 0.087 | 0.020 | 0.082 | 0.086 | 0.107 | 0.071 | 0.438 | 0.451 | 0.354 |
| M13 | 0.455 | 0.473 | 0.359 | 0.178 | 0.037 | 0.167 | 0.063 | 0.151 | 0.055 | 0.403 | 0.403 | 0.329 |
| M14 | 0.433 | 0.467 | 0.355 | 0.114 | 0.187 | 0.105 | -0.007 | 0.055 | -0.005 | 0.421 | 0.435 | 0.336 |
| M15 | 0.372 | 0.385 | 0.279 | 0.189 | 0.316 | 0.177 | 0.100 | 0.087 | 0.085 | 0.252 | 0.288 | 0.193 |
| M17 | 0.291 | 0.320 | 0.233 | -0.086 | -0.061 | -0.084 | 0.017 | 0.046 | 0.015 | 0.204 | 0.199 | 0.175 |
| M20 | 0.382 | 0.394 | 0.310 | 0.528 | 0.501 | 0.430 | 0.367 | 0.349 | 0.307 | 0.292 | 0.278 | 0.237 |
| M22 | 0.215 | 0.293 | 0.162 | -0.072 | -0.052 | -0.070 | -0.114 | -0.131 | -0.101 | 0.201 | 0.300 | 0.172 |
| M23 | 0.392 | 0.427 | 0.320 | 0.003 | 0.305 | 0.002 | -0.078 | -0.022 | -0.069 | 0.223 | 0.324 | 0.184 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.289 | 0.310 | 0.236 | 0.235 | 0.362 | 0.226 | 0.348 | 0.348 | 0.321 | 0.349 | 0.419 | 0.302 |
| M9 | 0.170 | 0.170 | 0.148 | 0.211 | 0.321 | 0.200 | 0.198 | 0.299 | 0.174 | 0.318 | 0.356 | 0.276 |
| M10 | 0.352 | 0.314 | 0.293 | 0.138 | 0.273 | 0.125 | 0.155 | 0.210 | 0.138 | 0.420 | 0.427 | 0.362 |
| M11 | 0.285 | 0.312 | 0.250 | 0.380 | 0.370 | 0.317 | 0.200 | 0.218 | 0.163 | 0.397 | 0.428 | 0.334 |
| M12 | 0.306 | 0.304 | 0.258 | -0.025 | -0.059 | -0.024 | 0.256 | 0.306 | 0.239 | 0.283 | 0.287 | 0.246 |
| M13 | 0.425 | 0.425 | 0.351 | 0.471 | 0.312 | 0.457 | 0.199 | 0.214 | 0.186 | 0.435 | 0.402 | 0.375 |
| M14 | 0.490 | 0.472 | 0.422 | 0.112 | 0.157 | 0.110 | 0.084 | 0.143 | 0.078 | 0.326 | 0.329 | 0.284 |
| M15 | 0.317 | 0.298 | 0.250 | 0.061 | 0.513 | 0.060 | 0.055 | 0.025 | 0.050 | 0.433 | 0.420 | 0.378 |
| M17 | 0.250 | 0.255 | 0.215 | -0.106 | -0.081 | -0.105 | -0.011 | 0.024 | -0.011 | 0.293 | 0.285 | 0.260 |
| M20 | 0.463 | 0.450 | 0.381 | 0.455 | 0.442 | 0.371 | 0.494 | 0.450 | 0.404 | 0.326 | 0.334 | 0.264 |
| M22 | 0.211 | 0.173 | 0.182 | -0.096 | -0.080 | -0.095 | -0.092 | -0.056 | -0.087 | 0.352 | 0.371 | 0.313 |
| M23 | 0.218 | 0.209 | 0.189 | 0.107 | 0.448 | 0.104 | -0.275 | -0.269 | -0.261 | 0.147 | 0.187 | 0.129 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||||||
| Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | |
| M8 | 2.43 | 3.70 | 4.56 | 3.29 | 4.21 | 4.70 | 4.77 | 4.65 | 2.51 | 3.76 | 4.61 | 4.79 | 3.56 | 4.33 | 4.71 | 3.55 |
| M9 | 1.80 | 3.51 | 4.41 | 2.38 | 3.93 | 4.69 | 4.94 | 4.67 | 2.16 | 3.45 | 4.15 | 4.50 | 3.51 | 4.33 | 4.82 | 3.52 |
| M10 | 2.05 | 3.49 | 4.05 | 2.73 | 3.77 | 4.52 | 4.60 | 4.25 | 2.23 | 3.43 | 3.92 | 4.42 | 3.45 | 4.25 | 4.62 | 3.38 |
| M11 | 1.70 | 2.63 | 2.93 | 2.28 | 2.36 | 4.11 | 3.72 | 3.27 | 1.71 | 2.66 | 2.77 | 3.65 | 2.92 | 4.06 | 3.87 | 3.15 |
| M12 | 2.75 | 3.92 | 4.75 | 3.60 | 4.46 | 4.75 | 5.00 | 4.96 | 2.59 | 3.84 | 4.72 | 4.85 | 3.89 | 4.40 | 4.95 | 3.85 |
| M13 | 2.91 | 3.99 | 4.76 | 3.44 | 4.40 | 4.73 | 4.89 | 4.82 | 2.69 | 3.93 | 4.69 | 4.86 | 3.90 | 4.41 | 4.85 | 3.83 |
| M14 | 2.38 | 3.87 | 4.54 | 3.20 | 4.25 | 4.70 | 4.99 | 4.90 | 2.42 | 3.83 | 4.53 | 4.74 | 3.67 | 4.32 | 4.84 | 3.63 |
| M15 | 2.65 | 3.81 | 4.58 | 3.35 | 4.33 | 4.75 | 4.90 | 4.95 | 2.60 | 3.84 | 4.61 | 4.80 | 3.78 | 4.32 | 4.82 | 3.67 |
| M17 | 3.05 | 4.11 | 4.89 | 4.00 | 4.84 | 4.87 | 4.97 | 4.93 | 3.29 | 4.08 | 4.79 | 4.93 | 4.30 | 4.43 | 4.97 | 4.23 |
| M20 | 2.00 | 2.95 | 2.99 | 3.63 | 2.66 | 3.57 | 3.51 | 3.40 | 2.22 | 3.02 | 3.05 | 3.97 | 2.73 | 3.78 | 3.47 | 3.30 |
| M22 | 3.54 | 4.09 | 4.92 | 4.18 | 4.83 | 4.82 | 4.97 | 4.94 | 3.57 | 4.06 | 4.94 | 4.90 | 4.41 | 4.43 | 4.95 | 4.25 |
| M23 | 3.00 | 4.14 | 4.88 | 4.16 | 4.74 | 4.84 | 4.94 | 4.91 | 3.08 | 4.05 | 4.76 | 4.88 | 4.10 | 4.45 | 4.96 | 4.26 |
| avg | 2.52 | 3.68 | 4.35 | 3.35 | 4.06 | 4.59 | 4.68 | 4.55 | 2.59 | 3.66 | 4.29 | 4.61 | 3.68 | 4.29 | 4.65 | 3.72 |
| Consistency | The summary is inconsistent with the article as it omits important details and repeats a phrase multiple times, resulting in a score of 1. |
| Consistency | The summary includes some relevant information but lacks coherence and omits important details, resulting in a score of 2.5 out of 5 for consistency. |
| Fluency | The summary is well-written and grammatically correct, but lacks important details from the article such as the potential link between antibiotics and weight gain, and the warnings about over-prescription of antibiotics and the development of superbugs. Score: 3/5. |
| Fluency | The summary lacks coherence and does not provide a clear understanding of the article, resulting in a score of 1 for disfluency. |
| Coherence | The summary is somewhat coherent but includes incorrect information, such as stating that Merson restarted his row with Burnley instead of Andros Townsend, and does not flow smoothly; therefore, I would give it a score of 3. |
| Consistency | The summary is consistent with the article as it mentions the helicopter airlifting patients to the hospital, but it leaves out important details such as the number of people involved and their injuries. Final score: 3. |
| Consistency | The summary accurately reflects the main points of the article, although it could have mentioned Falcao's unhappiness in France. Final score: 4. |
| Fluency | The summary is coherent and grammatically correct, but it does not provide enough information about the incident. Final score: 3. |
| Fluency | The summary is only one sentence and do not provide enough information about the article's content. Final score: 1. |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.429 | 0.446 | 0.362 | 0.449 | 0.597 | 0.433 | 0.412 | 0.409 | 0.385 | 0.425 | 0.489 | 0.365 |
| M9 | 0.288 | 0.346 | 0.243 | 0.349 | 0.350 | 0.331 | 0.462 | 0.465 | 0.407 | 0.413 | 0.490 | 0.358 |
| M10 | 0.495 | 0.480 | 0.403 | 0.518 | 0.683 | 0.469 | 0.406 | 0.579 | 0.360 | 0.510 | 0.537 | 0.429 |
| M11 | 0.584 | 0.572 | 0.474 | 0.529 | 0.536 | 0.439 | 0.448 | 0.436 | 0.358 | 0.500 | 0.519 | 0.405 |
| M12 | 0.245 | 0.360 | 0.209 | - | - | - | 0.387 | 0.506 | 0.372 | 0.169 | 0.106 | 0.149 |
| M13 | 0.271 | 0.271 | 0.230 | 0.535 | 0.444 | 0.518 | 0.093 | 0.337 | 0.089 | 0.237 | 0.337 | 0.206 |
| M14 | 0.263 | 0.312 | 0.222 | -0.040 | -0.026 | -0.040 | 0.311 | 0.432 | 0.291 | 0.277 | 0.315 | 0.241 |
| M15 | 0.403 | 0.418 | 0.342 | 0.158 | 0.504 | 0.155 | 0.240 | 0.242 | 0.225 | 0.307 | 0.402 | 0.267 |
| M17 | 0.285 | 0.240 | 0.246 | 0.390 | 0.748 | 0.385 | 0.091 | 0.131 | 0.088 | 0.210 | 0.253 | 0.185 |
| M20 | 0.427 | 0.418 | 0.334 | 0.378 | 0.362 | 0.313 | 0.482 | 0.468 | 0.402 | 0.479 | 0.491 | 0.391 |
| M22 | 0.026 | 0.004 | 0.023 | 0.346 | 0.608 | 0.343 | 0.248 | 0.141 | 0.242 | 0.143 | 0.084 | 0.126 |
| M23 | 0.320 | 0.330 | 0.283 | 0.360 | 0.168 | 0.354 | 0.267 | 0.136 | 0.251 | 0.005 | 0.003 | 0.005 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.464 | 0.455 | 0.407 | 0.557 | 0.613 | 0.526 | 0.391 | 0.367 | 0.341 | 0.517 | 0.514 | 0.471 |
| M9 | 0.436 | 0.415 | 0.416 | 0.215 | 0.342 | 0.195 | 0.578 | 0.571 | 0.517 | 0.450 | 0.485 | 0.412 |
| M10 | 0.374 | 0.345 | 0.340 | 0.300 | 0.297 | 0.270 | 0.631 | 0.585 | 0.551 | 0.506 | 0.507 | 0.459 |
| M11 | 0.326 | 0.428 | 0.307 | 0.437 | 0.421 | 0.393 | 0.354 | 0.509 | 0.303 | 0.467 | 0.486 | 0.419 |
| M12 | 0.500 | 0.469 | 0.444 | 0.446 | 0.486 | 0.424 | 0.273 | 0.320 | 0.241 | 0.588 | 0.591 | 0.542 |
| M13 | 0.462 | 0.450 | 0.408 | 0.470 | 0.510 | 0.445 | 0.325 | 0.345 | 0.282 | 0.515 | 0.498 | 0.475 |
| M14 | 0.524 | 0.460 | 0.463 | 0.325 | 0.293 | 0.307 | 0.216 | 0.257 | 0.194 | 0.501 | 0.476 | 0.455 |
| M15 | 0.534 | 0.499 | 0.470 | 0.266 | 0.326 | 0.251 | 0.383 | 0.369 | 0.337 | 0.584 | 0.550 | 0.535 |
| M17 | 0.312 | 0.294 | 0.274 | -0.016 | -0.008 | -0.015 | 0.251 | 0.261 | 0.225 | 0.530 | 0.479 | 0.504 |
| M20 | 0.636 | 0.623 | 0.566 | 0.733 | 0.664 | 0.644 | 0.566 | 0.556 | 0.492 | 0.570 | 0.565 | 0.496 |
| M22 | 0.314 | 0.338 | 0.276 | 0.245 | 0.151 | 0.241 | 0.371 | 0.341 | 0.330 | 0.471 | 0.412 | 0.451 |
| M23 | 0.380 | 0.386 | 0.337 | 0.158 | 0.496 | 0.153 | 0.296 | 0.275 | 0.259 | 0.450 | 0.505 | 0.421 |
| Metrics | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1* | 0.193 | 0.202 | 0.136 | 0.155 | 0.186 | 0.121 | 0.075 | 0.153 | 0.058 | 0.323 | 0.361 | 0.231 |
| ROUGE-2* | 0.145 | 0.146 | 0.101 | 0.137 | 0.156 | 0.107 | 0.053 | 0.095 | 0.041 | 0.255 | 0.262 | 0.181 |
| ROUGE-L* | 0.148 | 0.158 | 0.105 | 0.133 | 0.167 | 0.103 | 0.078 | 0.146 | 0.060 | 0.306 | 0.340 | 0.219 |
| BERTScore* | 0.375 | 0.383 | 0.265 | 0.163 | 0.182 | 0.127 | 0.167 | 0.229 | 0.130 | 0.396 | 0.414 | 0.285 |
| BARTScore* | 0.381 | 0.391 | 0.275 | 0.271 | 0.265 | 0.212 | 0.168 | 0.187 | 0.131 | 0.381 | 0.391 | 0.276 |
| BARTScore-CNN* | 0.461 | 0.480 | 0.332 | 0.389 | 0.413 | 0.305 | 0.310 | 0.378 | 0.241 | 0.425 | 0.450 | 0.309 |
| BARTScore-CNN-PARA* | 0.455 | 0.455 | 0.328 | 0.413 | 0.459 | 0.324 | 0.368 | 0.417 | 0.286 | 0.414 | 0.440 | 0.299 |
| ChatGPT-RTS | 0.388 | 0.399 | 0.312 | 0.423 | 0.532 | 0.378 | 0.285 | 0.302 | 0.240 | 0.448 | 0.463 | 0.357 |
| ChatGPT-MCQ | 0.424 | 0.416 | 0.350 | 0.343 | 0.487 | 0.320 | 0.343 | 0.431 | 0.305 | 0.384 | 0.395 | 0.329 |
| GPT-4-RTS | 0.427 | 0.461 | 0.361 | 0.556 | 0.618 | 0.522 | 0.498 | 0.600 | 0.452 | 0.448 | 0.428 | 0.373 |
| Metrics | Coh | Con | Flu | Rel | Avg |
| Random | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 |
| ROUGE-1 | 3/40 | 5/46 | 3/46 | 4/47 | 3.75/44.75 |
| ROUGE-2 | 2/38 | 7/48 | 3/45 | 4/46 | 4.00/44.25 |
| ROUGE-L | 2/31 | 5/37 | 4/39 | 6/41 | 4.25/37.00 |
| BERTScore | 4/48 | 5/44 | 4/44 | 6/46 | 4.75/45.50 |
| BARTScore | 8/46 | 7/48 | 5/45 | 6/46 | 6.50/46.25 |
| BARTScore-CNN | 9/53 | 5/53 | 5/54 | 4/53 | 5.75/53.25 |
| BARTScore-CNN-PARA | 9/49 | 8/54 | 6/52 | 5/51 | 7.00/51.50 |
| ChatGPT-RTS | 6/54 | 6/56 | 9/62 | 7/62 | 7.00/58.50 |
| ChatGPT-MCQ | 5/54 | 7/56 | 8/60 | 7/58 | 6.75/57.00 |
| ChatGPT-H2H | 8/- | 7/- | 7/- | 4/- | 6.50/- |
| GPT-4-RTS | 5/55 | 7/53 | 8/60 | 7/56 | 6.75/56.00 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1 | 0.0 | -0.032 | -0.091 | -0.490 | -0.527 | -0.303 | -0.420 | -0.518 | -0.273 | -0.420 | -0.387 | -0.273 |
| ROUGE-2 | 0.559 | 0.508 | 0.485 | -0.259 | -0.480 | -0.152 | -0.217 | -0.438 | -0.152 | -0.084 | -0.120 | -0.121 |
| ROUGE-L | 0.231 | 0.251 | 0.121 | -0.510 | -0.522 | -0.303 | -0.168 | -0.412 | -0.152 | -0.224 | -0.266 | -0.121 |
| BERTScore | -0.413 | -0.403 | -0.212 | -0.580 | -0.869 | -0.424 | -0.455 | -0.663 | -0.303 | -0.685 | -0.756 | -0.515 |
| BARTScore | -0.916 | -0.747 | -0.788 | -0.266 | -0.504 | -0.121 | 0.154 | 0.123 | 0.182 | -0.769 | -0.837 | -0.606 |
| BARTScore-CNN | -0.748 | -0.800 | -0.636 | -0.671 | -0.913 | -0.515 | -0.510 | -0.604 | -0.485 | -0.825 | -0.852 | -0.667 |
| BARTScore-CNN-PARA | -0.720 | -0.858 | -0.606 | -0.685 | -0.888 | -0.576 | -0.294 | -0.522 | -0.212 | -0.853 | -0.880 | -0.727 |
| ChatGPT-RTS | -0.042 | -0.072 | -0.121 | -0.811 | -0.751 | -0.636 | -0.748 | -0.728 | -0.606 | -0.559 | -0.473 | -0.394 |
| ChatGPT-MCQ | -0.175 | -0.11 | -0.182 | -0.818 | -0.411 | -0.636 | -0.622 | -0.484 | -0.394 | -0.350 | -0.622 | -0.212 |
| GPT-4-RTS | -0.531 | -0.678 | -0.424 | -0.600 | -0.103 | -0.236 | -0.839 | -0.520 | -0.515 | -0.958 | -0.880 | -0.848 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ρ(Ri, PRTS) | 0.343 | 0.198 | 0.091 | 0.685 | 0.506 | 0.576 | 0.727 | 0.797 | 0.545 | 0.168 | 0.128 | 0.000 |
| ρ(Ri, PMCQ) | 0.657 | 0.625 | 0.394 | 0.685 | 0.616 | 0.515 | 0.322 | 0.573 | 0.212 | 0.091 | -0.106 | 0.000 |
| Coh | Con | Flu | Rel | Avg | |
| ChatGPT-RTS2 | 5 | 6 | 7 | 6 | 6.00 |
| ChatGPT-MCQ2 | 4 | 7 | 6 | 4 | 5.25 |
| ChatGPT-StarEval | 7 | 7 | 7 | 6 | 6.75 |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| RTS2 | Human-Corr | 0.339 | 0.338 | 0.285 | 0.393 | 0.497 | 0.350 | 0.290 | 0.280 | 0.252 | 0.440 | 0.441 | 0.351 |
| Meta-Corr | -0.559 | -0.476 | -0.424 | -0.748 | -0.843 | -0.576 | -0.825 | -0.823 | -0.636 | -0.385 | -0.506 | -0.212 | |
| MCQ2 | Human-Corr | 0.430 | 0.423 | 0.355 | 0.327 | 0.483 | 0.306 | 0.258 | 0.396 | 0.229 | 0.240 | 0.258 | 0.206 |
| Meta-Corr | -0.308 | -0.275 | -0.212 | -0.545 | -0.446 | -0.364 | -0.811 | -0.615 | -0.636 | -0.217 | -0.707 | -0.182 | |
| StarEval | Human-Corr | 0.418 | 0.417 | 0.341 | 0.297 | 0.421 | 0.264 | 0.246 | 0.323 | 0.217 | 0.393 | 0.405 | 0.323 |
| Meta-Corr | -0.084 | -0.101 | 0.000 | -0.664 | -0.684 | -0.545 | -0.441 | -0.460 | -0.212 | -0.497 | -0.579 | -0.303 |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| 7B | Human-Corr | -0.001 | -0.000 | -0.001 | 0.114 | 0.130 | 0.104 | 0.043 | 0.038 | 0.039 | 0.067 | 0.064 | 0.057 |
| Meta-Corr | 0.245 | -0.258 | 0.182 | -0.524 | -0.720 | -0.424 | -0.042 | 0.463 | -0.03 | 0.238 | 0.169 | 0.212 | |
| 13B | Human-Corr | 0.153 | 0.165 | 0.123 | 0.180 | 0.209 | 0.162 | 0.187 | 0.179 | 0.167 | 0.234 | 0.266 | 0.192 |
| Meta-Corr | -0.287 | -0.425 | -0.182 | -0.580 | -0.495 | -0.424 | -0.455 | -0.233 | -0.303 | -0.049 | -0.406 | 0.000 | |
| 70B | Human-Corr | 0.234 | 0.254 | 0.186 | 0.357 | 0.395 | 0.319 | 0.155 | 0.161 | 0.134 | 0.248 | 0.285 | 0.200 |
| Meta-Corr | -0.420 | -0.407 | -0.273 | -0.811 | -0.690 | -0.667 | -0.322 | -0.319 | -0.182 | -0.238 | -0.123 | -0.182 |
| ID | Coh | Con | Flu | Rel |
| M8 | 0.58 | 0.09 | 0.13 | 0.51 |
| M9 | 0.65 | 0.15 | 0.34 | 0.54 |
| M10 | 0.58 | 0.22 | 0.20 | 0.54 |
| M11 | 0.59 | 0.34 | 0.40 | 0.49 |
| M12 | 0.65 | 0.03 | 0.14 | 0.54 |
| M13 | 0.62 | 0.07 | 0.14 | 0.51 |
| M14 | 0.65 | 0.09 | 0.18 | 0.54 |
| M15 | 0.60 | 0.04 | 0.13 | 0.53 |
| M17 | 0.58 | 0.05 | 0.08 | 0.52 |
| M20 | 0.48 | 0.29 | 0.24 | 0.49 |
| M22 | 0.58 | 0.02 | 0.14 | 0.54 |
| M23 | 0.58 | 0.05 | 0.12 | 0.54 |
| ID | Model Name | Coh | Con | Flu | Rel | Avg |
| M22 | BART | 4.18 | 4.94 | 4.90 | 4.25 | 4.57 |
| M23 | Pegasus (C4) | 4.16 | 4.91 | 4.88 | 4.26 | 4.55 |
| M17 | T5 | 4.00 | 4.93 | 4.93 | 4.23 | 4.52 |
| M12 | Unified-ext-abs | 3.60 | 4.96 | 4.85 | 3.85 | 4.32 |
| M13 | ROUGESal | 3.44 | 4.82 | 4.86 | 3.83 | 4.24 |
| M15 | Closed book decoder | 3.35 | 4.95 | 4.80 | 3.67 | 4.19 |
| M14 | Multi-task (Ent + QG) | 3.20 | 4.90 | 4.74 | 3.63 | 4.12 |
| M8 | Pointer Generator | 3.29 | 4.65 | 4.79 | 3.55 | 4.07 |
| M9 | Fast-abs-rl | 2.38 | 4.67 | 4.50 | 3.52 | 3.77 |
| M10 | Bottom-Up | 2.73 | 4.25 | 4.42 | 3.38 | 3.70 |
| M20 | GPT-2 (zero-shot) | 3.63 | 3.40 | 3.97 | 3.30 | 3.58 |
| M11 | Improve-abs | 2.28 | 3.27 | 3.65 | 3.15 | 3.09 |
| Model A | Model B | Coherence | Consistency | Fluency | Relevance | ||||||||
| LLM | Human | LLM | Human | LLM | Human | LLM | Human | ||||||
| M22 | M23 | 65.5 | 53.5 | ✓ | 55.75 | 52.5 | ✓ | 61 | 49.5 | × | 58.75 | 49.5 | × |
| M23 | M17 | 48.25 | 52.5 | × | 47 | 49 | ✓ | 49 | 45.5 | ✓ | 45 | 52 | × |
| M17 | M12 | 44 | 66.5 | × | 43.25 | 48.5 | ✓ | 40.5 | 54.5 | × | 49.25 | 72.5 | × |
| M12 | M13 | 58 | 51 | ✓ | 56.5 | 54.5 | ✓ | 56.75 | 50 | × | 58 | 45 | × |
| M13 | M15 | 45.5 | 49.5 | ✓ | 52 | 46 | × | 48.75 | 52 | × | 51.25 | 60.5 | ✓ |
| M15 | M14 | 57 | 54 | ✓ | 55 | 53.5 | ✓ | 56.5 | 52 | ✓ | 54 | 57 | ✓ |
| M14 | M8 | 49.25 | 44 | ✓ | 50 | 54.5 | × | 47 | 46.5 | ✓ | 49.25 | 53.5 | × |
| M8 | M9 | 77.5 | 82 | ✓ | 78.5 | 53 | ✓ | 80.5 | 63.5 | ✓ | 76 | 54 | ✓ |
| M9 | M10 | 45 | 36 | ✓ | 41.5 | 58 | × | 46 | 44.5 | ✓ | 41.5 | 56 | × |
| M10 | M20 | 58.25 | 24 | × | 61.75 | 64 | ✓ | 63.75 | 61.5 | ✓ | 61.5 | 54.5 | ✓ |
| M20 | M11 | 56.5 | 82 | ✓ | 50 | 53 | × | 51 | 58.5 | ✓ | 50 | 53 | × |
| #CP | 8 | 7 | 7 | 4 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.420 | 0.383 | 0.323 | 0.229 | 0.273 | 0.209 | 0.274 | 0.245 | 0.236 | 0.519 | 0.509 | 0.438 |
| M9 | 0.174 | 0.243 | 0.142 | 0.119 | 0.265 | 0.103 | 0.256 | 0.258 | 0.219 | 0.254 | 0.307 | 0.200 |
| M10 | 0.365 | 0.415 | 0.292 | 0.305 | 0.452 | 0.251 | 0.258 | 0.288 | 0.223 | 0.367 | 0.378 | 0.284 |
| M11 | 0.378 | 0.420 | 0.320 | 0.404 | 0.488 | 0.335 | 0.227 | 0.288 | 0.182 | 0.501 | 0.511 | 0.394 |
| M12 | 0.208 | 0.237 | 0.160 | 0.087 | 0.020 | 0.082 | 0.086 | 0.107 | 0.071 | 0.438 | 0.451 | 0.354 |
| M13 | 0.455 | 0.473 | 0.359 | 0.178 | 0.037 | 0.167 | 0.063 | 0.151 | 0.055 | 0.403 | 0.403 | 0.329 |
| M14 | 0.433 | 0.467 | 0.355 | 0.114 | 0.187 | 0.105 | -0.007 | 0.055 | -0.005 | 0.421 | 0.435 | 0.336 |
| M15 | 0.372 | 0.385 | 0.279 | 0.189 | 0.316 | 0.177 | 0.100 | 0.087 | 0.085 | 0.252 | 0.288 | 0.193 |
| M17 | 0.291 | 0.320 | 0.233 | -0.086 | -0.061 | -0.084 | 0.017 | 0.046 | 0.015 | 0.204 | 0.199 | 0.175 |
| M20 | 0.382 | 0.394 | 0.310 | 0.528 | 0.501 | 0.430 | 0.367 | 0.349 | 0.307 | 0.292 | 0.278 | 0.237 |
| M22 | 0.215 | 0.293 | 0.162 | -0.072 | -0.052 | -0.070 | -0.114 | -0.131 | -0.101 | 0.201 | 0.300 | 0.172 |
| M23 | 0.392 | 0.427 | 0.320 | 0.003 | 0.305 | 0.002 | -0.078 | -0.022 | -0.069 | 0.223 | 0.324 | 0.184 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.289 | 0.310 | 0.236 | 0.235 | 0.362 | 0.226 | 0.348 | 0.348 | 0.321 | 0.349 | 0.419 | 0.302 |
| M9 | 0.170 | 0.170 | 0.148 | 0.211 | 0.321 | 0.200 | 0.198 | 0.299 | 0.174 | 0.318 | 0.356 | 0.276 |
| M10 | 0.352 | 0.314 | 0.293 | 0.138 | 0.273 | 0.125 | 0.155 | 0.210 | 0.138 | 0.420 | 0.427 | 0.362 |
| M11 | 0.285 | 0.312 | 0.250 | 0.380 | 0.370 | 0.317 | 0.200 | 0.218 | 0.163 | 0.397 | 0.428 | 0.334 |
| M12 | 0.306 | 0.304 | 0.258 | -0.025 | -0.059 | -0.024 | 0.256 | 0.306 | 0.239 | 0.283 | 0.287 | 0.246 |
| M13 | 0.425 | 0.425 | 0.351 | 0.471 | 0.312 | 0.457 | 0.199 | 0.214 | 0.186 | 0.435 | 0.402 | 0.375 |
| M14 | 0.490 | 0.472 | 0.422 | 0.112 | 0.157 | 0.110 | 0.084 | 0.143 | 0.078 | 0.326 | 0.329 | 0.284 |
| M15 | 0.317 | 0.298 | 0.250 | 0.061 | 0.513 | 0.060 | 0.055 | 0.025 | 0.050 | 0.433 | 0.420 | 0.378 |
| M17 | 0.250 | 0.255 | 0.215 | -0.106 | -0.081 | -0.105 | -0.011 | 0.024 | -0.011 | 0.293 | 0.285 | 0.260 |
| M20 | 0.463 | 0.450 | 0.381 | 0.455 | 0.442 | 0.371 | 0.494 | 0.450 | 0.404 | 0.326 | 0.334 | 0.264 |
| M22 | 0.211 | 0.173 | 0.182 | -0.096 | -0.080 | -0.095 | -0.092 | -0.056 | -0.087 | 0.352 | 0.371 | 0.313 |
| M23 | 0.218 | 0.209 | 0.189 | 0.107 | 0.448 | 0.104 | -0.275 | -0.269 | -0.261 | 0.147 | 0.187 | 0.129 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||||||
| Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | |
| M8 | 2.43 | 3.70 | 4.56 | 3.29 | 4.21 | 4.70 | 4.77 | 4.65 | 2.51 | 3.76 | 4.61 | 4.79 | 3.56 | 4.33 | 4.71 | 3.55 |
| M9 | 1.80 | 3.51 | 4.41 | 2.38 | 3.93 | 4.69 | 4.94 | 4.67 | 2.16 | 3.45 | 4.15 | 4.50 | 3.51 | 4.33 | 4.82 | 3.52 |
| M10 | 2.05 | 3.49 | 4.05 | 2.73 | 3.77 | 4.52 | 4.60 | 4.25 | 2.23 | 3.43 | 3.92 | 4.42 | 3.45 | 4.25 | 4.62 | 3.38 |
| M11 | 1.70 | 2.63 | 2.93 | 2.28 | 2.36 | 4.11 | 3.72 | 3.27 | 1.71 | 2.66 | 2.77 | 3.65 | 2.92 | 4.06 | 3.87 | 3.15 |
| M12 | 2.75 | 3.92 | 4.75 | 3.60 | 4.46 | 4.75 | 5.00 | 4.96 | 2.59 | 3.84 | 4.72 | 4.85 | 3.89 | 4.40 | 4.95 | 3.85 |
| M13 | 2.91 | 3.99 | 4.76 | 3.44 | 4.40 | 4.73 | 4.89 | 4.82 | 2.69 | 3.93 | 4.69 | 4.86 | 3.90 | 4.41 | 4.85 | 3.83 |
| M14 | 2.38 | 3.87 | 4.54 | 3.20 | 4.25 | 4.70 | 4.99 | 4.90 | 2.42 | 3.83 | 4.53 | 4.74 | 3.67 | 4.32 | 4.84 | 3.63 |
| M15 | 2.65 | 3.81 | 4.58 | 3.35 | 4.33 | 4.75 | 4.90 | 4.95 | 2.60 | 3.84 | 4.61 | 4.80 | 3.78 | 4.32 | 4.82 | 3.67 |
| M17 | 3.05 | 4.11 | 4.89 | 4.00 | 4.84 | 4.87 | 4.97 | 4.93 | 3.29 | 4.08 | 4.79 | 4.93 | 4.30 | 4.43 | 4.97 | 4.23 |
| M20 | 2.00 | 2.95 | 2.99 | 3.63 | 2.66 | 3.57 | 3.51 | 3.40 | 2.22 | 3.02 | 3.05 | 3.97 | 2.73 | 3.78 | 3.47 | 3.30 |
| M22 | 3.54 | 4.09 | 4.92 | 4.18 | 4.83 | 4.82 | 4.97 | 4.94 | 3.57 | 4.06 | 4.94 | 4.90 | 4.41 | 4.43 | 4.95 | 4.25 |
| M23 | 3.00 | 4.14 | 4.88 | 4.16 | 4.74 | 4.84 | 4.94 | 4.91 | 3.08 | 4.05 | 4.76 | 4.88 | 4.10 | 4.45 | 4.96 | 4.26 |
| avg | 2.52 | 3.68 | 4.35 | 3.35 | 4.06 | 4.59 | 4.68 | 4.55 | 2.59 | 3.66 | 4.29 | 4.61 | 3.68 | 4.29 | 4.65 | 3.72 |
| Consistency | The summary is inconsistent with the article as it omits important details and repeats a phrase multiple times, resulting in a score of 1. |
| Consistency | The summary includes some relevant information but lacks coherence and omits important details, resulting in a score of 2.5 out of 5 for consistency. |
| Fluency | The summary is well-written and grammatically correct, but lacks important details from the article such as the potential link between antibiotics and weight gain, and the warnings about over-prescription of antibiotics and the development of superbugs. Score: 3/5. |
| Fluency | The summary lacks coherence and does not provide a clear understanding of the article, resulting in a score of 1 for disfluency. |
| Coherence | The summary is somewhat coherent but includes incorrect information, such as stating that Merson restarted his row with Burnley instead of Andros Townsend, and does not flow smoothly; therefore, I would give it a score of 3. |
| Consistency | The summary is consistent with the article as it mentions the helicopter airlifting patients to the hospital, but it leaves out important details such as the number of people involved and their injuries. Final score: 3. |
| Consistency | The summary accurately reflects the main points of the article, although it could have mentioned Falcao's unhappiness in France. Final score: 4. |
| Fluency | The summary is coherent and grammatically correct, but it does not provide enough information about the incident. Final score: 3. |
| Fluency | The summary is only one sentence and do not provide enough information about the article's content. Final score: 1. |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.429 | 0.446 | 0.362 | 0.449 | 0.597 | 0.433 | 0.412 | 0.409 | 0.385 | 0.425 | 0.489 | 0.365 |
| M9 | 0.288 | 0.346 | 0.243 | 0.349 | 0.350 | 0.331 | 0.462 | 0.465 | 0.407 | 0.413 | 0.490 | 0.358 |
| M10 | 0.495 | 0.480 | 0.403 | 0.518 | 0.683 | 0.469 | 0.406 | 0.579 | 0.360 | 0.510 | 0.537 | 0.429 |
| M11 | 0.584 | 0.572 | 0.474 | 0.529 | 0.536 | 0.439 | 0.448 | 0.436 | 0.358 | 0.500 | 0.519 | 0.405 |
| M12 | 0.245 | 0.360 | 0.209 | - | - | - | 0.387 | 0.506 | 0.372 | 0.169 | 0.106 | 0.149 |
| M13 | 0.271 | 0.271 | 0.230 | 0.535 | 0.444 | 0.518 | 0.093 | 0.337 | 0.089 | 0.237 | 0.337 | 0.206 |
| M14 | 0.263 | 0.312 | 0.222 | -0.040 | -0.026 | -0.040 | 0.311 | 0.432 | 0.291 | 0.277 | 0.315 | 0.241 |
| M15 | 0.403 | 0.418 | 0.342 | 0.158 | 0.504 | 0.155 | 0.240 | 0.242 | 0.225 | 0.307 | 0.402 | 0.267 |
| M17 | 0.285 | 0.240 | 0.246 | 0.390 | 0.748 | 0.385 | 0.091 | 0.131 | 0.088 | 0.210 | 0.253 | 0.185 |
| M20 | 0.427 | 0.418 | 0.334 | 0.378 | 0.362 | 0.313 | 0.482 | 0.468 | 0.402 | 0.479 | 0.491 | 0.391 |
| M22 | 0.026 | 0.004 | 0.023 | 0.346 | 0.608 | 0.343 | 0.248 | 0.141 | 0.242 | 0.143 | 0.084 | 0.126 |
| M23 | 0.320 | 0.330 | 0.283 | 0.360 | 0.168 | 0.354 | 0.267 | 0.136 | 0.251 | 0.005 | 0.003 | 0.005 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.464 | 0.455 | 0.407 | 0.557 | 0.613 | 0.526 | 0.391 | 0.367 | 0.341 | 0.517 | 0.514 | 0.471 |
| M9 | 0.436 | 0.415 | 0.416 | 0.215 | 0.342 | 0.195 | 0.578 | 0.571 | 0.517 | 0.450 | 0.485 | 0.412 |
| M10 | 0.374 | 0.345 | 0.340 | 0.300 | 0.297 | 0.270 | 0.631 | 0.585 | 0.551 | 0.506 | 0.507 | 0.459 |
| M11 | 0.326 | 0.428 | 0.307 | 0.437 | 0.421 | 0.393 | 0.354 | 0.509 | 0.303 | 0.467 | 0.486 | 0.419 |
| M12 | 0.500 | 0.469 | 0.444 | 0.446 | 0.486 | 0.424 | 0.273 | 0.320 | 0.241 | 0.588 | 0.591 | 0.542 |
| M13 | 0.462 | 0.450 | 0.408 | 0.470 | 0.510 | 0.445 | 0.325 | 0.345 | 0.282 | 0.515 | 0.498 | 0.475 |
| M14 | 0.524 | 0.460 | 0.463 | 0.325 | 0.293 | 0.307 | 0.216 | 0.257 | 0.194 | 0.501 | 0.476 | 0.455 |
| M15 | 0.534 | 0.499 | 0.470 | 0.266 | 0.326 | 0.251 | 0.383 | 0.369 | 0.337 | 0.584 | 0.550 | 0.535 |
| M17 | 0.312 | 0.294 | 0.274 | -0.016 | -0.008 | -0.015 | 0.251 | 0.261 | 0.225 | 0.530 | 0.479 | 0.504 |
| M20 | 0.636 | 0.623 | 0.566 | 0.733 | 0.664 | 0.644 | 0.566 | 0.556 | 0.492 | 0.570 | 0.565 | 0.496 |
| M22 | 0.314 | 0.338 | 0.276 | 0.245 | 0.151 | 0.241 | 0.371 | 0.341 | 0.330 | 0.471 | 0.412 | 0.451 |
| M23 | 0.380 | 0.386 | 0.337 | 0.158 | 0.496 | 0.153 | 0.296 | 0.275 | 0.259 | 0.450 | 0.505 | 0.421 |
| Metrics | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1* | 0.193 | 0.202 | 0.136 | 0.155 | 0.186 | 0.121 | 0.075 | 0.153 | 0.058 | 0.323 | 0.361 | 0.231 |
| ROUGE-2* | 0.145 | 0.146 | 0.101 | 0.137 | 0.156 | 0.107 | 0.053 | 0.095 | 0.041 | 0.255 | 0.262 | 0.181 |
| ROUGE-L* | 0.148 | 0.158 | 0.105 | 0.133 | 0.167 | 0.103 | 0.078 | 0.146 | 0.060 | 0.306 | 0.340 | 0.219 |
| BERTScore* | 0.375 | 0.383 | 0.265 | 0.163 | 0.182 | 0.127 | 0.167 | 0.229 | 0.130 | 0.396 | 0.414 | 0.285 |
| BARTScore* | 0.381 | 0.391 | 0.275 | 0.271 | 0.265 | 0.212 | 0.168 | 0.187 | 0.131 | 0.381 | 0.391 | 0.276 |
| BARTScore-CNN* | 0.461 | 0.480 | 0.332 | 0.389 | 0.413 | 0.305 | 0.310 | 0.378 | 0.241 | 0.425 | 0.450 | 0.309 |
| BARTScore-CNN-PARA* | 0.455 | 0.455 | 0.328 | 0.413 | 0.459 | 0.324 | 0.368 | 0.417 | 0.286 | 0.414 | 0.440 | 0.299 |
| ChatGPT-RTS | 0.388 | 0.399 | 0.312 | 0.423 | 0.532 | 0.378 | 0.285 | 0.302 | 0.240 | 0.448 | 0.463 | 0.357 |
| ChatGPT-MCQ | 0.424 | 0.416 | 0.350 | 0.343 | 0.487 | 0.320 | 0.343 | 0.431 | 0.305 | 0.384 | 0.395 | 0.329 |
| GPT-4-RTS | 0.427 | 0.461 | 0.361 | 0.556 | 0.618 | 0.522 | 0.498 | 0.600 | 0.452 | 0.448 | 0.428 | 0.373 |
| Metrics | Coh | Con | Flu | Rel | Avg |
| Random | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 |
| ROUGE-1 | 3/40 | 5/46 | 3/46 | 4/47 | 3.75/44.75 |
| ROUGE-2 | 2/38 | 7/48 | 3/45 | 4/46 | 4.00/44.25 |
| ROUGE-L | 2/31 | 5/37 | 4/39 | 6/41 | 4.25/37.00 |
| BERTScore | 4/48 | 5/44 | 4/44 | 6/46 | 4.75/45.50 |
| BARTScore | 8/46 | 7/48 | 5/45 | 6/46 | 6.50/46.25 |
| BARTScore-CNN | 9/53 | 5/53 | 5/54 | 4/53 | 5.75/53.25 |
| BARTScore-CNN-PARA | 9/49 | 8/54 | 6/52 | 5/51 | 7.00/51.50 |
| ChatGPT-RTS | 6/54 | 6/56 | 9/62 | 7/62 | 7.00/58.50 |
| ChatGPT-MCQ | 5/54 | 7/56 | 8/60 | 7/58 | 6.75/57.00 |
| ChatGPT-H2H | 8/- | 7/- | 7/- | 4/- | 6.50/- |
| GPT-4-RTS | 5/55 | 7/53 | 8/60 | 7/56 | 6.75/56.00 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1 | 0.0 | -0.032 | -0.091 | -0.490 | -0.527 | -0.303 | -0.420 | -0.518 | -0.273 | -0.420 | -0.387 | -0.273 |
| ROUGE-2 | 0.559 | 0.508 | 0.485 | -0.259 | -0.480 | -0.152 | -0.217 | -0.438 | -0.152 | -0.084 | -0.120 | -0.121 |
| ROUGE-L | 0.231 | 0.251 | 0.121 | -0.510 | -0.522 | -0.303 | -0.168 | -0.412 | -0.152 | -0.224 | -0.266 | -0.121 |
| BERTScore | -0.413 | -0.403 | -0.212 | -0.580 | -0.869 | -0.424 | -0.455 | -0.663 | -0.303 | -0.685 | -0.756 | -0.515 |
| BARTScore | -0.916 | -0.747 | -0.788 | -0.266 | -0.504 | -0.121 | 0.154 | 0.123 | 0.182 | -0.769 | -0.837 | -0.606 |
| BARTScore-CNN | -0.748 | -0.800 | -0.636 | -0.671 | -0.913 | -0.515 | -0.510 | -0.604 | -0.485 | -0.825 | -0.852 | -0.667 |
| BARTScore-CNN-PARA | -0.720 | -0.858 | -0.606 | -0.685 | -0.888 | -0.576 | -0.294 | -0.522 | -0.212 | -0.853 | -0.880 | -0.727 |
| ChatGPT-RTS | -0.042 | -0.072 | -0.121 | -0.811 | -0.751 | -0.636 | -0.748 | -0.728 | -0.606 | -0.559 | -0.473 | -0.394 |
| ChatGPT-MCQ | -0.175 | -0.11 | -0.182 | -0.818 | -0.411 | -0.636 | -0.622 | -0.484 | -0.394 | -0.350 | -0.622 | -0.212 |
| GPT-4-RTS | -0.531 | -0.678 | -0.424 | -0.600 | -0.103 | -0.236 | -0.839 | -0.520 | -0.515 | -0.958 | -0.880 | -0.848 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ρ(Ri, PRTS) | 0.343 | 0.198 | 0.091 | 0.685 | 0.506 | 0.576 | 0.727 | 0.797 | 0.545 | 0.168 | 0.128 | 0.000 |
| ρ(Ri, PMCQ) | 0.657 | 0.625 | 0.394 | 0.685 | 0.616 | 0.515 | 0.322 | 0.573 | 0.212 | 0.091 | -0.106 | 0.000 |
| Coh | Con | Flu | Rel | Avg | |
| ChatGPT-RTS2 | 5 | 6 | 7 | 6 | 6.00 |
| ChatGPT-MCQ2 | 4 | 7 | 6 | 4 | 5.25 |
| ChatGPT-StarEval | 7 | 7 | 7 | 6 | 6.75 |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| RTS2 | Human-Corr | 0.339 | 0.338 | 0.285 | 0.393 | 0.497 | 0.350 | 0.290 | 0.280 | 0.252 | 0.440 | 0.441 | 0.351 |
| Meta-Corr | -0.559 | -0.476 | -0.424 | -0.748 | -0.843 | -0.576 | -0.825 | -0.823 | -0.636 | -0.385 | -0.506 | -0.212 | |
| MCQ2 | Human-Corr | 0.430 | 0.423 | 0.355 | 0.327 | 0.483 | 0.306 | 0.258 | 0.396 | 0.229 | 0.240 | 0.258 | 0.206 |
| Meta-Corr | -0.308 | -0.275 | -0.212 | -0.545 | -0.446 | -0.364 | -0.811 | -0.615 | -0.636 | -0.217 | -0.707 | -0.182 | |
| StarEval | Human-Corr | 0.418 | 0.417 | 0.341 | 0.297 | 0.421 | 0.264 | 0.246 | 0.323 | 0.217 | 0.393 | 0.405 | 0.323 |
| Meta-Corr | -0.084 | -0.101 | 0.000 | -0.664 | -0.684 | -0.545 | -0.441 | -0.460 | -0.212 | -0.497 | -0.579 | -0.303 | |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| 7B | Human-Corr | -0.001 | -0.000 | -0.001 | 0.114 | 0.130 | 0.104 | 0.043 | 0.038 | 0.039 | 0.067 | 0.064 | 0.057 |
| Meta-Corr | 0.245 | -0.258 | 0.182 | -0.524 | -0.720 | -0.424 | -0.042 | 0.463 | -0.03 | 0.238 | 0.169 | 0.212 | |
| 13B | Human-Corr | 0.153 | 0.165 | 0.123 | 0.180 | 0.209 | 0.162 | 0.187 | 0.179 | 0.167 | 0.234 | 0.266 | 0.192 |
| Meta-Corr | -0.287 | -0.425 | -0.182 | -0.580 | -0.495 | -0.424 | -0.455 | -0.233 | -0.303 | -0.049 | -0.406 | 0.000 | |
| 70B | Human-Corr | 0.234 | 0.254 | 0.186 | 0.357 | 0.395 | 0.319 | 0.155 | 0.161 | 0.134 | 0.248 | 0.285 | 0.200 |
| Meta-Corr | -0.420 | -0.407 | -0.273 | -0.811 | -0.690 | -0.667 | -0.322 | -0.319 | -0.182 | -0.238 | -0.123 | -0.182 | |
| ID | Coh | Con | Flu | Rel |
| M8 | 0.58 | 0.09 | 0.13 | 0.51 |
| M9 | 0.65 | 0.15 | 0.34 | 0.54 |
| M10 | 0.58 | 0.22 | 0.20 | 0.54 |
| M11 | 0.59 | 0.34 | 0.40 | 0.49 |
| M12 | 0.65 | 0.03 | 0.14 | 0.54 |
| M13 | 0.62 | 0.07 | 0.14 | 0.51 |
| M14 | 0.65 | 0.09 | 0.18 | 0.54 |
| M15 | 0.60 | 0.04 | 0.13 | 0.53 |
| M17 | 0.58 | 0.05 | 0.08 | 0.52 |
| M20 | 0.48 | 0.29 | 0.24 | 0.49 |
| M22 | 0.58 | 0.02 | 0.14 | 0.54 |
| M23 | 0.58 | 0.05 | 0.12 | 0.54 |
| ID | Model Name | Coh | Con | Flu | Rel | Avg |
| M22 | BART | 4.18 | 4.94 | 4.90 | 4.25 | 4.57 |
| M23 | Pegasus (C4) | 4.16 | 4.91 | 4.88 | 4.26 | 4.55 |
| M17 | T5 | 4.00 | 4.93 | 4.93 | 4.23 | 4.52 |
| M12 | Unified-ext-abs | 3.60 | 4.96 | 4.85 | 3.85 | 4.32 |
| M13 | ROUGESal | 3.44 | 4.82 | 4.86 | 3.83 | 4.24 |
| M15 | Closed book decoder | 3.35 | 4.95 | 4.80 | 3.67 | 4.19 |
| M14 | Multi-task (Ent + QG) | 3.20 | 4.90 | 4.74 | 3.63 | 4.12 |
| M8 | Pointer Generator | 3.29 | 4.65 | 4.79 | 3.55 | 4.07 |
| M9 | Fast-abs-rl | 2.38 | 4.67 | 4.50 | 3.52 | 3.77 |
| M10 | Bottom-Up | 2.73 | 4.25 | 4.42 | 3.38 | 3.70 |
| M20 | GPT-2 (zero-shot) | 3.63 | 3.40 | 3.97 | 3.30 | 3.58 |
| M11 | Improve-abs | 2.28 | 3.27 | 3.65 | 3.15 | 3.09 |
| Model A | Model B | Coherence | Consistency | Fluency | Relevance | ||||||||
| LLM | Human | LLM | Human | LLM | Human | LLM | Human | ||||||
| M22 | M23 | 65.5 | 53.5 | ✓ | 55.75 | 52.5 | ✓ | 61 | 49.5 | × | 58.75 | 49.5 | × |
| M23 | M17 | 48.25 | 52.5 | × | 47 | 49 | ✓ | 49 | 45.5 | ✓ | 45 | 52 | × |
| M17 | M12 | 44 | 66.5 | × | 43.25 | 48.5 | ✓ | 40.5 | 54.5 | × | 49.25 | 72.5 | × |
| M12 | M13 | 58 | 51 | ✓ | 56.5 | 54.5 | ✓ | 56.75 | 50 | × | 58 | 45 | × |
| M13 | M15 | 45.5 | 49.5 | ✓ | 52 | 46 | × | 48.75 | 52 | × | 51.25 | 60.5 | ✓ |
| M15 | M14 | 57 | 54 | ✓ | 55 | 53.5 | ✓ | 56.5 | 52 | ✓ | 54 | 57 | ✓ |
| M14 | M8 | 49.25 | 44 | ✓ | 50 | 54.5 | × | 47 | 46.5 | ✓ | 49.25 | 53.5 | × |
| M8 | M9 | 77.5 | 82 | ✓ | 78.5 | 53 | ✓ | 80.5 | 63.5 | ✓ | 76 | 54 | ✓ |
| M9 | M10 | 45 | 36 | ✓ | 41.5 | 58 | × | 46 | 44.5 | ✓ | 41.5 | 56 | × |
| M10 | M20 | 58.25 | 24 | × | 61.75 | 64 | ✓ | 63.75 | 61.5 | ✓ | 61.5 | 54.5 | ✓ |
| M20 | M11 | 56.5 | 82 | ✓ | 50 | 53 | × | 51 | 58.5 | ✓ | 50 | 53 | × |
| #CP | 8 | 7 | 7 | 4 | |||||||||
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.420 | 0.383 | 0.323 | 0.229 | 0.273 | 0.209 | 0.274 | 0.245 | 0.236 | 0.519 | 0.509 | 0.438 |
| M9 | 0.174 | 0.243 | 0.142 | 0.119 | 0.265 | 0.103 | 0.256 | 0.258 | 0.219 | 0.254 | 0.307 | 0.200 |
| M10 | 0.365 | 0.415 | 0.292 | 0.305 | 0.452 | 0.251 | 0.258 | 0.288 | 0.223 | 0.367 | 0.378 | 0.284 |
| M11 | 0.378 | 0.420 | 0.320 | 0.404 | 0.488 | 0.335 | 0.227 | 0.288 | 0.182 | 0.501 | 0.511 | 0.394 |
| M12 | 0.208 | 0.237 | 0.160 | 0.087 | 0.020 | 0.082 | 0.086 | 0.107 | 0.071 | 0.438 | 0.451 | 0.354 |
| M13 | 0.455 | 0.473 | 0.359 | 0.178 | 0.037 | 0.167 | 0.063 | 0.151 | 0.055 | 0.403 | 0.403 | 0.329 |
| M14 | 0.433 | 0.467 | 0.355 | 0.114 | 0.187 | 0.105 | -0.007 | 0.055 | -0.005 | 0.421 | 0.435 | 0.336 |
| M15 | 0.372 | 0.385 | 0.279 | 0.189 | 0.316 | 0.177 | 0.100 | 0.087 | 0.085 | 0.252 | 0.288 | 0.193 |
| M17 | 0.291 | 0.320 | 0.233 | -0.086 | -0.061 | -0.084 | 0.017 | 0.046 | 0.015 | 0.204 | 0.199 | 0.175 |
| M20 | 0.382 | 0.394 | 0.310 | 0.528 | 0.501 | 0.430 | 0.367 | 0.349 | 0.307 | 0.292 | 0.278 | 0.237 |
| M22 | 0.215 | 0.293 | 0.162 | -0.072 | -0.052 | -0.070 | -0.114 | -0.131 | -0.101 | 0.201 | 0.300 | 0.172 |
| M23 | 0.392 | 0.427 | 0.320 | 0.003 | 0.305 | 0.002 | -0.078 | -0.022 | -0.069 | 0.223 | 0.324 | 0.184 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.289 | 0.310 | 0.236 | 0.235 | 0.362 | 0.226 | 0.348 | 0.348 | 0.321 | 0.349 | 0.419 | 0.302 |
| M9 | 0.170 | 0.170 | 0.148 | 0.211 | 0.321 | 0.200 | 0.198 | 0.299 | 0.174 | 0.318 | 0.356 | 0.276 |
| M10 | 0.352 | 0.314 | 0.293 | 0.138 | 0.273 | 0.125 | 0.155 | 0.210 | 0.138 | 0.420 | 0.427 | 0.362 |
| M11 | 0.285 | 0.312 | 0.250 | 0.380 | 0.370 | 0.317 | 0.200 | 0.218 | 0.163 | 0.397 | 0.428 | 0.334 |
| M12 | 0.306 | 0.304 | 0.258 | -0.025 | -0.059 | -0.024 | 0.256 | 0.306 | 0.239 | 0.283 | 0.287 | 0.246 |
| M13 | 0.425 | 0.425 | 0.351 | 0.471 | 0.312 | 0.457 | 0.199 | 0.214 | 0.186 | 0.435 | 0.402 | 0.375 |
| M14 | 0.490 | 0.472 | 0.422 | 0.112 | 0.157 | 0.110 | 0.084 | 0.143 | 0.078 | 0.326 | 0.329 | 0.284 |
| M15 | 0.317 | 0.298 | 0.250 | 0.061 | 0.513 | 0.060 | 0.055 | 0.025 | 0.050 | 0.433 | 0.420 | 0.378 |
| M17 | 0.250 | 0.255 | 0.215 | -0.106 | -0.081 | -0.105 | -0.011 | 0.024 | -0.011 | 0.293 | 0.285 | 0.260 |
| M20 | 0.463 | 0.450 | 0.381 | 0.455 | 0.442 | 0.371 | 0.494 | 0.450 | 0.404 | 0.326 | 0.334 | 0.264 |
| M22 | 0.211 | 0.173 | 0.182 | -0.096 | -0.080 | -0.095 | -0.092 | -0.056 | -0.087 | 0.352 | 0.371 | 0.313 |
| M23 | 0.218 | 0.209 | 0.189 | 0.107 | 0.448 | 0.104 | -0.275 | -0.269 | -0.261 | 0.147 | 0.187 | 0.129 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||||||
| Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | |
| M8 | 2.43 | 3.70 | 4.56 | 3.29 | 4.21 | 4.70 | 4.77 | 4.65 | 2.51 | 3.76 | 4.61 | 4.79 | 3.56 | 4.33 | 4.71 | 3.55 |
| M9 | 1.80 | 3.51 | 4.41 | 2.38 | 3.93 | 4.69 | 4.94 | 4.67 | 2.16 | 3.45 | 4.15 | 4.50 | 3.51 | 4.33 | 4.82 | 3.52 |
| M10 | 2.05 | 3.49 | 4.05 | 2.73 | 3.77 | 4.52 | 4.60 | 4.25 | 2.23 | 3.43 | 3.92 | 4.42 | 3.45 | 4.25 | 4.62 | 3.38 |
| M11 | 1.70 | 2.63 | 2.93 | 2.28 | 2.36 | 4.11 | 3.72 | 3.27 | 1.71 | 2.66 | 2.77 | 3.65 | 2.92 | 4.06 | 3.87 | 3.15 |
| M12 | 2.75 | 3.92 | 4.75 | 3.60 | 4.46 | 4.75 | 5.00 | 4.96 | 2.59 | 3.84 | 4.72 | 4.85 | 3.89 | 4.40 | 4.95 | 3.85 |
| M13 | 2.91 | 3.99 | 4.76 | 3.44 | 4.40 | 4.73 | 4.89 | 4.82 | 2.69 | 3.93 | 4.69 | 4.86 | 3.90 | 4.41 | 4.85 | 3.83 |
| M14 | 2.38 | 3.87 | 4.54 | 3.20 | 4.25 | 4.70 | 4.99 | 4.90 | 2.42 | 3.83 | 4.53 | 4.74 | 3.67 | 4.32 | 4.84 | 3.63 |
| M15 | 2.65 | 3.81 | 4.58 | 3.35 | 4.33 | 4.75 | 4.90 | 4.95 | 2.60 | 3.84 | 4.61 | 4.80 | 3.78 | 4.32 | 4.82 | 3.67 |
| M17 | 3.05 | 4.11 | 4.89 | 4.00 | 4.84 | 4.87 | 4.97 | 4.93 | 3.29 | 4.08 | 4.79 | 4.93 | 4.30 | 4.43 | 4.97 | 4.23 |
| M20 | 2.00 | 2.95 | 2.99 | 3.63 | 2.66 | 3.57 | 3.51 | 3.40 | 2.22 | 3.02 | 3.05 | 3.97 | 2.73 | 3.78 | 3.47 | 3.30 |
| M22 | 3.54 | 4.09 | 4.92 | 4.18 | 4.83 | 4.82 | 4.97 | 4.94 | 3.57 | 4.06 | 4.94 | 4.90 | 4.41 | 4.43 | 4.95 | 4.25 |
| M23 | 3.00 | 4.14 | 4.88 | 4.16 | 4.74 | 4.84 | 4.94 | 4.91 | 3.08 | 4.05 | 4.76 | 4.88 | 4.10 | 4.45 | 4.96 | 4.26 |
| avg | 2.52 | 3.68 | 4.35 | 3.35 | 4.06 | 4.59 | 4.68 | 4.55 | 2.59 | 3.66 | 4.29 | 4.61 | 3.68 | 4.29 | 4.65 | 3.72 |
| Consistency | The summary is inconsistent with the article as it omits important details and repeats a phrase multiple times, resulting in a score of 1. |
| Consistency | The summary includes some relevant information but lacks coherence and omits important details, resulting in a score of 2.5 out of 5 for consistency. |
| Fluency | The summary is well-written and grammatically correct, but lacks important details from the article such as the potential link between antibiotics and weight gain, and the warnings about over-prescription of antibiotics and the development of superbugs. Score: 3/5. |
| Fluency | The summary lacks coherence and does not provide a clear understanding of the article, resulting in a score of 1 for disfluency. |
| Coherence | The summary is somewhat coherent but includes incorrect information, such as stating that Merson restarted his row with Burnley instead of Andros Townsend, and does not flow smoothly; therefore, I would give it a score of 3. |
| Consistency | The summary is consistent with the article as it mentions the helicopter airlifting patients to the hospital, but it leaves out important details such as the number of people involved and their injuries. Final score: 3. |
| Consistency | The summary accurately reflects the main points of the article, although it could have mentioned Falcao's unhappiness in France. Final score: 4. |
| Fluency | The summary is coherent and grammatically correct, but it does not provide enough information about the incident. Final score: 3. |
| Fluency | The summary is only one sentence and do not provide enough information about the article's content. Final score: 1. |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.429 | 0.446 | 0.362 | 0.449 | 0.597 | 0.433 | 0.412 | 0.409 | 0.385 | 0.425 | 0.489 | 0.365 |
| M9 | 0.288 | 0.346 | 0.243 | 0.349 | 0.350 | 0.331 | 0.462 | 0.465 | 0.407 | 0.413 | 0.490 | 0.358 |
| M10 | 0.495 | 0.480 | 0.403 | 0.518 | 0.683 | 0.469 | 0.406 | 0.579 | 0.360 | 0.510 | 0.537 | 0.429 |
| M11 | 0.584 | 0.572 | 0.474 | 0.529 | 0.536 | 0.439 | 0.448 | 0.436 | 0.358 | 0.500 | 0.519 | 0.405 |
| M12 | 0.245 | 0.360 | 0.209 | - | - | - | 0.387 | 0.506 | 0.372 | 0.169 | 0.106 | 0.149 |
| M13 | 0.271 | 0.271 | 0.230 | 0.535 | 0.444 | 0.518 | 0.093 | 0.337 | 0.089 | 0.237 | 0.337 | 0.206 |
| M14 | 0.263 | 0.312 | 0.222 | -0.040 | -0.026 | -0.040 | 0.311 | 0.432 | 0.291 | 0.277 | 0.315 | 0.241 |
| M15 | 0.403 | 0.418 | 0.342 | 0.158 | 0.504 | 0.155 | 0.240 | 0.242 | 0.225 | 0.307 | 0.402 | 0.267 |
| M17 | 0.285 | 0.240 | 0.246 | 0.390 | 0.748 | 0.385 | 0.091 | 0.131 | 0.088 | 0.210 | 0.253 | 0.185 |
| M20 | 0.427 | 0.418 | 0.334 | 0.378 | 0.362 | 0.313 | 0.482 | 0.468 | 0.402 | 0.479 | 0.491 | 0.391 |
| M22 | 0.026 | 0.004 | 0.023 | 0.346 | 0.608 | 0.343 | 0.248 | 0.141 | 0.242 | 0.143 | 0.084 | 0.126 |
| M23 | 0.320 | 0.330 | 0.283 | 0.360 | 0.168 | 0.354 | 0.267 | 0.136 | 0.251 | 0.005 | 0.003 | 0.005 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.464 | 0.455 | 0.407 | 0.557 | 0.613 | 0.526 | 0.391 | 0.367 | 0.341 | 0.517 | 0.514 | 0.471 |
| M9 | 0.436 | 0.415 | 0.416 | 0.215 | 0.342 | 0.195 | 0.578 | 0.571 | 0.517 | 0.450 | 0.485 | 0.412 |
| M10 | 0.374 | 0.345 | 0.340 | 0.300 | 0.297 | 0.270 | 0.631 | 0.585 | 0.551 | 0.506 | 0.507 | 0.459 |
| M11 | 0.326 | 0.428 | 0.307 | 0.437 | 0.421 | 0.393 | 0.354 | 0.509 | 0.303 | 0.467 | 0.486 | 0.419 |
| M12 | 0.500 | 0.469 | 0.444 | 0.446 | 0.486 | 0.424 | 0.273 | 0.320 | 0.241 | 0.588 | 0.591 | 0.542 |
| M13 | 0.462 | 0.450 | 0.408 | 0.470 | 0.510 | 0.445 | 0.325 | 0.345 | 0.282 | 0.515 | 0.498 | 0.475 |
| M14 | 0.524 | 0.460 | 0.463 | 0.325 | 0.293 | 0.307 | 0.216 | 0.257 | 0.194 | 0.501 | 0.476 | 0.455 |
| M15 | 0.534 | 0.499 | 0.470 | 0.266 | 0.326 | 0.251 | 0.383 | 0.369 | 0.337 | 0.584 | 0.550 | 0.535 |
| M17 | 0.312 | 0.294 | 0.274 | -0.016 | -0.008 | -0.015 | 0.251 | 0.261 | 0.225 | 0.530 | 0.479 | 0.504 |
| M20 | 0.636 | 0.623 | 0.566 | 0.733 | 0.664 | 0.644 | 0.566 | 0.556 | 0.492 | 0.570 | 0.565 | 0.496 |
| M22 | 0.314 | 0.338 | 0.276 | 0.245 | 0.151 | 0.241 | 0.371 | 0.341 | 0.330 | 0.471 | 0.412 | 0.451 |
| M23 | 0.380 | 0.386 | 0.337 | 0.158 | 0.496 | 0.153 | 0.296 | 0.275 | 0.259 | 0.450 | 0.505 | 0.421 |
| Metrics | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1* | 0.193 | 0.202 | 0.136 | 0.155 | 0.186 | 0.121 | 0.075 | 0.153 | 0.058 | 0.323 | 0.361 | 0.231 |
| ROUGE-2* | 0.145 | 0.146 | 0.101 | 0.137 | 0.156 | 0.107 | 0.053 | 0.095 | 0.041 | 0.255 | 0.262 | 0.181 |
| ROUGE-L* | 0.148 | 0.158 | 0.105 | 0.133 | 0.167 | 0.103 | 0.078 | 0.146 | 0.060 | 0.306 | 0.340 | 0.219 |
| BERTScore* | 0.375 | 0.383 | 0.265 | 0.163 | 0.182 | 0.127 | 0.167 | 0.229 | 0.130 | 0.396 | 0.414 | 0.285 |
| BARTScore* | 0.381 | 0.391 | 0.275 | 0.271 | 0.265 | 0.212 | 0.168 | 0.187 | 0.131 | 0.381 | 0.391 | 0.276 |
| BARTScore-CNN* | 0.461 | 0.480 | 0.332 | 0.389 | 0.413 | 0.305 | 0.310 | 0.378 | 0.241 | 0.425 | 0.450 | 0.309 |
| BARTScore-CNN-PARA* | 0.455 | 0.455 | 0.328 | 0.413 | 0.459 | 0.324 | 0.368 | 0.417 | 0.286 | 0.414 | 0.440 | 0.299 |
| ChatGPT-RTS | 0.388 | 0.399 | 0.312 | 0.423 | 0.532 | 0.378 | 0.285 | 0.302 | 0.240 | 0.448 | 0.463 | 0.357 |
| ChatGPT-MCQ | 0.424 | 0.416 | 0.350 | 0.343 | 0.487 | 0.320 | 0.343 | 0.431 | 0.305 | 0.384 | 0.395 | 0.329 |
| GPT-4-RTS | 0.427 | 0.461 | 0.361 | 0.556 | 0.618 | 0.522 | 0.498 | 0.600 | 0.452 | 0.448 | 0.428 | 0.373 |
| Metrics | Coh | Con | Flu | Rel | Avg |
| Random | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 | 3.67/22 |
| ROUGE-1 | 3/40 | 5/46 | 3/46 | 4/47 | 3.75/44.75 |
| ROUGE-2 | 2/38 | 7/48 | 3/45 | 4/46 | 4.00/44.25 |
| ROUGE-L | 2/31 | 5/37 | 4/39 | 6/41 | 4.25/37.00 |
| BERTScore | 4/48 | 5/44 | 4/44 | 6/46 | 4.75/45.50 |
| BARTScore | 8/46 | 7/48 | 5/45 | 6/46 | 6.50/46.25 |
| BARTScore-CNN | 9/53 | 5/53 | 5/54 | 4/53 | 5.75/53.25 |
| BARTScore-CNN-PARA | 9/49 | 8/54 | 6/52 | 5/51 | 7.00/51.50 |
| ChatGPT-RTS | 6/54 | 6/56 | 9/62 | 7/62 | 7.00/58.50 |
| ChatGPT-MCQ | 5/54 | 7/56 | 8/60 | 7/58 | 6.75/57.00 |
| ChatGPT-H2H | 8/- | 7/- | 7/- | 4/- | 6.50/- |
| GPT-4-RTS | 5/55 | 7/53 | 8/60 | 7/56 | 6.75/56.00 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ROUGE-1 | 0.0 | -0.032 | -0.091 | -0.490 | -0.527 | -0.303 | -0.420 | -0.518 | -0.273 | -0.420 | -0.387 | -0.273 |
| ROUGE-2 | 0.559 | 0.508 | 0.485 | -0.259 | -0.480 | -0.152 | -0.217 | -0.438 | -0.152 | -0.084 | -0.120 | -0.121 |
| ROUGE-L | 0.231 | 0.251 | 0.121 | -0.510 | -0.522 | -0.303 | -0.168 | -0.412 | -0.152 | -0.224 | -0.266 | -0.121 |
| BERTScore | -0.413 | -0.403 | -0.212 | -0.580 | -0.869 | -0.424 | -0.455 | -0.663 | -0.303 | -0.685 | -0.756 | -0.515 |
| BARTScore | -0.916 | -0.747 | -0.788 | -0.266 | -0.504 | -0.121 | 0.154 | 0.123 | 0.182 | -0.769 | -0.837 | -0.606 |
| BARTScore-CNN | -0.748 | -0.800 | -0.636 | -0.671 | -0.913 | -0.515 | -0.510 | -0.604 | -0.485 | -0.825 | -0.852 | -0.667 |
| BARTScore-CNN-PARA | -0.720 | -0.858 | -0.606 | -0.685 | -0.888 | -0.576 | -0.294 | -0.522 | -0.212 | -0.853 | -0.880 | -0.727 |
| ChatGPT-RTS | -0.042 | -0.072 | -0.121 | -0.811 | -0.751 | -0.636 | -0.748 | -0.728 | -0.606 | -0.559 | -0.473 | -0.394 |
| ChatGPT-MCQ | -0.175 | -0.11 | -0.182 | -0.818 | -0.411 | -0.636 | -0.622 | -0.484 | -0.394 | -0.350 | -0.622 | -0.212 |
| GPT-4-RTS | -0.531 | -0.678 | -0.424 | -0.600 | -0.103 | -0.236 | -0.839 | -0.520 | -0.515 | -0.958 | -0.880 | -0.848 |
| Coherence | Consistency | Fluency | Relevance | |||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| ρ(Ri, PRTS) | 0.343 | 0.198 | 0.091 | 0.685 | 0.506 | 0.576 | 0.727 | 0.797 | 0.545 | 0.168 | 0.128 | 0.000 |
| ρ(Ri, PMCQ) | 0.657 | 0.625 | 0.394 | 0.685 | 0.616 | 0.515 | 0.322 | 0.573 | 0.212 | 0.091 | -0.106 | 0.000 |
| Coh | Con | Flu | Rel | Avg | |
| ChatGPT-RTS2 | 5 | 6 | 7 | 6 | 6.00 |
| ChatGPT-MCQ2 | 4 | 7 | 6 | 4 | 5.25 |
| ChatGPT-StarEval | 7 | 7 | 7 | 6 | 6.75 |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| RTS2 | Human-Corr | 0.339 | 0.338 | 0.285 | 0.393 | 0.497 | 0.350 | 0.290 | 0.280 | 0.252 | 0.440 | 0.441 | 0.351 |
| Meta-Corr | -0.559 | -0.476 | -0.424 | -0.748 | -0.843 | -0.576 | -0.825 | -0.823 | -0.636 | -0.385 | -0.506 | -0.212 | |
| MCQ2 | Human-Corr | 0.430 | 0.423 | 0.355 | 0.327 | 0.483 | 0.306 | 0.258 | 0.396 | 0.229 | 0.240 | 0.258 | 0.206 |
| Meta-Corr | -0.308 | -0.275 | -0.212 | -0.545 | -0.446 | -0.364 | -0.811 | -0.615 | -0.636 | -0.217 | -0.707 | -0.182 | |
| StarEval | Human-Corr | 0.418 | 0.417 | 0.341 | 0.297 | 0.421 | 0.264 | 0.246 | 0.323 | 0.217 | 0.393 | 0.405 | 0.323 |
| Meta-Corr | -0.084 | -0.101 | 0.000 | -0.664 | -0.684 | -0.545 | -0.441 | -0.460 | -0.212 | -0.497 | -0.579 | -0.303 |
| Coherence | Consistency | Fluency | Relevance | ||||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | ||
| 7B | Human-Corr | -0.001 | -0.000 | -0.001 | 0.114 | 0.130 | 0.104 | 0.043 | 0.038 | 0.039 | 0.067 | 0.064 | 0.057 |
| Meta-Corr | 0.245 | -0.258 | 0.182 | -0.524 | -0.720 | -0.424 | -0.042 | 0.463 | -0.03 | 0.238 | 0.169 | 0.212 | |
| 13B | Human-Corr | 0.153 | 0.165 | 0.123 | 0.180 | 0.209 | 0.162 | 0.187 | 0.179 | 0.167 | 0.234 | 0.266 | 0.192 |
| Meta-Corr | -0.287 | -0.425 | -0.182 | -0.580 | -0.495 | -0.424 | -0.455 | -0.233 | -0.303 | -0.049 | -0.406 | 0.000 | |
| 70B | Human-Corr | 0.234 | 0.254 | 0.186 | 0.357 | 0.395 | 0.319 | 0.155 | 0.161 | 0.134 | 0.248 | 0.285 | 0.200 |
| Meta-Corr | -0.420 | -0.407 | -0.273 | -0.811 | -0.690 | -0.667 | -0.322 | -0.319 | -0.182 | -0.238 | -0.123 | -0.182 |
| ID | Coh | Con | Flu | Rel |
| M8 | 0.58 | 0.09 | 0.13 | 0.51 |
| M9 | 0.65 | 0.15 | 0.34 | 0.54 |
| M10 | 0.58 | 0.22 | 0.20 | 0.54 |
| M11 | 0.59 | 0.34 | 0.40 | 0.49 |
| M12 | 0.65 | 0.03 | 0.14 | 0.54 |
| M13 | 0.62 | 0.07 | 0.14 | 0.51 |
| M14 | 0.65 | 0.09 | 0.18 | 0.54 |
| M15 | 0.60 | 0.04 | 0.13 | 0.53 |
| M17 | 0.58 | 0.05 | 0.08 | 0.52 |
| M20 | 0.48 | 0.29 | 0.24 | 0.49 |
| M22 | 0.58 | 0.02 | 0.14 | 0.54 |
| M23 | 0.58 | 0.05 | 0.12 | 0.54 |
| ID | Model Name | Coh | Con | Flu | Rel | Avg |
| M22 | BART | 4.18 | 4.94 | 4.90 | 4.25 | 4.57 |
| M23 | Pegasus (C4) | 4.16 | 4.91 | 4.88 | 4.26 | 4.55 |
| M17 | T5 | 4.00 | 4.93 | 4.93 | 4.23 | 4.52 |
| M12 | Unified-ext-abs | 3.60 | 4.96 | 4.85 | 3.85 | 4.32 |
| M13 | ROUGESal | 3.44 | 4.82 | 4.86 | 3.83 | 4.24 |
| M15 | Closed book decoder | 3.35 | 4.95 | 4.80 | 3.67 | 4.19 |
| M14 | Multi-task (Ent + QG) | 3.20 | 4.90 | 4.74 | 3.63 | 4.12 |
| M8 | Pointer Generator | 3.29 | 4.65 | 4.79 | 3.55 | 4.07 |
| M9 | Fast-abs-rl | 2.38 | 4.67 | 4.50 | 3.52 | 3.77 |
| M10 | Bottom-Up | 2.73 | 4.25 | 4.42 | 3.38 | 3.70 |
| M20 | GPT-2 (zero-shot) | 3.63 | 3.40 | 3.97 | 3.30 | 3.58 |
| M11 | Improve-abs | 2.28 | 3.27 | 3.65 | 3.15 | 3.09 |
| Model A | Model B | Coherence | Consistency | Fluency | Relevance | ||||||||
| LLM | Human | LLM | Human | LLM | Human | LLM | Human | ||||||
| M22 | M23 | 65.5 | 53.5 | ✓ | 55.75 | 52.5 | ✓ | 61 | 49.5 | × | 58.75 | 49.5 | × |
| M23 | M17 | 48.25 | 52.5 | × | 47 | 49 | ✓ | 49 | 45.5 | ✓ | 45 | 52 | × |
| M17 | M12 | 44 | 66.5 | × | 43.25 | 48.5 | ✓ | 40.5 | 54.5 | × | 49.25 | 72.5 | × |
| M12 | M13 | 58 | 51 | ✓ | 56.5 | 54.5 | ✓ | 56.75 | 50 | × | 58 | 45 | × |
| M13 | M15 | 45.5 | 49.5 | ✓ | 52 | 46 | × | 48.75 | 52 | × | 51.25 | 60.5 | ✓ |
| M15 | M14 | 57 | 54 | ✓ | 55 | 53.5 | ✓ | 56.5 | 52 | ✓ | 54 | 57 | ✓ |
| M14 | M8 | 49.25 | 44 | ✓ | 50 | 54.5 | × | 47 | 46.5 | ✓ | 49.25 | 53.5 | × |
| M8 | M9 | 77.5 | 82 | ✓ | 78.5 | 53 | ✓ | 80.5 | 63.5 | ✓ | 76 | 54 | ✓ |
| M9 | M10 | 45 | 36 | ✓ | 41.5 | 58 | × | 46 | 44.5 | ✓ | 41.5 | 56 | × |
| M10 | M20 | 58.25 | 24 | × | 61.75 | 64 | ✓ | 63.75 | 61.5 | ✓ | 61.5 | 54.5 | ✓ |
| M20 | M11 | 56.5 | 82 | ✓ | 50 | 53 | × | 51 | 58.5 | ✓ | 50 | 53 | × |
| #CP | 8 | 7 | 7 | 4 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.420 | 0.383 | 0.323 | 0.229 | 0.273 | 0.209 | 0.274 | 0.245 | 0.236 | 0.519 | 0.509 | 0.438 |
| M9 | 0.174 | 0.243 | 0.142 | 0.119 | 0.265 | 0.103 | 0.256 | 0.258 | 0.219 | 0.254 | 0.307 | 0.200 |
| M10 | 0.365 | 0.415 | 0.292 | 0.305 | 0.452 | 0.251 | 0.258 | 0.288 | 0.223 | 0.367 | 0.378 | 0.284 |
| M11 | 0.378 | 0.420 | 0.320 | 0.404 | 0.488 | 0.335 | 0.227 | 0.288 | 0.182 | 0.501 | 0.511 | 0.394 |
| M12 | 0.208 | 0.237 | 0.160 | 0.087 | 0.020 | 0.082 | 0.086 | 0.107 | 0.071 | 0.438 | 0.451 | 0.354 |
| M13 | 0.455 | 0.473 | 0.359 | 0.178 | 0.037 | 0.167 | 0.063 | 0.151 | 0.055 | 0.403 | 0.403 | 0.329 |
| M14 | 0.433 | 0.467 | 0.355 | 0.114 | 0.187 | 0.105 | -0.007 | 0.055 | -0.005 | 0.421 | 0.435 | 0.336 |
| M15 | 0.372 | 0.385 | 0.279 | 0.189 | 0.316 | 0.177 | 0.100 | 0.087 | 0.085 | 0.252 | 0.288 | 0.193 |
| M17 | 0.291 | 0.320 | 0.233 | -0.086 | -0.061 | -0.084 | 0.017 | 0.046 | 0.015 | 0.204 | 0.199 | 0.175 |
| M20 | 0.382 | 0.394 | 0.310 | 0.528 | 0.501 | 0.430 | 0.367 | 0.349 | 0.307 | 0.292 | 0.278 | 0.237 |
| M22 | 0.215 | 0.293 | 0.162 | -0.072 | -0.052 | -0.070 | -0.114 | -0.131 | -0.101 | 0.201 | 0.300 | 0.172 |
| M23 | 0.392 | 0.427 | 0.320 | 0.003 | 0.305 | 0.002 | -0.078 | -0.022 | -0.069 | 0.223 | 0.324 | 0.184 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.289 | 0.310 | 0.236 | 0.235 | 0.362 | 0.226 | 0.348 | 0.348 | 0.321 | 0.349 | 0.419 | 0.302 |
| M9 | 0.170 | 0.170 | 0.148 | 0.211 | 0.321 | 0.200 | 0.198 | 0.299 | 0.174 | 0.318 | 0.356 | 0.276 |
| M10 | 0.352 | 0.314 | 0.293 | 0.138 | 0.273 | 0.125 | 0.155 | 0.210 | 0.138 | 0.420 | 0.427 | 0.362 |
| M11 | 0.285 | 0.312 | 0.250 | 0.380 | 0.370 | 0.317 | 0.200 | 0.218 | 0.163 | 0.397 | 0.428 | 0.334 |
| M12 | 0.306 | 0.304 | 0.258 | -0.025 | -0.059 | -0.024 | 0.256 | 0.306 | 0.239 | 0.283 | 0.287 | 0.246 |
| M13 | 0.425 | 0.425 | 0.351 | 0.471 | 0.312 | 0.457 | 0.199 | 0.214 | 0.186 | 0.435 | 0.402 | 0.375 |
| M14 | 0.490 | 0.472 | 0.422 | 0.112 | 0.157 | 0.110 | 0.084 | 0.143 | 0.078 | 0.326 | 0.329 | 0.284 |
| M15 | 0.317 | 0.298 | 0.250 | 0.061 | 0.513 | 0.060 | 0.055 | 0.025 | 0.050 | 0.433 | 0.420 | 0.378 |
| M17 | 0.250 | 0.255 | 0.215 | -0.106 | -0.081 | -0.105 | -0.011 | 0.024 | -0.011 | 0.293 | 0.285 | 0.260 |
| M20 | 0.463 | 0.450 | 0.381 | 0.455 | 0.442 | 0.371 | 0.494 | 0.450 | 0.404 | 0.326 | 0.334 | 0.264 |
| M22 | 0.211 | 0.173 | 0.182 | -0.096 | -0.080 | -0.095 | -0.092 | -0.056 | -0.087 | 0.352 | 0.371 | 0.313 |
| M23 | 0.218 | 0.209 | 0.189 | 0.107 | 0.448 | 0.104 | -0.275 | -0.269 | -0.261 | 0.147 | 0.187 | 0.129 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||||||
| Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | Chat-RTS | Chat-MCQ | GPT-4 | human | |
| M8 | 2.43 | 3.70 | 4.56 | 3.29 | 4.21 | 4.70 | 4.77 | 4.65 | 2.51 | 3.76 | 4.61 | 4.79 | 3.56 | 4.33 | 4.71 | 3.55 |
| M9 | 1.80 | 3.51 | 4.41 | 2.38 | 3.93 | 4.69 | 4.94 | 4.67 | 2.16 | 3.45 | 4.15 | 4.50 | 3.51 | 4.33 | 4.82 | 3.52 |
| M10 | 2.05 | 3.49 | 4.05 | 2.73 | 3.77 | 4.52 | 4.60 | 4.25 | 2.23 | 3.43 | 3.92 | 4.42 | 3.45 | 4.25 | 4.62 | 3.38 |
| M11 | 1.70 | 2.63 | 2.93 | 2.28 | 2.36 | 4.11 | 3.72 | 3.27 | 1.71 | 2.66 | 2.77 | 3.65 | 2.92 | 4.06 | 3.87 | 3.15 |
| M12 | 2.75 | 3.92 | 4.75 | 3.60 | 4.46 | 4.75 | 5.00 | 4.96 | 2.59 | 3.84 | 4.72 | 4.85 | 3.89 | 4.40 | 4.95 | 3.85 |
| M13 | 2.91 | 3.99 | 4.76 | 3.44 | 4.40 | 4.73 | 4.89 | 4.82 | 2.69 | 3.93 | 4.69 | 4.86 | 3.90 | 4.41 | 4.85 | 3.83 |
| M14 | 2.38 | 3.87 | 4.54 | 3.20 | 4.25 | 4.70 | 4.99 | 4.90 | 2.42 | 3.83 | 4.53 | 4.74 | 3.67 | 4.32 | 4.84 | 3.63 |
| M15 | 2.65 | 3.81 | 4.58 | 3.35 | 4.33 | 4.75 | 4.90 | 4.95 | 2.60 | 3.84 | 4.61 | 4.80 | 3.78 | 4.32 | 4.82 | 3.67 |
| M17 | 3.05 | 4.11 | 4.89 | 4.00 | 4.84 | 4.87 | 4.97 | 4.93 | 3.29 | 4.08 | 4.79 | 4.93 | 4.30 | 4.43 | 4.97 | 4.23 |
| M20 | 2.00 | 2.95 | 2.99 | 3.63 | 2.66 | 3.57 | 3.51 | 3.40 | 2.22 | 3.02 | 3.05 | 3.97 | 2.73 | 3.78 | 3.47 | 3.30 |
| M22 | 3.54 | 4.09 | 4.92 | 4.18 | 4.83 | 4.82 | 4.97 | 4.94 | 3.57 | 4.06 | 4.94 | 4.90 | 4.41 | 4.43 | 4.95 | 4.25 |
| M23 | 3.00 | 4.14 | 4.88 | 4.16 | 4.74 | 4.84 | 4.94 | 4.91 | 3.08 | 4.05 | 4.76 | 4.88 | 4.10 | 4.45 | 4.96 | 4.26 |
| avg | 2.52 | 3.68 | 4.35 | 3.35 | 4.06 | 4.59 | 4.68 | 4.55 | 2.59 | 3.66 | 4.29 | 4.61 | 3.68 | 4.29 | 4.65 | 3.72 |
| Consistency | The summary is inconsistent with the article as it omits important details and repeats a phrase multiple times, resulting in a score of 1. |
| Consistency | The summary includes some relevant information but lacks coherence and omits important details, resulting in a score of 2.5 out of 5 for consistency. |
| Fluency | The summary is well-written and grammatically correct, but lacks important details from the article such as the potential link between antibiotics and weight gain, and the warnings about over-prescription of antibiotics and the development of superbugs. Score: 3/5. |
| Fluency | The summary lacks coherence and does not provide a clear understanding of the article, resulting in a score of 1 for disfluency. |
| Coherence | The summary is somewhat coherent but includes incorrect information, such as stating that Merson restarted his row with Burnley instead of Andros Townsend, and does not flow smoothly; therefore, I would give it a score of 3. |
| Consistency | The summary is consistent with the article as it mentions the helicopter airlifting patients to the hospital, but it leaves out important details such as the number of people involved and their injuries. Final score: 3. |
| Consistency | The summary accurately reflects the main points of the article, although it could have mentioned Falcao's unhappiness in France. Final score: 4. |
| Fluency | The summary is coherent and grammatically correct, but it does not provide enough information about the incident. Final score: 3. |
| Fluency | The summary is only one sentence and do not provide enough information about the article's content. Final score: 1. |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.429 | 0.446 | 0.362 | 0.449 | 0.597 | 0.433 | 0.412 | 0.409 | 0.385 | 0.425 | 0.489 | 0.365 |
| M9 | 0.288 | 0.346 | 0.243 | 0.349 | 0.350 | 0.331 | 0.462 | 0.465 | 0.407 | 0.413 | 0.490 | 0.358 |
| M10 | 0.495 | 0.480 | 0.403 | 0.518 | 0.683 | 0.469 | 0.406 | 0.579 | 0.360 | 0.510 | 0.537 | 0.429 |
| M11 | 0.584 | 0.572 | 0.474 | 0.529 | 0.536 | 0.439 | 0.448 | 0.436 | 0.358 | 0.500 | 0.519 | 0.405 |
| M12 | 0.245 | 0.360 | 0.209 | - | - | - | 0.387 | 0.506 | 0.372 | 0.169 | 0.106 | 0.149 |
| M13 | 0.271 | 0.271 | 0.230 | 0.535 | 0.444 | 0.518 | 0.093 | 0.337 | 0.089 | 0.237 | 0.337 | 0.206 |
| M14 | 0.263 | 0.312 | 0.222 | -0.040 | -0.026 | -0.040 | 0.311 | 0.432 | 0.291 | 0.277 | 0.315 | 0.241 |
| M15 | 0.403 | 0.418 | 0.342 | 0.158 | 0.504 | 0.155 | 0.240 | 0.242 | 0.225 | 0.307 | 0.402 | 0.267 |
| M17 | 0.285 | 0.240 | 0.246 | 0.390 | 0.748 | 0.385 | 0.091 | 0.131 | 0.088 | 0.210 | 0.253 | 0.185 |
| M20 | 0.427 | 0.418 | 0.334 | 0.378 | 0.362 | 0.313 | 0.482 | 0.468 | 0.402 | 0.479 | 0.491 | 0.391 |
| M22 | 0.026 | 0.004 | 0.023 | 0.346 | 0.608 | 0.343 | 0.248 | 0.141 | 0.242 | 0.143 | 0.084 | 0.126 |
| M23 | 0.320 | 0.330 | 0.283 | 0.360 | 0.168 | 0.354 | 0.267 | 0.136 | 0.251 | 0.005 | 0.003 | 0.005 |
| ID | Coherence | Consistency | Fluency | Relevance | ||||||||
| Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | Spear. | Pear. | Kend. | |
| M8 | 0.464 | 0.455 | 0.407 | 0.557 | 0.613 | 0.526 | 0.391 | 0.367 | 0.341 | 0.517 | 0.514 | 0.471 |
| M9 | 0.436 | 0.415 | 0.416 | 0.215 | 0.342 | 0.195 | 0.578 | 0.571 | 0.517 | 0.450 | 0.485 | 0.412 |
| M10 | 0.374 | 0.345 | 0.340 | 0.300 | 0.297 | 0.270 | 0.631 | 0.585 | 0.551 | 0.506 | 0.507 | 0.459 |
| M11 | 0.326 | 0.428 | 0.307 | 0.437 | 0.421 | 0.393 | 0.354 | 0.509 | 0.303 | 0.467 | 0.486 | 0.419 |
| M12 | 0.500 | 0.469 | 0.444 | 0.446 | 0.486 | 0.424 | 0.273 | 0.320 | 0.241 | 0.588 | 0.591 | 0.542 |
| M13 | 0.462 | 0.450 | 0.408 | 0.470 | 0.510 | 0.445 | 0.325 | 0.345 | 0.282 | 0.515 | 0.498 | 0.475 |
| M14 | 0.524 | 0.460 | 0.463 | 0.325 | 0.293 | 0.307 | 0.216 | 0.257 | 0.194 | 0.501 | 0.476 | 0.455 |
| M15 | 0.534 | 0.499 | 0.470 | 0.266 | 0.326 | 0.251 | 0.383 | 0.369 | 0.337 | 0.584 | 0.550 | 0.535 |
| M17 | 0.312 | 0.294 | 0.274 | -0.016 | -0.008 | -0.015 | 0.251 | 0.261 | 0.225 | 0.530 | 0.479 | 0.504 |
| M20 | 0.636 | 0.623 | 0.566 | 0.733 | 0.664 | 0.644 | 0.566 | 0.556 | 0.492 | 0.570 | 0.565 | 0.496 |
| M22 | 0.314 | 0.338 | 0.276 | 0.245 | 0.151 | 0.241 | 0.371 | 0.341 | 0.330 | 0.471 | 0.412 | 0.451 |
| M23 | 0.380 | 0.386 | 0.337 | 0.158 | 0.496 | 0.153 | 0.296 | 0.275 | 0.259 | 0.450 | 0.505 | 0.421 |
| KBP | Train | Valid | Test |
| # dialogues | 1,981 | 248 | 248 |
| # samples | 9,821 | 1,227 | 1,229 |
| # avg turns | 4.96 | 4.93 | 4.96 |
| # utterances | 19,642 | 2,454 | 2,458 |
| # avg length | 17.6 | 17.3 | 17.5 |
| # resp w/ persona | 86.1% | 84.4% | 85.3% |
| # resp w/ p_and_k | 76.3% | 74.2% | 75.1% |
| Model | NULL | Persona | Both |
| Supervised | |||
| BELLE-LLAMA-7B-2M | 42.67 (194) | 14.08 (17) | 83.77 (1018) |
| CHATGLM-6B | 47.10 (129) | 31.96 (69) | 86.59 (1031) |
| Unsupervised | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 28.55 (940) | 8.94 (54) | 32.47 (235) |
| CHATGLM-6B | 25.60 (1225) | 0.0 (0) | 0.43 (4) |
| CHATGPT | 11.45 (116) | 20.67 (233) | 74.88 (880) |
| In-context | |||
| BELLE-LLAMA-7B-2M | 9.22 (36) | 18.21 (1193) | 0.0 (0) |
| CHATGLM-6B | 25.67 (1190) | 1.49 (9) | 4.62 (30) |
| CHATGPT | 27.95 (699) | 23.14 (238) | 41.98 (292) |
| Model | Persona | Both | ||
| PERSONA | DOCUMENTS | DOCUMENTS† | ||
| BM25 | 36.80 | 48.97 | 15.05 | 11.37 |
| RocketQAv2 | 80.00 | 92.31 | 50.49 | 35.75 |
| DPR | 83.20 | 93.07 | 51.67 | 39.33 |
| Model | BLEU1 | Rouge-L | P.C | K.C |
| Supervised Setting | ||||
| BELLE-LLAMA-7B-2M | 30.48 | 34.61 | 75.34 | 46.62 |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| Unsupervised Setting | ||||
| Zero-shot | ||||
| BELLE-LLAMA-7B-2M | 11.84 | 19.24 | 30.59 | 27.34 |
| CHATGLM-6B | 6.18 | 14.50 | 14.73 | 24.73 |
| CHATGPT | 12.06 | 24.44 | 73.47 | 38.00 |
| In-context | ||||
| BELLE-LLAMA-7B-2M | 19.51 | 22.25 | 72.98 | 24.89 |
| CHATGLM-6B | 13.74 | 19.69 | 16.92 | 24.89 |
| CHATGPT | 16.03 | 25.62 | 46.38 | 35.56 |
| Model | BLEU1 | RougeL | P.C | K.C |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| + Ground Planning | 24.29 | 27.01 | 86.16 | 57.12 |
| + Ground Retrieval | 25.86 | 29.15 | 79.52 | 53.95 |
| + Ground P & R | 25.71 | 29.43 | 90.56 | 72.99 |
| - Dependency | 23.32 | 25.53 | 75.67 | 38.49 |
| - Documents | 23.06 | 25.34 | 75.91 | 36.53 |
| - Planning* | 23.51 | 25.98 | 72.90 | 24.89 |
| - Planning** | 23.69 | 26.81 | 71.60 | 34.91 |
| Number | Assembling | |||
| BLEU1 | RougeL | P.C | K.C | |
| 1 | 23.81 | 26.70 | 76.99 | 42.39 |
| 2 | 22.70 | 25.57 | 71.03 | 29.45 |
| 3 | 20.69 | 24.05 | 69.73 | 27.91 |
| Model | Coh. | Per.Cs (%) | Know.Cs (%) |
| Supervised Setting | |||
| BELLE-LLAMA-7B-2M | 4.38 | 72.0 | 63.8 |
| CHATGLM-6B | 4.06 | 68.0 | 59.1 |
| Unsupervised Setting | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 2.84 | 24.7 | 19.5 |
| CHATGLM-6B | 2.58 | 17.0 | 14.8 |
| CHATGPT | 4.00 | 63.4 | 33.3 |
| In-context | |||
| BELLE-LLAMA-7B-2M | 3.36 | 40.0 | 21.7 |
| CHATGLM-6B | 2.88 | 32.0 | 28.9 |
| CHATGPT | 4.03 | 54.0 | 48.8 |
| Attributes | Templates |
| nick_name, birth_place, ... rewards, members, ... | {head entity}’s {attribute} is {tail entity} The {attribute} of {head entity} contains {tail entity} |
| KBP | Train | Valid | Test |
| # dialogues | 1,981 | 248 | 248 |
| # samples | 9,821 | 1,227 | 1,229 |
| # avg turns | 4.96 | 4.93 | 4.96 |
| # utterances | 19,642 | 2,454 | 2,458 |
| # avg length | 17.6 | 17.3 | 17.5 |
| # resp w/ persona | 86.1% | 84.4% | 85.3% |
| # resp w/ p_and_k | 76.3% | 74.2% | 75.1% |
| Model | NULL | Persona | Both |
| Supervised | |||
| BELLE-LLAMA-7B-2M | 42.67 (194) | 14.08 (17) | 83.77 (1018) |
| CHATGLM-6B | 47.10 (129) | 31.96 (69) | 86.59 (1031) |
| Unsupervised | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 28.55 (940) | 8.94 (54) | 32.47 (235) |
| CHATGLM-6B | 25.60 (1225) | 0.0 (0) | 0.43 (4) |
| CHATGPT | 11.45 (116) | 20.67 (233) | 74.88 (880) |
| In-context | |||
| BELLE-LLAMA-7B-2M | 9.22 (36) | 18.21 (1193) | 0.0 (0) |
| CHATGLM-6B | 25.67 (1190) | 1.49 (9) | 4.62 (30) |
| CHATGPT | 27.95 (699) | 23.14 (238) | 41.98 (292) |
| Model | Persona | Both | ||
| PERSONA | DOCUMENTS | DOCUMENTS† | ||
| BM25 | 36.80 | 48.97 | 15.05 | 11.37 |
| RocketQAv2 | 80.00 | 92.31 | 50.49 | 35.75 |
| DPR | 83.20 | 93.07 | 51.67 | 39.33 |
| Model | BLEU1 | Rouge-L | P.C | K.C |
| Supervised Setting | ||||
| BELLE-LLAMA-7B-2M | 30.48 | 34.61 | 75.34 | 46.62 |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| Unsupervised Setting | ||||
| Zero-shot | ||||
| BELLE-LLAMA-7B-2M | 11.84 | 19.24 | 30.59 | 27.34 |
| CHATGLM-6B | 6.18 | 14.50 | 14.73 | 24.73 |
| CHATGPT | 12.06 | 24.44 | 73.47 | 38.00 |
| In-context | ||||
| BELLE-LLAMA-7B-2M | 19.51 | 22.25 | 72.98 | 24.89 |
| CHATGLM-6B | 13.74 | 19.69 | 16.92 | 24.89 |
| CHATGPT | 16.03 | 25.62 | 46.38 | 35.56 |
| Model | BLEU1 | RougeL | P.C | K.C |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| + Ground Planning | 24.29 | 27.01 | 86.16 | 57.12 |
| + Ground Retrieval | 25.86 | 29.15 | 79.52 | 53.95 |
| + Ground P & R | 25.71 | 29.43 | 90.56 | 72.99 |
| - Dependency | 23.32 | 25.53 | 75.67 | 38.49 |
| - Documents | 23.06 | 25.34 | 75.91 | 36.53 |
| - Planning* | 23.51 | 25.98 | 72.90 | 24.89 |
| - Planning** | 23.69 | 26.81 | 71.60 | 34.91 |
| Number | Assembling | |||
| BLEU1 | RougeL | P.C | K.C | |
| 1 | 23.81 | 26.70 | 76.99 | 42.39 |
| 2 | 22.70 | 25.57 | 71.03 | 29.45 |
| 3 | 20.69 | 24.05 | 69.73 | 27.91 |
| Model | Coh. | Per.Cs (%) | Know.Cs (%) |
| Supervised Setting | |||
| BELLE-LLAMA-7B-2M | 4.38 | 72.0 | 63.8 |
| CHATGLM-6B | 4.06 | 68.0 | 59.1 |
| Unsupervised Setting | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 2.84 | 24.7 | 19.5 |
| CHATGLM-6B | 2.58 | 17.0 | 14.8 |
| CHATGPT | 4.00 | 63.4 | 33.3 |
| In-context | |||
| BELLE-LLAMA-7B-2M | 3.36 | 40.0 | 21.7 |
| CHATGLM-6B | 2.88 | 32.0 | 28.9 |
| CHATGPT | 4.03 | 54.0 | 48.8 |
| Attributes | Templates |
| nick_name, birth_place, ... rewards, members, ... | {head entity}’s {attribute} is {tail entity} The {attribute} of {head entity} contains {tail entity} |
| KBP | Train | Valid | Test |
| # dialogues | 1,981 | 248 | 248 |
| # samples | 9,821 | 1,227 | 1,229 |
| # avg turns | 4.96 | 4.93 | 4.96 |
| # utterances | 19,642 | 2,454 | 2,458 |
| # avg length | 17.6 | 17.3 | 17.5 |
| # resp w/ persona | 86.1% | 84.4% | 85.3% |
| # resp w/ p_and_k | 76.3% | 74.2% | 75.1% |
| Model | NULL | Persona | Both |
| Supervised | |||
| BELLE-LLAMA-7B-2M | 42.67 (194) | 14.08 (17) | 83.77 (1018) |
| CHATGLM-6B | 47.10 (129) | 31.96 (69) | 86.59 (1031) |
| Unsupervised | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 28.55 (940) | 8.94 (54) | 32.47 (235) |
| CHATGLM-6B | 25.60 (1225) | 0.0 (0) | 0.43 (4) |
| CHATGPT | 11.45 (116) | 20.67 (233) | 74.88 (880) |
| In-context | |||
| BELLE-LLAMA-7B-2M | 9.22 (36) | 18.21 (1193) | 0.0 (0) |
| CHATGLM-6B | 25.67 (1190) | 1.49 (9) | 4.62 (30) |
| CHATGPT | 27.95 (699) | 23.14 (238) | 41.98 (292) |
| Model | Persona | Both | ||
| PERSONA | DOCUMENTS | DOCUMENTS† | ||
| BM25 | 36.80 | 48.97 | 15.05 | 11.37 |
| RocketQAv2 | 80.00 | 92.31 | 50.49 | 35.75 |
| DPR | 83.20 | 93.07 | 51.67 | 39.33 |
| Model | BLEU1 | Rouge-L | P.C | K.C |
| Supervised Setting | ||||
| BELLE-LLAMA-7B-2M | 30.48 | 34.61 | 75.34 | 46.62 |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| Unsupervised Setting | ||||
| Zero-shot | ||||
| BELLE-LLAMA-7B-2M | 11.84 | 19.24 | 30.59 | 27.34 |
| CHATGLM-6B | 6.18 | 14.50 | 14.73 | 24.73 |
| CHATGPT | 12.06 | 24.44 | 73.47 | 38.00 |
| In-context | ||||
| BELLE-LLAMA-7B-2M | 19.51 | 22.25 | 72.98 | 24.89 |
| CHATGLM-6B | 13.74 | 19.69 | 16.92 | 24.89 |
| CHATGPT | 16.03 | 25.62 | 46.38 | 35.56 |
| Model | BLEU1 | RougeL | P.C | K.C |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| + Ground Planning | 24.29 | 27.01 | 86.16 | 57.12 |
| + Ground Retrieval | 25.86 | 29.15 | 79.52 | 53.95 |
| + Ground P & R | 25.71 | 29.43 | 90.56 | 72.99 |
| - Dependency | 23.32 | 25.53 | 75.67 | 38.49 |
| - Documents | 23.06 | 25.34 | 75.91 | 36.53 |
| - Planning* | 23.51 | 25.98 | 72.90 | 24.89 |
| - Planning** | 23.69 | 26.81 | 71.60 | 34.91 |
| Number | Assembling | |||
| BLEU1 | RougeL | P.C | K.C | |
| 1 | 23.81 | 26.70 | 76.99 | 42.39 |
| 2 | 22.70 | 25.57 | 71.03 | 29.45 |
| 3 | 20.69 | 24.05 | 69.73 | 27.91 |
| Model | Coh. | Per.Cs (%) | Know.Cs (%) |
| Supervised Setting | |||
| BELLE-LLAMA-7B-2M | 4.38 | 72.0 | 63.8 |
| CHATGLM-6B | 4.06 | 68.0 | 59.1 |
| Unsupervised Setting | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 2.84 | 24.7 | 19.5 |
| CHATGLM-6B | 2.58 | 17.0 | 14.8 |
| CHATGPT | 4.00 | 63.4 | 33.3 |
| In-context | |||
| BELLE-LLAMA-7B-2M | 3.36 | 40.0 | 21.7 |
| CHATGLM-6B | 2.88 | 32.0 | 28.9 |
| CHATGPT | 4.03 | 54.0 | 48.8 |
| Attributes | Templates |
| nick_name, birth_place, ... rewards, members, ... | {head entity}’s {attribute} is {tail entity} The {attribute} of {head entity} contains {tail entity} |
| KBP | Train | Valid | Test |
| # dialogues | 1,981 | 248 | 248 |
| # samples | 9,821 | 1,227 | 1,229 |
| # avg turns | 4.96 | 4.93 | 4.96 |
| # utterances | 19,642 | 2,454 | 2,458 |
| # avg length | 17.6 | 17.3 | 17.5 |
| # resp w/ persona | 86.1% | 84.4% | 85.3% |
| # resp w/ p_and_k | 76.3% | 74.2% | 75.1% |
| Model | NULL | Persona | Both |
| Supervised | |||
| BELLE-LLAMA-7B-2M | 42.67 (194) | 14.08 (17) | 83.77 (1018) |
| CHATGLM-6B | 47.10 (129) | 31.96 (69) | 86.59 (1031) |
| Unsupervised | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 28.55 (940) | 8.94 (54) | 32.47 (235) |
| CHATGLM-6B | 25.60 (1225) | 0.0 (0) | 0.43 (4) |
| CHATGPT | 11.45 (116) | 20.67 (233) | 74.88 (880) |
| In-context | |||
| BELLE-LLAMA-7B-2M | 9.22 (36) | 18.21 (1193) | 0.0 (0) |
| CHATGLM-6B | 25.67 (1190) | 1.49 (9) | 4.62 (30) |
| CHATGPT | 27.95 (699) | 23.14 (238) | 41.98 (292) |
| Model | Persona | Both | ||
| PERSONA | DOCUMENTS | DOCUMENTS† | ||
| BM25 | 36.80 | 48.97 | 15.05 | 11.37 |
| RocketQAv2 | 80.00 | 92.31 | 50.49 | 35.75 |
| DPR | 83.20 | 93.07 | 51.67 | 39.33 |
| Model | BLEU1 | Rouge-L | P.C | K.C |
| Supervised Setting | ||||
| BELLE-LLAMA-7B-2M | 30.48 | 34.61 | 75.34 | 46.62 |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| Unsupervised Setting | ||||
| Zero-shot | ||||
| BELLE-LLAMA-7B-2M | 11.84 | 19.24 | 30.59 | 27.34 |
| CHATGLM-6B | 6.18 | 14.50 | 14.73 | 24.73 |
| CHATGPT | 12.06 | 24.44 | 73.47 | 38.00 |
| In-context | ||||
| BELLE-LLAMA-7B-2M | 19.51 | 22.25 | 72.98 | 24.89 |
| CHATGLM-6B | 13.74 | 19.69 | 16.92 | 24.89 |
| CHATGPT | 16.03 | 25.62 | 46.38 | 35.56 |
| Model | BLEU1 | RougeL | P.C | K.C |
| CHATGLM-6B | 23.81 | 26.70 | 76.99 | 42.39 |
| + Ground Planning | 24.29 | 27.01 | 86.16 | 57.12 |
| + Ground Retrieval | 25.86 | 29.15 | 79.52 | 53.95 |
| + Ground P & R | 25.71 | 29.43 | 90.56 | 72.99 |
| - Dependency | 23.32 | 25.53 | 75.67 | 38.49 |
| - Documents | 23.06 | 25.34 | 75.91 | 36.53 |
| - Planning* | 23.51 | 25.98 | 72.90 | 24.89 |
| - Planning** | 23.69 | 26.81 | 71.60 | 34.91 |
| Number | Assembling | |||
| BLEU1 | RougeL | P.C | K.C | |
| 1 | 23.81 | 26.70 | 76.99 | 42.39 |
| 2 | 22.70 | 25.57 | 71.03 | 29.45 |
| 3 | 20.69 | 24.05 | 69.73 | 27.91 |
| Model | Coh. | Per.Cs (%) | Know.Cs (%) |
| Supervised Setting | |||
| BELLE-LLAMA-7B-2M | 4.38 | 72.0 | 63.8 |
| CHATGLM-6B | 4.06 | 68.0 | 59.1 |
| Unsupervised Setting | |||
| Zero-shot | |||
| BELLE-LLAMA-7B-2M | 2.84 | 24.7 | 19.5 |
| CHATGLM-6B | 2.58 | 17.0 | 14.8 |
| CHATGPT | 4.00 | 63.4 | 33.3 |
| In-context | |||
| BELLE-LLAMA-7B-2M | 3.36 | 40.0 | 21.7 |
| CHATGLM-6B | 2.88 | 32.0 | 28.9 |
| CHATGPT | 4.03 | 54.0 | 48.8 |
| Attributes | Templates |
| nick_name, birth_place, ... rewards, members, ... | {head entity}’s {attribute} is {tail entity} The {attribute} of {head entity} contains {tail entity} |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Unsupervised MEANSUM† | 10.44 | 0.63 | 9.55 | 5.95 | 0.45 | 5.29 | 13.78 | 0.93 | 11.70 |
| LEXRANK† | 14.12 | 1.50 | 12.81 | 8.28 | 0.82 | 7.24 | 15.12 | 1.84 | 12.60 |
| COPYCAT† | 15.12 | 1.48 | 13.85 | 6.81 | 0.82 | 5.89 | 17.05 | 1.78 | 14.50 |
| EXTSUM† | 19.06 | 2.47 | 17.49 | 11.63 | 1.19 | 10.44 | 18.74 | 3.01 | 15.74 |
| Supervised SELSUM† | 21.29 | 4.00 | 19.39 | 14.96 | 2.60 | 13.07 | 24.33 | 5.29 | 18.84 |
| LONGFORMER | 22.40 | 4.71 | 15.36 | 14.68 | 2.53 | 11.62 | 22.56 | 4.83 | 17.08 |
| SELSUM* | 23.17 | 4.77 | 21.13 | 15.12 | 2.83 | 13.07 | 22.87 | 4.85 | 18.05 |
| BRIO | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| Zero-Shot QG | 18.40 | 2.21 | 16.02 | 13.27 | 1.39 | 11.61 | 14.75 | 1.23 | 12.06 |
| CHATGPT | 18.53 | 2.37 | 14.68 | 13.92 | 1.78 | 11.93 | 22.88 | 3.50 | 19.79 |
| Method | R-1 | R-2 | R-L |
| Unsupervised | |||
| W2VCENT† | 13.93 | 2.10 | 10.81 |
| LEXRANK† | 14.88 | 1.94 | 10.50 |
| OPINOSIST | 14.98 | 3.07 | 12.19 |
| MEANSUM† | 15.79 | 1.94 | 12.26 |
| SNCENT† | 15.90 | 2.01 | 11.74 |
| BERTCENT† | 17.65 | 2.78 | 12.78 |
| DENOISESUM† | 21.26 | 4.61 | 16.27 |
| Weakly Supervised | |||
| PLANSUM† | 21.77 | 6.18 | 16.98 |
| Supervised | |||
| BRIO | 23.72 | 5.16 | 18.05 |
| LONGFORMER | 24.96 | 6.34 | 18.40 |
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| Zero-Shot | |||
| QG | 18.14 | 2.34 | 14.28 |
| CHATGPT | 22.73 | 4.21 | 17.52 |
| Summary | Pros | Cons | Verdict | ||||||
| Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | |
| GOLD | 19.7 | 5.7 | 6.4 | 0 | 0 | 4 | 6.9 | 5.2 | 10.2 |
| SELSUM | -20.2 | 2.9 | 11.8 | -6.3 | -0.6 | -0.9 | -7.2 | 6.5 | 0.6 |
| BRIO | -7.5 | -25.6 | -23.5 | -5.1 | -7.3 | -10.7 | -8.4 | -17.2 | -14.7 |
| SUBSUMM | 8 | 17 | 5.3 | 11.4 | 7.9 | 7.6 | 8.6 | 5.5 | 3.9 |
| Method | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| SUBSUMM | 0.66 | 0.67 | 0.89 | 0.70 |
| CHATGPT | 0.34 | 0.33 | 0.11 | 0.30 |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Sampling Strategy | |||||||||
| RAND | 21.83 | 4.28 | 19.91 | 14.06 | 2.55 | 12.35 | 22.67 | 4.66 | 17.81 |
| SENTI-RAND | 22.33 | 4.41 | 20.45 | 14.92 | 2.75 | 13.13 | 23.06 | 4.79 | 17.98 |
| SENTIINFO | 22.82 | 4.60 | 20.80 | 14.42 | 2.48 | 12.59 | 23.51 | 5.03 | 18.16 |
| SENTI-RANDINFO | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| Training Scheme | |||||||||
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| w/o Stage I | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| w/o Stage II | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| RAND in Stage I | 26.03 | 4.93 | 23.96 | 16.66 | 2.97 | 14.78 | 24.71 | 4.93 | 19.43 |
| RAND in Stage II | 25.95 | 4.95 | 24.02 | 16.50 | 2.77 | 14.54 | 24.60 | 4.96 | 19.45 |
| AmaSum | Train | Dev | Test |
| Entity | 26,660 | 3,302 | 3,362 |
| Reviews per Ent. | 77.78 | 77.76 | 76.80 |
| Tokens per Rev. | 53.98 | 54.13 | 53.97 |
| Tokens per Sum. | |||
| - Pros | 39.13 | 39.21 | 38.81 |
| - Cons | 18.34 | 18.31 | 17.81 |
| - Verdict | 21.72 | 21.97 | 21.59 |
| Rotten Tomatoes | Train | Dev | Test |
| Entity | 2,453 | 536 | 737 |
| Reviews per Ent. | 74.39 | 74.85 | 73.26 |
| Tokens per Rev. | 26.52 | 26.32 | 26.41 |
| Tokens per Sum. | 23.80 | 23.59 | 23.82 |
| Hyperparameter | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| bsz I | 16 | 16 | 16 | 8 |
| bsz II | 16 | 16 | 16 | 16 |
| warmup I | 5,000 | 5,000 | 5,000 | 500 |
| warmup II | 3,000 | 3,000 | 3,000 | 300 |
| γ (Eq. 13) | 0.1 | 1.0 | 0.1 | 0.1 |
| lenpen | 0.5 | 0.5 | 1.0 | 1.0 |
| minlen | 35 | 25 | 25 | 30 |
| Value | R-1 | R-2 | R-L |
| K=6 | 21.36 | 4.30 | 15.87 |
| K=8 | 22.16 | 4.89 | 16.62 |
| K=10 | 23.20 | 5.56 | 17.28 |
| K=12 | 22.54 | 5.40 | 16.58 |
| Method | R-1 | R-2 | R-L |
| Sampling Strategy | |||
| RAND | 22.73 | 5.27 | 16.80 |
| SENTI-RAND | 22.76 | 5.36 | 16.81 |
| SENTIINFO | 23.44 | 5.84 | 17.25 |
| SENTI-RANDINFO | 23.99 | 6.09 | 17.54 |
| Training Scheme | |||
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| w/o Stage I | 23.72 | 5.16 | 18.05 |
| w/o Stage II | 23.99 | 6.09 | 17.54 |
| RAND in Stage I | 25.40 | 6.74 | 19.13 |
| RAND in Stage II | 25.46 | 6.77 | 18.90 |
| Method | Pros | Cons | Verdict |
| GOLD | A quality pumice stone crafted from 100% pure volcanic lava. Has a good shape, making it easy to hold and apply pressure to hard skin. Use the handy rope loop to hang the stone out to dry after use. | Some users find the coarse-ness rough and takes time to get used to. | Comfortable to hold and use, this all natural vol- canic pumice stone works hard on calluses and tough skin. |
| SELSUM | Made of natural materi- als that are easy to clean. Comes with a rope at- tached to hang in the shower for easy use. | Some users find it difficult to remove dead skin while it's in use. May not re- move all dead skin on the feet, but it's a good choice for those with dry skin. | A good choice if you're looking for a natural pumice stone that's easy to hold and use. |
| BRIO | A before and after, it was a bit embarrassing, and my feet were a few deep cracks for a few scrubs my feet are gone and a couple of scrubs | Users report that the pumice stone is not as smooth as they would like. Some reports of the stone breaking. | Pumice stone is the real deal and I appreciate that. using it for the first time is a great experience. |
| CHATGPT | Pumice Valley's natural pumice stone is effective in removing dead skin, comes with a handy rope for easy use, and has a fine grain that doesn't rip | The stone is not sharp or rough enough to re- move dry skin, it may be chipped or in poor shape, it can | The pumice stone is effec- tive in removing dead skin and smoothing heels, but may not work well on ex- isting calluses. |
| SUBSUMM | Natural lava stone that's easy to use and easy to clean. Comes with a rope for hanging in the shower. Comes in a variety of sizes. Made from lava. | May be too rough for some users. Some users find it difficult to remove the calluses from the stone. May not remove all cal- luses. | You're looking for a pumice stone that's easy to use and clean, this is the one to buy. |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Unsupervised MEANSUM† | 10.44 | 0.63 | 9.55 | 5.95 | 0.45 | 5.29 | 13.78 | 0.93 | 11.70 |
| LEXRANK† | 14.12 | 1.50 | 12.81 | 8.28 | 0.82 | 7.24 | 15.12 | 1.84 | 12.60 |
| COPYCAT† | 15.12 | 1.48 | 13.85 | 6.81 | 0.82 | 5.89 | 17.05 | 1.78 | 14.50 |
| EXTSUM† | 19.06 | 2.47 | 17.49 | 11.63 | 1.19 | 10.44 | 18.74 | 3.01 | 15.74 |
| Supervised SELSUM† | 21.29 | 4.00 | 19.39 | 14.96 | 2.60 | 13.07 | 24.33 | 5.29 | 18.84 |
| LONGFORMER | 22.40 | 4.71 | 15.36 | 14.68 | 2.53 | 11.62 | 22.56 | 4.83 | 17.08 |
| SELSUM* | 23.17 | 4.77 | 21.13 | 15.12 | 2.83 | 13.07 | 22.87 | 4.85 | 18.05 |
| BRIO | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| Zero-Shot QG | 18.40 | 2.21 | 16.02 | 13.27 | 1.39 | 11.61 | 14.75 | 1.23 | 12.06 |
| CHATGPT | 18.53 | 2.37 | 14.68 | 13.92 | 1.78 | 11.93 | 22.88 | 3.50 | 19.79 |
| Method | R-1 | R-2 | R-L |
| Unsupervised | |||
| W2VCENT† | 13.93 | 2.10 | 10.81 |
| LEXRANK† | 14.88 | 1.94 | 10.50 |
| OPINOSIST | 14.98 | 3.07 | 12.19 |
| MEANSUM† | 15.79 | 1.94 | 12.26 |
| SNCENT† | 15.90 | 2.01 | 11.74 |
| BERTCENT† | 17.65 | 2.78 | 12.78 |
| DENOISESUM† | 21.26 | 4.61 | 16.27 |
| Weakly Supervised | |||
| PLANSUM† | 21.77 | 6.18 | 16.98 |
| Supervised | |||
| BRIO | 23.72 | 5.16 | 18.05 |
| LONGFORMER | 24.96 | 6.34 | 18.40 |
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| Zero-Shot | |||
| QG | 18.14 | 2.34 | 14.28 |
| CHATGPT | 22.73 | 4.21 | 17.52 |
| Summary | Pros | Cons | Verdict | ||||||
| Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | |
| GOLD | 19.7 | 5.7 | 6.4 | 0 | 0 | 4 | 6.9 | 5.2 | 10.2 |
| SELSUM | -20.2 | 2.9 | 11.8 | -6.3 | -0.6 | -0.9 | -7.2 | 6.5 | 0.6 |
| BRIO | -7.5 | -25.6 | -23.5 | -5.1 | -7.3 | -10.7 | -8.4 | -17.2 | -14.7 |
| SUBSUMM | 8 | 17 | 5.3 | 11.4 | 7.9 | 7.6 | 8.6 | 5.5 | 3.9 |
| Method | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| SUBSUMM | 0.66 | 0.67 | 0.89 | 0.70 |
| CHATGPT | 0.34 | 0.33 | 0.11 | 0.30 |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Sampling Strategy | |||||||||
| RAND | 21.83 | 4.28 | 19.91 | 14.06 | 2.55 | 12.35 | 22.67 | 4.66 | 17.81 |
| SENTI-RAND | 22.33 | 4.41 | 20.45 | 14.92 | 2.75 | 13.13 | 23.06 | 4.79 | 17.98 |
| SENTIINFO | 22.82 | 4.60 | 20.80 | 14.42 | 2.48 | 12.59 | 23.51 | 5.03 | 18.16 |
| SENTI-RANDINFO | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| Training Scheme | |||||||||
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| w/o Stage I | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| w/o Stage II | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| RAND in Stage I | 26.03 | 4.93 | 23.96 | 16.66 | 2.97 | 14.78 | 24.71 | 4.93 | 19.43 |
| RAND in Stage II | 25.95 | 4.95 | 24.02 | 16.50 | 2.77 | 14.54 | 24.60 | 4.96 | 19.45 |
| AmaSum | Train | Dev | Test |
| Entity | 26,660 | 3,302 | 3,362 |
| Reviews per Ent. | 77.78 | 77.76 | 76.80 |
| Tokens per Rev. | 53.98 | 54.13 | 53.97 |
| Tokens per Sum. | |||
| - Pros | 39.13 | 39.21 | 38.81 |
| - Cons | 18.34 | 18.31 | 17.81 |
| - Verdict | 21.72 | 21.97 | 21.59 |
| Rotten Tomatoes | Train | Dev | Test |
| Entity | 2,453 | 536 | 737 |
| Reviews per Ent. | 74.39 | 74.85 | 73.26 |
| Tokens per Rev. | 26.52 | 26.32 | 26.41 |
| Tokens per Sum. | 23.80 | 23.59 | 23.82 |
| Hyperparameter | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| bsz I | 16 | 16 | 16 | 8 |
| bsz II | 16 | 16 | 16 | 16 |
| warmup I | 5,000 | 5,000 | 5,000 | 500 |
| warmup II | 3,000 | 3,000 | 3,000 | 300 |
| γ (Eq. 13) | 0.1 | 1.0 | 0.1 | 0.1 |
| lenpen | 0.5 | 0.5 | 1.0 | 1.0 |
| minlen | 35 | 25 | 25 | 30 |
| Value | R-1 | R-2 | R-L |
| K=6 | 21.36 | 4.30 | 15.87 |
| K=8 | 22.16 | 4.89 | 16.62 |
| K=10 | 23.20 | 5.56 | 17.28 |
| K=12 | 22.54 | 5.40 | 16.58 |
| Method | R-1 | R-2 | R-L |
| Sampling Strategy | |||
| RAND | 22.73 | 5.27 | 16.80 |
| SENTI-RAND | 22.76 | 5.36 | 16.81 |
| SENTIINFO | 23.44 | 5.84 | 17.25 |
| SENTI-RANDINFO | 23.99 | 6.09 | 17.54 |
| Training Scheme | |||
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| w/o Stage I | 23.72 | 5.16 | 18.05 |
| w/o Stage II | 23.99 | 6.09 | 17.54 |
| RAND in Stage I | 25.40 | 6.74 | 19.13 |
| RAND in Stage II | 25.46 | 6.77 | 18.90 |
| Method | Pros | Cons | Verdict |
| GOLD | A quality pumice stone crafted from 100% pure volcanic lava. Has a good shape, making it easy to hold and apply pressure to hard skin. Use the handy rope loop to hang the stone out to dry after use. | Some users find the coarse-ness rough and takes time to get used to. | Comfortable to hold and use, this all natural vol- canic pumice stone works hard on calluses and tough skin. |
| SELSUM | Made of natural materi- als that are easy to clean. Comes with a rope at- tached to hang in the shower for easy use. | Some users find it difficult to remove dead skin while it's in use. May not re- move all dead skin on the feet, but it's a good choice for those with dry skin. | A good choice if you're looking for a natural pumice stone that's easy to hold and use. |
| BRIO | A before and after, it was a bit embarrassing, and my feet were a few deep cracks for a few scrubs my feet are gone and a couple of scrubs | Users report that the pumice stone is not as smooth as they would like. Some reports of the stone breaking. | Pumice stone is the real deal and I appreciate that. using it for the first time is a great experience. |
| CHATGPT | Pumice Valley's natural pumice stone is effective in removing dead skin, comes with a handy rope for easy use, and has a fine grain that doesn't rip | The stone is not sharp or rough enough to re- move dry skin, it may be chipped or in poor shape, it can | The pumice stone is effec- tive in removing dead skin and smoothing heels, but may not work well on ex- isting calluses. |
| SUBSUMM | Natural lava stone that's easy to use and easy to clean. Comes with a rope for hanging in the shower. Comes in a variety of sizes. Made from lava. | May be too rough for some users. Some users find it difficult to remove the calluses from the stone. May not remove all cal- luses. | You're looking for a pumice stone that's easy to use and clean, this is the one to buy. |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Unsupervised MEANSUM† | 10.44 | 0.63 | 9.55 | 5.95 | 0.45 | 5.29 | 13.78 | 0.93 | 11.70 |
| LEXRANK† | 14.12 | 1.50 | 12.81 | 8.28 | 0.82 | 7.24 | 15.12 | 1.84 | 12.60 |
| COPYCAT† | 15.12 | 1.48 | 13.85 | 6.81 | 0.82 | 5.89 | 17.05 | 1.78 | 14.50 |
| EXTSUM† | 19.06 | 2.47 | 17.49 | 11.63 | 1.19 | 10.44 | 18.74 | 3.01 | 15.74 |
| Supervised SELSUM† | 21.29 | 4.00 | 19.39 | 14.96 | 2.60 | 13.07 | 24.33 | 5.29 | 18.84 |
| LONGFORMER | 22.40 | 4.71 | 15.36 | 14.68 | 2.53 | 11.62 | 22.56 | 4.83 | 17.08 |
| SELSUM* | 23.17 | 4.77 | 21.13 | 15.12 | 2.83 | 13.07 | 22.87 | 4.85 | 18.05 |
| BRIO | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| Zero-Shot QG | 18.40 | 2.21 | 16.02 | 13.27 | 1.39 | 11.61 | 14.75 | 1.23 | 12.06 |
| CHATGPT | 18.53 | 2.37 | 14.68 | 13.92 | 1.78 | 11.93 | 22.88 | 3.50 | 19.79 |
| Method | R-1 | R-2 | R-L |
| Unsupervised | |||
| W2VCENT† | 13.93 | 2.10 | 10.81 |
| LEXRANK† | 14.88 | 1.94 | 10.50 |
| OPINOSIST | 14.98 | 3.07 | 12.19 |
| MEANSUM† | 15.79 | 1.94 | 12.26 |
| SNCENT† | 15.90 | 2.01 | 11.74 |
| BERTCENT† | 17.65 | 2.78 | 12.78 |
| DENOISESUM† | 21.26 | 4.61 | 16.27 |
| Weakly Supervised | |||
| PLANSUM† | 21.77 | 6.18 | 16.98 |
| Supervised | |||
| BRIO | 23.72 | 5.16 | 18.05 |
| LONGFORMER | 24.96 | 6.34 | 18.40 |
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| Zero-Shot | |||
| QG | 18.14 | 2.34 | 14.28 |
| CHATGPT | 22.73 | 4.21 | 17.52 |
| Summary | Pros | Cons | Verdict | ||||||
| Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | |
| GOLD | 19.7 | 5.7 | 6.4 | 0 | 0 | 4 | 6.9 | 5.2 | 10.2 |
| SELSUM | -20.2 | 2.9 | 11.8 | -6.3 | -0.6 | -0.9 | -7.2 | 6.5 | 0.6 |
| BRIO | -7.5 | -25.6 | -23.5 | -5.1 | -7.3 | -10.7 | -8.4 | -17.2 | -14.7 |
| SUBSUMM | 8 | 17 | 5.3 | 11.4 | 7.9 | 7.6 | 8.6 | 5.5 | 3.9 |
| Method | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| SUBSUMM | 0.66 | 0.67 | 0.89 | 0.70 |
| CHATGPT | 0.34 | 0.33 | 0.11 | 0.30 |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Sampling Strategy | |||||||||
| RAND | 21.83 | 4.28 | 19.91 | 14.06 | 2.55 | 12.35 | 22.67 | 4.66 | 17.81 |
| SENTI-RAND | 22.33 | 4.41 | 20.45 | 14.92 | 2.75 | 13.13 | 23.06 | 4.79 | 17.98 |
| SENTIINFO | 22.82 | 4.60 | 20.80 | 14.42 | 2.48 | 12.59 | 23.51 | 5.03 | 18.16 |
| SENTI-RANDINFO | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| Training Scheme | |||||||||
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| w/o Stage I | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| w/o Stage II | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| RAND in Stage I | 26.03 | 4.93 | 23.96 | 16.66 | 2.97 | 14.78 | 24.71 | 4.93 | 19.43 |
| RAND in Stage II | 25.95 | 4.95 | 24.02 | 16.50 | 2.77 | 14.54 | 24.60 | 4.96 | 19.45 |
| AmaSum | Train | Dev | Test |
| Entity | 26,660 | 3,302 | 3,362 |
| Reviews per Ent. | 77.78 | 77.76 | 76.80 |
| Tokens per Rev. | 53.98 | 54.13 | 53.97 |
| Tokens per Sum. | |||
| - Pros | 39.13 | 39.21 | 38.81 |
| - Cons | 18.34 | 18.31 | 17.81 |
| - Verdict | 21.72 | 21.97 | 21.59 |
| Rotten Tomatoes | Train | Dev | Test |
| Entity | 2,453 | 536 | 737 |
| Reviews per Ent. | 74.39 | 74.85 | 73.26 |
| Tokens per Rev. | 26.52 | 26.32 | 26.41 |
| Tokens per Sum. | 23.80 | 23.59 | 23.82 |
| Hyperparameter | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| bsz I | 16 | 16 | 16 | 8 |
| bsz II | 16 | 16 | 16 | 16 |
| warmup I | 5,000 | 5,000 | 5,000 | 500 |
| warmup II | 3,000 | 3,000 | 3,000 | 300 |
| γ (Eq. 13) | 0.1 | 1.0 | 0.1 | 0.1 |
| lenpen | 0.5 | 0.5 | 1.0 | 1.0 |
| minlen | 35 | 25 | 25 | 30 |
| Value | R-1 | R-2 | R-L |
| K=6 | 21.36 | 4.30 | 15.87 |
| K=8 | 22.16 | 4.89 | 16.62 |
| K=10 | 23.20 | 5.56 | 17.28 |
| K=12 | 22.54 | 5.40 | 16.58 |
| Method | R-1 | R-2 | R-L |
| Sampling Strategy | |||
| RAND | 22.73 | 5.27 | 16.80 |
| SENTI-RAND | 22.76 | 5.36 | 16.81 |
| SENTIINFO | 23.44 | 5.84 | 17.25 |
| SENTI-RANDINFO | 23.99 | 6.09 | 17.54 |
| Training Scheme | |||
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| w/o Stage I | 23.72 | 5.16 | 18.05 |
| w/o Stage II | 23.99 | 6.09 | 17.54 |
| RAND in Stage I | 25.40 | 6.74 | 19.13 |
| RAND in Stage II | 25.46 | 6.77 | 18.90 |
| Method | Pros | Cons | Verdict |
| GOLD | A quality pumice stone crafted from 100% pure volcanic lava. Has a good shape, making it easy to hold and apply pressure to hard skin. Use the handy rope loop to hang the stone out to dry after use. | Some users find the coarse-ness rough and takes time to get used to. | Comfortable to hold and use, this all natural vol- canic pumice stone works hard on calluses and tough skin. |
| SELSUM | Made of natural materi- als that are easy to clean. Comes with a rope at- tached to hang in the shower for easy use. | Some users find it difficult to remove dead skin while it's in use. May not re- move all dead skin on the feet, but it's a good choice for those with dry skin. | A good choice if you're looking for a natural pumice stone that's easy to hold and use. |
| BRIO | A before and after, it was a bit embarrassing, and my feet were a few deep cracks for a few scrubs my feet are gone and a couple of scrubs | Users report that the pumice stone is not as smooth as they would like. Some reports of the stone breaking. | Pumice stone is the real deal and I appreciate that. using it for the first time is a great experience. |
| CHATGPT | Pumice Valley's natural pumice stone is effective in removing dead skin, comes with a handy rope for easy use, and has a fine grain that doesn't rip | The stone is not sharp or rough enough to re- move dry skin, it may be chipped or in poor shape, it can | The pumice stone is effec- tive in removing dead skin and smoothing heels, but may not work well on ex- isting calluses. |
| SUBSUMM | Natural lava stone that's easy to use and easy to clean. Comes with a rope for hanging in the shower. Comes in a variety of sizes. Made from lava. | May be too rough for some users. Some users find it difficult to remove the calluses from the stone. May not remove all cal- luses. | You're looking for a pumice stone that's easy to use and clean, this is the one to buy. |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Unsupervised MEANSUM† | 10.44 | 0.63 | 9.55 | 5.95 | 0.45 | 5.29 | 13.78 | 0.93 | 11.70 |
| LEXRANK† | 14.12 | 1.50 | 12.81 | 8.28 | 0.82 | 7.24 | 15.12 | 1.84 | 12.60 |
| COPYCAT† | 15.12 | 1.48 | 13.85 | 6.81 | 0.82 | 5.89 | 17.05 | 1.78 | 14.50 |
| EXTSUM† | 19.06 | 2.47 | 17.49 | 11.63 | 1.19 | 10.44 | 18.74 | 3.01 | 15.74 |
| Supervised SELSUM† | 21.29 | 4.00 | 19.39 | 14.96 | 2.60 | 13.07 | 24.33 | 5.29 | 18.84 |
| LONGFORMER | 22.40 | 4.71 | 15.36 | 14.68 | 2.53 | 11.62 | 22.56 | 4.83 | 17.08 |
| SELSUM* | 23.17 | 4.77 | 21.13 | 15.12 | 2.83 | 13.07 | 22.87 | 4.85 | 18.05 |
| BRIO | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| Zero-Shot QG | 18.40 | 2.21 | 16.02 | 13.27 | 1.39 | 11.61 | 14.75 | 1.23 | 12.06 |
| CHATGPT | 18.53 | 2.37 | 14.68 | 13.92 | 1.78 | 11.93 | 22.88 | 3.50 | 19.79 |
| Method | R-1 | R-2 | R-L |
| Unsupervised | |||
| W2VCENT† | 13.93 | 2.10 | 10.81 |
| LEXRANK† | 14.88 | 1.94 | 10.50 |
| OPINOSIST | 14.98 | 3.07 | 12.19 |
| MEANSUM† | 15.79 | 1.94 | 12.26 |
| SNCENT† | 15.90 | 2.01 | 11.74 |
| BERTCENT† | 17.65 | 2.78 | 12.78 |
| DENOISESUM† | 21.26 | 4.61 | 16.27 |
| Weakly Supervised | |||
| PLANSUM† | 21.77 | 6.18 | 16.98 |
| Supervised | |||
| BRIO | 23.72 | 5.16 | 18.05 |
| LONGFORMER | 24.96 | 6.34 | 18.40 |
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| Zero-Shot | |||
| QG | 18.14 | 2.34 | 14.28 |
| CHATGPT | 22.73 | 4.21 | 17.52 |
| Summary | Pros | Cons | Verdict | ||||||
| Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | Info↑ | Coh↑ | N-R↑ | |
| GOLD | 19.7 | 5.7 | 6.4 | 0 | 0 | 4 | 6.9 | 5.2 | 10.2 |
| SELSUM | -20.2 | 2.9 | 11.8 | -6.3 | -0.6 | -0.9 | -7.2 | 6.5 | 0.6 |
| BRIO | -7.5 | -25.6 | -23.5 | -5.1 | -7.3 | -10.7 | -8.4 | -17.2 | -14.7 |
| SUBSUMM | 8 | 17 | 5.3 | 11.4 | 7.9 | 7.6 | 8.6 | 5.5 | 3.9 |
| Method | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| SUBSUMM | 0.66 | 0.67 | 0.89 | 0.70 |
| CHATGPT | 0.34 | 0.33 | 0.11 | 0.30 |
| Method | Pros | Cons | Verdict | ||||||
| R-1 | R-2 | R-L | R-1 | R-2 | R-L | R-1 | R-2 | R-L | |
| Sampling Strategy | |||||||||
| RAND | 21.83 | 4.28 | 19.91 | 14.06 | 2.55 | 12.35 | 22.67 | 4.66 | 17.81 |
| SENTI-RAND | 22.33 | 4.41 | 20.45 | 14.92 | 2.75 | 13.13 | 23.06 | 4.79 | 17.98 |
| SENTIINFO | 22.82 | 4.60 | 20.80 | 14.42 | 2.48 | 12.59 | 23.51 | 5.03 | 18.16 |
| SENTI-RANDINFO | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| Training Scheme | |||||||||
| SUBSUMM | 26.25 | 4.96 | 24.18 | 16.72 | 3.00 | 14.80 | 25.36 | 5.04 | 19.58 |
| w/o Stage I | 25.48 | 4.58 | 23.50 | 16.65 | 2.94 | 14.60 | 24.93 | 4.78 | 19.44 |
| w/o Stage II | 23.01 | 4.76 | 21.02 | 15.46 | 2.92 | 13.60 | 23.51 | 5.04 | 18.32 |
| RAND in Stage I | 26.03 | 4.93 | 23.96 | 16.66 | 2.97 | 14.78 | 24.71 | 4.93 | 19.43 |
| RAND in Stage II | 25.95 | 4.95 | 24.02 | 16.50 | 2.77 | 14.54 | 24.60 | 4.96 | 19.45 |
| AmaSum | Train | Dev | Test |
| Entity | 26,660 | 3,302 | 3,362 |
| Reviews per Ent. | 77.78 | 77.76 | 76.80 |
| Tokens per Rev. | 53.98 | 54.13 | 53.97 |
| Tokens per Sum. | |||
| - Pros | 39.13 | 39.21 | 38.81 |
| - Cons | 18.34 | 18.31 | 17.81 |
| - Verdict | 21.72 | 21.97 | 21.59 |
| Rotten Tomatoes | Train | Dev | Test |
| Entity | 2,453 | 536 | 737 |
| Reviews per Ent. | 74.39 | 74.85 | 73.26 |
| Tokens per Rev. | 26.52 | 26.32 | 26.41 |
| Tokens per Sum. | 23.80 | 23.59 | 23.82 |
| Hyperparameter | AmaSum | RT | ||
| Pros | Cons | Verdict | ||
| bsz I | 16 | 16 | 16 | 8 |
| bsz II | 16 | 16 | 16 | 16 |
| warmup I | 5,000 | 5,000 | 5,000 | 500 |
| warmup II | 3,000 | 3,000 | 3,000 | 300 |
| γ (Eq. 13) | 0.1 | 1.0 | 0.1 | 0.1 |
| lenpen | 0.5 | 0.5 | 1.0 | 1.0 |
| minlen | 35 | 25 | 25 | 30 |
| Value | R-1 | R-2 | R-L |
| K=6 | 21.36 | 4.30 | 15.87 |
| K=8 | 22.16 | 4.89 | 16.62 |
| K=10 | 23.20 | 5.56 | 17.28 |
| K=12 | 22.54 | 5.40 | 16.58 |
| Method | R-1 | R-2 | R-L |
| Sampling Strategy | |||
| RAND | 22.73 | 5.27 | 16.80 |
| SENTI-RAND | 22.76 | 5.36 | 16.81 |
| SENTIINFO | 23.44 | 5.84 | 17.25 |
| SENTI-RANDINFO | 23.99 | 6.09 | 17.54 |
| Training Scheme | |||
| SUBSUMM | 24.96 | 6.66 | 19.08 |
| w/o Stage I | 23.72 | 5.16 | 18.05 |
| w/o Stage II | 23.99 | 6.09 | 17.54 |
| RAND in Stage I | 25.40 | 6.74 | 19.13 |
| RAND in Stage II | 25.46 | 6.77 | 18.90 |
| Method | Pros | Cons | Verdict |
| GOLD | A quality pumice stone crafted from 100% pure volcanic lava. Has a good shape, making it easy to hold and apply pressure to hard skin. Use the handy rope loop to hang the stone out to dry after use. | Some users find the coarse-ness rough and takes time to get used to. | Comfortable to hold and use, this all natural vol- canic pumice stone works hard on calluses and tough skin. |
| SELSUM | Made of natural materi- als that are easy to clean. Comes with a rope at- tached to hang in the shower for easy use. | Some users find it difficult to remove dead skin while it's in use. May not re- move all dead skin on the feet, but it's a good choice for those with dry skin. | A good choice if you're looking for a natural pumice stone that's easy to hold and use. |
| BRIO | A before and after, it was a bit embarrassing, and my feet were a few deep cracks for a few scrubs my feet are gone and a couple of scrubs | Users report that the pumice stone is not as smooth as they would like. Some reports of the stone breaking. | Pumice stone is the real deal and I appreciate that. using it for the first time is a great experience. |
| CHATGPT | Pumice Valley's natural pumice stone is effective in removing dead skin, comes with a handy rope for easy use, and has a fine grain that doesn't rip | The stone is not sharp or rough enough to re- move dry skin, it may be chipped or in poor shape, it can | The pumice stone is effec- tive in removing dead skin and smoothing heels, but may not work well on ex- isting calluses. |
| SUBSUMM | Natural lava stone that's easy to use and easy to clean. Comes with a rope for hanging in the shower. Comes in a variety of sizes. Made from lava. | May be too rough for some users. Some users find it difficult to remove the calluses from the stone. May not remove all cal- luses. | You're looking for a pumice stone that's easy to use and clean, this is the one to buy. |
| Hindi | Romanized |
| न्दय (new) | naya nga |
| रैशटम (end) | khatam khtm khatm |
| बस (sufficient) | bas bs bus |
| Sentiment | En-Es | En-Hi |
| Positive | 433 | 3202 |
| Negative | 379 | 2319 |
| Neutral | 711 | 4559 |
| Total | 1523 | 10080 |
| Sentiment | En-Es | En-Hi |
| Positive | 55 | 399 |
| Negative | 46 | 283 |
| Neutral | 83 | 578 |
| Total | 184 | 1260 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 16 | 5 |
| mBERT | 4e-5 | 1e-8 | 16 | 5 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 32 | 3 |
| mBERT | 4e-5 | 1e-8 | 32 | 3 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Hi-BERT | 5e-5 | 1e-8 | 32 | 3 |
| Model | F1 |
| mBERT (Khanuja et al., 2020) | 63.02% |
| BERT | 65.05% |
| XLM-R | 69.42% |
| mBERT - BERT (Late Fusion) | 69.40% |
| XLM-R - BERT (Late Fusion) | 70.9% |
| XLM-R - mBERT (Late Fusion) | 73.66% |
| Model | F1 |
| XLM-R | 28.5% |
| Hi-BERT | 55.90% |
| mBERT (Khanuja et al., 2020) | 57.51% |
| BERT | 59.05% |
| BERT - Hi-BERT (Late Fusion) | 59.73% |
| mBERT - Hi-BERT (Late Fusion) | 59.97% |
| mBERT - BERT (Late Fusion) | 61.24% |
| Hindi | Romanized |
| न्दय (new) | naya nga |
| रैशटम (end) | khatam khtm khatm |
| बस (sufficient) | bas bs bus |
| Sentiment | En-Es | En-Hi |
| Positive | 433 | 3202 |
| Negative | 379 | 2319 |
| Neutral | 711 | 4559 |
| Total | 1523 | 10080 |
| Sentiment | En-Es | En-Hi |
| Positive | 55 | 399 |
| Negative | 46 | 283 |
| Neutral | 83 | 578 |
| Total | 184 | 1260 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 16 | 5 |
| mBERT | 4e-5 | 1e-8 | 16 | 5 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 32 | 3 |
| mBERT | 4e-5 | 1e-8 | 32 | 3 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Hi-BERT | 5e-5 | 1e-8 | 32 | 3 |
| Model | F1 |
| mBERT (Khanuja et al., 2020) | 63.02% |
| BERT | 65.05% |
| XLM-R | 69.42% |
| mBERT - BERT (Late Fusion) | 69.40% |
| XLM-R - BERT (Late Fusion) | 70.9% |
| XLM-R - mBERT (Late Fusion) | 73.66% |
| Model | F1 |
| XLM-R | 28.5% |
| Hi-BERT | 55.90% |
| mBERT (Khanuja et al., 2020) | 57.51% |
| BERT | 59.05% |
| BERT - Hi-BERT (Late Fusion) | 59.73% |
| mBERT - Hi-BERT (Late Fusion) | 59.97% |
| mBERT - BERT (Late Fusion) | 61.24% |
| Hindi | Romanized |
| न्दय (new) | naya nga |
| रैशटम (end) | khatam khtm khatm |
| बस (sufficient) | bas bs bus |
| Sentiment | En-Es | En-Hi |
| Positive | 433 | 3202 |
| Negative | 379 | 2319 |
| Neutral | 711 | 4559 |
| Total | 1523 | 10080 |
| Sentiment | En-Es | En-Hi |
| Positive | 55 | 399 |
| Negative | 46 | 283 |
| Neutral | 83 | 578 |
| Total | 184 | 1260 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 16 | 5 |
| mBERT | 4e-5 | 1e-8 | 16 | 5 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 32 | 3 |
| mBERT | 4e-5 | 1e-8 | 32 | 3 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Hi-BERT | 5e-5 | 1e-8 | 32 | 3 |
| Model | F1 |
| mBERT (Khanuja et al., 2020) | 63.02% |
| BERT | 65.05% |
| XLM-R | 69.42% |
| mBERT - BERT (Late Fusion) | 69.40% |
| XLM-R - BERT (Late Fusion) | 70.9% |
| XLM-R - mBERT (Late Fusion) | 73.66% |
| Model | F1 |
| XLM-R | 28.5% |
| Hi-BERT | 55.90% |
| mBERT (Khanuja et al., 2020) | 57.51% |
| BERT | 59.05% |
| BERT - Hi-BERT (Late Fusion) | 59.73% |
| mBERT - Hi-BERT (Late Fusion) | 59.97% |
| mBERT - BERT (Late Fusion) | 61.24% |
| Hindi | Romanized |
| न्दय (new) | naya nga |
| रैशटम (end) | khatam khtm khatm |
| बस (sufficient) | bas bs bus |
| Sentiment | En-Es | En-Hi |
| Positive | 433 | 3202 |
| Negative | 379 | 2319 |
| Neutral | 711 | 4559 |
| Total | 1523 | 10080 |
| Sentiment | En-Es | En-Hi |
| Positive | 55 | 399 |
| Negative | 46 | 283 |
| Neutral | 83 | 578 |
| Total | 184 | 1260 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 16 | 5 |
| mBERT | 4e-5 | 1e-8 | 16 | 5 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Model | gain | epsilon | batch | epoch |
| BERT | 5e-5 | 1e-8 | 32 | 3 |
| mBERT | 4e-5 | 1e-8 | 32 | 3 |
| XLM-R | 3e-5 | 1e-8 | 16 | 5 |
| Hi-BERT | 5e-5 | 1e-8 | 32 | 3 |
| Model | F1 |
| mBERT (Khanuja et al., 2020) | 63.02% |
| BERT | 65.05% |
| XLM-R | 69.42% |
| mBERT - BERT (Late Fusion) | 69.40% |
| XLM-R - BERT (Late Fusion) | 70.9% |
| XLM-R - mBERT (Late Fusion) | 73.66% |
| Model | F1 |
| XLM-R | 28.5% |
| Hi-BERT | 55.90% |
| mBERT (Khanuja et al., 2020) | 57.51% |
| BERT | 59.05% |
| BERT - Hi-BERT (Late Fusion) | 59.73% |
| mBERT - Hi-BERT (Late Fusion) | 59.97% |
| mBERT - BERT (Late Fusion) | 61.24% |
| Dataset | ReadingBank | DITrans | ||
| Domain | General | Political Report | Newspaper | Advertisement |
| # Page | 410K | 902 | 396 | 377 |
| # Word Avg. | 196 | 245 | 219 | 123 |
| # Sent. Avg. | 21 | 24 | 16 | 16 |
| Train/Test (# Page) | 400K/10K | 790/112 | 347/49 | 330/47 |
| Train/Test (# Sent.) | 8,467K/211K | 19,210/2,698 | 5,501/929 | 5,303/815 |
| # | Method | End-to-End | # Params (M) | BLEU |
| 1 | DocHandler-1(Sable et al., 2023) | - | 142 | 30.47 |
| 2 | DocHandler-2(Sable et al., 2023) | - | 172 | 37.75 |
| 3 | LayoutLM-Dec | ✓ | 232 | 45.54 |
| 4 | LiLT-Dec | ✓ | 250 | 45.79 |
| 5 | LayoutDIT-Cascade | - | 293 | 48.29 |
| 6 | LayoutDIT | ✓ | 206 | 48.20 |
| # | Method | End-to-End | # Params (M) | Multi-Domain Learning | Zero-Shot Cross-Domain Transfer | ||||
| Political Report | Newspaper | Advertisement | Political Report (SD) | Newspaper (TD.1) | Advertisement (TD.2) | ||||
| 1 | DocHandler-1 (Sable et al., 2023) | - | 142 | 21.47 | 26.28 | 23.14 | 21.80 | 16.87 | 13.55 |
| 2 | DocHandler-2 (Sable et al., 2023) | - | 172 | 24.80 | 27.34 | 24.99 | 25.19 | 16.88 | 13.60 |
| 3 | MGTrans-DETR (Hinami et al., 2021) | - | 212 | 28.30 | 27.61 | 25.76 | 27.34 | 16.26 | 13.09 |
| 4 | MGTrans-Conv (Hinami et al., 2021) | - | 238 | 28.99 | 28.84 | 25.63 | 29.51 | 17.78 | 14.07 |
| 5 | LayoutLM-Dec | ✓ | 232 | 45.86 | 40.66 | 36.90 | 45.71 | 27.56 | 21.40 |
| 6 | LiLT-Dec | ✓ | 250 | 46.11 | 40.37 | 36.22 | 46.61 | 23.71 | 18.14 |
| 7 | LayoutDIT-Cascade | - | 293 | 36.78 | 34.56 | 35.36 | 39.85 | 21.89 | 16.62 |
| 8 | LayoutDIT | ✓ | 206 | 46.97 | 38.82 | 45.72 | 46.70 | 29.01 | 29.46 |
| # | Model | BLEU |
| 1 | LayoutDIT | 48.20 |
| 2a | w/o 2D layout Embedding | 43.87 |
| 2b | w/o 1D position Embedding | 39.40 |
| 3a | w/o reading order decoding loss | 3.15 |
| 3b | w/o decoding processes decomposition | 13.99 |
| 4 | w/ SSD's hidden states as context memory | 45.85 |
| Method | ReadingBank En-De | ReadingBank En-Fr |
| DocHandler-2 | 17.15 | 18.23 |
| LiLT-Dec | 30.60 | 36.00 |
| LayoutDIT | 33.25 | 40.09 |
| Dataset | ReadingBank | DITrans | ||
| Domain | General | Political Report | Newspaper | Advertisement |
| # Page | 410K | 902 | 396 | 377 |
| # Word Avg. | 196 | 245 | 219 | 123 |
| # Sent. Avg. | 21 | 24 | 16 | 16 |
| Train/Test (# Page) | 400K/10K | 790/112 | 347/49 | 330/47 |
| Train/Test (# Sent.) | 8,467K/211K | 19,210/2,698 | 5,501/929 | 5,303/815 |
| # | Method | End-to-End | # Params (M) | BLEU |
| 1 | DocHandler-1(Sable et al., 2023) | - | 142 | 30.47 |
| 2 | DocHandler-2(Sable et al., 2023) | - | 172 | 37.75 |
| 3 | LayoutLM-Dec | ✓ | 232 | 45.54 |
| 4 | LiLT-Dec | ✓ | 250 | 45.79 |
| 5 | LayoutDIT-Cascade | - | 293 | 48.29 |
| 6 | LayoutDIT | ✓ | 206 | 48.20 |
| # | Method | End-to-End | # Params (M) | Multi-Domain Learning | Zero-Shot Cross-Domain Transfer | ||||
| Political Report | Newspaper | Advertisement | Political Report (SD) | Newspaper (TD.1) | Advertisement (TD.2) | ||||
| 1 | DocHandler-1 (Sable et al., 2023) | - | 142 | 21.47 | 26.28 | 23.14 | 21.80 | 16.87 | 13.55 |
| 2 | DocHandler-2 (Sable et al., 2023) | - | 172 | 24.80 | 27.34 | 24.99 | 25.19 | 16.88 | 13.60 |
| 3 | MGTrans-DETR (Hinami et al., 2021) | - | 212 | 28.30 | 27.61 | 25.76 | 27.34 | 16.26 | 13.09 |
| 4 | MGTrans-Conv (Hinami et al., 2021) | - | 238 | 28.99 | 28.84 | 25.63 | 29.51 | 17.78 | 14.07 |
| 5 | LayoutLM-Dec | ✓ | 232 | 45.86 | 40.66 | 36.90 | 45.71 | 27.56 | 21.40 |
| 6 | LiLT-Dec | ✓ | 250 | 46.11 | 40.37 | 36.22 | 46.61 | 23.71 | 18.14 |
| 7 | LayoutDIT-Cascade | - | 293 | 36.78 | 34.56 | 35.36 | 39.85 | 21.89 | 16.62 |
| 8 | LayoutDIT | ✓ | 206 | 46.97 | 38.82 | 45.72 | 46.70 | 29.01 | 29.46 |
| # | Model | BLEU |
| 1 | LayoutDIT | 48.20 |
| 2a | w/o 2D layout Embedding | 43.87 |
| 2b | w/o 1D position Embedding | 39.40 |
| 3a | w/o reading order decoding loss | 3.15 |
| 3b | w/o decoding processes decomposition | 13.99 |
| 4 | w/ SSD's hidden states as context memory | 45.85 |
| Method | ReadingBank En-De | ReadingBank En-Fr |
| DocHandler-2 | 17.15 | 18.23 |
| LiLT-Dec | 30.60 | 36.00 |
| LayoutDIT | 33.25 | 40.09 |
| Dataset | ReadingBank | DITrans | ||
| Domain | General | Political Report | Newspaper | Advertisement |
| # Page | 410K | 902 | 396 | 377 |
| # Word Avg. | 196 | 245 | 219 | 123 |
| # Sent. Avg. | 21 | 24 | 16 | 16 |
| Train/Test (# Page) | 400K/10K | 790/112 | 347/49 | 330/47 |
| Train/Test (# Sent.) | 8,467K/211K | 19,210/2,698 | 5,501/929 | 5,303/815 |
| # | Method | End-to-End | # Params (M) | BLEU |
| 1 | DocHandler-1(Sable et al., 2023) | - | 142 | 30.47 |
| 2 | DocHandler-2(Sable et al., 2023) | - | 172 | 37.75 |
| 3 | LayoutLM-Dec | ✓ | 232 | 45.54 |
| 4 | LiLT-Dec | ✓ | 250 | 45.79 |
| 5 | LayoutDIT-Cascade | - | 293 | 48.29 |
| 6 | LayoutDIT | ✓ | 206 | 48.20 |
| # | Method | End-to-End | # Params (M) | Multi-Domain Learning | Zero-Shot Cross-Domain Transfer | ||||
| Political Report | Newspaper | Advertisement | Political Report (SD) | Newspaper (TD.1) | Advertisement (TD.2) | ||||
| 1 | DocHandler-1 (Sable et al., 2023) | - | 142 | 21.47 | 26.28 | 23.14 | 21.80 | 16.87 | 13.55 |
| 2 | DocHandler-2 (Sable et al., 2023) | - | 172 | 24.80 | 27.34 | 24.99 | 25.19 | 16.88 | 13.60 |
| 3 | MGTrans-DETR (Hinami et al., 2021) | - | 212 | 28.30 | 27.61 | 25.76 | 27.34 | 16.26 | 13.09 |
| 4 | MGTrans-Conv (Hinami et al., 2021) | - | 238 | 28.99 | 28.84 | 25.63 | 29.51 | 17.78 | 14.07 |
| 5 | LayoutLM-Dec | ✓ | 232 | 45.86 | 40.66 | 36.90 | 45.71 | 27.56 | 21.40 |
| 6 | LiLT-Dec | ✓ | 250 | 46.11 | 40.37 | 36.22 | 46.61 | 23.71 | 18.14 |
| 7 | LayoutDIT-Cascade | - | 293 | 36.78 | 34.56 | 35.36 | 39.85 | 21.89 | 16.62 |
| 8 | LayoutDIT | ✓ | 206 | 46.97 | 38.82 | 45.72 | 46.70 | 29.01 | 29.46 |
| # | Model | BLEU |
| 1 | LayoutDIT | 48.20 |
| 2a | w/o 2D layout Embedding | 43.87 |
| 2b | w/o 1D position Embedding | 39.40 |
| 3a | w/o reading order decoding loss | 3.15 |
| 3b | w/o decoding processes decomposition | 13.99 |
| 4 | w/ SSD's hidden states as context memory | 45.85 |
| Method | ReadingBank En-De | ReadingBank En-Fr |
| DocHandler-2 | 17.15 | 18.23 |
| LiLT-Dec | 30.60 | 36.00 |
| LayoutDIT | 33.25 | 40.09 |
| Dataset | ReadingBank | DITrans | ||
| Domain | General | Political Report | Newspaper | Advertisement |
| # Page | 410K | 902 | 396 | 377 |
| # Word Avg. | 196 | 245 | 219 | 123 |
| # Sent. Avg. | 21 | 24 | 16 | 16 |
| Train/Test (# Page) | 400K/10K | 790/112 | 347/49 | 330/47 |
| Train/Test (# Sent.) | 8,467K/211K | 19,210/2,698 | 5,501/929 | 5,303/815 |
| # | Method | End-to-End | # Params (M) | BLEU |
| 1 | DocHandler-1(Sable et al., 2023) | - | 142 | 30.47 |
| 2 | DocHandler-2(Sable et al., 2023) | - | 172 | 37.75 |
| 3 | LayoutLM-Dec | ✓ | 232 | 45.54 |
| 4 | LiLT-Dec | ✓ | 250 | 45.79 |
| 5 | LayoutDIT-Cascade | - | 293 | 48.29 |
| 6 | LayoutDIT | ✓ | 206 | 48.20 |
| # | Method | End-to-End | # Params (M) | Multi-Domain Learning | Zero-Shot Cross-Domain Transfer | ||||
| Political Report | Newspaper | Advertisement | Political Report (SD) | Newspaper (TD.1) | Advertisement (TD.2) | ||||
| 1 | DocHandler-1 (Sable et al., 2023) | - | 142 | 21.47 | 26.28 | 23.14 | 21.80 | 16.87 | 13.55 |
| 2 | DocHandler-2 (Sable et al., 2023) | - | 172 | 24.80 | 27.34 | 24.99 | 25.19 | 16.88 | 13.60 |
| 3 | MGTrans-DETR (Hinami et al., 2021) | - | 212 | 28.30 | 27.61 | 25.76 | 27.34 | 16.26 | 13.09 |
| 4 | MGTrans-Conv (Hinami et al., 2021) | - | 238 | 28.99 | 28.84 | 25.63 | 29.51 | 17.78 | 14.07 |
| 5 | LayoutLM-Dec | ✓ | 232 | 45.86 | 40.66 | 36.90 | 45.71 | 27.56 | 21.40 |
| 6 | LiLT-Dec | ✓ | 250 | 46.11 | 40.37 | 36.22 | 46.61 | 23.71 | 18.14 |
| 7 | LayoutDIT-Cascade | - | 293 | 36.78 | 34.56 | 35.36 | 39.85 | 21.89 | 16.62 |
| 8 | LayoutDIT | ✓ | 206 | 46.97 | 38.82 | 45.72 | 46.70 | 29.01 | 29.46 |
| # | Model | BLEU |
| 1 | LayoutDIT | 48.20 |
| 2a | w/o 2D layout Embedding | 43.87 |
| 2b | w/o 1D position Embedding | 39.40 |
| 3a | w/o reading order decoding loss | 3.15 |
| 3b | w/o decoding processes decomposition | 13.99 |
| 4 | w/ SSD's hidden states as context memory | 45.85 |
| Method | ReadingBank En-De | ReadingBank En-Fr |
| DocHandler-2 | 17.15 | 18.23 |
| LiLT-Dec | 30.60 | 36.00 |
| LayoutDIT | 33.25 | 40.09 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 77.43±0.22/85.67±0.10 | 93.19±0.09 |
| Distilbert6-cosD | 77.82±0.29/85.92±0.23 | 93.66±0.08 |
| Distilbert6-MSE | 77.72±0.21/85.86±0.12 | 93.49±0.08 |
| Distilbert6-ours | 78.79±0.12/86.95±0.06 | 93.87±0.01 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 74.32±0.37/83.69±0.31 | 92.71±0.08 |
| Distilbert6-cosD | 75.78±0.27/84.57±0.11 | 93.90±0.08 |
| Distilbert6-MSE | 75.17±0.18/84.25±0.15 | 93.82±0.07 |
| Distilbert6-ours | 79.46±0.08/87.48±0.05 | 93.96±0.01 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Distilbert6-FT | 84.48 | 56.17 | 44.28 | 89.54 | 85.40 | 80.41 | 86.71 | 87.79 | 76.85 |
| Distilbert6-CosD | 85.38 | 63.75 | 46.82 | 89.79 | 85.59 | 80.51 | 88.62 | 87.78 | 78.53 |
| Distilbert6-MSE | 86.04 | 62.38 | 47.35 | 89.79 | 85.42 | 80.12 | 88.13 | 87.90 | 78.39 |
| Distilbert6-PKD | 86.06 | 62.24 | 47.28 | 90.05 | 85.77 | 81.62 | 87.16 | 88.33 | 78.56 |
| Distilbert6-KD | 84.56 | 56.53 | 45.86 | 89.63 | 85.50 | 80.51 | 87.71 | 87.60 | 77.23 |
| Distilbert6-ours | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 82.76 | 89.54 | 88.04 | 79.70 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Ours | 86.03 | 64.19 | 53.21 | 89.91 | 85.88 | 82.29 | 89.09 | 88.28 | 79.84 |
| Ours-Masked(80%) | 85.79 | 64.55 | 53.16 | 90.18 | 85.94 | 82.41 | 89.44 | 88.32 | 79.97 |
| Ours-Masked(70%) | 86.18 | 64.84 | 52.71 | 89.63 | 85.97 | 82.28 | 89.37 | 88.32 | 79.91 |
| Ours-Masked(Best) | 86.45 | 65.05 | 53.16 | 90.18 | 85.97 | 82.42 | 89.44 | 88.37 | 80.13 |
| approach | MRPC | RTE | CoLA | SST-2 | STS-B | QNLI | MNLI-m | QQP | AVG |
| LCE+l | 86.85 | 63.61 | 48.24 | 90.87 | 85.47 | 89.63 | 81.85 | 87.60 | 79.26 |
| LCE+R | 75.85 | 53.79 | 34.71 | 85.25 | 85.25 | 85.66 | 74.43 | 87.69 | 68.49 |
| LCE+R+l | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 89.54 | 82.76 | 88.04 | 79.70 |
| approach | QNLI(%) | RTE(%) |
| SRRL | 88.04 | 63.68 |
| SRRL-Ours | 88.68 | 64.10 |
| approach | MNLI-m (H) | QNLI(H) | SST-2(H) |
| PKD | 3.00 | 0.73 | 0.53 |
| Ours | 2.13 | 0.63 | 0.47 |
| Batch-size | MNLI-m (%) | QNLI (%) | SST-2 (%) | RTE (%) |
| 4 | 82.41 | 89.43 | 90.18 | 63.24 |
| 8 | 82.54 | 89.53 | 90.44 | 65.85 |
| 16 | 82.76 | 89.54 | 90.44 | 63.83 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 77.43±0.22/85.67±0.10 | 93.19±0.09 |
| Distilbert6-cosD | 77.82±0.29/85.92±0.23 | 93.66±0.08 |
| Distilbert6-MSE | 77.72±0.21/85.86±0.12 | 93.49±0.08 |
| Distilbert6-ours | 78.79±0.12/86.95±0.06 | 93.87±0.01 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 74.32±0.37/83.69±0.31 | 92.71±0.08 |
| Distilbert6-cosD | 75.78±0.27/84.57±0.11 | 93.90±0.08 |
| Distilbert6-MSE | 75.17±0.18/84.25±0.15 | 93.82±0.07 |
| Distilbert6-ours | 79.46±0.08/87.48±0.05 | 93.96±0.01 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Distilbert6-FT | 84.48 | 56.17 | 44.28 | 89.54 | 85.40 | 80.41 | 86.71 | 87.79 | 76.85 |
| Distilbert6-CosD | 85.38 | 63.75 | 46.82 | 89.79 | 85.59 | 80.51 | 88.62 | 87.78 | 78.53 |
| Distilbert6-MSE | 86.04 | 62.38 | 47.35 | 89.79 | 85.42 | 80.12 | 88.13 | 87.90 | 78.39 |
| Distilbert6-PKD | 86.06 | 62.24 | 47.28 | 90.05 | 85.77 | 81.62 | 87.16 | 88.33 | 78.56 |
| Distilbert6-KD | 84.56 | 56.53 | 45.86 | 89.63 | 85.50 | 80.51 | 87.71 | 87.60 | 77.23 |
| Distilbert6-ours | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 82.76 | 89.54 | 88.04 | 79.70 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Ours | 86.03 | 64.19 | 53.21 | 89.91 | 85.88 | 82.29 | 89.09 | 88.28 | 79.84 |
| Ours-Masked(80%) | 85.79 | 64.55 | 53.16 | 90.18 | 85.94 | 82.41 | 89.44 | 88.32 | 79.97 |
| Ours-Masked(70%) | 86.18 | 64.84 | 52.71 | 89.63 | 85.97 | 82.28 | 89.37 | 88.32 | 79.91 |
| Ours-Masked(Best) | 86.45 | 65.05 | 53.16 | 90.18 | 85.97 | 82.42 | 89.44 | 88.37 | 80.13 |
| approach | MRPC | RTE | CoLA | SST-2 | STS-B | QNLI | MNLI-m | QQP | AVG |
| LCE+l | 86.85 | 63.61 | 48.24 | 90.87 | 85.47 | 89.63 | 81.85 | 87.60 | 79.26 |
| LCE+R | 75.85 | 53.79 | 34.71 | 85.25 | 85.25 | 85.66 | 74.43 | 87.69 | 68.49 |
| LCE+R+l | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 89.54 | 82.76 | 88.04 | 79.70 |
| approach | QNLI(%) | RTE(%) |
| SRRL | 88.04 | 63.68 |
| SRRL-Ours | 88.68 | 64.10 |
| approach | MNLI-m (H) | QNLI(H) | SST-2(H) |
| PKD | 3.00 | 0.73 | 0.53 |
| Ours | 2.13 | 0.63 | 0.47 |
| Batch-size | MNLI-m (%) | QNLI (%) | SST-2 (%) | RTE (%) |
| 4 | 82.41 | 89.43 | 90.18 | 63.24 |
| 8 | 82.54 | 89.53 | 90.44 | 65.85 |
| 16 | 82.76 | 89.54 | 90.44 | 63.83 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 77.43±0.22/85.67±0.10 | 93.19±0.09 |
| Distilbert6-cosD | 77.82±0.29/85.92±0.23 | 93.66±0.08 |
| Distilbert6-MSE | 77.72±0.21/85.86±0.12 | 93.49±0.08 |
| Distilbert6-ours | 78.79±0.12/86.95±0.06 | 93.87±0.01 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 74.32±0.37/83.69±0.31 | 92.71±0.08 |
| Distilbert6-cosD | 75.78±0.27/84.57±0.11 | 93.90±0.08 |
| Distilbert6-MSE | 75.17±0.18/84.25±0.15 | 93.82±0.07 |
| Distilbert6-ours | 79.46±0.08/87.48±0.05 | 93.96±0.01 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Distilbert6-FT | 84.48 | 56.17 | 44.28 | 89.54 | 85.40 | 80.41 | 86.71 | 87.79 | 76.85 |
| Distilbert6-CosD | 85.38 | 63.75 | 46.82 | 89.79 | 85.59 | 80.51 | 88.62 | 87.78 | 78.53 |
| Distilbert6-MSE | 86.04 | 62.38 | 47.35 | 89.79 | 85.42 | 80.12 | 88.13 | 87.90 | 78.39 |
| Distilbert6-PKD | 86.06 | 62.24 | 47.28 | 90.05 | 85.77 | 81.62 | 87.16 | 88.33 | 78.56 |
| Distilbert6-KD | 84.56 | 56.53 | 45.86 | 89.63 | 85.50 | 80.51 | 87.71 | 87.60 | 77.23 |
| Distilbert6-ours | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 82.76 | 89.54 | 88.04 | 79.70 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Ours | 86.03 | 64.19 | 53.21 | 89.91 | 85.88 | 82.29 | 89.09 | 88.28 | 79.84 |
| Ours-Masked(80%) | 85.79 | 64.55 | 53.16 | 90.18 | 85.94 | 82.41 | 89.44 | 88.32 | 79.97 |
| Ours-Masked(70%) | 86.18 | 64.84 | 52.71 | 89.63 | 85.97 | 82.28 | 89.37 | 88.32 | 79.91 |
| Ours-Masked(Best) | 86.45 | 65.05 | 53.16 | 90.18 | 85.97 | 82.42 | 89.44 | 88.37 | 80.13 |
| approach | MRPC | RTE | CoLA | SST-2 | STS-B | QNLI | MNLI-m | QQP | AVG |
| LCE+l | 86.85 | 63.61 | 48.24 | 90.87 | 85.47 | 89.63 | 81.85 | 87.60 | 79.26 |
| LCE+R | 75.85 | 53.79 | 34.71 | 85.25 | 85.25 | 85.66 | 74.43 | 87.69 | 68.49 |
| LCE+R+l | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 89.54 | 82.76 | 88.04 | 79.70 |
| approach | QNLI(%) | RTE(%) |
| SRRL | 88.04 | 63.68 |
| SRRL-Ours | 88.68 | 64.10 |
| approach | MNLI-m (H) | QNLI(H) | SST-2(H) |
| PKD | 3.00 | 0.73 | 0.53 |
| Ours | 2.13 | 0.63 | 0.47 |
| Batch-size | MNLI-m (%) | QNLI (%) | SST-2 (%) | RTE (%) |
| 4 | 82.41 | 89.43 | 90.18 | 63.24 |
| 8 | 82.54 | 89.53 | 90.44 | 65.85 |
| 16 | 82.76 | 89.54 | 90.44 | 63.83 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 77.43±0.22/85.67±0.10 | 93.19±0.09 |
| Distilbert6-cosD | 77.82±0.29/85.92±0.23 | 93.66±0.08 |
| Distilbert6-MSE | 77.72±0.21/85.86±0.12 | 93.49±0.08 |
| Distilbert6-ours | 78.79±0.12/86.95±0.06 | 93.87±0.01 |
| Approach | SQUAD-V1 (%) | IMDB (%) |
| BERT12 (teacher) | 80.36/88.13 | 94.06 |
| Distilbert6-FT | 74.32±0.37/83.69±0.31 | 92.71±0.08 |
| Distilbert6-cosD | 75.78±0.27/84.57±0.11 | 93.90±0.08 |
| Distilbert6-MSE | 75.17±0.18/84.25±0.15 | 93.82±0.07 |
| Distilbert6-ours | 79.46±0.08/87.48±0.05 | 93.96±0.01 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Distilbert6-FT | 84.48 | 56.17 | 44.28 | 89.54 | 85.40 | 80.41 | 86.71 | 87.79 | 76.85 |
| Distilbert6-CosD | 85.38 | 63.75 | 46.82 | 89.79 | 85.59 | 80.51 | 88.62 | 87.78 | 78.53 |
| Distilbert6-MSE | 86.04 | 62.38 | 47.35 | 89.79 | 85.42 | 80.12 | 88.13 | 87.90 | 78.39 |
| Distilbert6-PKD | 86.06 | 62.24 | 47.28 | 90.05 | 85.77 | 81.62 | 87.16 | 88.33 | 78.56 |
| Distilbert6-KD | 84.56 | 56.53 | 45.86 | 89.63 | 85.50 | 80.51 | 87.71 | 87.60 | 77.23 |
| Distilbert6-ours | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 82.76 | 89.54 | 88.04 | 79.70 |
| Approach | MRPC | RTE | CoLA | SST-2 | STS-B | MNLI-m | QNLI | QQP | AVG |
| BERT12(teacher) | 87.86 | 66.79 | 54.84 | 90.02 | 89.30 | 82.85 | 90.70 | 88.04 | 81.30 |
| Ours | 86.03 | 64.19 | 53.21 | 89.91 | 85.88 | 82.29 | 89.09 | 88.28 | 79.84 |
| Ours-Masked(80%) | 85.79 | 64.55 | 53.16 | 90.18 | 85.94 | 82.41 | 89.44 | 88.32 | 79.97 |
| Ours-Masked(70%) | 86.18 | 64.84 | 52.71 | 89.63 | 85.97 | 82.28 | 89.37 | 88.32 | 79.91 |
| Ours-Masked(Best) | 86.45 | 65.05 | 53.16 | 90.18 | 85.97 | 82.42 | 89.44 | 88.37 | 80.13 |
| approach | MRPC | RTE | CoLA | SST-2 | STS-B | QNLI | MNLI-m | QQP | AVG |
| LCE+l | 86.85 | 63.61 | 48.24 | 90.87 | 85.47 | 89.63 | 81.85 | 87.60 | 79.26 |
| LCE+R | 75.85 | 53.79 | 34.71 | 85.25 | 85.25 | 85.66 | 74.43 | 87.69 | 68.49 |
| LCE+R+l | 86.57 | 63.83 | 50.73 | 90.44 | 85.66 | 89.54 | 82.76 | 88.04 | 79.70 |
| approach | QNLI(%) | RTE(%) |
| SRRL | 88.04 | 63.68 |
| SRRL-Ours | 88.68 | 64.10 |
| approach | MNLI-m (H) | QNLI(H) | SST-2(H) |
| PKD | 3.00 | 0.73 | 0.53 |
| Ours | 2.13 | 0.63 | 0.47 |
| Batch-size | MNLI-m (%) | QNLI (%) | SST-2 (%) | RTE (%) |
| 4 | 82.41 | 89.43 | 90.18 | 63.24 |
| 8 | 82.54 | 89.53 | 90.44 | 65.85 |
| 16 | 82.76 | 89.54 | 90.44 | 63.83 |
| Efficiency | Expressivity | Accuracy | |
| Subword \nByte/ \nChar | High | Low | Mid |
| Low | High | Low | |
| eByte/ \neChar | Mid | Mid | High |
| Dataset | Size (MBs) | Words (Mil.) | Chars/Word |
| English | 4.7 | 1.34 | 5.46 |
| French | 5.1 | 1.55 | 5.18 |
| Russian | 7.5 | 1.18 | 6.39 |
| Numeracy | 6.6 | 1.35 | 5.09 |
| Tokenizer | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) |
| Language | English | French | Russian | ||||||
| Subword | 14.37 | 0.55 | 76.8 | 41.20 | 1.50 | 76.8 | 8.31 | 1.49 | 76.8 |
| Byte | 13.69 | 0.53 | 25.7 | 17.39 | 0.54 | 25.7 | 12.76 | 0.53 | 25.7 |
| Char | 13.68 | 0.54 | 26.3 | 16.95 | 0.53 | 25.7 | 10.01 | 0.54 | 26.1 |
| eByte | 44.17 | 3.84 | 38.7 | 46.44 | 6.01 | 38.7 | 35.00 | 4.92 | 38.7 |
| eChar | 42.94 | 2.94 | 39.2 | 47.06 | 3.59 | 38.7 | 37.15 | 3.95 | 39.0 |
| Lang | Tok | 1 CLS | 4 CLS | Δ% | Δ Mem |
| en | eByte | 31 | 44 | 42% | 0.02 |
| en | eChar | 29 | 43 | 48% | 0.02 |
| fr | eByte | 31 | 46 | 48% | 0.02 |
| fr | eChar | 34 | 47 | 38% | 0.02 |
| ru | eByte | 26 | 35 | 35% | 0.02 |
| ru | eChar | 29 | 37 | 28% | 0.02 |
| Tokenizer | Rare | Frequent |
| Subword | 0.11 | 7.20 |
| Byte | 0.00 | 4.36 |
| Char | 0.28 | 9.84 |
| eByte | 5.90 | 42.90 |
| eChar | 6.78 | 44.17 |
| Tokenizer | % Num ↑ | EAcc ↑ | MdAPE ↓ |
| Subword | 20.0 | 44.8 | 95.72 |
| Byte | 39.9 | 40.5 | 99.00 |
| Char | 42.8 | 46.6 | 92.5 |
| eByte | 47.5 | 49.9 | 88.37 |
| eChar | 46.7 | 45.6 | 90.0 |
| Dataset | En | Fr | Ru | Ru (rare) | Ru (freq) | Numeracy [3] | |
| Metric | Next Word Prediction Accuracy | EAcc↑ | MdAPE↓ | ||||
| Subword | 14.37 | 41.20 | 8.31 | 0.11 | 7.20 | 44.8 | 95.7 |
| Byte | 13.69 | 17.39 | 12.76 | 0.00 | 4.36 | 40.5 | 99.0 |
| Char | 13.68 | 16.95 | 10.01 | 0.28 | 9.84 | 46.6 | 92.5 |
| eByte | 44.17 | 46.44 | 35.00 | 5.90 | 42.90 | 49.9 | 88.4 |
| eChar | 42.94 | 47.06 | 37.15 | 6.78 | 44.17 | 45.6 | 90.0 |
| Method | Citation | Compress? | Generate? | Learnt? | Word level? |
| GPT | Radford et al. (2018) | Lookup | Yes | No | Yes |
| ByT5 | Xue et al. (2022) | None | Yes | No | No |
| MANTa | Godey et al. (2022) | Segment | Yes | Yes | No |
| RetVec | Bursztein et al. (2023) | Conv. | No | Yes | Yes |
| FastText | Bojanowski et al. (2017) | Conv. | No | Yes | Yes |
| ELMo | Peters et al. (2018) | Conv. | No | Yes | Yes |
| CharBERT | El Boukkouri et al. (2020) | Conv. | No | Yes | Yes |
| CharFormer | Tay et al. (2021) | Conv. | No | Yes | No |
| LOBEF-nCF | Sreedhar et al. (2022) | None | Yes | Yes | No |
| LOBEF-WSF | Sreedhar et al. (2022) | None | Yes | Yes | Yes |
| CANINE | Clark et al. (2022) | Conv. | Yes | Yes | No |
| MegaByte | Yu et al. (2023) | Dense | Yes | Yes | No |
| Ours | Attn. | Yes | Yes | Yes |
| Efficiency | Expressivity | Accuracy | |
| Subword \nByte/ \nChar | High | Low | Mid |
| Low | High | Low | |
| eByte/ \neChar | Mid | Mid | High |
| Dataset | Size (MBs) | Words (Mil.) | Chars/Word |
| English | 4.7 | 1.34 | 5.46 |
| French | 5.1 | 1.55 | 5.18 |
| Russian | 7.5 | 1.18 | 6.39 |
| Numeracy | 6.6 | 1.35 | 5.09 |
| Tokenizer | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) |
| Language | English | French | Russian | ||||||
| Subword | 14.37 | 0.55 | 76.8 | 41.20 | 1.50 | 76.8 | 8.31 | 1.49 | 76.8 |
| Byte | 13.69 | 0.53 | 25.7 | 17.39 | 0.54 | 25.7 | 12.76 | 0.53 | 25.7 |
| Char | 13.68 | 0.54 | 26.3 | 16.95 | 0.53 | 25.7 | 10.01 | 0.54 | 26.1 |
| eByte | 44.17 | 3.84 | 38.7 | 46.44 | 6.01 | 38.7 | 35.00 | 4.92 | 38.7 |
| eChar | 42.94 | 2.94 | 39.2 | 47.06 | 3.59 | 38.7 | 37.15 | 3.95 | 39.0 |
| Lang | Tok | 1 CLS | 4 CLS | Δ% | Δ Mem |
| en | eByte | 31 | 44 | 42% | 0.02 |
| en | eChar | 29 | 43 | 48% | 0.02 |
| fr | eByte | 31 | 46 | 48% | 0.02 |
| fr | eChar | 34 | 47 | 38% | 0.02 |
| ru | eByte | 26 | 35 | 35% | 0.02 |
| ru | eChar | 29 | 37 | 28% | 0.02 |
| Tokenizer | Rare | Frequent |
| Subword | 0.11 | 7.20 |
| Byte | 0.00 | 4.36 |
| Char | 0.28 | 9.84 |
| eByte | 5.90 | 42.90 |
| eChar | 6.78 | 44.17 |
| Tokenizer | % Num ↑ | EAcc ↑ | MdAPE ↓ |
| Subword | 20.0 | 44.8 | 95.72 |
| Byte | 39.9 | 40.5 | 99.00 |
| Char | 42.8 | 46.6 | 92.5 |
| eByte | 47.5 | 49.9 | 88.37 |
| eChar | 46.7 | 45.6 | 90.0 |
| Dataset | En | Fr | Ru | Ru (rare) | Ru (freq) | Numeracy [3] | |
| Metric | Next Word Prediction Accuracy | EAcc↑ | MdAPE↓ | ||||
| Subword | 14.37 | 41.20 | 8.31 | 0.11 | 7.20 | 44.8 | 95.7 |
| Byte | 13.69 | 17.39 | 12.76 | 0.00 | 4.36 | 40.5 | 99.0 |
| Char | 13.68 | 16.95 | 10.01 | 0.28 | 9.84 | 46.6 | 92.5 |
| eByte | 44.17 | 46.44 | 35.00 | 5.90 | 42.90 | 49.9 | 88.4 |
| eChar | 42.94 | 47.06 | 37.15 | 6.78 | 44.17 | 45.6 | 90.0 |
| Method | Citation | Compress? | Generate? | Learnt? | Word level? |
| GPT | Radford et al. (2018) | Lookup | Yes | No | Yes |
| ByT5 | Xue et al. (2022) | None | Yes | No | No |
| MANTa | Godey et al. (2022) | Segment | Yes | Yes | No |
| RetVec | Bursztein et al. (2023) | Conv. | No | Yes | Yes |
| FastText | Bojanowski et al. (2017) | Conv. | No | Yes | Yes |
| ELMo | Peters et al. (2018) | Conv. | No | Yes | Yes |
| CharBERT | El Boukkouri et al. (2020) | Conv. | No | Yes | Yes |
| CharFormer | Tay et al. (2021) | Conv. | No | Yes | No |
| LOBEF-nCF | Sreedhar et al. (2022) | None | Yes | Yes | No |
| LOBEF-WSF | Sreedhar et al. (2022) | None | Yes | Yes | Yes |
| CANINE | Clark et al. (2022) | Conv. | Yes | Yes | No |
| MegaByte | Yu et al. (2023) | Dense | Yes | Yes | No |
| Ours | Attn. | Yes | Yes | Yes |
| Efficiency | Expressivity | Accuracy | |
| Subword +Byte/ +Char | High | Low | Mid |
| Low | High | Low | |
| eByte/ +eChar | Mid | Mid | High |
| Dataset | Size (MBs) | Words (Mil.) | Chars/Word |
| English | 4.7 | 1.34 | 5.46 |
| French | 5.1 | 1.55 | 5.18 |
| Russian | 7.5 | 1.18 | 6.39 |
| Numeracy | 6.6 | 1.35 | 5.09 |
| Tokenizer | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) |
| Language | English | French | Russian | ||||||
| Subword | 14.37 | 0.55 | 76.8 | 41.20 | 1.50 | 76.8 | 8.31 | 1.49 | 76.8 |
| Byte | 13.69 | 0.53 | 25.7 | 17.39 | 0.54 | 25.7 | 12.76 | 0.53 | 25.7 |
| Char | 13.68 | 0.54 | 26.3 | 16.95 | 0.53 | 25.7 | 10.01 | 0.54 | 26.1 |
| eByte | 44.17 | 3.84 | 38.7 | 46.44 | 6.01 | 38.7 | 35.00 | 4.92 | 38.7 |
| eChar | 42.94 | 2.94 | 39.2 | 47.06 | 3.59 | 38.7 | 37.15 | 3.95 | 39.0 |
| Lang | Tok | 1 CLS | 4 CLS | Δ% | Δ Mem |
| en | eByte | 31 | 44 | 42% | 0.02 |
| en | eChar | 29 | 43 | 48% | 0.02 |
| fr | eByte | 31 | 46 | 48% | 0.02 |
| fr | eChar | 34 | 47 | 38% | 0.02 |
| ru | eByte | 26 | 35 | 35% | 0.02 |
| ru | eChar | 29 | 37 | 28% | 0.02 |
| Tokenizer | Rare | Frequent |
| Subword | 0.11 | 7.20 |
| Byte | 0.00 | 4.36 |
| Char | 0.28 | 9.84 |
| eByte | 5.90 | 42.90 |
| eChar | 6.78 | 44.17 |
| Tokenizer | % Num ↑ | EAcc ↑ | MdAPE ↓ |
| Subword | 20.0 | 44.8 | 95.72 |
| Byte | 39.9 | 40.5 | 99.00 |
| Char | 42.8 | 46.6 | 92.5 |
| eByte | 47.5 | 49.9 | 88.37 |
| eChar | 46.7 | 45.6 | 90.0 |
| Dataset | En | Fr | Ru | Ru (rare) | Ru (freq) | Numeracy [3] | |
| Metric | Next Word Prediction Accuracy | EAcc↑ | MdAPE↓ | ||||
| Subword | 14.37 | 41.20 | 8.31 | 0.11 | 7.20 | 44.8 | 95.7 |
| Byte | 13.69 | 17.39 | 12.76 | 0.00 | 4.36 | 40.5 | 99.0 |
| Char | 13.68 | 16.95 | 10.01 | 0.28 | 9.84 | 46.6 | 92.5 |
| eByte | 44.17 | 46.44 | 35.00 | 5.90 | 42.90 | 49.9 | 88.4 |
| eChar | 42.94 | 47.06 | 37.15 | 6.78 | 44.17 | 45.6 | 90.0 |
| Method | Citation | Compress? | Generate? | Learnt? | Word level? |
| GPT | Radford et al. (2018) | Lookup | Yes | No | Yes |
| ByT5 | Xue et al. (2022) | None | Yes | No | No |
| MANTa | Godey et al. (2022) | Segment | Yes | Yes | No |
| RetVec | Bursztein et al. (2023) | Conv. | No | Yes | Yes |
| FastText | Bojanowski et al. (2017) | Conv. | No | Yes | Yes |
| ELMo | Peters et al. (2018) | Conv. | No | Yes | Yes |
| CharBERT | El Boukkouri et al. (2020) | Conv. | No | Yes | Yes |
| CharFormer | Tay et al. (2021) | Conv. | No | Yes | No |
| LOBEF-nCF | Sreedhar et al. (2022) | None | Yes | Yes | No |
| LOBEF-WSF | Sreedhar et al. (2022) | None | Yes | Yes | Yes |
| CANINE | Clark et al. (2022) | Conv. | Yes | Yes | No |
| MegaByte | Yu et al. (2023) | Dense | Yes | Yes | No |
| Ours | Attn. | Yes | Yes | Yes |
| Efficiency | Expressivity | Accuracy | |
| Subword \nByte/ \nChar | High | Low | Mid |
| Low | High | Low | |
| eByte/ \neChar | Mid | Mid | High |
| Dataset | Size (MBs) | Words (Mil.) | Chars/Word |
| English | 4.7 | 1.34 | 5.46 |
| French | 5.1 | 1.55 | 5.18 |
| Russian | 7.5 | 1.18 | 6.39 |
| Numeracy | 6.6 | 1.35 | 5.09 |
| Tokenizer | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) | Acc (%) | Mem (GBs) | Params (Mil.) |
| Language | English | French | Russian | ||||||
| Subword | 14.37 | 0.55 | 76.8 | 41.20 | 1.50 | 76.8 | 8.31 | 1.49 | 76.8 |
| Byte | 13.69 | 0.53 | 25.7 | 17.39 | 0.54 | 25.7 | 12.76 | 0.53 | 25.7 |
| Char | 13.68 | 0.54 | 26.3 | 16.95 | 0.53 | 25.7 | 10.01 | 0.54 | 26.1 |
| eByte | 44.17 | 3.84 | 38.7 | 46.44 | 6.01 | 38.7 | 35.00 | 4.92 | 38.7 |
| eChar | 42.94 | 2.94 | 39.2 | 47.06 | 3.59 | 38.7 | 37.15 | 3.95 | 39.0 |
| Lang | Tok | 1 CLS | 4 CLS | Δ% | Δ Mem |
| en | eByte | 31 | 44 | 42% | 0.02 |
| en | eChar | 29 | 43 | 48% | 0.02 |
| fr | eByte | 31 | 46 | 48% | 0.02 |
| fr | eChar | 34 | 47 | 38% | 0.02 |
| ru | eByte | 26 | 35 | 35% | 0.02 |
| ru | eChar | 29 | 37 | 28% | 0.02 |
| Tokenizer | Rare | Frequent |
| Subword | 0.11 | 7.20 |
| Byte | 0.00 | 4.36 |
| Char | 0.28 | 9.84 |
| eByte | 5.90 | 42.90 |
| eChar | 6.78 | 44.17 |
| Tokenizer | % Num ↑ | EAcc ↑ | MdAPE ↓ |
| Subword | 20.0 | 44.8 | 95.72 |
| Byte | 39.9 | 40.5 | 99.00 |
| Char | 42.8 | 46.6 | 92.5 |
| eByte | 47.5 | 49.9 | 88.37 |
| eChar | 46.7 | 45.6 | 90.0 |
| Dataset | En | Fr | Ru | Ru (rare) | Ru (freq) | Numeracy [3] | |
| Metric | Next Word Prediction Accuracy | EAcc↑ | MdAPE↓ | ||||
| Subword | 14.37 | 41.20 | 8.31 | 0.11 | 7.20 | 44.8 | 95.7 |
| Byte | 13.69 | 17.39 | 12.76 | 0.00 | 4.36 | 40.5 | 99.0 |
| Char | 13.68 | 16.95 | 10.01 | 0.28 | 9.84 | 46.6 | 92.5 |
| eByte | 44.17 | 46.44 | 35.00 | 5.90 | 42.90 | 49.9 | 88.4 |
| eChar | 42.94 | 47.06 | 37.15 | 6.78 | 44.17 | 45.6 | 90.0 |
| Method | Citation | Compress? | Generate? | Learnt? | Word level? |
| GPT | Radford et al. (2018) | Lookup | Yes | No | Yes |
| ByT5 | Xue et al. (2022) | None | Yes | No | No |
| MANTa | Godey et al. (2022) | Segment | Yes | Yes | No |
| RetVec | Bursztein et al. (2023) | Conv. | No | Yes | Yes |
| FastText | Bojanowski et al. (2017) | Conv. | No | Yes | Yes |
| ELMo | Peters et al. (2018) | Conv. | No | Yes | Yes |
| CharBERT | El Boukkouri et al. (2020) | Conv. | No | Yes | Yes |
| CharFormer | Tay et al. (2021) | Conv. | No | Yes | No |
| LOBEF-nCF | Sreedhar et al. (2022) | None | Yes | Yes | No |
| LOBEF-WSF | Sreedhar et al. (2022) | None | Yes | Yes | Yes |
| CANINE | Clark et al. (2022) | Conv. | Yes | Yes | No |
| MegaByte | Yu et al. (2023) | Dense | Yes | Yes | No |
| Ours | Attn. | Yes | Yes | Yes |
| RST-DT | SciDTB | CDTB | |
| #doc (train) | 312 | 743 | 1599 |
| #doc (dev) | 30 | 154 | 350 |
| #doc (test) | 38 | 152 | 374 |
| #rel (coarse) | 19 | 17 | 18 |
| #rel (fine) | 111 | 26 | - |
| #para/docu | 10.03 | 1 | 1 |
| #sent/para | 2.01 | 5.33 | 2.03 |
| #edu/sent | 2.76 | 2.63 | 2.23 |
| #token/edu | 10.85 | 11.10 | 22.76 |
| stage | L | H | bs | lr | epochs |
| structure(E2S) | 2 | 3 | 16 | 2e-5 | 5 |
| structure(S2P&P2D) | 3 | 6 | 8 | 2e-5 | 5 |
| relation | 3 | 3 | 32 | * | 10 |
| UAS | LAS | ||
| SciDTB | Zhou22-rep | 79.3 | 65.0 |
| Zhou22-run | 78.08 ± 1.06 | 62.42 ± 0.72 | |
| Our model | 81.28 ± 0.34 | 67.31 ± 0.44 | |
| CDTB | Zhou22-rep | 82.2 | 64.8 |
| Zhou22-run | 81.76 ± 1.18 | 62.63 ± 0.87 | |
| Our model | 84.36 ± 0.05 | 67.61 ± 0.36 |
| SciDTB (c) | RST-DT (c) | |||
| UAS | LAS | UAS | LAS | |
| Zhang21 | 79.1 | 65.0 | 70.2 | 51.8 |
| Our model | 81.28 | 71.23 | 71.41 | 55.64 |
| Intra | No-Graph | 87.45 ± 0.86 |
| Ours (Star Graph) | 88.85 ± 0.70 | |
| Ours (Full Graph) | 89.09 ± 0.35 | |
| Inter_p | No-Graph | 63.21 ± 1.49 |
| Ours (Star Graph) | 65.36 ± 1.28 | |
| Ours (Full Graph) | 67.81 ± 0.88 | |
| Inter_g | No-Graph | 66.54 ± 1.30 |
| Ours (Star Graph) | 68.73 ± 1.13 | |
| Ours (Full Graph) | 71.16 ± 0.66 |
| Intra_g | Inter_g | LAS_g | |
| Seq-label | 82.4 | 62.2 | 77.4 |
| Ours (Node+Edge) | 83.5 | 64.4 | 78.6 |
| Ours (EGAT) | 83.3 | 66.2 | 79.0 |
| N_edu | No-Graph | Star-Graph | Full-Graph |
| 2 (24) | 66.67 | 65.28 | 64.58 (-0.70) |
| 3 (98) | 58.50 | 62.24 | 62.41 (+0.17) |
| 4 (148) | 65.99 | 64.98 | 66.55 (+0.57) |
| 5 (157) | 63.06 | 63.27 | 65.50 (+2.23) |
| 6 (108) | 66.36 | 68.21 | 72.53 (+4.32) |
| 7 (69) | 58.94 | 59.66 | 64.01 (+4.35) |
| 8 (30) | 64.44 | 63.89 | 73.33 (+9.44) |
| >8 (32) | 58.31 | 58.84 | 62.50 (+3.66) |
| No-Graph | Star-Graph | Full-Graph | |
| P2D_p | 25.6 | 22.4 | 30.0 |
| P2D_g | 36.3 | 33.3 | 41.1 |
| Dis_edu | No-Graph | Star-Graph | Full-Graph |
| 1~4 (130) | 45.4 | 28.5 | 50.0 |
| 5~8 (110) | 31.0 | 28.2 | 41.0 |
| 9~18 (106) | 22.6 | 20.8 | 35.9 |
| >18 (93) | 21.5 | 35.5 | 12.9 |
| RST-DT | SciDTB | CDTB | |
| #doc (train) | 312 | 743 | 1599 |
| #doc (dev) | 30 | 154 | 350 |
| #doc (test) | 38 | 152 | 374 |
| #rel (coarse) | 19 | 17 | 18 |
| #rel (fine) | 111 | 26 | - |
| #para/docu | 10.03 | 1 | 1 |
| #sent/para | 2.01 | 5.33 | 2.03 |
| #edu/sent | 2.76 | 2.63 | 2.23 |
| #token/edu | 10.85 | 11.10 | 22.76 |
| stage | L | H | bs | lr | epochs |
| structure(E2S) | 2 | 3 | 16 | 2e-5 | 5 |
| structure(S2P&P2D) | 3 | 6 | 8 | 2e-5 | 5 |
| relation | 3 | 3 | 32 | * | 10 |
| UAS | LAS | ||
| SciDTB | Zhou22-rep | 79.3 | 65.0 |
| Zhou22-run | 78.08 ± 1.06 | 62.42 ± 0.72 | |
| Our model | 81.28 ± 0.34 | 67.31 ± 0.44 | |
| CDTB | Zhou22-rep | 82.2 | 64.8 |
| Zhou22-run | 81.76 ± 1.18 | 62.63 ± 0.87 | |
| Our model | 84.36 ± 0.05 | 67.61 ± 0.36 |
| SciDTB (c) | RST-DT (c) | |||
| UAS | LAS | UAS | LAS | |
| Zhang21 | 79.1 | 65.0 | 70.2 | 51.8 |
| Our model | 81.28 | 71.23 | 71.41 | 55.64 |
| Intra | No-Graph | 87.45 ± 0.86 |
| Ours (Star Graph) | 88.85 ± 0.70 | |
| Ours (Full Graph) | 89.09 ± 0.35 | |
| Inter_p | No-Graph | 63.21 ± 1.49 |
| Ours (Star Graph) | 65.36 ± 1.28 | |
| Ours (Full Graph) | 67.81 ± 0.88 | |
| Inter_g | No-Graph | 66.54 ± 1.30 |
| Ours (Star Graph) | 68.73 ± 1.13 | |
| Ours (Full Graph) | 71.16 ± 0.66 |
| Intra_g | Inter_g | LAS_g | |
| Seq-label | 82.4 | 62.2 | 77.4 |
| Ours (Node+Edge) | 83.5 | 64.4 | 78.6 |
| Ours (EGAT) | 83.3 | 66.2 | 79.0 |
| N_edu | No-Graph | Star-Graph | Full-Graph |
| 2 (24) | 66.67 | 65.28 | 64.58 (-0.70) |
| 3 (98) | 58.50 | 62.24 | 62.41 (+0.17) |
| 4 (148) | 65.99 | 64.98 | 66.55 (+0.57) |
| 5 (157) | 63.06 | 63.27 | 65.50 (+2.23) |
| 6 (108) | 66.36 | 68.21 | 72.53 (+4.32) |
| 7 (69) | 58.94 | 59.66 | 64.01 (+4.35) |
| 8 (30) | 64.44 | 63.89 | 73.33 (+9.44) |
| >8 (32) | 58.31 | 58.84 | 62.50 (+3.66) |
| No-Graph | Star-Graph | Full-Graph | |
| P2D_p | 25.6 | 22.4 | 30.0 |
| P2D_g | 36.3 | 33.3 | 41.1 |
| Dis_edu | No-Graph | Star-Graph | Full-Graph |
| 1~4 (130) | 45.4 | 28.5 | 50.0 |
| 5~8 (110) | 31.0 | 28.2 | 41.0 |
| 9~18 (106) | 22.6 | 20.8 | 35.9 |
| >18 (93) | 21.5 | 35.5 | 12.9 |
| RST-DT | SciDTB | CDTB | |
| #doc (train) | 312 | 743 | 1599 |
| #doc (dev) | 30 | 154 | 350 |
| #doc (test) | 38 | 152 | 374 |
| #rel (coarse) | 19 | 17 | 18 |
| #rel (fine) | 111 | 26 | - |
| #para/docu | 10.03 | 1 | 1 |
| #sent/para | 2.01 | 5.33 | 2.03 |
| #edu/sent | 2.76 | 2.63 | 2.23 |
| #token/edu | 10.85 | 11.10 | 22.76 |
| stage | L | H | bs | lr | epochs |
| structure(E2S) | 2 | 3 | 16 | 2e-5 | 5 |
| structure(S2P&P2D) | 3 | 6 | 8 | 2e-5 | 5 |
| relation | 3 | 3 | 32 | * | 10 |
| UAS | LAS | ||
| SciDTB | Zhou22-rep | 79.3 | 65.0 |
| Zhou22-run | 78.08 ± 1.06 | 62.42 ± 0.72 | |
| Our model | 81.28 ± 0.34 | 67.31 ± 0.44 | |
| CDTB | Zhou22-rep | 82.2 | 64.8 |
| Zhou22-run | 81.76 ± 1.18 | 62.63 ± 0.87 | |
| Our model | 84.36 ± 0.05 | 67.61 ± 0.36 |
| SciDTB (c) | RST-DT (c) | |||
| UAS | LAS | UAS | LAS | |
| Zhang21 | 79.1 | 65.0 | 70.2 | 51.8 |
| Our model | 81.28 | 71.23 | 71.41 | 55.64 |
| Intra | No-Graph | 87.45 ± 0.86 |
| Ours (Star Graph) | 88.85 ± 0.70 | |
| Ours (Full Graph) | 89.09 ± 0.35 | |
| Inter_p | No-Graph | 63.21 ± 1.49 |
| Ours (Star Graph) | 65.36 ± 1.28 | |
| Ours (Full Graph) | 67.81 ± 0.88 | |
| Inter_g | No-Graph | 66.54 ± 1.30 |
| Ours (Star Graph) | 68.73 ± 1.13 | |
| Ours (Full Graph) | 71.16 ± 0.66 |
| Intra_g | Inter_g | LAS_g | |
| Seq-label | 82.4 | 62.2 | 77.4 |
| Ours (Node+Edge) | 83.5 | 64.4 | 78.6 |
| Ours (EGAT) | 83.3 | 66.2 | 79.0 |
| N_edu | No-Graph | Star-Graph | Full-Graph |
| 2 (24) | 66.67 | 65.28 | 64.58 (-0.70) |
| 3 (98) | 58.50 | 62.24 | 62.41 (+0.17) |
| 4 (148) | 65.99 | 64.98 | 66.55 (+0.57) |
| 5 (157) | 63.06 | 63.27 | 65.50 (+2.23) |
| 6 (108) | 66.36 | 68.21 | 72.53 (+4.32) |
| 7 (69) | 58.94 | 59.66 | 64.01 (+4.35) |
| 8 (30) | 64.44 | 63.89 | 73.33 (+9.44) |
| >8 (32) | 58.31 | 58.84 | 62.50 (+3.66) |
| No-Graph | Star-Graph | Full-Graph | |
| P2D_p | 25.6 | 22.4 | 30.0 |
| P2D_g | 36.3 | 33.3 | 41.1 |
| Dis_edu | No-Graph | Star-Graph | Full-Graph |
| 1~4 (130) | 45.4 | 28.5 | 50.0 |
| 5~8 (110) | 31.0 | 28.2 | 41.0 |
| 9~18 (106) | 22.6 | 20.8 | 35.9 |
| >18 (93) | 21.5 | 35.5 | 12.9 |
| RST-DT | SciDTB | CDTB | |
| #doc (train) | 312 | 743 | 1599 |
| #doc (dev) | 30 | 154 | 350 |
| #doc (test) | 38 | 152 | 374 |
| #rel (coarse) | 19 | 17 | 18 |
| #rel (fine) | 111 | 26 | - |
| #para/docu | 10.03 | 1 | 1 |
| #sent/para | 2.01 | 5.33 | 2.03 |
| #edu/sent | 2.76 | 2.63 | 2.23 |
| #token/edu | 10.85 | 11.10 | 22.76 |
| stage | L | H | bs | lr | epochs |
| structure(E2S) | 2 | 3 | 16 | 2e-5 | 5 |
| structure(S2P&P2D) | 3 | 6 | 8 | 2e-5 | 5 |
| relation | 3 | 3 | 32 | * | 10 |
| UAS | LAS | ||
| SciDTB | Zhou22-rep | 79.3 | 65.0 |
| Zhou22-run | 78.08 ± 1.06 | 62.42 ± 0.72 | |
| Our model | 81.28 ± 0.34 | 67.31 ± 0.44 | |
| CDTB | Zhou22-rep | 82.2 | 64.8 |
| Zhou22-run | 81.76 ± 1.18 | 62.63 ± 0.87 | |
| Our model | 84.36 ± 0.05 | 67.61 ± 0.36 |
| SciDTB (c) | RST-DT (c) | |||
| UAS | LAS | UAS | LAS | |
| Zhang21 | 79.1 | 65.0 | 70.2 | 51.8 |
| Our model | 81.28 | 71.23 | 71.41 | 55.64 |
| Intra | No-Graph | 87.45 ± 0.86 |
| Ours (Star Graph) | 88.85 ± 0.70 | |
| Ours (Full Graph) | 89.09 ± 0.35 | |
| Inter_p | No-Graph | 63.21 ± 1.49 |
| Ours (Star Graph) | 65.36 ± 1.28 | |
| Ours (Full Graph) | 67.81 ± 0.88 | |
| Inter_g | No-Graph | 66.54 ± 1.30 |
| Ours (Star Graph) | 68.73 ± 1.13 | |
| Ours (Full Graph) | 71.16 ± 0.66 |
| Intra_g | Inter_g | LAS_g | |
| Seq-label | 82.4 | 62.2 | 77.4 |
| Ours (Node+Edge) | 83.5 | 64.4 | 78.6 |
| Ours (EGAT) | 83.3 | 66.2 | 79.0 |
| N_edu | No-Graph | Star-Graph | Full-Graph |
| 2 (24) | 66.67 | 65.28 | 64.58 (-0.70) |
| 3 (98) | 58.50 | 62.24 | 62.41 (+0.17) |
| 4 (148) | 65.99 | 64.98 | 66.55 (+0.57) |
| 5 (157) | 63.06 | 63.27 | 65.50 (+2.23) |
| 6 (108) | 66.36 | 68.21 | 72.53 (+4.32) |
| 7 (69) | 58.94 | 59.66 | 64.01 (+4.35) |
| 8 (30) | 64.44 | 63.89 | 73.33 (+9.44) |
| >8 (32) | 58.31 | 58.84 | 62.50 (+3.66) |
| No-Graph | Star-Graph | Full-Graph | |
| P2D_p | 25.6 | 22.4 | 30.0 |
| P2D_g | 36.3 | 33.3 | 41.1 |
| Dis_edu | No-Graph | Star-Graph | Full-Graph |
| 1~4 (130) | 45.4 | 28.5 | 50.0 |
| 5~8 (110) | 31.0 | 28.2 | 41.0 |
| 9~18 (106) | 22.6 | 20.8 | 35.9 |
| >18 (93) | 21.5 | 35.5 | 12.9 |
| Dataset | Task Type | # of Train |
| Source Tasks | ||
| MNLI | Natural Language Inference (NLI) | 393K |
| QNLI | Natural Language Inference (NLI) | 105K |
| QQP | Paraphrase Identification (PI) | 364K |
| SST-2 | Sentiment Analysis (SA) | 66K |
| Yelp-2 | Sentiment Analysis (SA) | 540K |
| ReCoRD | Question Answering (QA) | 101K |
| WinoGrande | Commonsense Reasoning (CR) | 40K |
| Target Tasks | ||
| RTE | Natural Language Inference (NLI) | 2.5K |
| MRPC | Paraphrase Identification (PI) | 3.7K |
| CR | Sentiment Analysis (SA) | 2.3K |
| MR | Sentiment Analysis (SA) | 6.4K |
| MPQA | Sentiment Analysis (SA) | 6.4K |
| BoolQ | Question Answering (QA) | 9.4K |
| MultiRC | Question Answering (QA) | 27K |
| CosmosQA | Commonsense Reasoning (CR) | 25K |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Upper bound (computationally expensive) | ||||||||||
| Fine-tuning the whole language model | X | 86.38 | 88.01 | 93.59 | 91.19 | 91.93 | 85.28 | 80.80 | 77.77 | 86.87 |
| Lower bound, without source task pre-training | ||||||||||
| Fine-tuning with frozen language model | ✓ | 58.59 | 77.28 | 91.66 | 87.99 | 90.29 | 70.53 | 68.74 | 57.27 | 75.29 |
| With source task pre-training | ||||||||||
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training (ours) | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| - Removing at most 1 task | ✓ | 86.76 | 86.21 | 92.05 | 90.64 | 91.28 | 81.65 | 76.12 | 71.11 | 84.48 |
| - Removing at most 2 tasks | ✓ | 87.30 | 86.67 | 93.05 | 90.76 | 91.70 | 82.18 | 76.12 | 72.00 | 84.97 |
| Dataset | Task Type | # of Train |
| Source Tasks | ||
| MNLI | Natural Language Inference (NLI) | 393K |
| QNLI | Natural Language Inference (NLI) | 105K |
| QQP | Paraphrase Identification (PI) | 364K |
| SST-2 | Sentiment Analysis (SA) | 66K |
| Yelp-2 | Sentiment Analysis (SA) | 540K |
| ReCoRD | Question Answering (QA) | 101K |
| WinoGrande | Commonsense Reasoning (CR) | 40K |
| Target Tasks | ||
| RTE | Natural Language Inference (NLI) | 2.5K |
| MRPC | Paraphrase Identification (PI) | 3.7K |
| CR | Sentiment Analysis (SA) | 2.3K |
| MR | Sentiment Analysis (SA) | 6.4K |
| MPQA | Sentiment Analysis (SA) | 6.4K |
| BoolQ | Question Answering (QA) | 9.4K |
| MultiRC | Question Answering (QA) | 27K |
| CosmosQA | Commonsense Reasoning (CR) | 25K |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Upper bound (computationally expensive) | ||||||||||
| Fine-tuning the whole language model | X | 86.38 | 88.01 | 93.59 | 91.19 | 91.93 | 85.28 | 80.80 | 77.77 | 86.87 |
| Lower bound, without source task pre-training | ||||||||||
| Fine-tuning with frozen language model | ✓ | 58.59 | 77.28 | 91.66 | 87.99 | 90.29 | 70.53 | 68.74 | 57.27 | 75.29 |
| With source task pre-training | ||||||||||
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training (ours) | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| - Removing at most 1 task | ✓ | 86.76 | 86.21 | 92.05 | 90.64 | 91.28 | 81.65 | 76.12 | 71.11 | 84.48 |
| - Removing at most 2 tasks | ✓ | 87.30 | 86.67 | 93.05 | 90.76 | 91.70 | 82.18 | 76.12 | 72.00 | 84.97 |
| Dataset | Task Type | # of Train |
| Source Tasks | ||
| MNLI | Natural Language Inference (NLI) | 393K |
| QNLI | Natural Language Inference (NLI) | 105K |
| QQP | Paraphrase Identification (PI) | 364K |
| SST-2 | Sentiment Analysis (SA) | 66K |
| Yelp-2 | Sentiment Analysis (SA) | 540K |
| ReCoRD | Question Answering (QA) | 101K |
| WinoGrande | Commonsense Reasoning (CR) | 40K |
| Target Tasks | ||
| RTE | Natural Language Inference (NLI) | 2.5K |
| MRPC | Paraphrase Identification (PI) | 3.7K |
| CR | Sentiment Analysis (SA) | 2.3K |
| MR | Sentiment Analysis (SA) | 6.4K |
| MPQA | Sentiment Analysis (SA) | 6.4K |
| BoolQ | Question Answering (QA) | 9.4K |
| MultiRC | Question Answering (QA) | 27K |
| CosmosQA | Commonsense Reasoning (CR) | 25K |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Upper bound (computationally expensive) | ||||||||||
| Fine-tuning the whole language model | X | 86.38 | 88.01 | 93.59 | 91.19 | 91.93 | 85.28 | 80.80 | 77.77 | 86.87 |
| Lower bound, without source task pre-training | ||||||||||
| Fine-tuning with frozen language model | ✓ | 58.59 | 77.28 | 91.66 | 87.99 | 90.29 | 70.53 | 68.74 | 57.27 | 75.29 |
| With source task pre-training | ||||||||||
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training (ours) | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| - Removing at most 1 task | ✓ | 86.76 | 86.21 | 92.05 | 90.64 | 91.28 | 81.65 | 76.12 | 71.11 | 84.48 |
| - Removing at most 2 tasks | ✓ | 87.30 | 86.67 | 93.05 | 90.76 | 91.70 | 82.18 | 76.12 | 72.00 | 84.97 |
| Dataset | Task Type | # of Train |
| Source Tasks | ||
| MNLI | Natural Language Inference (NLI) | 393K |
| QNLI | Natural Language Inference (NLI) | 105K |
| QQP | Paraphrase Identification (PI) | 364K |
| SST-2 | Sentiment Analysis (SA) | 66K |
| Yelp-2 | Sentiment Analysis (SA) | 540K |
| ReCoRD | Question Answering (QA) | 101K |
| WinoGrande | Commonsense Reasoning (CR) | 40K |
| Target Tasks | ||
| RTE | Natural Language Inference (NLI) | 2.5K |
| MRPC | Paraphrase Identification (PI) | 3.7K |
| CR | Sentiment Analysis (SA) | 2.3K |
| MR | Sentiment Analysis (SA) | 6.4K |
| MPQA | Sentiment Analysis (SA) | 6.4K |
| BoolQ | Question Answering (QA) | 9.4K |
| MultiRC | Question Answering (QA) | 27K |
| CosmosQA | Commonsense Reasoning (CR) | 25K |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Upper bound (computationally expensive) | ||||||||||
| Fine-tuning the whole language model | X | 86.38 | 88.01 | 93.59 | 91.19 | 91.93 | 85.28 | 80.80 | 77.77 | 86.87 |
| Lower bound, without source task pre-training | ||||||||||
| Fine-tuning with frozen language model | ✓ | 58.59 | 77.28 | 91.66 | 87.99 | 90.29 | 70.53 | 68.74 | 57.27 | 75.29 |
| With source task pre-training | ||||||||||
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training (ours) | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| Method | Freeze | RTE | MRPC | CR | MR | MPQA | BoolQ | MultiRC | CosmosQA | Avg. |
| Multi-tasking training | ✓ | 82.16 | 84.90 | 91.03 | 90.60 | 90.83 | 79.44 | 74.00 | 66.53 | 82.44 |
| Prefix-based training | ✓ | 84.86 | 85.37 | 92.03 | 90.08 | 91.05 | 80.84 | 75.66 | 70.90 | 83.85 |
| - Removing at most 1 task | ✓ | 86.76 | 86.21 | 92.05 | 90.64 | 91.28 | 81.65 | 76.12 | 71.11 | 84.48 |
| - Removing at most 2 tasks | ✓ | 87.30 | 86.67 | 93.05 | 90.76 | 91.70 | 82.18 | 76.12 | 72.00 | 84.97 |
| # of Reddit Authors | 1,000 |
| # of Reddit Posts | 10,000 |
| # of Interpretable Style Attributes | 1,255,874 |
| # of (Text, Style Attribute) labeled pairs | 5,490,847 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | The author uses a negative tone. | 0.788 |
| IMDB Large Movie Review Dataset | 0.665 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | The author is expressing {{emotion}}. | 0.542 |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 |
| Text | Top 5 LISA Vector Dimensions |
| Lol right on | 1. 1.00 – The author is being polite.\n2. 1.00 – The author is writing in a cheerful manner.\n3. 1.00 – The author is using a lighthearted tone.\n4. 1.00 – The author is laughing.\n5. 1.00 – The author is complimentary. |
| There is the Toyota GT86 R3 ;) http://www.toyota-motorsport.com/motorsport/downloads/com_droppics/59/DSF4311.jpg | 1. 0.99 – The author is simply describing a product. * \n2. 0.99 – The author is providing a visual cue to the reader.\n3. 0.98 – The author simply provides information.\n4. 0.97 – The author is using an emoji.\n5. 0.97 – The author provides information. |
| Every time i watched this episode as a kid i was always like "WTF, JAMES, A POKEBALL ISNT EVEN A POKE-MON?! GET YOUR ACT TOGETHER, SON!!" | 1. 1.00 – The author is discussing a television show. * \n2. 0.99 – The author has a hint of nostalgia.\n3. 0.99 – The author is making a humorous comment.\n4. 0.99 – The author is animated in their writing.\n5. 0.99 – The author is laughing. |
| 13 POT MORDE BABY WOOOOOOOOOOOOOO | 1. 1.00 – The author is using an elongated word.\n2. 1.00 – The author is using only English words.\n3. 1.00 – The author is using a single word.\n4. 1.00 – The author is using swear words.\n5. 1.00 – The author uses two exclamation marks. |
| No wonder everyone resorts to performing a murder spree eventually. | 1. 1.00 – The author is scornful.\n2. 1.00 – The author is ungenerous.\n3. 0.99 – The author is expressing antisocial behaviors.\n4. 0.99 – The author is uncaring.\n5. 0.99 – The author is dramatic. |
| Podcast originally refers to an iPod, and before that there was definitely TWiT, which still calls itself a Netcast | 1. 0.97 – The author uses a variety of words to describe the same concept.\n2. 0.97 – The author is simply describing a product. * \n3. 0.95 – The author uses specific terms related to the topic.\n4. 0.94 – The author has a deep understanding of the topic.\n5. 0.94 – The author is using words focusing on the past. |
| Model | Formal | Complex | Numb3r | C'tion | Avg | Interpretable |
| Random Baseline | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | |
| Content-Aware Representations | ||||||
| SBERT | 0.78/0.00 | 0.54/0.01 | 0.81/0.04 | 0.86/0.00 | 0.75/0.01 | × |
| LUAR | 0.80/0.14 | 0.67/0.00 | 0.74/0.03 | 0.77/0.00 | 0.75/0.04 | × |
| Content-Independent Style Representations | ||||||
| LIWC | 0.52/- | 0.52/- | 0.50/- | 0.99/- | 0.63/- | ✓ |
| Wegmann et al. (2022) | 0.84/0.69 | 0.59/0.26 | 0.56/0.03 | 0.96/0.02 | 0.74/0.25 | × |
| LISA | 0.69/0.07 | 0.57/0.01 | 0.80/0.03 | 0.77/0.00 | 0.71/0.03 | ✓ |
| LISA (Wegmann + w) | 0.72/0.07 | 0.61/0.03 | 0.81/0.08 | 0.68/0.00 | 0.71/0.05 | ✓ |
| LISA (Wegmann + W) | 0.66/0.03 | 0.56/0.01 | 0.70/0.01 | 0.87/0.00 | 0.70/0.01 | ✓ |
| LISA (LUAR + w) | 0.73/0.05 | 0.65/0.00 | 0.85/0.03 | 0.92/0.00 | 0.79/0.02 | ✓ |
| LISA (LUAR + W) | 0.81/0.07 | 0.56/0.01 | 0.74/0.03 | 0.82/0.00 | 0.73/0.03 | ✓ |
| Text | Style Attribute |
| Subscribed. Interesting idea. I would like to see some advanced stats on hitting percentages to different locations on the court. For example, having the court broken up into maybe 12 zones and then hitting percentages from each position to those zones. I remember seeing an article that did this years ago and I have never been able to find anything online. I said recently on this sub that the deep angle shot from the left side or right side was the highest percentage shot in volleyball, but I was not able to back up my claim with any sources or anything. Anyways, I am a VB nerd, no doubt. Interested to see what direction you take this. Cheers! | The author is being polite. |
| Yeah I also work in QA, and seeing this kind of stuff get released is maddening. About a year ago working on a new platform we were seeing bugs in the hundreds each week, we pushed back the release of the product 3 months because basically it didn't work. If it was up to the devs, they'd have released it on time, because the stuff they'd written code for worked. Thorough doesn't even cover the work we go through every 3 months, and Niantic's approach seems completely amateur from this side. They're putting bandaids on problems and hiding things like the 3 step problem behind curtains without seemingly fixing anything, although I do have to say their balance tweaks to battling have been a big step in the right direction. | The author is using a personal anecdote to illustrate their point. |
| Thank you. I'd be interested in reading more about your experiences, in addition to the "American Wedding" story. Are you watching the stream? I wish there was a way to find out how many people in the world are watching it. The music is lovely, huh? God damn. He's got his bunny Fair Isle sweater on, drinking Dunkin' Donuts coffee. I would have thought him a Starbucks man. :-) | The author is using an emoji. |
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot.\nThis is why devices that have no switches (primarily wall warts) need not have polarized plugs.\nSame Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood.\nDifferent Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) – The author is using a scientific approach.\n(0.96, 0.98) – The author is using a combination of technical terms and everyday language.\n(0.91, 0.84) – The author is using formal and professional language.\n(0.06, 0.99) – The author is using words related to the game they are discussing. *\n(0.00, 0.88) – The author is using an emoji.\n(0.02, 0.87) – The author uses an emoticon at the end. |
| Prompts Used to Generate STYLEGENOME | Validation F1 |
| Open-ended Prompts | 0.865 |
| Targeted Prompts | 0.898 |
| Open-ended Prompts & Targeted Prompts | 0.920 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | 0.788 | |
| IMDB Large Movie Review Dataset | 0.665 | ||
| Amazon Customer Reviews Dataset | The author uses a negative tone. | 0.432 | |
| Rotten Tomatoes Movie Review Data | 0.463 | ||
| App Reviews | 0.350 | ||
| Twitter Sentiment Analysis Training Corpus | 0.299 | ||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | 0.542 | |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| GoEmotions (Disappointment vs. Admiration) | 0.571 | ||
| GoEmotions (Pride vs. Embarrassment) | 0.419 | ||
| GoEmotions (Nervousness vs. Optimism) | 0.447 | ||
| GoEmotions (Disapproval vs. Approval) | 0.432 | ||
| GoEmotions (Admiration vs. Neutral) | 0.578 | ||
| GoEmotions (Amusement vs. Neutral) | 0.508 | ||
| GoEmotions (Anger vs. Neutral) | 0.372 | ||
| GoEmotions (Annoyance vs. Neutral) | 0.355 | ||
| GoEmotions (Approval vs. Neutral) | 0.257 | ||
| GoEmotions (Caring vs. Neutral) | 0.303 | ||
| GoEmotions (Confusion vs. Neutral) | 0.343 | ||
| GoEmotions (Curiosity vs. Neutral) | 0.464 | ||
| GoEmotions (Desire vs. Neutral) | 0.315 | ||
| GoEmotions (Disappointment vs. Neutral) | The author is expressing {{emotion}}. | 0.317 | |
| GoEmotions (Disapproval vs. Neutral) | 0.284 | ||
| GoEmotions (Disgust vs. Neutral) | 0.307 | ||
| GoEmotions (Embarrassment vs. Neutral) | 0.173 | ||
| GoEmotions (Excitement vs. Neutral) | 0.295 | ||
| GoEmotions (Fear vs. Neutral) | 0.309 | ||
| GoEmotions (Gratitude vs. Neutral) | 0.626 | ||
| GoEmotions (Grief vs. Neutral) | 0.093 | ||
| GoEmotions (Joy vs. Neutral) | 0.390 | ||
| GoEmotions (Love vs. Neutral) | 0.497 | ||
| GoEmotions (Nervousness vs. Neutral) | 0.171 | ||
| GoEmotions (Optimism vs. Neutral) | 0.407 | ||
| GoEmotions (Pride vs. Neutral) | 0.097 | ||
| GoEmotions (Realization vs. Neutral) | 0.121 | ||
| GoEmotions (Relief vs. Neutral) | 0.123 | ||
| GoEmotions (Remorse vs. Neutral) | 0.276 | ||
| GoEmotions (Sadness vs. Neutral) | 0.413 | ||
| GoEmotions (Surprise vs. Neutral) | 0.317 | ||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 |
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot.\nThis is why devices that have no switches (primarily wall warts) need not have polarized plugs.\nSame Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood.\nDifferent Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) - The author is using a scientific approach.\n(0.96, 0.98) - The author is using a combination of technical terms and everyday language.\n(0.91, 0.84) - The author is using formal and professional language. |
| (0.06, 0.99) - The author is using words related to the game they are discussing. *\n(0.00, 0.88) - The author is using an emoji.\n(0.02, 0.87) - The author uses an emoticon at the end. | |
| Anchor: Not sure what the income tax is in Germany, but in the Netherlands the income can be up 50% for the higher income classes.\nSame Author: The salaries in the US alway blow my mind. A software developer in Amsterdam gets like €40.000/year, maybe €50.000/year if your good, and maybe €60.000/year if you're some kind of manager. Anything position over €100.000/year is basically running the entire company.\nDifferent Author: How would you even test this software? The setup would be just insane. | (0.84, 0.90) - The author is using words indicating poverty.\n(0.87, 1.00) - The author is using words indicating wealth.\n(0.80, 0.93) - The author is using words related to money.\n(0.00, 1.00) - The author is comfortable with technology.\n(0.00, 0.85) - The author is discussing a product. *\n(0.13, 0.92) - The author is using formal and professional language. |
| Anchor: If only there was something he could have done to avoid this backlash. Like maybe not acting like a complete d**khead.\nSame Author: I take issue with a faster landing being marked as less skilled. By that logic the slowest, smoothest possible landing would be the most skilled and that is plain wrong. Maybe war machine intentionally does faster and harder landings.\nDifferent Author: She was the Ronald Reagan of the UK in the same time period. | (1.00, 0.96) - The author is emphasizing the contrast between the two ideas.\n(0.78, 0.89) - The author is able to draw conclusions.\n(0.97, 0.86) - The author is using an all-or-none thinking style.\n(0.83, 0.05) - The author is describing sexual content. *\n(0.38, 0.94) - The author is using words related to politics. *\n(0.74, 0.38) - The author is using parentheticals. |
| # of Reddit Authors | 1,000 |
| # of Reddit Posts | 10,000 |
| # of Interpretable Style Attributes | 1,255,874 |
| # of (Text, Style Attribute) labeled pairs | 5,490,847 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | The author uses a negative tone. | 0.788 |
| IMDB Large Movie Review Dataset | 0.665 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | The author is expressing {{emotion}}. | 0.542 |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 |
| Text | Top 5 LISA Vector Dimensions |
| Lol right on | 1. 1.00 – The author is being polite.\n2. 1.00 – The author is writing in a cheerful manner.\n3. 1.00 – The author is using a lighthearted tone.\n4. 1.00 – The author is laughing.\n5. 1.00 – The author is complimentary. |
| There is the Toyota GT86 R3 ;) http://www.toyota-motorsport.com/motorsport/downloads/com_droppics/59/DSF4311.jpg | 1. 0.99 – The author is simply describing a product. * \n2. 0.99 – The author is providing a visual cue to the reader.\n3. 0.98 – The author simply provides information.\n4. 0.97 – The author is using an emoji.\n5. 0.97 – The author provides information. |
| Every time i watched this episode as a kid i was always like "WTF, JAMES, A POKEBALL ISNT EVEN A POKE-MON?! GET YOUR ACT TOGETHER, SON!!" | 1. 1.00 – The author is discussing a television show. * \n2. 0.99 – The author has a hint of nostalgia.\n3. 0.99 – The author is making a humorous comment.\n4. 0.99 – The author is animated in their writing.\n5. 0.99 – The author is laughing. |
| 13 POT MORDE BABY WOOOOOOOOOOOOOO | 1. 1.00 – The author is using an elongated word.\n2. 1.00 – The author is using only English words.\n3. 1.00 – The author is using a single word.\n4. 1.00 – The author is using swear words.\n5. 1.00 – The author uses two exclamation marks. |
| No wonder everyone resorts to performing a murder spree eventually. | 1. 1.00 – The author is scornful.\n2. 1.00 – The author is ungenerous.\n3. 0.99 – The author is expressing antisocial behaviors.\n4. 0.99 – The author is uncaring.\n5. 0.99 – The author is dramatic. |
| Podcast originally refers to an iPod, and before that there was definitely TWiT, which still calls itself a Netcast | 1. 0.97 – The author uses a variety of words to describe the same concept.\n2. 0.97 – The author is simply describing a product. * \n3. 0.95 – The author uses specific terms related to the topic.\n4. 0.94 – The author has a deep understanding of the topic.\n5. 0.94 – The author is using words focusing on the past. |
| Model | Formal | Complex | Numb3r | C'tion | Avg | Interpretable |
| Random Baseline | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | |
| Content-Aware Representations | ||||||
| SBERT | 0.78/0.00 | 0.54/0.01 | 0.81/0.04 | 0.86/0.00 | 0.75/0.01 | × |
| LUAR | 0.80/0.14 | 0.67/0.00 | 0.74/0.03 | 0.77/0.00 | 0.75/0.04 | × |
| Content-Independent Style Representations | ||||||
| LIWC | 0.52/- | 0.52/- | 0.50/- | 0.99/- | 0.63/- | ✓ |
| Wegmann et al. (2022) | 0.84/0.69 | 0.59/0.26 | 0.56/0.03 | 0.96/0.02 | 0.74/0.25 | × |
| LISA | 0.69/0.07 | 0.57/0.01 | 0.80/0.03 | 0.77/0.00 | 0.71/0.03 | ✓ |
| LISA (Wegmann + w) | 0.72/0.07 | 0.61/0.03 | 0.81/0.08 | 0.68/0.00 | 0.71/0.05 | ✓ |
| LISA (Wegmann + W) | 0.66/0.03 | 0.56/0.01 | 0.70/0.01 | 0.87/0.00 | 0.70/0.01 | ✓ |
| LISA (LUAR + w) | 0.73/0.05 | 0.65/0.00 | 0.85/0.03 | 0.92/0.00 | 0.79/0.02 | ✓ |
| LISA (LUAR + W) | 0.81/0.07 | 0.56/0.01 | 0.74/0.03 | 0.82/0.00 | 0.73/0.03 | ✓ |
| Text | Style Attribute |
| Subscribed. Interesting idea. I would like to see some advanced stats on hitting percentages to different locations on the court. For example, having the court broken up into maybe 12 zones and then hitting percentages from each position to those zones. I remember seeing an article that did this years ago and I have never been able to find anything online. I said recently on this sub that the deep angle shot from the left side or right side was the highest percentage shot in volleyball, but I was not able to back up my claim with any sources or anything. Anyways, I am a VB nerd, no doubt. Interested to see what direction you take this. Cheers! | The author is being polite. |
| Yeah I also work in QA, and seeing this kind of stuff get released is maddening. About a year ago working on a new platform we were seeing bugs in the hundreds each week, we pushed back the release of the product 3 months because basically it didn't work. If it was up to the devs, they'd have released it on time, because the stuff they'd written code for worked. Thorough doesn't even cover the work we go through every 3 months, and Niantic's approach seems completely amateur from this side. They're putting bandaids on problems and hiding things like the 3 step problem behind curtains without seemingly fixing anything, although I do have to say their balance tweaks to battling have been a big step in the right direction. | The author is using a personal anecdote to illustrate their point. |
| Thank you. I'd be interested in reading more about your experiences, in addition to the "American Wedding" story. Are you watching the stream? I wish there was a way to find out how many people in the world are watching it. The music is lovely, huh? God damn. He's got his bunny Fair Isle sweater on, drinking Dunkin' Donuts coffee. I would have thought him a Starbucks man. :-) | The author is using an emoji. |
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot.\nThis is why devices that have no switches (primarily wall warts) need not have polarized plugs.\nSame Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood.\nDifferent Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) – The author is using a scientific approach.\n(0.96, 0.98) – The author is using a combination of technical terms and everyday language.\n(0.91, 0.84) – The author is using formal and professional language.\n(0.06, 0.99) – The author is using words related to the game they are discussing. *\n(0.00, 0.88) – The author is using an emoji.\n(0.02, 0.87) – The author uses an emoticon at the end. |
| Prompts Used to Generate STYLEGENOME | Validation F1 |
| Open-ended Prompts | 0.865 |
| Targeted Prompts | 0.898 |
| Open-ended Prompts & Targeted Prompts | 0.920 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | 0.788 | |
| IMDB Large Movie Review Dataset | 0.665 | ||
| Amazon Customer Reviews Dataset | The author uses a negative tone. | 0.432 | |
| Rotten Tomatoes Movie Review Data | 0.463 | ||
| App Reviews | 0.350 | ||
| Twitter Sentiment Analysis Training Corpus | 0.299 | ||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | 0.542 | |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| GoEmotions (Disappointment vs. Admiration) | 0.571 | ||
| GoEmotions (Pride vs. Embarrassment) | 0.419 | ||
| GoEmotions (Nervousness vs. Optimism) | 0.447 | ||
| GoEmotions (Disapproval vs. Approval) | 0.432 | ||
| GoEmotions (Admiration vs. Neutral) | 0.578 | ||
| GoEmotions (Amusement vs. Neutral) | 0.508 | ||
| GoEmotions (Anger vs. Neutral) | 0.372 | ||
| GoEmotions (Annoyance vs. Neutral) | 0.355 | ||
| GoEmotions (Approval vs. Neutral) | 0.257 | ||
| GoEmotions (Caring vs. Neutral) | 0.303 | ||
| GoEmotions (Confusion vs. Neutral) | 0.343 | ||
| GoEmotions (Curiosity vs. Neutral) | 0.464 | ||
| GoEmotions (Desire vs. Neutral) | 0.315 | ||
| GoEmotions (Disappointment vs. Neutral) | The author is expressing {{emotion}}. | 0.317 | |
| GoEmotions (Disapproval vs. Neutral) | 0.284 | ||
| GoEmotions (Disgust vs. Neutral) | 0.307 | ||
| GoEmotions (Embarrassment vs. Neutral) | 0.173 | ||
| GoEmotions (Excitement vs. Neutral) | 0.295 | ||
| GoEmotions (Fear vs. Neutral) | 0.309 | ||
| GoEmotions (Gratitude vs. Neutral) | 0.626 | ||
| GoEmotions (Grief vs. Neutral) | 0.093 | ||
| GoEmotions (Joy vs. Neutral) | 0.390 | ||
| GoEmotions (Love vs. Neutral) | 0.497 | ||
| GoEmotions (Nervousness vs. Neutral) | 0.171 | ||
| GoEmotions (Optimism vs. Neutral) | 0.407 | ||
| GoEmotions (Pride vs. Neutral) | 0.097 | ||
| GoEmotions (Realization vs. Neutral) | 0.121 | ||
| GoEmotions (Relief vs. Neutral) | 0.123 | ||
| GoEmotions (Remorse vs. Neutral) | 0.276 | ||
| GoEmotions (Sadness vs. Neutral) | 0.413 | ||
| GoEmotions (Surprise vs. Neutral) | 0.317 | ||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 |
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot.\nThis is why devices that have no switches (primarily wall warts) need not have polarized plugs.\nSame Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood.\nDifferent Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) - The author is using a scientific approach.\n(0.96, 0.98) - The author is using a combination of technical terms and everyday language.\n(0.91, 0.84) - The author is using formal and professional language. |
| (0.06, 0.99) - The author is using words related to the game they are discussing. *\n(0.00, 0.88) - The author is using an emoji.\n(0.02, 0.87) - The author uses an emoticon at the end. | |
| Anchor: Not sure what the income tax is in Germany, but in the Netherlands the income can be up 50% for the higher income classes.\nSame Author: The salaries in the US alway blow my mind. A software developer in Amsterdam gets like €40.000/year, maybe €50.000/year if your good, and maybe €60.000/year if you're some kind of manager. Anything position over €100.000/year is basically running the entire company.\nDifferent Author: How would you even test this software? The setup would be just insane. | (0.84, 0.90) - The author is using words indicating poverty.\n(0.87, 1.00) - The author is using words indicating wealth.\n(0.80, 0.93) - The author is using words related to money.\n(0.00, 1.00) - The author is comfortable with technology.\n(0.00, 0.85) - The author is discussing a product. *\n(0.13, 0.92) - The author is using formal and professional language. |
| Anchor: If only there was something he could have done to avoid this backlash. Like maybe not acting like a complete d**khead.\nSame Author: I take issue with a faster landing being marked as less skilled. By that logic the slowest, smoothest possible landing would be the most skilled and that is plain wrong. Maybe war machine intentionally does faster and harder landings.\nDifferent Author: She was the Ronald Reagan of the UK in the same time period. | (1.00, 0.96) - The author is emphasizing the contrast between the two ideas.\n(0.78, 0.89) - The author is able to draw conclusions.\n(0.97, 0.86) - The author is using an all-or-none thinking style.\n(0.83, 0.05) - The author is describing sexual content. *\n(0.38, 0.94) - The author is using words related to politics. *\n(0.74, 0.38) - The author is using parentheticals. |
| # of Reddit Authors | 1,000 |
| # of Reddit Posts | 10,000 |
| # of Interpretable Style Attributes | 1,255,874 |
| # of (Text, Style Attribute) labeled pairs | 5,490,847 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | The author uses a negative tone. | 0.788 |
| IMDB Large Movie Review Dataset | 0.665 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | The author is expressing {{emotion}}. | 0.542 |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 | ||
| Text | Top 5 LISA Vector Dimensions |
| Lol right on | 1. 1.00 – The author is being polite. +2. 1.00 – The author is writing in a cheerful manner. +3. 1.00 – The author is using a lighthearted tone. +4. 1.00 – The author is laughing. +5. 1.00 – The author is complimentary. |
| There is the Toyota GT86 R3 ;) http://www.toyota-motorsport.com/motorsport/downloads/com_droppics/59/DSF4311.jpg | 1. 0.99 – The author is simply describing a product. * +2. 0.99 – The author is providing a visual cue to the reader. +3. 0.98 – The author simply provides information. +4. 0.97 – The author is using an emoji. +5. 0.97 – The author provides information. |
| Every time i watched this episode as a kid i was always like "WTF, JAMES, A POKEBALL ISNT EVEN A POKE-MON?! GET YOUR ACT TOGETHER, SON!!" | 1. 1.00 – The author is discussing a television show. * +2. 0.99 – The author has a hint of nostalgia. +3. 0.99 – The author is making a humorous comment. +4. 0.99 – The author is animated in their writing. +5. 0.99 – The author is laughing. |
| 13 POT MORDE BABY WOOOOOOOOOOOOOO | 1. 1.00 – The author is using an elongated word. +2. 1.00 – The author is using only English words. +3. 1.00 – The author is using a single word. +4. 1.00 – The author is using swear words. +5. 1.00 – The author uses two exclamation marks. |
| No wonder everyone resorts to performing a murder spree eventually. | 1. 1.00 – The author is scornful. +2. 1.00 – The author is ungenerous. +3. 0.99 – The author is expressing antisocial behaviors. +4. 0.99 – The author is uncaring. +5. 0.99 – The author is dramatic. |
| Podcast originally refers to an iPod, and before that there was definitely TWiT, which still calls itself a Netcast | 1. 0.97 – The author uses a variety of words to describe the same concept. +2. 0.97 – The author is simply describing a product. * +3. 0.95 – The author uses specific terms related to the topic. +4. 0.94 – The author has a deep understanding of the topic. +5. 0.94 – The author is using words focusing on the past. |
| Model | Formal | Complex | Numb3r | C'tion | Avg | Interpretable |
| Random Baseline | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | |
| Content-Aware Representations | ||||||
| SBERT | 0.78/0.00 | 0.54/0.01 | 0.81/0.04 | 0.86/0.00 | 0.75/0.01 | × |
| LUAR | 0.80/0.14 | 0.67/0.00 | 0.74/0.03 | 0.77/0.00 | 0.75/0.04 | × |
| Content-Independent Style Representations | ||||||
| LIWC | 0.52/- | 0.52/- | 0.50/- | 0.99/- | 0.63/- | ✓ |
| Wegmann et al. (2022) | 0.84/0.69 | 0.59/0.26 | 0.56/0.03 | 0.96/0.02 | 0.74/0.25 | × |
| LISA | 0.69/0.07 | 0.57/0.01 | 0.80/0.03 | 0.77/0.00 | 0.71/0.03 | ✓ |
| LISA (Wegmann + w) | 0.72/0.07 | 0.61/0.03 | 0.81/0.08 | 0.68/0.00 | 0.71/0.05 | ✓ |
| LISA (Wegmann + W) | 0.66/0.03 | 0.56/0.01 | 0.70/0.01 | 0.87/0.00 | 0.70/0.01 | ✓ |
| LISA (LUAR + w) | 0.73/0.05 | 0.65/0.00 | 0.85/0.03 | 0.92/0.00 | 0.79/0.02 | ✓ |
| LISA (LUAR + W) | 0.81/0.07 | 0.56/0.01 | 0.74/0.03 | 0.82/0.00 | 0.73/0.03 | ✓ |
| Text | Style Attribute |
| Subscribed. Interesting idea. I would like to see some advanced stats on hitting percentages to different locations on the court. For example, having the court broken up into maybe 12 zones and then hitting percentages from each position to those zones. I remember seeing an article that did this years ago and I have never been able to find anything online. I said recently on this sub that the deep angle shot from the left side or right side was the highest percentage shot in volleyball, but I was not able to back up my claim with any sources or anything. Anyways, I am a VB nerd, no doubt. Interested to see what direction you take this. Cheers! | The author is being polite. |
| Yeah I also work in QA, and seeing this kind of stuff get released is maddening. About a year ago working on a new platform we were seeing bugs in the hundreds each week, we pushed back the release of the product 3 months because basically it didn't work. If it was up to the devs, they'd have released it on time, because the stuff they'd written code for worked. Thorough doesn't even cover the work we go through every 3 months, and Niantic's approach seems completely amateur from this side. They're putting bandaids on problems and hiding things like the 3 step problem behind curtains without seemingly fixing anything, although I do have to say their balance tweaks to battling have been a big step in the right direction. | The author is using a personal anecdote to illustrate their point. |
| Thank you. I'd be interested in reading more about your experiences, in addition to the "American Wedding" story. Are you watching the stream? I wish there was a way to find out how many people in the world are watching it. The music is lovely, huh? God damn. He's got his bunny Fair Isle sweater on, drinking Dunkin' Donuts coffee. I would have thought him a Starbucks man. :-) | The author is using an emoji. |
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot. +This is why devices that have no switches (primarily wall warts) need not have polarized plugs. +Same Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood. +Different Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) – The author is using a scientific approach. +(0.96, 0.98) – The author is using a combination of technical terms and everyday language. +(0.91, 0.84) – The author is using formal and professional language. +(0.06, 0.99) – The author is using words related to the game they are discussing. * +(0.00, 0.88) – The author is using an emoji. +(0.02, 0.87) – The author uses an emoticon at the end. |
| Prompts Used to Generate STYLEGENOME | Validation F1 |
| Open-ended Prompts | 0.865 |
| Targeted Prompts | 0.898 |
| Open-ended Prompts & Targeted Prompts | 0.920 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | 0.788 | |
| IMDB Large Movie Review Dataset | 0.665 | ||
| Amazon Customer Reviews Dataset | The author uses a negative tone. | 0.432 | |
| Rotten Tomatoes Movie Review Data | 0.463 | ||
| App Reviews | 0.350 | ||
| Twitter Sentiment Analysis Training Corpus | 0.299 | ||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | 0.542 | |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| GoEmotions (Disappointment vs. Admiration) | 0.571 | ||
| GoEmotions (Pride vs. Embarrassment) | 0.419 | ||
| GoEmotions (Nervousness vs. Optimism) | 0.447 | ||
| GoEmotions (Disapproval vs. Approval) | 0.432 | ||
| GoEmotions (Admiration vs. Neutral) | 0.578 | ||
| GoEmotions (Amusement vs. Neutral) | 0.508 | ||
| GoEmotions (Anger vs. Neutral) | 0.372 | ||
| GoEmotions (Annoyance vs. Neutral) | 0.355 | ||
| GoEmotions (Approval vs. Neutral) | 0.257 | ||
| GoEmotions (Caring vs. Neutral) | 0.303 | ||
| GoEmotions (Confusion vs. Neutral) | 0.343 | ||
| GoEmotions (Curiosity vs. Neutral) | 0.464 | ||
| GoEmotions (Desire vs. Neutral) | 0.315 | ||
| GoEmotions (Disappointment vs. Neutral) | The author is expressing {{emotion}}. | 0.317 | |
| GoEmotions (Disapproval vs. Neutral) | 0.284 | ||
| GoEmotions (Disgust vs. Neutral) | 0.307 | ||
| GoEmotions (Embarrassment vs. Neutral) | 0.173 | ||
| GoEmotions (Excitement vs. Neutral) | 0.295 | ||
| GoEmotions (Fear vs. Neutral) | 0.309 | ||
| GoEmotions (Gratitude vs. Neutral) | 0.626 | ||
| GoEmotions (Grief vs. Neutral) | 0.093 | ||
| GoEmotions (Joy vs. Neutral) | 0.390 | ||
| GoEmotions (Love vs. Neutral) | 0.497 | ||
| GoEmotions (Nervousness vs. Neutral) | 0.171 | ||
| GoEmotions (Optimism vs. Neutral) | 0.407 | ||
| GoEmotions (Pride vs. Neutral) | 0.097 | ||
| GoEmotions (Realization vs. Neutral) | 0.121 | ||
| GoEmotions (Relief vs. Neutral) | 0.123 | ||
| GoEmotions (Remorse vs. Neutral) | 0.276 | ||
| GoEmotions (Sadness vs. Neutral) | 0.413 | ||
| GoEmotions (Surprise vs. Neutral) | 0.317 | ||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 | ||
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot. +This is why devices that have no switches (primarily wall warts) need not have polarized plugs. +Same Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood. +Different Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) - The author is using a scientific approach. +(0.96, 0.98) - The author is using a combination of technical terms and everyday language. +(0.91, 0.84) - The author is using formal and professional language. |
| (0.06, 0.99) - The author is using words related to the game they are discussing. * +(0.00, 0.88) - The author is using an emoji. +(0.02, 0.87) - The author uses an emoticon at the end. | |
| Anchor: Not sure what the income tax is in Germany, but in the Netherlands the income can be up 50% for the higher income classes. +Same Author: The salaries in the US alway blow my mind. A software developer in Amsterdam gets like €40.000/year, maybe €50.000/year if your good, and maybe €60.000/year if you're some kind of manager. Anything position over €100.000/year is basically running the entire company. +Different Author: How would you even test this software? The setup would be just insane. | (0.84, 0.90) - The author is using words indicating poverty. +(0.87, 1.00) - The author is using words indicating wealth. +(0.80, 0.93) - The author is using words related to money. +(0.00, 1.00) - The author is comfortable with technology. +(0.00, 0.85) - The author is discussing a product. * +(0.13, 0.92) - The author is using formal and professional language. |
| Anchor: If only there was something he could have done to avoid this backlash. Like maybe not acting like a complete d**khead. +Same Author: I take issue with a faster landing being marked as less skilled. By that logic the slowest, smoothest possible landing would be the most skilled and that is plain wrong. Maybe war machine intentionally does faster and harder landings. +Different Author: She was the Ronald Reagan of the UK in the same time period. | (1.00, 0.96) - The author is emphasizing the contrast between the two ideas. +(0.78, 0.89) - The author is able to draw conclusions. +(0.97, 0.86) - The author is using an all-or-none thinking style. +(0.83, 0.05) - The author is describing sexual content. * +(0.38, 0.94) - The author is using words related to politics. * +(0.74, 0.38) - The author is using parentheticals. |
| # of Reddit Authors | 1,000 |
| # of Reddit Posts | 10,000 |
| # of Interpretable Style Attributes | 1,255,874 |
| # of (Text, Style Attribute) labeled pairs | 5,490,847 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | The author uses a negative tone. | 0.788 |
| IMDB Large Movie Review Dataset | 0.665 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | The author is expressing {{emotion}}. | 0.542 |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| ...abbreviated for space, see Appendix F for full results | |||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 |
| Text | Top 5 LISA Vector Dimensions |
| Lol right on | 1. 1.00 – The author is being polite.\n2. 1.00 – The author is writing in a cheerful manner.\n3. 1.00 – The author is using a lighthearted tone.\n4. 1.00 – The author is laughing.\n5. 1.00 – The author is complimentary. |
| There is the Toyota GT86 R3 ;) http://www.toyota-motorsport.com/motorsport/downloads/com_droppics/59/DSF4311.jpg | 1. 0.99 – The author is simply describing a product. * \n2. 0.99 – The author is providing a visual cue to the reader.\n3. 0.98 – The author simply provides information.\n4. 0.97 – The author is using an emoji.\n5. 0.97 – The author provides information. |
| Every time i watched this episode as a kid i was always like "WTF, JAMES, A POKEBALL ISNT EVEN A POKE-MON?! GET YOUR ACT TOGETHER, SON!!" | 1. 1.00 – The author is discussing a television show. * \n2. 0.99 – The author has a hint of nostalgia.\n3. 0.99 – The author is making a humorous comment.\n4. 0.99 – The author is animated in their writing.\n5. 0.99 – The author is laughing. |
| 13 POT MORDE BABY WOOOOOOOOOOOOOO | 1. 1.00 – The author is using an elongated word.\n2. 1.00 – The author is using only English words.\n3. 1.00 – The author is using a single word.\n4. 1.00 – The author is using swear words.\n5. 1.00 – The author uses two exclamation marks. |
| No wonder everyone resorts to performing a murder spree eventually. | 1. 1.00 – The author is scornful.\n2. 1.00 – The author is ungenerous.\n3. 0.99 – The author is expressing antisocial behaviors.\n4. 0.99 – The author is uncaring.\n5. 0.99 – The author is dramatic. |
| Podcast originally refers to an iPod, and before that there was definitely TWiT, which still calls itself a Netcast | 1. 0.97 – The author uses a variety of words to describe the same concept.\n2. 0.97 – The author is simply describing a product. * \n3. 0.95 – The author uses specific terms related to the topic.\n4. 0.94 – The author has a deep understanding of the topic.\n5. 0.94 – The author is using words focusing on the past. |
| Model | Formal | Complex | Numb3r | C'tion | Avg | Interpretable |
| Random Baseline | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | 0.50/0.50 | |
| Content-Aware Representations | ||||||
| SBERT | 0.78/0.00 | 0.54/0.01 | 0.81/0.04 | 0.86/0.00 | 0.75/0.01 | × |
| LUAR | 0.80/0.14 | 0.67/0.00 | 0.74/0.03 | 0.77/0.00 | 0.75/0.04 | × |
| Content-Independent Style Representations | ||||||
| LIWC | 0.52/- | 0.52/- | 0.50/- | 0.99/- | 0.63/- | ✓ |
| Wegmann et al. (2022) | 0.84/0.69 | 0.59/0.26 | 0.56/0.03 | 0.96/0.02 | 0.74/0.25 | × |
| LISA | 0.69/0.07 | 0.57/0.01 | 0.80/0.03 | 0.77/0.00 | 0.71/0.03 | ✓ |
| LISA (Wegmann + w) | 0.72/0.07 | 0.61/0.03 | 0.81/0.08 | 0.68/0.00 | 0.71/0.05 | ✓ |
| LISA (Wegmann + W) | 0.66/0.03 | 0.56/0.01 | 0.70/0.01 | 0.87/0.00 | 0.70/0.01 | ✓ |
| LISA (LUAR + w) | 0.73/0.05 | 0.65/0.00 | 0.85/0.03 | 0.92/0.00 | 0.79/0.02 | ✓ |
| LISA (LUAR + W) | 0.81/0.07 | 0.56/0.01 | 0.74/0.03 | 0.82/0.00 | 0.73/0.03 | ✓ |
| Text | Style Attribute |
| Subscribed. Interesting idea. I would like to see some advanced stats on hitting percentages to different locations on the court. For example, having the court broken up into maybe 12 zones and then hitting percentages from each position to those zones. I remember seeing an article that did this years ago and I have never been able to find anything online. I said recently on this sub that the deep angle shot from the left side or right side was the highest percentage shot in volleyball, but I was not able to back up my claim with any sources or anything. Anyways, I am a VB nerd, no doubt. Interested to see what direction you take this. Cheers! | The author is being polite. |
| Yeah I also work in QA, and seeing this kind of stuff get released is maddening. About a year ago working on a new platform we were seeing bugs in the hundreds each week, we pushed back the release of the product 3 months because basically it didn't work. If it was up to the devs, they'd have released it on time, because the stuff they'd written code for worked. Thorough doesn't even cover the work we go through every 3 months, and Niantic's approach seems completely amateur from this side. They're putting bandaids on problems and hiding things like the 3 step problem behind curtains without seemingly fixing anything, although I do have to say their balance tweaks to battling have been a big step in the right direction. | The author is using a personal anecdote to illustrate their point. |
| Thank you. I'd be interested in reading more about your experiences, in addition to the "American Wedding" story. Are you watching the stream? I wish there was a way to find out how many people in the world are watching it. The music is lovely, huh? God damn. He's got his bunny Fair Isle sweater on, drinking Dunkin' Donuts coffee. I would have thought him a Starbucks man. :-) | The author is using an emoji. |
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot.\nThis is why devices that have no switches (primarily wall warts) need not have polarized plugs.\nSame Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood.\nDifferent Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) – The author is using a scientific approach.\n(0.96, 0.98) – The author is using a combination of technical terms and everyday language.\n(0.91, 0.84) – The author is using formal and professional language.\n(0.06, 0.99) – The author is using words related to the game they are discussing. *\n(0.00, 0.88) – The author is using an emoji.\n(0.02, 0.87) – The author uses an emoticon at the end. |
| Prompts Used to Generate STYLEGENOME | Validation F1 |
| Open-ended Prompts | 0.865 |
| Targeted Prompts | 0.898 |
| Open-ended Prompts & Targeted Prompts | 0.920 |
| Type | Dataset | Style Attribute | Spearman Correlation (ρ) |
| Formality | Formality in Online Communication | The author uses informal language. | 0.599 |
| Grammarly's Yahoo Answers Formality Corpus | 0.200 | ||
| Sentiment | Yelp Reviews | 0.788 | |
| IMDB Large Movie Review Dataset | 0.665 | ||
| Amazon Customer Reviews Dataset | The author uses a negative tone. | 0.432 | |
| Rotten Tomatoes Movie Review Data | 0.463 | ||
| App Reviews | 0.350 | ||
| Twitter Sentiment Analysis Training Corpus | 0.299 | ||
| Emotion | DAIR.AI Emotion (Love vs. Anger) | 0.542 | |
| DAIR.AI Emotion (Joy vs. Sad) | 0.531 | ||
| GoEmotions (Love vs. Anger) | 0.769 | ||
| GoEmotions (Joy vs. Sadness) | 0.639 | ||
| GoEmotions (Disgust vs. Desire) | 0.630 | ||
| GoEmotions (Disappointment vs. Admiration) | 0.571 | ||
| GoEmotions (Pride vs. Embarrassment) | 0.419 | ||
| GoEmotions (Nervousness vs. Optimism) | 0.447 | ||
| GoEmotions (Disapproval vs. Approval) | 0.432 | ||
| GoEmotions (Admiration vs. Neutral) | 0.578 | ||
| GoEmotions (Amusement vs. Neutral) | 0.508 | ||
| GoEmotions (Anger vs. Neutral) | 0.372 | ||
| GoEmotions (Annoyance vs. Neutral) | 0.355 | ||
| GoEmotions (Approval vs. Neutral) | 0.257 | ||
| GoEmotions (Caring vs. Neutral) | 0.303 | ||
| GoEmotions (Confusion vs. Neutral) | 0.343 | ||
| GoEmotions (Curiosity vs. Neutral) | 0.464 | ||
| GoEmotions (Desire vs. Neutral) | 0.315 | ||
| GoEmotions (Disappointment vs. Neutral) | The author is expressing {{emotion}}. | 0.317 | |
| GoEmotions (Disapproval vs. Neutral) | 0.284 | ||
| GoEmotions (Disgust vs. Neutral) | 0.307 | ||
| GoEmotions (Embarrassment vs. Neutral) | 0.173 | ||
| GoEmotions (Excitement vs. Neutral) | 0.295 | ||
| GoEmotions (Fear vs. Neutral) | 0.309 | ||
| GoEmotions (Gratitude vs. Neutral) | 0.626 | ||
| GoEmotions (Grief vs. Neutral) | 0.093 | ||
| GoEmotions (Joy vs. Neutral) | 0.390 | ||
| GoEmotions (Love vs. Neutral) | 0.497 | ||
| GoEmotions (Nervousness vs. Neutral) | 0.171 | ||
| GoEmotions (Optimism vs. Neutral) | 0.407 | ||
| GoEmotions (Pride vs. Neutral) | 0.097 | ||
| GoEmotions (Realization vs. Neutral) | 0.121 | ||
| GoEmotions (Relief vs. Neutral) | 0.123 | ||
| GoEmotions (Remorse vs. Neutral) | 0.276 | ||
| GoEmotions (Sadness vs. Neutral) | 0.413 | ||
| GoEmotions (Surprise vs. Neutral) | 0.317 | ||
| Author Profiling | Political Slant | The author is a Democrat. | 0.005 |
| Twitter User Gender Classification | The author is female. | 0.166 | |
| African-American Vernacular English | The author uses African-American Vernacular English. | 0.238 | |
| Shakespeare (Early Modern English) | The author uses Early Modern English. | 0.108 | |
| Wikipedia Bias | The author has a biased point of view. | 0.014 | |
| Harmful Speech | HateSpeech18 | The author's writing contains hate speech. | 0.229 |
| Offensive Social Media | The author uses offensive language. | 0.401 | |
| Text Simplification | Simple Wikipedia | The author uses simple language. | 0.043 |
| ASSET | 0.050 | ||
| Linguistic Acceptability | CoLA | The author uses incorrect grammar. | 0.078 |
| BLiMP | 0.020 | ||
| Average | 0.342 |
| Texts | Top 3 Common/Distinct Style Attributes |
| Anchor: Devices that use two pronged instead of three pronged plugs are required to meet certain safe design requirements. Among other things, if a device has a switch, the switched line MUST BE hot, not neutral. The polarized plugs make sure that the right prong/wire is hot.\nThis is why devices that have no switches (primarily wall warts) need not have polarized plugs.\nSame Author: Your diaphragm would be trying to contract against the air pressure in your lungs. That's why deep sea diving requires regulators, to match the pressure of the air supply to the pressure surrounding your rib cage. You can breathe against a maximum of about 1/2 PSI, which is not enough pressure to adequately oxygenate your blood.\nDifferent Author: That's great! I'm glad it seems to be finding its' niche. Now if they could just make a Star Wars version of this game, I'd happily swallow that fat learning curve and overcome my frustrations with the combat system. ;) | (0.89, 1.00) - The author is using a scientific approach.\n(0.96, 0.98) - The author is using a combination of technical terms and everyday language.\n(0.91, 0.84) - The author is using formal and professional language. |
| (0.06, 0.99) - The author is using words related to the game they are discussing. *\n(0.00, 0.88) - The author is using an emoji.\n(0.02, 0.87) - The author uses an emoticon at the end. | |
| Anchor: Not sure what the income tax is in Germany, but in the Netherlands the income can be up 50% for the higher income classes.\nSame Author: The salaries in the US alway blow my mind. A software developer in Amsterdam gets like €40.000/year, maybe €50.000/year if your good, and maybe €60.000/year if you're some kind of manager. Anything position over €100.000/year is basically running the entire company.\nDifferent Author: How would you even test this software? The setup would be just insane. | (0.84, 0.90) - The author is using words indicating poverty.\n(0.87, 1.00) - The author is using words indicating wealth.\n(0.80, 0.93) - The author is using words related to money.\n(0.00, 1.00) - The author is comfortable with technology.\n(0.00, 0.85) - The author is discussing a product. *\n(0.13, 0.92) - The author is using formal and professional language. |
| Anchor: If only there was something he could have done to avoid this backlash. Like maybe not acting like a complete d**khead.\nSame Author: I take issue with a faster landing being marked as less skilled. By that logic the slowest, smoothest possible landing would be the most skilled and that is plain wrong. Maybe war machine intentionally does faster and harder landings.\nDifferent Author: She was the Ronald Reagan of the UK in the same time period. | (1.00, 0.96) - The author is emphasizing the contrast between the two ideas.\n(0.78, 0.89) - The author is able to draw conclusions.\n(0.97, 0.86) - The author is using an all-or-none thinking style.\n(0.83, 0.05) - The author is describing sexual content. *\n(0.38, 0.94) - The author is using words related to politics. *\n(0.74, 0.38) - The author is using parentheticals. |
| η | σ | Admissible alignments (yV - yP) |
| price-54.4 | 01 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| price-54.4 | 02 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| admire-31.2 | 02 | Experiencer-Arg0; Stimulus-Arg1 |
| Data | Train\nFine-tune×2 | Inf\nSEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | IWCS2021 (gold η & σ) | - | - | - | 88.2 | 88.7 | - | 83.0 | 82.8 | - | |
| Multitask | X | 89.51.09 | 87.91.17 | 3.68.06 | 90.74.23 | 89.33.15 | 3.43.50 | 83.691.54 | 81.73.98 | 8.71.74 | |
| Joint | ✓ | 90.20.22 | 87.58.18 | 0 | 91.42.11 | 89.12.16 | 0 | 85.901.11 | 82.56.52 | 0 | |
| Semi | Multitask | X | 89.56.09 | 88.30.42 | 4.82.35 | 90.45.09 | 89.34.74 | 4.08.37 | 84.28.85 | 82.51.70 | 10.48.16 |
| Joint+CRFpB | ✓ | 90.73.32 | 88.42.11 | 0 | 91.52.36 | 89.30.29 | 0 | 85.72.13 | 82.33.31 | 0 | |
| Marginal | ✓ | 90.91.27 | 88.68.11 | 0 | 91.74.22 | 89.51.28 | 0 | 85.391.74 | 82.461.06 | 0 | |
| MarginalSEML | ✓ | 90.87.32 | 88.59.24 | 0 | 91.55.28 | 89.49.18 | 0 | 86.39.17 | 83.04.37 | 0 |
| Data | Train\nFine-tune×2 | Inf\nSEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | Joint | X | 89.28 | 87.70 | 1.58 | 90.73 | 89.10 | 1.59 | 84.43 | 82.48 | 5.99 |
| Joint | ✓ | 90.20 | 87.58 | 0 | 91.42 | 89.12 | 0 | 85.90 | 82.56 | 0 | |
| Semi | Joint+CRFpB | X | 89.66 | 88.31 | 2.74 | 90.61 | 89.21 | 2.37 | 83.53 | 81.87 | 7.59 |
| Joint+CRFpB | ✓ | 90.73 | 88.42 | 0 | 91.52 | 89.30 | 0 | 85.72 | 82.33 | 0 | |
| Marginal | X | 89.79 | 88.50 | 2.92 | 90.73 | 89.30 | 2.39 | 83.23 | 81.88 | 6.74 | |
| Marginal | ✓ | 90.91 | 88.68 | 0 | 91.74 | 89.51 | 0 | 85.39 | 82.46 | 0 | |
| MarginalSEML | X | 89.53 | 88.55 | 2.29 | 90.70 | 89.46 | 1.98 | 83.64 | 82.71 | 7.77 | |
| MarginalSEML | ✓ | 90.87 | 88.59 | 0 | 91.55 | 89.49 | 0 | 86.39 | 83.04 | 0 |
| Data | Model | WSJ VN | Brown VN |
| xWP | |||
| Joint | Multitask | 83.14 | 86.26 |
| Joint | 99.83 | 98.12 | |
| Semi | Multitask | 93.08 | 85.71 |
| Joint+CRFpB | 99.81 | 97.88 | |
| Marginal | 99.76 | 97.45 | |
| MarginalSEML | 99.68 | 97.71 | |
| xCOMP | |||
| Joint | Joint | 99.85 | 99.02 |
| Joint | WSJ | Brown | |||
| SEMLINK | VN | PB | VB | PB | |
| 1 | × | 91.42 | 89.12 | 85.90 | 82.56 |
| 2 | ✓ | 91.42 | 89.21 | 85.41 | 82.46 |
| η corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 89.45 | 83.64 | 82.90 | 79.02 |
| PB F1 | 88.82 | 88.23 | 87.32 | 85.56 | 84.66 |
| σ corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 90.35 | 89.55 | 87.18 | 84.94 |
| PB F1 | 88.82 | 88.17 | 87.21 | 84.92 | 83.11 |
| Data | Model | λ |
| Joint | Multitask | 1 × 10-4 |
| Joint | 1 × 10-4 | |
| Semi | Multitask | 1 × 10-4 |
| Joint+CRFpB | 1 × 10-4 | |
| Marginal | 3 × 10-5 | |
| MarginalSEML | 1 × 10-5 |
| η | σ | Admissible alignments (yV - yP) |
| price-54.4 | 01 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| price-54.4 | 02 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| admire-31.2 | 02 | Experiencer-Arg0; Stimulus-Arg1 |
| Data | Train\nFine-tune×2 | Inf\nSEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | IWCS2021 (gold η & σ) | - | - | - | 88.2 | 88.7 | - | 83.0 | 82.8 | - | |
| Multitask | X | 89.51.09 | 87.91.17 | 3.68.06 | 90.74.23 | 89.33.15 | 3.43.50 | 83.691.54 | 81.73.98 | 8.71.74 | |
| Joint | ✓ | 90.20.22 | 87.58.18 | 0 | 91.42.11 | 89.12.16 | 0 | 85.901.11 | 82.56.52 | 0 | |
| Semi | Multitask | X | 89.56.09 | 88.30.42 | 4.82.35 | 90.45.09 | 89.34.74 | 4.08.37 | 84.28.85 | 82.51.70 | 10.48.16 |
| Joint+CRFpB | ✓ | 90.73.32 | 88.42.11 | 0 | 91.52.36 | 89.30.29 | 0 | 85.72.13 | 82.33.31 | 0 | |
| Marginal | ✓ | 90.91.27 | 88.68.11 | 0 | 91.74.22 | 89.51.28 | 0 | 85.391.74 | 82.461.06 | 0 | |
| MarginalSEML | ✓ | 90.87.32 | 88.59.24 | 0 | 91.55.28 | 89.49.18 | 0 | 86.39.17 | 83.04.37 | 0 |
| Data | Train\nFine-tune×2 | Inf\nSEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | Joint | X | 89.28 | 87.70 | 1.58 | 90.73 | 89.10 | 1.59 | 84.43 | 82.48 | 5.99 |
| Joint | ✓ | 90.20 | 87.58 | 0 | 91.42 | 89.12 | 0 | 85.90 | 82.56 | 0 | |
| Semi | Joint+CRFpB | X | 89.66 | 88.31 | 2.74 | 90.61 | 89.21 | 2.37 | 83.53 | 81.87 | 7.59 |
| Joint+CRFpB | ✓ | 90.73 | 88.42 | 0 | 91.52 | 89.30 | 0 | 85.72 | 82.33 | 0 | |
| Marginal | X | 89.79 | 88.50 | 2.92 | 90.73 | 89.30 | 2.39 | 83.23 | 81.88 | 6.74 | |
| Marginal | ✓ | 90.91 | 88.68 | 0 | 91.74 | 89.51 | 0 | 85.39 | 82.46 | 0 | |
| MarginalSEML | X | 89.53 | 88.55 | 2.29 | 90.70 | 89.46 | 1.98 | 83.64 | 82.71 | 7.77 | |
| MarginalSEML | ✓ | 90.87 | 88.59 | 0 | 91.55 | 89.49 | 0 | 86.39 | 83.04 | 0 |
| Data | Model | WSJ VN | Brown VN |
| xWP | |||
| Joint | Multitask | 83.14 | 86.26 |
| Joint | 99.83 | 98.12 | |
| Semi | Multitask | 93.08 | 85.71 |
| Joint+CRFpB | 99.81 | 97.88 | |
| Marginal | 99.76 | 97.45 | |
| MarginalSEML | 99.68 | 97.71 | |
| xCOMP | |||
| Joint | Joint | 99.85 | 99.02 |
| Joint | WSJ | Brown | |||
| SEMLINK | VN | PB | VB | PB | |
| 1 | × | 91.42 | 89.12 | 85.90 | 82.56 |
| 2 | ✓ | 91.42 | 89.21 | 85.41 | 82.46 |
| η corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 89.45 | 83.64 | 82.90 | 79.02 |
| PB F1 | 88.82 | 88.23 | 87.32 | 85.56 | 84.66 |
| σ corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 90.35 | 89.55 | 87.18 | 84.94 |
| PB F1 | 88.82 | 88.17 | 87.21 | 84.92 | 83.11 |
| Data | Model | λ |
| Joint | Multitask | 1 × 10-4 |
| Joint | 1 × 10-4 | |
| Semi | Multitask | 1 × 10-4 |
| Joint+CRFpB | 1 × 10-4 | |
| Marginal | 3 × 10-5 | |
| MarginalSEML | 1 × 10-5 |
| η | σ | Admissible alignments (yV - yP) |
| price-54.4 | 01 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| price-54.4 | 02 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| admire-31.2 | 02 | Experiencer-Arg0; Stimulus-Arg1 |
| Data | Train +Fine-tune×2 | Inf +SEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | IWCS2021 (gold η & σ) | - | - | - | 88.2 | 88.7 | - | 83.0 | 82.8 | - | |
| Multitask | X | 89.51.09 | 87.91.17 | 3.68.06 | 90.74.23 | 89.33.15 | 3.43.50 | 83.691.54 | 81.73.98 | 8.71.74 | |
| Joint | ✓ | 90.20.22 | 87.58.18 | 0 | 91.42.11 | 89.12.16 | 0 | 85.901.11 | 82.56.52 | 0 | |
| Semi | Multitask | X | 89.56.09 | 88.30.42 | 4.82.35 | 90.45.09 | 89.34.74 | 4.08.37 | 84.28.85 | 82.51.70 | 10.48.16 |
| Joint+CRFpB | ✓ | 90.73.32 | 88.42.11 | 0 | 91.52.36 | 89.30.29 | 0 | 85.72.13 | 82.33.31 | 0 | |
| Marginal | ✓ | 90.91.27 | 88.68.11 | 0 | 91.74.22 | 89.51.28 | 0 | 85.391.74 | 82.461.06 | 0 | |
| MarginalSEML | ✓ | 90.87.32 | 88.59.24 | 0 | 91.55.28 | 89.49.18 | 0 | 86.39.17 | 83.04.37 | 0 | |
| Data | Train +Fine-tune×2 | Inf +SEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | Joint | X | 89.28 | 87.70 | 1.58 | 90.73 | 89.10 | 1.59 | 84.43 | 82.48 | 5.99 |
| Joint | ✓ | 90.20 | 87.58 | 0 | 91.42 | 89.12 | 0 | 85.90 | 82.56 | 0 | |
| Semi | Joint+CRFpB | X | 89.66 | 88.31 | 2.74 | 90.61 | 89.21 | 2.37 | 83.53 | 81.87 | 7.59 |
| Joint+CRFpB | ✓ | 90.73 | 88.42 | 0 | 91.52 | 89.30 | 0 | 85.72 | 82.33 | 0 | |
| Marginal | X | 89.79 | 88.50 | 2.92 | 90.73 | 89.30 | 2.39 | 83.23 | 81.88 | 6.74 | |
| Marginal | ✓ | 90.91 | 88.68 | 0 | 91.74 | 89.51 | 0 | 85.39 | 82.46 | 0 | |
| MarginalSEML | X | 89.53 | 88.55 | 2.29 | 90.70 | 89.46 | 1.98 | 83.64 | 82.71 | 7.77 | |
| MarginalSEML | ✓ | 90.87 | 88.59 | 0 | 91.55 | 89.49 | 0 | 86.39 | 83.04 | 0 | |
| Data | Model | WSJ VN | Brown VN |
| xWP | |||
| Joint | Multitask | 83.14 | 86.26 |
| Joint | 99.83 | 98.12 | |
| Semi | Multitask | 93.08 | 85.71 |
| Joint+CRFpB | 99.81 | 97.88 | |
| Marginal | 99.76 | 97.45 | |
| MarginalSEML | 99.68 | 97.71 | |
| xCOMP | |||
| Joint | Joint | 99.85 | 99.02 |
| Joint | WSJ | Brown | |||
| SEMLINK | VN | PB | VB | PB | |
| 1 | × | 91.42 | 89.12 | 85.90 | 82.56 |
| 2 | ✓ | 91.42 | 89.21 | 85.41 | 82.46 |
| η corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 89.45 | 83.64 | 82.90 | 79.02 |
| PB F1 | 88.82 | 88.23 | 87.32 | 85.56 | 84.66 |
| σ corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 90.35 | 89.55 | 87.18 | 84.94 |
| PB F1 | 88.82 | 88.17 | 87.21 | 84.92 | 83.11 |
| Data | Model | λ |
| Joint | Multitask | 1 × 10-4 |
| Joint | 1 × 10-4 | |
| Semi | Multitask | 1 × 10-4 |
| Joint+CRFpB | 1 × 10-4 | |
| Marginal | 3 × 10-5 | |
| MarginalSEML | 1 × 10-5 |
| η | σ | Admissible alignments (yV - yP) |
| price-54.4 | 01 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| price-54.4 | 02 | Agent-Arg0; Theme-Arg1; Value-Arg2 |
| admire-31.2 | 02 | Experiencer-Arg0; Stimulus-Arg1 |
| Data | Train\nFine-tune×2 | Inf\nSEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | IWCS2021 (gold η & σ) | - | - | - | 88.2 | 88.7 | - | 83.0 | 82.8 | - | |
| Multitask | X | 89.51.09 | 87.91.17 | 3.68.06 | 90.74.23 | 89.33.15 | 3.43.50 | 83.691.54 | 81.73.98 | 8.71.74 | |
| Joint | ✓ | 90.20.22 | 87.58.18 | 0 | 91.42.11 | 89.12.16 | 0 | 85.901.11 | 82.56.52 | 0 | |
| Semi | Multitask | X | 89.56.09 | 88.30.42 | 4.82.35 | 90.45.09 | 89.34.74 | 4.08.37 | 84.28.85 | 82.51.70 | 10.48.16 |
| Joint+CRFpB | ✓ | 90.73.32 | 88.42.11 | 0 | 91.52.36 | 89.30.29 | 0 | 85.72.13 | 82.33.31 | 0 | |
| Marginal | ✓ | 90.91.27 | 88.68.11 | 0 | 91.74.22 | 89.51.28 | 0 | 85.391.74 | 82.461.06 | 0 | |
| MarginalSEML | ✓ | 90.87.32 | 88.59.24 | 0 | 91.55.28 | 89.49.18 | 0 | 86.39.17 | 83.04.37 | 0 |
| Data | Train\nFine-tune×2 | Inf\nSEML | Dev | WSJ | Brown | ||||||
| VN | PB | ρ↓ | VN | PB | ρ↓ | VN | PB | ρ↓ | |||
| Joint | Joint | X | 89.28 | 87.70 | 1.58 | 90.73 | 89.10 | 1.59 | 84.43 | 82.48 | 5.99 |
| Joint | ✓ | 90.20 | 87.58 | 0 | 91.42 | 89.12 | 0 | 85.90 | 82.56 | 0 | |
| Semi | Joint+CRFpB | X | 89.66 | 88.31 | 2.74 | 90.61 | 89.21 | 2.37 | 83.53 | 81.87 | 7.59 |
| Joint+CRFpB | ✓ | 90.73 | 88.42 | 0 | 91.52 | 89.30 | 0 | 85.72 | 82.33 | 0 | |
| Marginal | X | 89.79 | 88.50 | 2.92 | 90.73 | 89.30 | 2.39 | 83.23 | 81.88 | 6.74 | |
| Marginal | ✓ | 90.91 | 88.68 | 0 | 91.74 | 89.51 | 0 | 85.39 | 82.46 | 0 | |
| MarginalSEML | X | 89.53 | 88.55 | 2.29 | 90.70 | 89.46 | 1.98 | 83.64 | 82.71 | 7.77 | |
| MarginalSEML | ✓ | 90.87 | 88.59 | 0 | 91.55 | 89.49 | 0 | 86.39 | 83.04 | 0 |
| Data | Model | WSJ VN | Brown VN |
| xWP | |||
| Joint | Multitask | 83.14 | 86.26 |
| Joint | 99.83 | 98.12 | |
| Semi | Multitask | 93.08 | 85.71 |
| Joint+CRFpB | 99.81 | 97.88 | |
| Marginal | 99.76 | 97.45 | |
| MarginalSEML | 99.68 | 97.71 | |
| xCOMP | |||
| Joint | Joint | 99.85 | 99.02 |
| Joint | WSJ | Brown | |||
| SEMLINK | VN | PB | VB | PB | |
| 1 | × | 91.42 | 89.12 | 85.90 | 82.56 |
| 2 | ✓ | 91.42 | 89.21 | 85.41 | 82.46 |
| η corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 89.45 | 83.64 | 82.90 | 79.02 |
| PB F1 | 88.82 | 88.23 | 87.32 | 85.56 | 84.66 |
| σ corruption (%) | 0 | 5 | 10 | 20 | 30 |
| VN F1 | 91.23 | 90.35 | 89.55 | 87.18 | 84.94 |
| PB F1 | 88.82 | 88.17 | 87.21 | 84.92 | 83.11 |
| Data | Model | λ |
| Joint | Multitask | 1 × 10-4 |
| Joint | 1 × 10-4 | |
| Semi | Multitask | 1 × 10-4 |
| Joint+CRFpB | 1 × 10-4 | |
| Marginal | 3 × 10-5 | |
| MarginalSEML | 1 × 10-5 |
| P | R | F1 | |
| Transformer | 95.51 | 56.51 | 64.52 |
| NVIB | 85.23 | 79.02 | 78.86 |
| Task | Transformer | NVIB |
| Computer science | 42.33 | 44.47 |
| Mathematics | 44.02 | 47.13 |
| Physics | 48.83 | 52.32 |
| Average | 45.06 | 47.97 |
| Transformer | NVIB | |
| NVIB layers | - | 3 |
| λG | - | 1e-2 |
| λD | - | 1 |
| αΔ | - | 0.25 |
| Training Steps | 2.5K | 8K |
| Val. CE | 0.19 | 0.19 |
| Layer | CoordInv | ObjNum | TreeDepth | TopConst | BShift | Tense | SubjNum | |
| Chance | 0.5 | 0.5 | 0.125 | 0.05 | 0.5 | 0.5 | 0.5 | |
| Transformer | 1 | 0.5023 | 0.6498 | 0.2200 | 0.2880 | 0.5006 | 0.7306 | 0.6500 |
| 2 | 0.5144 | 0.7255 | 0.2350 | 0.3724 | 0.4994 | 0.7891 | 0.7131 | |
| 3 | 0.5190 | 0.7547 | 0.2594 | 0.4261 | 0.5055 | 0.8263 | 0.7297 | |
| 4 | 0.5196 | 0.7687 | 0.2692 | 0.4368 | 0.5108 | 0.8114 | 0.7545 | |
| 5 | 0.5196 | 0.7737 | 0.2736 | 0.4369 | 0.5304 | 0.8320 | 0.7435 | |
| 6 | 0.5227 | 0.7756 | 0.2736 | 0.4212 | 0.5465 | 0.8384 | 0.7683 | |
| NVIB | 1 | 0.5037 | 0.7646 | 0.2349 | 0.3323 | 0.5007 | 0.8344 | 0.7285 |
| 2 | 0.5069 | 0.7859 | 0.2511 | 0.4243 | 0.5108 | 0.8379 | 0.7777 | |
| 3 | 0.5110 | 0.7963 | 0.2589 | 0.4453 | 0.5466 | 0.8606 | 0.7844 | |
| 4 | 0.5111 | 0.7879 | 0.2655 | 0.5290 | 0.5361 | 0.8481 | 0.7943 | |
| 5 | 0.5299 | 0.7660 | 0.2651 | 0.5283 | 0.5571 | 0.8371 | 0.7793 | |
| 6 | 0.5523 | 0.8207 | 0.2923 | 0.5766 | 0.6075 | 0.8531 | 0.8038 |
| P | R | F1 | |
| Transformer | 95.51 | 56.51 | 64.52 |
| NVIB | 85.23 | 79.02 | 78.86 |
| Task | Transformer | NVIB |
| Computer science | 42.33 | 44.47 |
| Mathematics | 44.02 | 47.13 |
| Physics | 48.83 | 52.32 |
| Average | 45.06 | 47.97 |
| Transformer | NVIB | |
| NVIB layers | - | 3 |
| λG | - | 1e-2 |
| λD | - | 1 |
| αΔ | - | 0.25 |
| Training Steps | 2.5K | 8K |
| Val. CE | 0.19 | 0.19 |
| Layer | CoordInv | ObjNum | TreeDepth | TopConst | BShift | Tense | SubjNum | |
| Chance | 0.5 | 0.5 | 0.125 | 0.05 | 0.5 | 0.5 | 0.5 | |
| Transformer | 1 | 0.5023 | 0.6498 | 0.2200 | 0.2880 | 0.5006 | 0.7306 | 0.6500 |
| 2 | 0.5144 | 0.7255 | 0.2350 | 0.3724 | 0.4994 | 0.7891 | 0.7131 | |
| 3 | 0.5190 | 0.7547 | 0.2594 | 0.4261 | 0.5055 | 0.8263 | 0.7297 | |
| 4 | 0.5196 | 0.7687 | 0.2692 | 0.4368 | 0.5108 | 0.8114 | 0.7545 | |
| 5 | 0.5196 | 0.7737 | 0.2736 | 0.4369 | 0.5304 | 0.8320 | 0.7435 | |
| 6 | 0.5227 | 0.7756 | 0.2736 | 0.4212 | 0.5465 | 0.8384 | 0.7683 | |
| NVIB | 1 | 0.5037 | 0.7646 | 0.2349 | 0.3323 | 0.5007 | 0.8344 | 0.7285 |
| 2 | 0.5069 | 0.7859 | 0.2511 | 0.4243 | 0.5108 | 0.8379 | 0.7777 | |
| 3 | 0.5110 | 0.7963 | 0.2589 | 0.4453 | 0.5466 | 0.8606 | 0.7844 | |
| 4 | 0.5111 | 0.7879 | 0.2655 | 0.5290 | 0.5361 | 0.8481 | 0.7943 | |
| 5 | 0.5299 | 0.7660 | 0.2651 | 0.5283 | 0.5571 | 0.8371 | 0.7793 | |
| 6 | 0.5523 | 0.8207 | 0.2923 | 0.5766 | 0.6075 | 0.8531 | 0.8038 |
| P | R | F1 | |
| Transformer | 95.51 | 56.51 | 64.52 |
| NVIB | 85.23 | 79.02 | 78.86 |
| Task | Transformer | NVIB |
| Computer science | 42.33 | 44.47 |
| Mathematics | 44.02 | 47.13 |
| Physics | 48.83 | 52.32 |
| Average | 45.06 | 47.97 |
| Transformer | NVIB | |
| NVIB layers | - | 3 |
| λG | - | 1e-2 |
| λD | - | 1 |
| αΔ | - | 0.25 |
| Training Steps | 2.5K | 8K |
| Val. CE | 0.19 | 0.19 |
| Layer | CoordInv | ObjNum | TreeDepth | TopConst | BShift | Tense | SubjNum | |
| Chance | 0.5 | 0.5 | 0.125 | 0.05 | 0.5 | 0.5 | 0.5 | |
| Transformer | 1 | 0.5023 | 0.6498 | 0.2200 | 0.2880 | 0.5006 | 0.7306 | 0.6500 |
| 2 | 0.5144 | 0.7255 | 0.2350 | 0.3724 | 0.4994 | 0.7891 | 0.7131 | |
| 3 | 0.5190 | 0.7547 | 0.2594 | 0.4261 | 0.5055 | 0.8263 | 0.7297 | |
| 4 | 0.5196 | 0.7687 | 0.2692 | 0.4368 | 0.5108 | 0.8114 | 0.7545 | |
| 5 | 0.5196 | 0.7737 | 0.2736 | 0.4369 | 0.5304 | 0.8320 | 0.7435 | |
| 6 | 0.5227 | 0.7756 | 0.2736 | 0.4212 | 0.5465 | 0.8384 | 0.7683 | |
| NVIB | 1 | 0.5037 | 0.7646 | 0.2349 | 0.3323 | 0.5007 | 0.8344 | 0.7285 |
| 2 | 0.5069 | 0.7859 | 0.2511 | 0.4243 | 0.5108 | 0.8379 | 0.7777 | |
| 3 | 0.5110 | 0.7963 | 0.2589 | 0.4453 | 0.5466 | 0.8606 | 0.7844 | |
| 4 | 0.5111 | 0.7879 | 0.2655 | 0.5290 | 0.5361 | 0.8481 | 0.7943 | |
| 5 | 0.5299 | 0.7660 | 0.2651 | 0.5283 | 0.5571 | 0.8371 | 0.7793 | |
| 6 | 0.5523 | 0.8207 | 0.2923 | 0.5766 | 0.6075 | 0.8531 | 0.8038 |
| P | R | F1 | |
| Transformer | 95.51 | 56.51 | 64.52 |
| NVIB | 85.23 | 79.02 | 78.86 |
| Task | Transformer | NVIB |
| Computer science | 42.33 | 44.47 |
| Mathematics | 44.02 | 47.13 |
| Physics | 48.83 | 52.32 |
| Average | 45.06 | 47.97 |
| Transformer | NVIB | |
| NVIB layers | - | 3 |
| λG | - | 1e-2 |
| λD | - | 1 |
| αΔ | - | 0.25 |
| Training Steps | 2.5K | 8K |
| Val. CE | 0.19 | 0.19 |
| Layer | CoordInv | ObjNum | TreeDepth | TopConst | BShift | Tense | SubjNum | |
| Chance | 0.5 | 0.5 | 0.125 | 0.05 | 0.5 | 0.5 | 0.5 | |
| Transformer | 1 | 0.5023 | 0.6498 | 0.2200 | 0.2880 | 0.5006 | 0.7306 | 0.6500 |
| 2 | 0.5144 | 0.7255 | 0.2350 | 0.3724 | 0.4994 | 0.7891 | 0.7131 | |
| 3 | 0.5190 | 0.7547 | 0.2594 | 0.4261 | 0.5055 | 0.8263 | 0.7297 | |
| 4 | 0.5196 | 0.7687 | 0.2692 | 0.4368 | 0.5108 | 0.8114 | 0.7545 | |
| 5 | 0.5196 | 0.7737 | 0.2736 | 0.4369 | 0.5304 | 0.8320 | 0.7435 | |
| 6 | 0.5227 | 0.7756 | 0.2736 | 0.4212 | 0.5465 | 0.8384 | 0.7683 | |
| NVIB | 1 | 0.5037 | 0.7646 | 0.2349 | 0.3323 | 0.5007 | 0.8344 | 0.7285 |
| 2 | 0.5069 | 0.7859 | 0.2511 | 0.4243 | 0.5108 | 0.8379 | 0.7777 | |
| 3 | 0.5110 | 0.7963 | 0.2589 | 0.4453 | 0.5466 | 0.8606 | 0.7844 | |
| 4 | 0.5111 | 0.7879 | 0.2655 | 0.5290 | 0.5361 | 0.8481 | 0.7943 | |
| 5 | 0.5299 | 0.7660 | 0.2651 | 0.5283 | 0.5571 | 0.8371 | 0.7793 | |
| 6 | 0.5523 | 0.8207 | 0.2923 | 0.5766 | 0.6075 | 0.8531 | 0.8038 |
| Model | #Params | Compound Translation Error Rate (CTER) ↓ | BLEU ↑ | ||||
| NP | VP | PP | Total | Δ | |||
| Transformer | 35M | 24.7%/55.2% | 24.8%/59.5% | 35.7%/73.9% | 28.4%/62.9% | -/- | 59.5 |
| Transformer-Rela | 35M | 30.1%/58.1% | 27.6%/61.2% | 38.5%/74.1% | 32.1%/64.5% | +3.7%/+1.6% | 59.1 |
| Transformer-Small | 25M | 25.1%/56.9% | 25.6%/60.3% | 39.1%/75.0% | 29.9%/64.5% | +1.5%/+1.6% | 59.0 |
| Transformer-Deep | 40M | 23.3%/51.6% | 24.1%/58.0% | 33.8%/72.6% | 27.0%/60.7% | -1.4%/-2.0% | 60.1 |
| Bow | 35M | 22.2%47.9% | 24.8%/55.6% | 35.0%/73.2% | 27.3%/58.9% | -1.1%/-3.0% | - |
| SeqMix | 35M | 24.5%/49.7% | 26.9%/58.9% | 34.4%/73.1% | 28.6%/60.6% | +0.2%/-2.3% | - |
| Dangle | 35M | -/- | -/- | -/- | 24.4%/55.5% | -5.0%/-7.4% | 59.7 |
| Proto-Transformer | 42M | 14.1%/36.5% | 22.1%/50.9% | 28.9%/68.2% | 21.7%/51.8% | -6.7%/-11.1% | 60.1 |
| Transformer+CReg | 25M | -/- | -/- | -/- | 20.2%/48.3% | -8.2%/-14.6% | 61.3 |
| R-Dangle sep | 70M | -/- | -/- | -/- | 16.0%/42.1% | -12.4%/-20.8% | 63.4 |
| DLCL | 35M | -/- | -/- | -/- | 28.4%/67.9% | +0.0%/+5.0% | 59.2 |
| COMPOSITION | 35M | 10.0%/32.6% | 22.1%/54.8% | 29.2%/68.5% | 20.4%/52.0% | -8.0%/-10.9% | 61.5 |
| COMPOSITION-Rela | 35M | 15.5%/39.2% | 22.4%/54.0% | 29.1%/67.3% | 22.3%/53.5% | -6.1%/-9.4% | 61.6 |
| COMPOSITION-Small | 25M | 14.3%/40.3% | 24.4%/58.1% | 34.5%/73.4% | 24.4%/57.3% | -4.0%/-5.6% | 60.1 |
| COMPOSITION-Deep | 40M | 11.4%/34.7% | 19.5%/50.4% | 26.7%/65.6% | 19.2%/50.2% | -9.2%/-12.7% | 62.0 |
| Model | MCD1 | MCD2 | MCD3 | Mean |
| LSTM+attention | 28.9 | 5.0 | 10.8 | 14.9 |
| Transformer | 34.9 | 8.2 | 10.6 | 17.9 |
| Universal Transformer | 37.4 | 8.1 | 11.3 | 18.9 |
| Evolved Transformer | 42.4 | 9.3 | 10.8 | 20.8 |
| CGPS | 13.2 | 1.6 | 6.6 | 7.1 |
| NSEN | 5.1 | 0.9 | 2.3 | 2.8 |
| T5-11B | 61.4 | 30.1 | 31.2 | 40.9 |
| T5-11B-mod | 61.6 | 31.3 | 33.3 | 42.1 |
| RoBERTa | 60.6 | 33.6 | 36.0 | 43.4 |
| HPD | 72.0 | 66.1 | 63.9 | 67.3 |
| Dangle | 78.3 | 59.5 | 60.4 | 66.1 |
| RoBERTa+CReg | 74.8 | 53.3 | 58.3 | 62.1 |
| COMPOSITION | 72.8 | 53.2 | 52.2 | 59.4 |
| Model | Alleviate | CTERInst ↓ | CTERAggr ↓ | |
| E | K, V | |||
| Transformer | X | X | 28.4% | 62.9% |
| COMPOSITION | ✓ | X | 22.6% (-5.8%) | 55.1% (-7.8%) |
| COMPOSITION | ✓ | ✓ | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-SA | 22.2% (-6.2%) | 53.8% (-9.1%) |
| COMPOSITION-FF | 22.6% (-5.8%) | 55.6% (-7.3%) |
| COMPOSITION-SA & FF | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Source | Transformer | COMPOSITION |
| The waiter he liked wore each other's clothes. | 他喜欢穿对方的衣服。(He liked to wear each other's clothes.) | 他喜欢的服务员穿着彼此的衣服。(The waiter he liked wore each other's clothes.) |
| The waiter he liked came by and chased the bully off. | 服务员来了，把那个恶霸赶走了。(The waiter came by and chased the bully off.) | 他喜欢的服务员过来把那个恶霸赶走了。(The waiter he liked came by and chased the bully off.) |
| The waiter he liked picked up his mail. | 服务员喜欢拿起他的邮件。(The waiter liked to pick up his mail.) | 他喜欢的服务员拿起了他的邮件。(The waiter he liked picked up his mail.) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| Transformer-accu | failed | failed |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-Half | 27.0% (-1.4%) | 61.3% (-1.6%) |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | #Params | Compound Translation Error Rate (CTER) ↓ | BLEU ↑ | ||||
| NP | VP | PP | Total | Δ | |||
| Transformer | 35M | 24.7%/55.2% | 24.8%/59.5% | 35.7%/73.9% | 28.4%/62.9% | -/- | 59.5 |
| Transformer-Rela | 35M | 30.1%/58.1% | 27.6%/61.2% | 38.5%/74.1% | 32.1%/64.5% | +3.7%/+1.6% | 59.1 |
| Transformer-Small | 25M | 25.1%/56.9% | 25.6%/60.3% | 39.1%/75.0% | 29.9%/64.5% | +1.5%/+1.6% | 59.0 |
| Transformer-Deep | 40M | 23.3%/51.6% | 24.1%/58.0% | 33.8%/72.6% | 27.0%/60.7% | -1.4%/-2.0% | 60.1 |
| Bow | 35M | 22.2%47.9% | 24.8%/55.6% | 35.0%/73.2% | 27.3%/58.9% | -1.1%/-3.0% | - |
| SeqMix | 35M | 24.5%/49.7% | 26.9%/58.9% | 34.4%/73.1% | 28.6%/60.6% | +0.2%/-2.3% | - |
| Dangle | 35M | -/- | -/- | -/- | 24.4%/55.5% | -5.0%/-7.4% | 59.7 |
| Proto-Transformer | 42M | 14.1%/36.5% | 22.1%/50.9% | 28.9%/68.2% | 21.7%/51.8% | -6.7%/-11.1% | 60.1 |
| Transformer+CReg | 25M | -/- | -/- | -/- | 20.2%/48.3% | -8.2%/-14.6% | 61.3 |
| R-Dangle sep | 70M | -/- | -/- | -/- | 16.0%/42.1% | -12.4%/-20.8% | 63.4 |
| DLCL | 35M | -/- | -/- | -/- | 28.4%/67.9% | +0.0%/+5.0% | 59.2 |
| COMPOSITION | 35M | 10.0%/32.6% | 22.1%/54.8% | 29.2%/68.5% | 20.4%/52.0% | -8.0%/-10.9% | 61.5 |
| COMPOSITION-Rela | 35M | 15.5%/39.2% | 22.4%/54.0% | 29.1%/67.3% | 22.3%/53.5% | -6.1%/-9.4% | 61.6 |
| COMPOSITION-Small | 25M | 14.3%/40.3% | 24.4%/58.1% | 34.5%/73.4% | 24.4%/57.3% | -4.0%/-5.6% | 60.1 |
| COMPOSITION-Deep | 40M | 11.4%/34.7% | 19.5%/50.4% | 26.7%/65.6% | 19.2%/50.2% | -9.2%/-12.7% | 62.0 |
| Model | MCD1 | MCD2 | MCD3 | Mean |
| LSTM+attention | 28.9 | 5.0 | 10.8 | 14.9 |
| Transformer | 34.9 | 8.2 | 10.6 | 17.9 |
| Universal Transformer | 37.4 | 8.1 | 11.3 | 18.9 |
| Evolved Transformer | 42.4 | 9.3 | 10.8 | 20.8 |
| CGPS | 13.2 | 1.6 | 6.6 | 7.1 |
| NSEN | 5.1 | 0.9 | 2.3 | 2.8 |
| T5-11B | 61.4 | 30.1 | 31.2 | 40.9 |
| T5-11B-mod | 61.6 | 31.3 | 33.3 | 42.1 |
| RoBERTa | 60.6 | 33.6 | 36.0 | 43.4 |
| HPD | 72.0 | 66.1 | 63.9 | 67.3 |
| Dangle | 78.3 | 59.5 | 60.4 | 66.1 |
| RoBERTa+CReg | 74.8 | 53.3 | 58.3 | 62.1 |
| COMPOSITION | 72.8 | 53.2 | 52.2 | 59.4 |
| Model | Alleviate | CTERInst ↓ | CTERAggr ↓ | |
| E | K, V | |||
| Transformer | X | X | 28.4% | 62.9% |
| COMPOSITION | ✓ | X | 22.6% (-5.8%) | 55.1% (-7.8%) |
| COMPOSITION | ✓ | ✓ | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-SA | 22.2% (-6.2%) | 53.8% (-9.1%) |
| COMPOSITION-FF | 22.6% (-5.8%) | 55.6% (-7.3%) |
| COMPOSITION-SA & FF | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Source | Transformer | COMPOSITION |
| The waiter he liked wore each other's clothes. | 他喜欢穿对方的衣服。(He liked to wear each other's clothes.) | 他喜欢的服务员穿着彼此的衣服。(The waiter he liked wore each other's clothes.) |
| The waiter he liked came by and chased the bully off. | 服务员来了，把那个恶霸赶走了。(The waiter came by and chased the bully off.) | 他喜欢的服务员过来把那个恶霸赶走了。(The waiter he liked came by and chased the bully off.) |
| The waiter he liked picked up his mail. | 服务员喜欢拿起他的邮件。(The waiter liked to pick up his mail.) | 他喜欢的服务员拿起了他的邮件。(The waiter he liked picked up his mail.) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| Transformer-accu | failed | failed |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-Half | 27.0% (-1.4%) | 61.3% (-1.6%) |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | #Params | Compound Translation Error Rate (CTER) ↓ | BLEU ↑ | ||||
| NP | VP | PP | Total | Δ | |||
| Transformer | 35M | 24.7%/55.2% | 24.8%/59.5% | 35.7%/73.9% | 28.4%/62.9% | -/- | 59.5 |
| Transformer-Rela | 35M | 30.1%/58.1% | 27.6%/61.2% | 38.5%/74.1% | 32.1%/64.5% | +3.7%/+1.6% | 59.1 |
| Transformer-Small | 25M | 25.1%/56.9% | 25.6%/60.3% | 39.1%/75.0% | 29.9%/64.5% | +1.5%/+1.6% | 59.0 |
| Transformer-Deep | 40M | 23.3%/51.6% | 24.1%/58.0% | 33.8%/72.6% | 27.0%/60.7% | -1.4%/-2.0% | 60.1 |
| Bow | 35M | 22.2%47.9% | 24.8%/55.6% | 35.0%/73.2% | 27.3%/58.9% | -1.1%/-3.0% | - |
| SeqMix | 35M | 24.5%/49.7% | 26.9%/58.9% | 34.4%/73.1% | 28.6%/60.6% | +0.2%/-2.3% | - |
| Dangle | 35M | -/- | -/- | -/- | 24.4%/55.5% | -5.0%/-7.4% | 59.7 |
| Proto-Transformer | 42M | 14.1%/36.5% | 22.1%/50.9% | 28.9%/68.2% | 21.7%/51.8% | -6.7%/-11.1% | 60.1 |
| Transformer+CReg | 25M | -/- | -/- | -/- | 20.2%/48.3% | -8.2%/-14.6% | 61.3 |
| R-Dangle sep | 70M | -/- | -/- | -/- | 16.0%/42.1% | -12.4%/-20.8% | 63.4 |
| DLCL | 35M | -/- | -/- | -/- | 28.4%/67.9% | +0.0%/+5.0% | 59.2 |
| COMPOSITION | 35M | 10.0%/32.6% | 22.1%/54.8% | 29.2%/68.5% | 20.4%/52.0% | -8.0%/-10.9% | 61.5 |
| COMPOSITION-Rela | 35M | 15.5%/39.2% | 22.4%/54.0% | 29.1%/67.3% | 22.3%/53.5% | -6.1%/-9.4% | 61.6 |
| COMPOSITION-Small | 25M | 14.3%/40.3% | 24.4%/58.1% | 34.5%/73.4% | 24.4%/57.3% | -4.0%/-5.6% | 60.1 |
| COMPOSITION-Deep | 40M | 11.4%/34.7% | 19.5%/50.4% | 26.7%/65.6% | 19.2%/50.2% | -9.2%/-12.7% | 62.0 |
| Model | MCD1 | MCD2 | MCD3 | Mean |
| LSTM+attention | 28.9 | 5.0 | 10.8 | 14.9 |
| Transformer | 34.9 | 8.2 | 10.6 | 17.9 |
| Universal Transformer | 37.4 | 8.1 | 11.3 | 18.9 |
| Evolved Transformer | 42.4 | 9.3 | 10.8 | 20.8 |
| CGPS | 13.2 | 1.6 | 6.6 | 7.1 |
| NSEN | 5.1 | 0.9 | 2.3 | 2.8 |
| T5-11B | 61.4 | 30.1 | 31.2 | 40.9 |
| T5-11B-mod | 61.6 | 31.3 | 33.3 | 42.1 |
| RoBERTa | 60.6 | 33.6 | 36.0 | 43.4 |
| HPD | 72.0 | 66.1 | 63.9 | 67.3 |
| Dangle | 78.3 | 59.5 | 60.4 | 66.1 |
| RoBERTa+CReg | 74.8 | 53.3 | 58.3 | 62.1 |
| COMPOSITION | 72.8 | 53.2 | 52.2 | 59.4 |
| Model | Alleviate | CTERInst ↓ | CTERAggr ↓ | |
| E | K, V | |||
| Transformer | X | X | 28.4% | 62.9% |
| COMPOSITION | ✓ | X | 22.6% (-5.8%) | 55.1% (-7.8%) |
| COMPOSITION | ✓ | ✓ | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-SA | 22.2% (-6.2%) | 53.8% (-9.1%) |
| COMPOSITION-FF | 22.6% (-5.8%) | 55.6% (-7.3%) |
| COMPOSITION-SA & FF | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Source | Transformer | COMPOSITION |
| The waiter he liked wore each other's clothes. | 他喜欢穿对方的衣服。(He liked to wear each other's clothes.) | 他喜欢的服务员穿着彼此的衣服。(The waiter he liked wore each other's clothes.) |
| The waiter he liked came by and chased the bully off. | 服务员来了，把那个恶霸赶走了。(The waiter came by and chased the bully off.) | 他喜欢的服务员过来把那个恶霸赶走了。(The waiter he liked came by and chased the bully off.) |
| The waiter he liked picked up his mail. | 服务员喜欢拿起他的邮件。(The waiter liked to pick up his mail.) | 他喜欢的服务员拿起了他的邮件。(The waiter he liked picked up his mail.) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| Transformer-accu | failed | failed |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-Half | 27.0% (-1.4%) | 61.3% (-1.6%) |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | #Params | Compound Translation Error Rate (CTER) ↓ | BLEU ↑ | ||||
| NP | VP | PP | Total | Δ | |||
| Transformer | 35M | 24.7%/55.2% | 24.8%/59.5% | 35.7%/73.9% | 28.4%/62.9% | -/- | 59.5 |
| Transformer-Rela | 35M | 30.1%/58.1% | 27.6%/61.2% | 38.5%/74.1% | 32.1%/64.5% | +3.7%/+1.6% | 59.1 |
| Transformer-Small | 25M | 25.1%/56.9% | 25.6%/60.3% | 39.1%/75.0% | 29.9%/64.5% | +1.5%/+1.6% | 59.0 |
| Transformer-Deep | 40M | 23.3%/51.6% | 24.1%/58.0% | 33.8%/72.6% | 27.0%/60.7% | -1.4%/-2.0% | 60.1 |
| Bow | 35M | 22.2%47.9% | 24.8%/55.6% | 35.0%/73.2% | 27.3%/58.9% | -1.1%/-3.0% | - |
| SeqMix | 35M | 24.5%/49.7% | 26.9%/58.9% | 34.4%/73.1% | 28.6%/60.6% | +0.2%/-2.3% | - |
| Dangle | 35M | -/- | -/- | -/- | 24.4%/55.5% | -5.0%/-7.4% | 59.7 |
| Proto-Transformer | 42M | 14.1%/36.5% | 22.1%/50.9% | 28.9%/68.2% | 21.7%/51.8% | -6.7%/-11.1% | 60.1 |
| Transformer+CReg | 25M | -/- | -/- | -/- | 20.2%/48.3% | -8.2%/-14.6% | 61.3 |
| R-Dangle sep | 70M | -/- | -/- | -/- | 16.0%/42.1% | -12.4%/-20.8% | 63.4 |
| DLCL | 35M | -/- | -/- | -/- | 28.4%/67.9% | +0.0%/+5.0% | 59.2 |
| COMPOSITION | 35M | 10.0%/32.6% | 22.1%/54.8% | 29.2%/68.5% | 20.4%/52.0% | -8.0%/-10.9% | 61.5 |
| COMPOSITION-Rela | 35M | 15.5%/39.2% | 22.4%/54.0% | 29.1%/67.3% | 22.3%/53.5% | -6.1%/-9.4% | 61.6 |
| COMPOSITION-Small | 25M | 14.3%/40.3% | 24.4%/58.1% | 34.5%/73.4% | 24.4%/57.3% | -4.0%/-5.6% | 60.1 |
| COMPOSITION-Deep | 40M | 11.4%/34.7% | 19.5%/50.4% | 26.7%/65.6% | 19.2%/50.2% | -9.2%/-12.7% | 62.0 |
| Model | MCD1 | MCD2 | MCD3 | Mean |
| LSTM+attention | 28.9 | 5.0 | 10.8 | 14.9 |
| Transformer | 34.9 | 8.2 | 10.6 | 17.9 |
| Universal Transformer | 37.4 | 8.1 | 11.3 | 18.9 |
| Evolved Transformer | 42.4 | 9.3 | 10.8 | 20.8 |
| CGPS | 13.2 | 1.6 | 6.6 | 7.1 |
| NSEN | 5.1 | 0.9 | 2.3 | 2.8 |
| T5-11B | 61.4 | 30.1 | 31.2 | 40.9 |
| T5-11B-mod | 61.6 | 31.3 | 33.3 | 42.1 |
| RoBERTa | 60.6 | 33.6 | 36.0 | 43.4 |
| HPD | 72.0 | 66.1 | 63.9 | 67.3 |
| Dangle | 78.3 | 59.5 | 60.4 | 66.1 |
| RoBERTa+CReg | 74.8 | 53.3 | 58.3 | 62.1 |
| COMPOSITION | 72.8 | 53.2 | 52.2 | 59.4 |
| Model | Alleviate | CTERInst ↓ | CTERAggr ↓ | |
| E | K, V | |||
| Transformer | X | X | 28.4% | 62.9% |
| COMPOSITION | ✓ | X | 22.6% (-5.8%) | 55.1% (-7.8%) |
| COMPOSITION | ✓ | ✓ | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-SA | 22.2% (-6.2%) | 53.8% (-9.1%) |
| COMPOSITION-FF | 22.6% (-5.8%) | 55.6% (-7.3%) |
| COMPOSITION-SA & FF | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Source | Transformer | COMPOSITION |
| The waiter he liked wore each other's clothes. | 他喜欢穿对方的衣服。(He liked to wear each other's clothes.) | 他喜欢的服务员穿着彼此的衣服。(The waiter he liked wore each other's clothes.) |
| The waiter he liked came by and chased the bully off. | 服务员来了，把那个恶霸赶走了。(The waiter came by and chased the bully off.) | 他喜欢的服务员过来把那个恶霸赶走了。(The waiter he liked came by and chased the bully off.) |
| The waiter he liked picked up his mail. | 服务员喜欢拿起他的邮件。(The waiter liked to pick up his mail.) | 他喜欢的服务员拿起了他的邮件。(The waiter he liked picked up his mail.) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| Transformer-accu | failed | failed |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Model | CTERInst ↓ | CTERAggr ↓ |
| Transformer | 28.4% | 62.9% |
| COMPOSITION-Half | 27.0% (-1.4%) | 61.3% (-1.6%) |
| COMPOSITION | 20.4% (-8.0%) | 52.0% (-10.9%) |
| Datasets | WiKi | OntoNotes | Ultra-Fine |
| hierarchy depth | 2 | 3 | 1 |
| entity types | 112 | 89 | 2519 |
| mentions-train | 2009898 | 253241 | 2000 |
| mentions-dev | - | 2202 | 2000 |
| mentions-test | 563 | 8963 | 2000 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 55.3 | 71.2 | 64.6 |
| NFETChier(2018) | 60.2±0.2 | 76.4±0.1 | 70.2±0.2 |
| ML-L2R(2020b) | 58.7 | 73.0 | 68.1 |
| LRN(2021b) | 56.6 | 77.6 | 71.8 |
| NFETC-ARhier(2021)† | 64.0±0.3 | 78.8±0.3 | 73.0±0.3 |
| FCLChier(2022) | 65.3±0.2 | 79.6±0.3 | 74.0±0.3 |
| DenoiseFET(2022)† | 59.2±0.2 | 81.3±0.3 | 75.3±0.4 |
| Baseline | 54.8±0.7 | 78.1±1.2 | 71.2±0.9 |
| Baseline (corrected) | 58.8±0.8 | 83.7±0.2 | 76.3±0.3 |
| w/ co-predict | 57.0±0.5 | 83.1±0.1 | 75.2±0.2 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 64.4±0.7 | 85.8±0.3 | 80.5±0.5 |
| Baseline (corrected) | 64.6±0.7 | 87.1±0.1 | 81.7±0.1 |
| w/ co-predict | 64.3±0.6 | 86.5±0.2 | 81.3±0.1 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 53.3 | 69.3 | 66.4 |
| NFETChier(2018) | 68.9±0.6 | 81.9±0.7 | 79.0±0.7 |
| ML-L2R(2020b) | 69.1 | 82.6 | 80.8 |
| NFETC-ARhier(2021)† | 70.1±0.9 | 83.2±0.7 | 80.1±0.6 |
| FCLChier(2022) | 71.3±1.1 | 82.2±0.7 | 81.1±0.6 |
| DenoiseFET(2022)†* | 67.5±0.3 | 82.3±0.6 | 78.5±0.6 |
| Baseline | 58.9±0.5 | 81.1±0.2 | 76.7±0.2 |
| Baseline (corrected) | 66.4±0.2 | 84.3±0.2 | 80.1±0.2 |
| w/ co-predict | 66.9±0.4 | 84.2±0.1 | 79.9±0.1 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 62.2±0.6 | 80.8±0.1 | 77.5±0.3 |
| Baseline (corrected) | 65.8±0.8 | 82.6±0.4 | 79.1±0.4 |
| w/ co-predict | 65.1±0.7 | 82.4±0.3 | 78.6±0.4 |
| Model | P | R | F1 |
| With Crowdsourced Annotated Training Samples | |||
| UFET(2018) | 47.1 | 24.2 | 32.0 |
| LDET(2019)† | 51.5 | 33.0 | 40.2 |
| LRN(2021b) | 54.5 | 38.9 | 45.4 |
| MLMET(2021) | 53.6 | 45.3 | 49.1 |
| UNIST(2022) | 49.2 | 49.4 | 49.3 |
| DenoiseFET(2022)† | 55.6±0.4 | 44.7±0.3 | 49.5±0.1 |
| PKL(2023) | 52.7 | 49.2 | 50.9 |
| Baseline | 57.0±0.7 | 41.2±0.3 | 48.0±0.2 |
| Baseline (corrected) | 53.2±0.7 | 48.1±0.5 | 50.5±0.1 |
| w/ co-predict | 52.7±0.6 | 48.2±0.5 | 50.3±0.1 |
| Dataset | Model | P | R | F1 |
| With Distantly Annotated Training Samples | ||||
| OntoNotes | Baseline | 87.8±0.9 | 80.0±0.6 | 83.7±0.2 |
| w/o unl. | 89.3±0.2 | 76.0±1.0 | 81.8±0.3 | |
| w/o ina. | 83.9±0.6 | 82.1±0.8 | 83.0±0.2 | |
| WiKi | Baseline | 82.9±0.1 | 86.1±0.2 | 84.3±0.2 |
| w/o unl. | 83.5±0.1 | 84.2±0.3 | 83.8±0.1 | |
| w/o ina. | 70.7±0.4 | 90.9±0.4 | 79.5±0.2 | |
| With ChatGPT Annotated Training Samples | ||||
| OntoNotes | Baseline | 88.8±0.9 | 85.4±0.9 | 87.1±0.1 |
| w/o unl. | 90.2±0.5 | 80.1±0.6 | 84.9±0.2 | |
| w/o ina. | 85.4±0.6 | 86.1±0.2 | 85.7±0.3 | |
| WiKi | Baseline | 82.2±0.5 | 83.1±0.2 | 82.6±0.4 |
| w/o unl. | 83.0±0.4 | 79.8±0.4 | 81.3±0.1 | |
| w/o ina. | 80.5±0.2 | 83.3±0.2 | 81.8±0.1 | |
| With Crowdsourced Annotated Training Samples | ||||
| Ultra-Fine | Baseline | 53.2±0.7 | 48.1±0.5 | 50.5±0.2 |
| w/o unl. | 59.5±1.2 | 41.0±0.7 | 48.5±0.3 | |
| w/o ina. | 51.5±0.6 | 48.8±0.4 | 50.1±0.1 |
| Parameters | OntoNotes | WiKi | Ultra-Fine |
| Batch size | 16 | 16 | 16 |
| Learning Rate | 3e-6 | 3e-6 | 2e-5 |
| Adam epsilon | 1e-8 | 1e-8 | 1e-8 |
| Warmup ratio | 0.0 | 0.0 | 0.0 |
| Emb. dropout | 0.2 | 0.2 | 0.2 |
| Weight decay | 0.01 | 0.01 | 0.01 |
| Clipping grad | 0.1 | 0.1 | 0.1 |
| Loss weight γ | 0.1 | 0.1 | 0.005 |
| Threshold ε | 0.2 | 0.05 | 0.8 |
| ε on ChatGPT | 0.3 | 0.5 | - |
| Datasets | WiKi | OntoNotes | Ultra-Fine |
| hierarchy depth | 2 | 3 | 1 |
| entity types | 112 | 89 | 2519 |
| mentions-train | 2009898 | 253241 | 2000 |
| mentions-dev | - | 2202 | 2000 |
| mentions-test | 563 | 8963 | 2000 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 55.3 | 71.2 | 64.6 |
| NFETChier(2018) | 60.2±0.2 | 76.4±0.1 | 70.2±0.2 |
| ML-L2R(2020b) | 58.7 | 73.0 | 68.1 |
| LRN(2021b) | 56.6 | 77.6 | 71.8 |
| NFETC-ARhier(2021)† | 64.0±0.3 | 78.8±0.3 | 73.0±0.3 |
| FCLChier(2022) | 65.3±0.2 | 79.6±0.3 | 74.0±0.3 |
| DenoiseFET(2022)† | 59.2±0.2 | 81.3±0.3 | 75.3±0.4 |
| Baseline | 54.8±0.7 | 78.1±1.2 | 71.2±0.9 |
| Baseline (corrected) | 58.8±0.8 | 83.7±0.2 | 76.3±0.3 |
| w/ co-predict | 57.0±0.5 | 83.1±0.1 | 75.2±0.2 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 64.4±0.7 | 85.8±0.3 | 80.5±0.5 |
| Baseline (corrected) | 64.6±0.7 | 87.1±0.1 | 81.7±0.1 |
| w/ co-predict | 64.3±0.6 | 86.5±0.2 | 81.3±0.1 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 53.3 | 69.3 | 66.4 |
| NFETChier(2018) | 68.9±0.6 | 81.9±0.7 | 79.0±0.7 |
| ML-L2R(2020b) | 69.1 | 82.6 | 80.8 |
| NFETC-ARhier(2021)† | 70.1±0.9 | 83.2±0.7 | 80.1±0.6 |
| FCLChier(2022) | 71.3±1.1 | 82.2±0.7 | 81.1±0.6 |
| DenoiseFET(2022)†* | 67.5±0.3 | 82.3±0.6 | 78.5±0.6 |
| Baseline | 58.9±0.5 | 81.1±0.2 | 76.7±0.2 |
| Baseline (corrected) | 66.4±0.2 | 84.3±0.2 | 80.1±0.2 |
| w/ co-predict | 66.9±0.4 | 84.2±0.1 | 79.9±0.1 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 62.2±0.6 | 80.8±0.1 | 77.5±0.3 |
| Baseline (corrected) | 65.8±0.8 | 82.6±0.4 | 79.1±0.4 |
| w/ co-predict | 65.1±0.7 | 82.4±0.3 | 78.6±0.4 |
| Model | P | R | F1 |
| With Crowdsourced Annotated Training Samples | |||
| UFET(2018) | 47.1 | 24.2 | 32.0 |
| LDET(2019)† | 51.5 | 33.0 | 40.2 |
| LRN(2021b) | 54.5 | 38.9 | 45.4 |
| MLMET(2021) | 53.6 | 45.3 | 49.1 |
| UNIST(2022) | 49.2 | 49.4 | 49.3 |
| DenoiseFET(2022)† | 55.6±0.4 | 44.7±0.3 | 49.5±0.1 |
| PKL(2023) | 52.7 | 49.2 | 50.9 |
| Baseline | 57.0±0.7 | 41.2±0.3 | 48.0±0.2 |
| Baseline (corrected) | 53.2±0.7 | 48.1±0.5 | 50.5±0.1 |
| w/ co-predict | 52.7±0.6 | 48.2±0.5 | 50.3±0.1 |
| Dataset | Model | P | R | F1 |
| With Distantly Annotated Training Samples | ||||
| OntoNotes | Baseline | 87.8±0.9 | 80.0±0.6 | 83.7±0.2 |
| w/o unl. | 89.3±0.2 | 76.0±1.0 | 81.8±0.3 | |
| w/o ina. | 83.9±0.6 | 82.1±0.8 | 83.0±0.2 | |
| WiKi | Baseline | 82.9±0.1 | 86.1±0.2 | 84.3±0.2 |
| w/o unl. | 83.5±0.1 | 84.2±0.3 | 83.8±0.1 | |
| w/o ina. | 70.7±0.4 | 90.9±0.4 | 79.5±0.2 | |
| With ChatGPT Annotated Training Samples | ||||
| OntoNotes | Baseline | 88.8±0.9 | 85.4±0.9 | 87.1±0.1 |
| w/o unl. | 90.2±0.5 | 80.1±0.6 | 84.9±0.2 | |
| w/o ina. | 85.4±0.6 | 86.1±0.2 | 85.7±0.3 | |
| WiKi | Baseline | 82.2±0.5 | 83.1±0.2 | 82.6±0.4 |
| w/o unl. | 83.0±0.4 | 79.8±0.4 | 81.3±0.1 | |
| w/o ina. | 80.5±0.2 | 83.3±0.2 | 81.8±0.1 | |
| With Crowdsourced Annotated Training Samples | ||||
| Ultra-Fine | Baseline | 53.2±0.7 | 48.1±0.5 | 50.5±0.2 |
| w/o unl. | 59.5±1.2 | 41.0±0.7 | 48.5±0.3 | |
| w/o ina. | 51.5±0.6 | 48.8±0.4 | 50.1±0.1 |
| Parameters | OntoNotes | WiKi | Ultra-Fine |
| Batch size | 16 | 16 | 16 |
| Learning Rate | 3e-6 | 3e-6 | 2e-5 |
| Adam epsilon | 1e-8 | 1e-8 | 1e-8 |
| Warmup ratio | 0.0 | 0.0 | 0.0 |
| Emb. dropout | 0.2 | 0.2 | 0.2 |
| Weight decay | 0.01 | 0.01 | 0.01 |
| Clipping grad | 0.1 | 0.1 | 0.1 |
| Loss weight γ | 0.1 | 0.1 | 0.005 |
| Threshold ε | 0.2 | 0.05 | 0.8 |
| ε on ChatGPT | 0.3 | 0.5 | - |
| Datasets | WiKi | OntoNotes | Ultra-Fine |
| hierarchy depth | 2 | 3 | 1 |
| entity types | 112 | 89 | 2519 |
| mentions-train | 2009898 | 253241 | 2000 |
| mentions-dev | - | 2202 | 2000 |
| mentions-test | 563 | 8963 | 2000 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 55.3 | 71.2 | 64.6 |
| NFETChier(2018) | 60.2±0.2 | 76.4±0.1 | 70.2±0.2 |
| ML-L2R(2020b) | 58.7 | 73.0 | 68.1 |
| LRN(2021b) | 56.6 | 77.6 | 71.8 |
| NFETC-ARhier(2021)† | 64.0±0.3 | 78.8±0.3 | 73.0±0.3 |
| FCLChier(2022) | 65.3±0.2 | 79.6±0.3 | 74.0±0.3 |
| DenoiseFET(2022)† | 59.2±0.2 | 81.3±0.3 | 75.3±0.4 |
| Baseline | 54.8±0.7 | 78.1±1.2 | 71.2±0.9 |
| Baseline (corrected) | 58.8±0.8 | 83.7±0.2 | 76.3±0.3 |
| w/ co-predict | 57.0±0.5 | 83.1±0.1 | 75.2±0.2 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 64.4±0.7 | 85.8±0.3 | 80.5±0.5 |
| Baseline (corrected) | 64.6±0.7 | 87.1±0.1 | 81.7±0.1 |
| w/ co-predict | 64.3±0.6 | 86.5±0.2 | 81.3±0.1 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 53.3 | 69.3 | 66.4 |
| NFETChier(2018) | 68.9±0.6 | 81.9±0.7 | 79.0±0.7 |
| ML-L2R(2020b) | 69.1 | 82.6 | 80.8 |
| NFETC-ARhier(2021)† | 70.1±0.9 | 83.2±0.7 | 80.1±0.6 |
| FCLChier(2022) | 71.3±1.1 | 82.2±0.7 | 81.1±0.6 |
| DenoiseFET(2022)†* | 67.5±0.3 | 82.3±0.6 | 78.5±0.6 |
| Baseline | 58.9±0.5 | 81.1±0.2 | 76.7±0.2 |
| Baseline (corrected) | 66.4±0.2 | 84.3±0.2 | 80.1±0.2 |
| w/ co-predict | 66.9±0.4 | 84.2±0.1 | 79.9±0.1 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 62.2±0.6 | 80.8±0.1 | 77.5±0.3 |
| Baseline (corrected) | 65.8±0.8 | 82.6±0.4 | 79.1±0.4 |
| w/ co-predict | 65.1±0.7 | 82.4±0.3 | 78.6±0.4 |
| Model | P | R | F1 |
| With Crowdsourced Annotated Training Samples | |||
| UFET(2018) | 47.1 | 24.2 | 32.0 |
| LDET(2019)† | 51.5 | 33.0 | 40.2 |
| LRN(2021b) | 54.5 | 38.9 | 45.4 |
| MLMET(2021) | 53.6 | 45.3 | 49.1 |
| UNIST(2022) | 49.2 | 49.4 | 49.3 |
| DenoiseFET(2022)† | 55.6±0.4 | 44.7±0.3 | 49.5±0.1 |
| PKL(2023) | 52.7 | 49.2 | 50.9 |
| Baseline | 57.0±0.7 | 41.2±0.3 | 48.0±0.2 |
| Baseline (corrected) | 53.2±0.7 | 48.1±0.5 | 50.5±0.1 |
| w/ co-predict | 52.7±0.6 | 48.2±0.5 | 50.3±0.1 |
| Dataset | Model | P | R | F1 |
| With Distantly Annotated Training Samples | ||||
| OntoNotes | Baseline | 87.8±0.9 | 80.0±0.6 | 83.7±0.2 |
| w/o unl. | 89.3±0.2 | 76.0±1.0 | 81.8±0.3 | |
| w/o ina. | 83.9±0.6 | 82.1±0.8 | 83.0±0.2 | |
| WiKi | Baseline | 82.9±0.1 | 86.1±0.2 | 84.3±0.2 |
| w/o unl. | 83.5±0.1 | 84.2±0.3 | 83.8±0.1 | |
| w/o ina. | 70.7±0.4 | 90.9±0.4 | 79.5±0.2 | |
| With ChatGPT Annotated Training Samples | ||||
| OntoNotes | Baseline | 88.8±0.9 | 85.4±0.9 | 87.1±0.1 |
| w/o unl. | 90.2±0.5 | 80.1±0.6 | 84.9±0.2 | |
| w/o ina. | 85.4±0.6 | 86.1±0.2 | 85.7±0.3 | |
| WiKi | Baseline | 82.2±0.5 | 83.1±0.2 | 82.6±0.4 |
| w/o unl. | 83.0±0.4 | 79.8±0.4 | 81.3±0.1 | |
| w/o ina. | 80.5±0.2 | 83.3±0.2 | 81.8±0.1 | |
| With Crowdsourced Annotated Training Samples | ||||
| Ultra-Fine | Baseline | 53.2±0.7 | 48.1±0.5 | 50.5±0.2 |
| w/o unl. | 59.5±1.2 | 41.0±0.7 | 48.5±0.3 | |
| w/o ina. | 51.5±0.6 | 48.8±0.4 | 50.1±0.1 | |
| Parameters | OntoNotes | WiKi | Ultra-Fine |
| Batch size | 16 | 16 | 16 |
| Learning Rate | 3e-6 | 3e-6 | 2e-5 |
| Adam epsilon | 1e-8 | 1e-8 | 1e-8 |
| Warmup ratio | 0.0 | 0.0 | 0.0 |
| Emb. dropout | 0.2 | 0.2 | 0.2 |
| Weight decay | 0.01 | 0.01 | 0.01 |
| Clipping grad | 0.1 | 0.1 | 0.1 |
| Loss weight γ | 0.1 | 0.1 | 0.005 |
| Threshold ε | 0.2 | 0.05 | 0.8 |
| ε on ChatGPT | 0.3 | 0.5 | - |
| Datasets | WiKi | OntoNotes | Ultra-Fine |
| hierarchy depth | 2 | 3 | 1 |
| entity types | 112 | 89 | 2519 |
| mentions-train | 2009898 | 253241 | 2000 |
| mentions-dev | - | 2202 | 2000 |
| mentions-test | 563 | 8963 | 2000 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 55.3 | 71.2 | 64.6 |
| NFETChier(2018) | 60.2±0.2 | 76.4±0.1 | 70.2±0.2 |
| ML-L2R(2020b) | 58.7 | 73.0 | 68.1 |
| LRN(2021b) | 56.6 | 77.6 | 71.8 |
| NFETC-ARhier(2021)† | 64.0±0.3 | 78.8±0.3 | 73.0±0.3 |
| FCLChier(2022) | 65.3±0.2 | 79.6±0.3 | 74.0±0.3 |
| DenoiseFET(2022)† | 59.2±0.2 | 81.3±0.3 | 75.3±0.4 |
| Baseline | 54.8±0.7 | 78.1±1.2 | 71.2±0.9 |
| Baseline (corrected) | 58.8±0.8 | 83.7±0.2 | 76.3±0.3 |
| w/ co-predict | 57.0±0.5 | 83.1±0.1 | 75.2±0.2 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 64.4±0.7 | 85.8±0.3 | 80.5±0.5 |
| Baseline (corrected) | 64.6±0.7 | 87.1±0.1 | 81.7±0.1 |
| w/ co-predict | 64.3±0.6 | 86.5±0.2 | 81.3±0.1 |
| Model | Acc | Macro F1 | Micro F1 |
| With Distantly Annotated Training Samples | |||
| AFET(2016) | 53.3 | 69.3 | 66.4 |
| NFETChier(2018) | 68.9±0.6 | 81.9±0.7 | 79.0±0.7 |
| ML-L2R(2020b) | 69.1 | 82.6 | 80.8 |
| NFETC-ARhier(2021)† | 70.1±0.9 | 83.2±0.7 | 80.1±0.6 |
| FCLChier(2022) | 71.3±1.1 | 82.2±0.7 | 81.1±0.6 |
| DenoiseFET(2022)†* | 67.5±0.3 | 82.3±0.6 | 78.5±0.6 |
| Baseline | 58.9±0.5 | 81.1±0.2 | 76.7±0.2 |
| Baseline (corrected) | 66.4±0.2 | 84.3±0.2 | 80.1±0.2 |
| w/ co-predict | 66.9±0.4 | 84.2±0.1 | 79.9±0.1 |
| With ChatGPT Annotated Training Samples | |||
| Baseline | 62.2±0.6 | 80.8±0.1 | 77.5±0.3 |
| Baseline (corrected) | 65.8±0.8 | 82.6±0.4 | 79.1±0.4 |
| w/ co-predict | 65.1±0.7 | 82.4±0.3 | 78.6±0.4 |
| Model | P | R | F1 |
| With Crowdsourced Annotated Training Samples | |||
| UFET(2018) | 47.1 | 24.2 | 32.0 |
| LDET(2019)† | 51.5 | 33.0 | 40.2 |
| LRN(2021b) | 54.5 | 38.9 | 45.4 |
| MLMET(2021) | 53.6 | 45.3 | 49.1 |
| UNIST(2022) | 49.2 | 49.4 | 49.3 |
| DenoiseFET(2022)† | 55.6±0.4 | 44.7±0.3 | 49.5±0.1 |
| PKL(2023) | 52.7 | 49.2 | 50.9 |
| Baseline | 57.0±0.7 | 41.2±0.3 | 48.0±0.2 |
| Baseline (corrected) | 53.2±0.7 | 48.1±0.5 | 50.5±0.1 |
| w/ co-predict | 52.7±0.6 | 48.2±0.5 | 50.3±0.1 |
| Dataset | Model | P | R | F1 |
| With Distantly Annotated Training Samples | ||||
| OntoNotes | Baseline | 87.8±0.9 | 80.0±0.6 | 83.7±0.2 |
| w/o unl. | 89.3±0.2 | 76.0±1.0 | 81.8±0.3 | |
| w/o ina. | 83.9±0.6 | 82.1±0.8 | 83.0±0.2 | |
| WiKi | Baseline | 82.9±0.1 | 86.1±0.2 | 84.3±0.2 |
| w/o unl. | 83.5±0.1 | 84.2±0.3 | 83.8±0.1 | |
| w/o ina. | 70.7±0.4 | 90.9±0.4 | 79.5±0.2 | |
| With ChatGPT Annotated Training Samples | ||||
| OntoNotes | Baseline | 88.8±0.9 | 85.4±0.9 | 87.1±0.1 |
| w/o unl. | 90.2±0.5 | 80.1±0.6 | 84.9±0.2 | |
| w/o ina. | 85.4±0.6 | 86.1±0.2 | 85.7±0.3 | |
| WiKi | Baseline | 82.2±0.5 | 83.1±0.2 | 82.6±0.4 |
| w/o unl. | 83.0±0.4 | 79.8±0.4 | 81.3±0.1 | |
| w/o ina. | 80.5±0.2 | 83.3±0.2 | 81.8±0.1 | |
| With Crowdsourced Annotated Training Samples | ||||
| Ultra-Fine | Baseline | 53.2±0.7 | 48.1±0.5 | 50.5±0.2 |
| w/o unl. | 59.5±1.2 | 41.0±0.7 | 48.5±0.3 | |
| w/o ina. | 51.5±0.6 | 48.8±0.4 | 50.1±0.1 |
| Parameters | OntoNotes | WiKi | Ultra-Fine |
| Batch size | 16 | 16 | 16 |
| Learning Rate | 3e-6 | 3e-6 | 2e-5 |
| Adam epsilon | 1e-8 | 1e-8 | 1e-8 |
| Warmup ratio | 0.0 | 0.0 | 0.0 |
| Emb. dropout | 0.2 | 0.2 | 0.2 |
| Weight decay | 0.01 | 0.01 | 0.01 |
| Clipping grad | 0.1 | 0.1 | 0.1 |
| Loss weight γ | 0.1 | 0.1 | 0.005 |
| Threshold ε | 0.2 | 0.05 | 0.8 |
| ε on ChatGPT | 0.3 | 0.5 | - |
| Dataset Type | Total Pairs |
| Instr-Pix2Pix (In-domain) | 200 |
| Magic-Brush (Out-domain) | 200 |
| Visual-Metaphors (Out-domain) | 65 |
| System | TIFA | Hscore-I | Hscore-O |
| Instruct-X-Decoder | 58.10 | 29.1 | 17.5 |
| Grounded Inpainting | 56.74 | 40.3 | 22.3 |
| InstrP2P | 62.24 | 49.3 | 25.2 |
| InstrP2P+BOUNDINGBOX | 63.39 | 62.8 | 37.7 |
| InstrP2P+ENTITYMASK | 65.84 | 69.8 | 65.1 |
| You will be given an input caption of an image and an instruction to transform it by an image editor. Sometimes, however, the instruction does not make sense as the resulting transformation would result in a nonsensical image. Based on the input caption and instruction, reason how the resulting image would look like and whether the resulting image would be possible to imagine. Provide your verdict on whether the transformation is possible. If the verdict is true, also state the entity. First, provide your reasoning, starting with the words “The resulting image would show...”. Then, return the verdict and the entity in JSON format. |
| Caption: Barn In Autumn Smoky Mountains by David Chasey Instruction: Change the barn to a castle The resulting image would show a castle in the mountains, which is a sensible image. { "verdict": "true", "entity": "barn"} |
| Caption: Sligachan Bridge by English Landscape Prints, Instruction: change the bridge to a wooden ship, The resulting image would show a ship up in the air which does not make logical sense. { "verdict": "false", "entity": "none"} |
| Caption:...... |
| Caption:...... |
| Caption:...... |
| You are given an image description and the corresponding entities present in the caption. Generate a question per entity to check whether the image aligns with the text. |
| Caption: handsome man wearing a tuxedo and top hat in casual style clothes over blurred mountain background. Entity: Handsome man, tuxedo and top hat, casual style clothes, blurred mountain background\nEntity: Handsome man\nQuestion: Is the person in the picture a handsome man?\nAnswer: Yes\nEntity: Tuxedo and top hat\nQuestion: Is the man wearing a tuxedo and a top hat?\nAnswer: Yes\nEntity: Casual style clothes\nQuestion: Is the man dressed in casual style clothes?\nAnswer: Yes\nEntity: Blurred mountain background\nQuestion: Is there a blurred mountain background in the picture?\nAnswer: Yes |
| Caption:...... |
| Caption:...... |
| Input | Instruction | Output | Explanation |
| Change the fashion to medieval | This is a good edit because the dress looks like something that could come out of medieval times ( armor) | ||
| Add street lamps | This edit is partially good because it adds the street lamps but it removes the benches from the picture | ||
| Make the hair of the boy spiky | This edit makes the hair more spiky but also changes the face. | ||
| Put carrots on the styrofoam tray | It put carrots on the styrofoam tray but it took off the other items from the tray which wasn’t a command |
| Dataset Type | Total Pairs |
| Instr-Pix2Pix (In-domain) | 200 |
| Magic-Brush (Out-domain) | 200 |
| Visual-Metaphors (Out-domain) | 65 |
| System | TIFA | Hscore-I | Hscore-O |
| Instruct-X-Decoder | 58.10 | 29.1 | 17.5 |
| Grounded Inpainting | 56.74 | 40.3 | 22.3 |
| InstrP2P | 62.24 | 49.3 | 25.2 |
| InstrP2P+BOUNDINGBOX | 63.39 | 62.8 | 37.7 |
| InstrP2P+ENTITYMASK | 65.84 | 69.8 | 65.1 |
| You will be given an input caption of an image and an instruction to transform it by an image editor. Sometimes, however, the instruction does not make sense as the resulting transformation would result in a nonsensical image. Based on the input caption and instruction, reason how the resulting image would look like and whether the resulting image would be possible to imagine. Provide your verdict on whether the transformation is possible. If the verdict is true, also state the entity. First, provide your reasoning, starting with the words “The resulting image would show...”. Then, return the verdict and the entity in JSON format. |
| Caption: Barn In Autumn Smoky Mountains by David Chasey Instruction: Change the barn to a castle The resulting image would show a castle in the mountains, which is a sensible image. { "verdict": "true", "entity": "barn"} |
| Caption: Sligachan Bridge by English Landscape Prints, Instruction: change the bridge to a wooden ship, The resulting image would show a ship up in the air which does not make logical sense. { "verdict": "false", "entity": "none"} |
| Caption:...... |
| Caption:...... |
| Caption:...... |
| You are given an image description and the corresponding entities present in the caption. Generate a question per entity to check whether the image aligns with the text. |
| Caption: handsome man wearing a tuxedo and top hat in casual style clothes over blurred mountain background. Entity: Handsome man, tuxedo and top hat, casual style clothes, blurred mountain background\nEntity: Handsome man\nQuestion: Is the person in the picture a handsome man?\nAnswer: Yes\nEntity: Tuxedo and top hat\nQuestion: Is the man wearing a tuxedo and a top hat?\nAnswer: Yes\nEntity: Casual style clothes\nQuestion: Is the man dressed in casual style clothes?\nAnswer: Yes\nEntity: Blurred mountain background\nQuestion: Is there a blurred mountain background in the picture?\nAnswer: Yes |
| Caption:...... |
| Caption:...... |
| Input | Instruction | Output | Explanation |
| Change the fashion to medieval | This is a good edit because the dress looks like something that could come out of medieval times ( armor) | ||
| Add street lamps | This edit is partially good because it adds the street lamps but it removes the benches from the picture | ||
| Make the hair of the boy spiky | This edit makes the hair more spiky but also changes the face. | ||
| Put carrots on the styrofoam tray | It put carrots on the styrofoam tray but it took off the other items from the tray which wasn’t a command |
| Dataset Type | Total Pairs |
| Instr-Pix2Pix (In-domain) | 200 |
| Magic-Brush (Out-domain) | 200 |
| Visual-Metaphors (Out-domain) | 65 |
| System | TIFA | Hscore-I | Hscore-O |
| Instruct-X-Decoder | 58.10 | 29.1 | 17.5 |
| Grounded Inpainting | 56.74 | 40.3 | 22.3 |
| InstrP2P | 62.24 | 49.3 | 25.2 |
| InstrP2P+BOUNDINGBOX | 63.39 | 62.8 | 37.7 |
| InstrP2P+ENTITYMASK | 65.84 | 69.8 | 65.1 |
| You will be given an input caption of an image and an instruction to transform it by an image editor. Sometimes, however, the instruction does not make sense as the resulting transformation would result in a nonsensical image. Based on the input caption and instruction, reason how the resulting image would look like and whether the resulting image would be possible to imagine. Provide your verdict on whether the transformation is possible. If the verdict is true, also state the entity. First, provide your reasoning, starting with the words “The resulting image would show...”. Then, return the verdict and the entity in JSON format. |
| Caption: Barn In Autumn Smoky Mountains by David Chasey Instruction: Change the barn to a castle The resulting image would show a castle in the mountains, which is a sensible image. { "verdict": "true", "entity": "barn"} |
| Caption: Sligachan Bridge by English Landscape Prints, Instruction: change the bridge to a wooden ship, The resulting image would show a ship up in the air which does not make logical sense. { "verdict": "false", "entity": "none"} |
| Caption:...... |
| Caption:...... |
| Caption:...... |
| You are given an image description and the corresponding entities present in the caption. Generate a question per entity to check whether the image aligns with the text. |
| Caption: handsome man wearing a tuxedo and top hat in casual style clothes over blurred mountain background. Entity: Handsome man, tuxedo and top hat, casual style clothes, blurred mountain background +Entity: Handsome man +Question: Is the person in the picture a handsome man? +Answer: Yes +Entity: Tuxedo and top hat +Question: Is the man wearing a tuxedo and a top hat? +Answer: Yes +Entity: Casual style clothes +Question: Is the man dressed in casual style clothes? +Answer: Yes +Entity: Blurred mountain background +Question: Is there a blurred mountain background in the picture? +Answer: Yes |
| Caption:...... |
| Caption:...... |
| Input | Instruction | Output | Explanation |
| Change the fashion to medieval | This is a good edit because the dress looks like something that could come out of medieval times ( armor) | ||
| Add street lamps | This edit is partially good because it adds the street lamps but it removes the benches from the picture | ||
| Make the hair of the boy spiky | This edit makes the hair more spiky but also changes the face. | ||
| Put carrots on the styrofoam tray | It put carrots on the styrofoam tray but it took off the other items from the tray which wasn’t a command |
| Dataset Type | Total Pairs |
| Instr-Pix2Pix (In-domain) | 200 |
| Magic-Brush (Out-domain) | 200 |
| Visual-Metaphors (Out-domain) | 65 |
| System | TIFA | Hscore-I | Hscore-O |
| Instruct-X-Decoder | 58.10 | 29.1 | 17.5 |
| Grounded Inpainting | 56.74 | 40.3 | 22.3 |
| InstrP2P | 62.24 | 49.3 | 25.2 |
| InstrP2P+BOUNDINGBOX | 63.39 | 62.8 | 37.7 |
| InstrP2P+ENTITYMASK | 65.84 | 69.8 | 65.1 |
| You will be given an input caption of an image and an instruction to transform it by an image editor. Sometimes, however, the instruction does not make sense as the resulting transformation would result in a nonsensical image. Based on the input caption and instruction, reason how the resulting image would look like and whether the resulting image would be possible to imagine. Provide your verdict on whether the transformation is possible. If the verdict is true, also state the entity. First, provide your reasoning, starting with the words “The resulting image would show...”. Then, return the verdict and the entity in JSON format. |
| Caption: Barn In Autumn Smoky Mountains by David Chasey Instruction: Change the barn to a castle The resulting image would show a castle in the mountains, which is a sensible image. { "verdict": "true", "entity": "barn"} |
| Caption: Sligachan Bridge by English Landscape Prints, Instruction: change the bridge to a wooden ship, The resulting image would show a ship up in the air which does not make logical sense. { "verdict": "false", "entity": "none"} |
| Caption:...... |
| Caption:...... |
| Caption:...... |
| You are given an image description and the corresponding entities present in the caption. Generate a question per entity to check whether the image aligns with the text. |
| Caption: handsome man wearing a tuxedo and top hat in casual style clothes over blurred mountain background. Entity: Handsome man, tuxedo and top hat, casual style clothes, blurred mountain background\nEntity: Handsome man\nQuestion: Is the person in the picture a handsome man?\nAnswer: Yes\nEntity: Tuxedo and top hat\nQuestion: Is the man wearing a tuxedo and a top hat?\nAnswer: Yes\nEntity: Casual style clothes\nQuestion: Is the man dressed in casual style clothes?\nAnswer: Yes\nEntity: Blurred mountain background\nQuestion: Is there a blurred mountain background in the picture?\nAnswer: Yes |
| Caption:...... |
| Caption:...... |
| Input | Instruction | Output | Explanation |
| Change the fashion to medieval | This is a good edit because the dress looks like something that could come out of medieval times ( armor) | ||
| Add street lamps | This edit is partially good because it adds the street lamps but it removes the benches from the picture | ||
| Make the hair of the boy spiky | This edit makes the hair more spiky but also changes the face. | ||
| Put carrots on the styrofoam tray | It put carrots on the styrofoam tray but it took off the other items from the tray which wasn’t a command |
| Grader 1 | Grader 2 | Grader 3 | |
| Essay 1 | 1 | 0 | |
| Essay 2 | 1 | 0 | |
| Essay 3 | 0 | 0 | |
| Essay 4 | 0 | 1 | |
| Essay 5 | 0 | 1 |
| 0 | 1 | |
| user 1 | 0.91 | 0.19 |
| user 2 | 0.35 | 0.65 |
| : | ||
| user m | 0.76 | 0.24 |
| 0 | 1 | |
| user 1 | 0.14 | 0.86 |
| user 2 | 0.73 | 0.27 |
| : | ||
| user m | 0.28 | 0.72 |
| Grader 1 | Grader 2 | Grader 3 | |
| Essay 1 | 1 | 0 | |
| Essay 2 | 1 | 0 | |
| Essay 3 | 0 | 0 | |
| Essay 4 | 0 | 1 | |
| Essay 5 | 0 | 1 |
| 0 | 1 | |
| user 1 | 0.91 | 0.19 |
| user 2 | 0.35 | 0.65 |
| : | ||
| user m | 0.76 | 0.24 |
| 0 | 1 | |
| user 1 | 0.14 | 0.86 |
| user 2 | 0.73 | 0.27 |
| : | ||
| user m | 0.28 | 0.72 |
| Grader 1 | Grader 2 | Grader 3 | |
| Essay 1 | 1 | 0 | |
| Essay 2 | 1 | 0 | |
| Essay 3 | 0 | 0 | |
| Essay 4 | 0 | 1 | |
| Essay 5 | 0 | 1 |
| 0 | 1 | |
| user 1 | 0.91 | 0.19 |
| user 2 | 0.35 | 0.65 |
| : | ||
| user m | 0.76 | 0.24 |
| 0 | 1 | |
| user 1 | 0.14 | 0.86 |
| user 2 | 0.73 | 0.27 |
| : | ||
| user m | 0.28 | 0.72 |
| Grader 1 | Grader 2 | Grader 3 | |
| Essay 1 | 1 | 0 | |
| Essay 2 | 1 | 0 | |
| Essay 3 | 0 | 0 | |
| Essay 4 | 0 | 1 | |
| Essay 5 | 0 | 1 |
| 0 | 1 | |
| user 1 | 0.91 | 0.19 |
| user 2 | 0.35 | 0.65 |
| : | ||
| user m | 0.76 | 0.24 |
| 0 | 1 | |
| user 1 | 0.14 | 0.86 |
| user 2 | 0.73 | 0.27 |
| : | ||
| user m | 0.28 | 0.72 |
| Base Model | Formulation | Dataset | |||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | Medical QP | ||||||||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | ||
| Longformer | DLLP | 34.17 | 58.46 | 43.13 | 86.06 | 85.89 | 85.87 | 64.66 | 55.25 | 46.71 | 51.45 | 50.41 | 40.41 |
| LLPSelfCLR | 79.94 | 75.38 | 72.99 | 90.39 | 90.38 | 90.38 | 83.87 | 83.42 | 83.37 | 53.18 | 53.03 | 52.55 | |
| Ours | 79.92 | 80.00 | 79.94 | 92.49 | 94.48 | 92.47 | 85.61 | 85.07 | 85.01 | 53.19 | 53.03 | 52.52 | |
| Oracle | 95.72 | 95.38 | 95.33 | 95.49 | 95.48 | 95.48 | 88.63 | 88.59 | 88.58 | 86.91 | 86.37 | 86.32 | |
| RoBERTa | DLLP | 34.17 | 58.46 | 43.13 | 89.03 | 89.01 | 89.01 | 78.85 | 69.29 | 66.39 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 54.34 | 53.84 | 54.04 | 90.08 | 90.08 | 90.08 | 77.42 | 77.23 | 77.19 | 50.47 | 49.18 | 49.31 | |
| Ours | 61.06 | 61.53 | 56.21 | 90.14 | 90.13 | 90.13 | 85.41 | 85.30 | 85.29 | 53.70 | 53.53 | 53.03 | |
| Oracle | 69.61 | 64.61 | 64.39 | 93.58 | 93.58 | 93.57 | 88.47 | 88.45 | 88.44 | 83.62 | 82.92 | 82.83 | |
| BERT | DLLP | 55.10 | 58.46 | 45.69 | 84.33 | 84.33 | 84.33 | 75.04 | 74.13 | 73.89 | 56.36 | 56.32 | 56.22 |
| LLPSelfCLR | 50.71 | 56.92 | 46.95 | 86.85 | 86.82 | 86.83 | 75.04 | 74.13 | 73.89 | 66.08 | 66.00 | 65.96 | |
| Ours | 77.36 | 63.07 | 52.73 | 87.58 | 87.50 | 87.49 | 74.94 | 73.80 | 73.49 | 64.54 | 64.53 | 64.52 | |
| Oracle | 87.70 | 87.69 | 87.61 | 91.59 | 91.57 | 91.57 | 86.00 | 85.86 | 85.85 | 81.44 | 81.28 | 81.25 | |
| DistilBERT | DLLP | 34.17 | 58.46 | 43.13 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 67.92 | 67.69 | 65.60 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 52.73 | 52.70 | 52.63 | |
| Ours | 70.21 | 69.23 | 66.93 | 87.64 | 87.62 | 87.62 | 79.69 | 77.46 | 77.03 | 55.17 | 55.17 | 55.17 | |
| Oracle | 92.56 | 92.30 | 92.22 | 92.81 | 92.82 | 92.81 | 83.76 | 83.75 | 83.75 | 77.83 | 76.35 | 76.04 |
| Formulation | Longformer | RoBERTa | ||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 69.97 | 70.18 | 65.85 | 76.97 | 73.91 | 72.44 |
| LLPSelfCLR | 80.62 | 78.57 | 78.30 | 78.62 | 75.67 | 74.18 |
| Ours | 81.38 | 79.50 | 79.06 | 80.64 | 77.43 | 76.82 |
| Oracle | 82.69 | 80.53 | 80.43 | 82.64 | 80.84 | 81.18 |
| BERT | DistilBERT | |||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 63.32 | 61.90 | 58.91 | 74.14 | 68.42 | 63.92 |
| LLPSelfCLR | 73.92 | 70.80 | 70.96 | 72.43 | 73.08 | 71.80 |
| Ours | 78.02 | 75.67 | 73.88 | 76.58 | 75.46 | 72.75 |
| Oracle | 79.67 | 77.01 | 76.51 | 80.90 | 77.63 | 78.12 |
| Dataset | W-P | W-R | W-F1 | |
| Hyperpartisan | DLLP | 34.17 | 58.46 | 43.13 |
| LαTV* | 76.80 | 61.53 | 49.72 | |
| Overall | 79.92 | 80.00 | 79.94 | |
| Rotten Tomatoes | DLLP | 64.66 | 55.25 | 46.71 |
| LαTV* | 84.37 | 84.32 | 84.31 | |
| Overall | 85.61 | 85.07 | 85.01 |
| Dataset | W-P | W-R | W-F1 | |
| IMDB | DLLP | 86.06 | 85.89 | 85.87 |
| LαTV* | 88.40 | 88.36 | 88.36 | |
| Overall | 92.49 | 94.48 | 92.47 | |
| Financial Phrasebank | DLLP | 69.97 | 70.18 | 65.85 |
| LαTV* | 76.36 | 74.94 | 71.48 | |
| Overall | 81.38 | 79.50 | 79.06 |
| Dataset | # Train | # Val | # Test | Average # Words |
| Hyperpartisan | 516 | 64 | 65 | 565 |
| IMDB | 22500 | 2500 | 25000 | 231 |
| Rotten Tomatoes | 7464 | 1068 | 2130 | 21 |
| Medical QP | 2134 | 305 | 609 | 40 |
| Fin. Phrasebank | 3394 | 486 | 966 | 23 |
| Model | Dataset | |||||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | ||||||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1.0 | 0.5 | 32 | 10 | 2e-4 | 1.0 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 2.5 | 32 | 10 | 3e-3 | 0.0 | 1.0 | 32 | 10 | 2e-4 | 1.0 |
| BERT | 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1e-4 | 3.5 | 32 | 10 | 2e-3 | 1e-3 |
| DistilBERT | 16 | 10 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 0.0 | 4.0 | 32 | 10 | 2e-4 | 1e-6 |
| Model | Dataset | ||||||||
| Medical QP | Financial Phrasebank | ||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 10 | 3e-5 | 1.0 | 1.0 | 16 | 10 | 2e-4 | 1e-2 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 5.0 | 16 | 10 | 2e-4 | 1e-2 |
| BERT | 16 | 10 | 3e-5 | 1e-2 | 0.5 | 16 | 10 | 2e-3 | 0.1 |
| DistilBERT | 16 | 10 | 3e-3 | 1.0 | 0.33 | 16 | 10 | 3e-3 | 1e-2 |
| Base Model | Formulation | Dataset | |||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | Medical QP | ||||||||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | ||
| Longformer | DLLP | 34.17 | 58.46 | 43.13 | 86.06 | 85.89 | 85.87 | 64.66 | 55.25 | 46.71 | 51.45 | 50.41 | 40.41 |
| LLPSelfCLR | 79.94 | 75.38 | 72.99 | 90.39 | 90.38 | 90.38 | 83.87 | 83.42 | 83.37 | 53.18 | 53.03 | 52.55 | |
| Ours | 79.92 | 80.00 | 79.94 | 92.49 | 94.48 | 92.47 | 85.61 | 85.07 | 85.01 | 53.19 | 53.03 | 52.52 | |
| Oracle | 95.72 | 95.38 | 95.33 | 95.49 | 95.48 | 95.48 | 88.63 | 88.59 | 88.58 | 86.91 | 86.37 | 86.32 | |
| RoBERTa | DLLP | 34.17 | 58.46 | 43.13 | 89.03 | 89.01 | 89.01 | 78.85 | 69.29 | 66.39 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 54.34 | 53.84 | 54.04 | 90.08 | 90.08 | 90.08 | 77.42 | 77.23 | 77.19 | 50.47 | 49.18 | 49.31 | |
| Ours | 61.06 | 61.53 | 56.21 | 90.14 | 90.13 | 90.13 | 85.41 | 85.30 | 85.29 | 53.70 | 53.53 | 53.03 | |
| Oracle | 69.61 | 64.61 | 64.39 | 93.58 | 93.58 | 93.57 | 88.47 | 88.45 | 88.44 | 83.62 | 82.92 | 82.83 | |
| BERT | DLLP | 55.10 | 58.46 | 45.69 | 84.33 | 84.33 | 84.33 | 75.04 | 74.13 | 73.89 | 56.36 | 56.32 | 56.22 |
| LLPSelfCLR | 50.71 | 56.92 | 46.95 | 86.85 | 86.82 | 86.83 | 75.04 | 74.13 | 73.89 | 66.08 | 66.00 | 65.96 | |
| Ours | 77.36 | 63.07 | 52.73 | 87.58 | 87.50 | 87.49 | 74.94 | 73.80 | 73.49 | 64.54 | 64.53 | 64.52 | |
| Oracle | 87.70 | 87.69 | 87.61 | 91.59 | 91.57 | 91.57 | 86.00 | 85.86 | 85.85 | 81.44 | 81.28 | 81.25 | |
| DistilBERT | DLLP | 34.17 | 58.46 | 43.13 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 67.92 | 67.69 | 65.60 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 52.73 | 52.70 | 52.63 | |
| Ours | 70.21 | 69.23 | 66.93 | 87.64 | 87.62 | 87.62 | 79.69 | 77.46 | 77.03 | 55.17 | 55.17 | 55.17 | |
| Oracle | 92.56 | 92.30 | 92.22 | 92.81 | 92.82 | 92.81 | 83.76 | 83.75 | 83.75 | 77.83 | 76.35 | 76.04 |
| Formulation | Longformer | RoBERTa | ||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 69.97 | 70.18 | 65.85 | 76.97 | 73.91 | 72.44 |
| LLPSelfCLR | 80.62 | 78.57 | 78.30 | 78.62 | 75.67 | 74.18 |
| Ours | 81.38 | 79.50 | 79.06 | 80.64 | 77.43 | 76.82 |
| Oracle | 82.69 | 80.53 | 80.43 | 82.64 | 80.84 | 81.18 |
| BERT | DistilBERT | |||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 63.32 | 61.90 | 58.91 | 74.14 | 68.42 | 63.92 |
| LLPSelfCLR | 73.92 | 70.80 | 70.96 | 72.43 | 73.08 | 71.80 |
| Ours | 78.02 | 75.67 | 73.88 | 76.58 | 75.46 | 72.75 |
| Oracle | 79.67 | 77.01 | 76.51 | 80.90 | 77.63 | 78.12 |
| Dataset | W-P | W-R | W-F1 | |
| Hyperpartisan | DLLP | 34.17 | 58.46 | 43.13 |
| LαTV* | 76.80 | 61.53 | 49.72 | |
| Overall | 79.92 | 80.00 | 79.94 | |
| Rotten Tomatoes | DLLP | 64.66 | 55.25 | 46.71 |
| LαTV* | 84.37 | 84.32 | 84.31 | |
| Overall | 85.61 | 85.07 | 85.01 |
| Dataset | W-P | W-R | W-F1 | |
| IMDB | DLLP | 86.06 | 85.89 | 85.87 |
| LαTV* | 88.40 | 88.36 | 88.36 | |
| Overall | 92.49 | 94.48 | 92.47 | |
| Financial Phrasebank | DLLP | 69.97 | 70.18 | 65.85 |
| LαTV* | 76.36 | 74.94 | 71.48 | |
| Overall | 81.38 | 79.50 | 79.06 |
| Dataset | # Train | # Val | # Test | Average # Words |
| Hyperpartisan | 516 | 64 | 65 | 565 |
| IMDB | 22500 | 2500 | 25000 | 231 |
| Rotten Tomatoes | 7464 | 1068 | 2130 | 21 |
| Medical QP | 2134 | 305 | 609 | 40 |
| Fin. Phrasebank | 3394 | 486 | 966 | 23 |
| Model | Dataset | |||||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | ||||||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1.0 | 0.5 | 32 | 10 | 2e-4 | 1.0 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 2.5 | 32 | 10 | 3e-3 | 0.0 | 1.0 | 32 | 10 | 2e-4 | 1.0 |
| BERT | 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1e-4 | 3.5 | 32 | 10 | 2e-3 | 1e-3 |
| DistilBERT | 16 | 10 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 0.0 | 4.0 | 32 | 10 | 2e-4 | 1e-6 |
| Model | Dataset | ||||||||
| Medical QP | Financial Phrasebank | ||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 10 | 3e-5 | 1.0 | 1.0 | 16 | 10 | 2e-4 | 1e-2 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 5.0 | 16 | 10 | 2e-4 | 1e-2 |
| BERT | 16 | 10 | 3e-5 | 1e-2 | 0.5 | 16 | 10 | 2e-3 | 0.1 |
| DistilBERT | 16 | 10 | 3e-3 | 1.0 | 0.33 | 16 | 10 | 3e-3 | 1e-2 |
| Base Model | Formulation | Dataset | |||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | Medical QP | ||||||||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | ||
| Longformer | DLLP | 34.17 | 58.46 | 43.13 | 86.06 | 85.89 | 85.87 | 64.66 | 55.25 | 46.71 | 51.45 | 50.41 | 40.41 |
| LLPSelfCLR | 79.94 | 75.38 | 72.99 | 90.39 | 90.38 | 90.38 | 83.87 | 83.42 | 83.37 | 53.18 | 53.03 | 52.55 | |
| Ours | 79.92 | 80.00 | 79.94 | 92.49 | 94.48 | 92.47 | 85.61 | 85.07 | 85.01 | 53.19 | 53.03 | 52.52 | |
| Oracle | 95.72 | 95.38 | 95.33 | 95.49 | 95.48 | 95.48 | 88.63 | 88.59 | 88.58 | 86.91 | 86.37 | 86.32 | |
| RoBERTa | DLLP | 34.17 | 58.46 | 43.13 | 89.03 | 89.01 | 89.01 | 78.85 | 69.29 | 66.39 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 54.34 | 53.84 | 54.04 | 90.08 | 90.08 | 90.08 | 77.42 | 77.23 | 77.19 | 50.47 | 49.18 | 49.31 | |
| Ours | 61.06 | 61.53 | 56.21 | 90.14 | 90.13 | 90.13 | 85.41 | 85.30 | 85.29 | 53.70 | 53.53 | 53.03 | |
| Oracle | 69.61 | 64.61 | 64.39 | 93.58 | 93.58 | 93.57 | 88.47 | 88.45 | 88.44 | 83.62 | 82.92 | 82.83 | |
| BERT | DLLP | 55.10 | 58.46 | 45.69 | 84.33 | 84.33 | 84.33 | 75.04 | 74.13 | 73.89 | 56.36 | 56.32 | 56.22 |
| LLPSelfCLR | 50.71 | 56.92 | 46.95 | 86.85 | 86.82 | 86.83 | 75.04 | 74.13 | 73.89 | 66.08 | 66.00 | 65.96 | |
| Ours | 77.36 | 63.07 | 52.73 | 87.58 | 87.50 | 87.49 | 74.94 | 73.80 | 73.49 | 64.54 | 64.53 | 64.52 | |
| Oracle | 87.70 | 87.69 | 87.61 | 91.59 | 91.57 | 91.57 | 86.00 | 85.86 | 85.85 | 81.44 | 81.28 | 81.25 | |
| DistilBERT | DLLP | 34.17 | 58.46 | 43.13 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 67.92 | 67.69 | 65.60 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 52.73 | 52.70 | 52.63 | |
| Ours | 70.21 | 69.23 | 66.93 | 87.64 | 87.62 | 87.62 | 79.69 | 77.46 | 77.03 | 55.17 | 55.17 | 55.17 | |
| Oracle | 92.56 | 92.30 | 92.22 | 92.81 | 92.82 | 92.81 | 83.76 | 83.75 | 83.75 | 77.83 | 76.35 | 76.04 | |
| Formulation | Longformer | RoBERTa | ||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 69.97 | 70.18 | 65.85 | 76.97 | 73.91 | 72.44 |
| LLPSelfCLR | 80.62 | 78.57 | 78.30 | 78.62 | 75.67 | 74.18 |
| Ours | 81.38 | 79.50 | 79.06 | 80.64 | 77.43 | 76.82 |
| Oracle | 82.69 | 80.53 | 80.43 | 82.64 | 80.84 | 81.18 |
| BERT | DistilBERT | |||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 63.32 | 61.90 | 58.91 | 74.14 | 68.42 | 63.92 |
| LLPSelfCLR | 73.92 | 70.80 | 70.96 | 72.43 | 73.08 | 71.80 |
| Ours | 78.02 | 75.67 | 73.88 | 76.58 | 75.46 | 72.75 |
| Oracle | 79.67 | 77.01 | 76.51 | 80.90 | 77.63 | 78.12 |
| Dataset | W-P | W-R | W-F1 | |
| Hyperpartisan | DLLP | 34.17 | 58.46 | 43.13 |
| LαTV* | 76.80 | 61.53 | 49.72 | |
| Overall | 79.92 | 80.00 | 79.94 | |
| Rotten Tomatoes | DLLP | 64.66 | 55.25 | 46.71 |
| LαTV* | 84.37 | 84.32 | 84.31 | |
| Overall | 85.61 | 85.07 | 85.01 |
| Dataset | W-P | W-R | W-F1 | |
| IMDB | DLLP | 86.06 | 85.89 | 85.87 |
| LαTV* | 88.40 | 88.36 | 88.36 | |
| Overall | 92.49 | 94.48 | 92.47 | |
| Financial Phrasebank | DLLP | 69.97 | 70.18 | 65.85 |
| LαTV* | 76.36 | 74.94 | 71.48 | |
| Overall | 81.38 | 79.50 | 79.06 |
| Dataset | # Train | # Val | # Test | Average # Words |
| Hyperpartisan | 516 | 64 | 65 | 565 |
| IMDB | 22500 | 2500 | 25000 | 231 |
| Rotten Tomatoes | 7464 | 1068 | 2130 | 21 |
| Medical QP | 2134 | 305 | 609 | 40 |
| Fin. Phrasebank | 3394 | 486 | 966 | 23 |
| Model | Dataset | |||||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | ||||||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1.0 | 0.5 | 32 | 10 | 2e-4 | 1.0 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 2.5 | 32 | 10 | 3e-3 | 0.0 | 1.0 | 32 | 10 | 2e-4 | 1.0 |
| BERT | 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1e-4 | 3.5 | 32 | 10 | 2e-3 | 1e-3 |
| DistilBERT | 16 | 10 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 0.0 | 4.0 | 32 | 10 | 2e-4 | 1e-6 |
| Model | Dataset | ||||||||
| Medical QP | Financial Phrasebank | ||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 10 | 3e-5 | 1.0 | 1.0 | 16 | 10 | 2e-4 | 1e-2 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 5.0 | 16 | 10 | 2e-4 | 1e-2 |
| BERT | 16 | 10 | 3e-5 | 1e-2 | 0.5 | 16 | 10 | 2e-3 | 0.1 |
| DistilBERT | 16 | 10 | 3e-3 | 1.0 | 0.33 | 16 | 10 | 3e-3 | 1e-2 |
| Base Model | Formulation | Dataset | |||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | Medical QP | ||||||||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | W-P | W-R | W-F1 | ||
| Longformer | DLLP | 34.17 | 58.46 | 43.13 | 86.06 | 85.89 | 85.87 | 64.66 | 55.25 | 46.71 | 51.45 | 50.41 | 40.41 |
| LLPSelfCLR | 79.94 | 75.38 | 72.99 | 90.39 | 90.38 | 90.38 | 83.87 | 83.42 | 83.37 | 53.18 | 53.03 | 52.55 | |
| Ours | 79.92 | 80.00 | 79.94 | 92.49 | 94.48 | 92.47 | 85.61 | 85.07 | 85.01 | 53.19 | 53.03 | 52.52 | |
| Oracle | 95.72 | 95.38 | 95.33 | 95.49 | 95.48 | 95.48 | 88.63 | 88.59 | 88.58 | 86.91 | 86.37 | 86.32 | |
| RoBERTa | DLLP | 34.17 | 58.46 | 43.13 | 89.03 | 89.01 | 89.01 | 78.85 | 69.29 | 66.39 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 54.34 | 53.84 | 54.04 | 90.08 | 90.08 | 90.08 | 77.42 | 77.23 | 77.19 | 50.47 | 49.18 | 49.31 | |
| Ours | 61.06 | 61.53 | 56.21 | 90.14 | 90.13 | 90.13 | 85.41 | 85.30 | 85.29 | 53.70 | 53.53 | 53.03 | |
| Oracle | 69.61 | 64.61 | 64.39 | 93.58 | 93.58 | 93.57 | 88.47 | 88.45 | 88.44 | 83.62 | 82.92 | 82.83 | |
| BERT | DLLP | 55.10 | 58.46 | 45.69 | 84.33 | 84.33 | 84.33 | 75.04 | 74.13 | 73.89 | 56.36 | 56.32 | 56.22 |
| LLPSelfCLR | 50.71 | 56.92 | 46.95 | 86.85 | 86.82 | 86.83 | 75.04 | 74.13 | 73.89 | 66.08 | 66.00 | 65.96 | |
| Ours | 77.36 | 63.07 | 52.73 | 87.58 | 87.50 | 87.49 | 74.94 | 73.80 | 73.49 | 64.54 | 64.53 | 64.52 | |
| Oracle | 87.70 | 87.69 | 87.61 | 91.59 | 91.57 | 91.57 | 86.00 | 85.86 | 85.85 | 81.44 | 81.28 | 81.25 | |
| DistilBERT | DLLP | 34.17 | 58.46 | 43.13 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 25.08 | 50.08 | 33.42 |
| LLPSelfCLR | 67.92 | 67.69 | 65.60 | 86.89 | 86.86 | 86.85 | 79.63 | 77.41 | 76.98 | 52.73 | 52.70 | 52.63 | |
| Ours | 70.21 | 69.23 | 66.93 | 87.64 | 87.62 | 87.62 | 79.69 | 77.46 | 77.03 | 55.17 | 55.17 | 55.17 | |
| Oracle | 92.56 | 92.30 | 92.22 | 92.81 | 92.82 | 92.81 | 83.76 | 83.75 | 83.75 | 77.83 | 76.35 | 76.04 |
| Formulation | Longformer | RoBERTa | ||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 69.97 | 70.18 | 65.85 | 76.97 | 73.91 | 72.44 |
| LLPSelfCLR | 80.62 | 78.57 | 78.30 | 78.62 | 75.67 | 74.18 |
| Ours | 81.38 | 79.50 | 79.06 | 80.64 | 77.43 | 76.82 |
| Oracle | 82.69 | 80.53 | 80.43 | 82.64 | 80.84 | 81.18 |
| BERT | DistilBERT | |||||
| W-P | W-R | W-F1 | W-P | W-R | W-F1 | |
| DLLP | 63.32 | 61.90 | 58.91 | 74.14 | 68.42 | 63.92 |
| LLPSelfCLR | 73.92 | 70.80 | 70.96 | 72.43 | 73.08 | 71.80 |
| Ours | 78.02 | 75.67 | 73.88 | 76.58 | 75.46 | 72.75 |
| Oracle | 79.67 | 77.01 | 76.51 | 80.90 | 77.63 | 78.12 |
| Dataset | W-P | W-R | W-F1 | |
| Hyperpartisan | DLLP | 34.17 | 58.46 | 43.13 |
| LαTV* | 76.80 | 61.53 | 49.72 | |
| Overall | 79.92 | 80.00 | 79.94 | |
| Rotten Tomatoes | DLLP | 64.66 | 55.25 | 46.71 |
| LαTV* | 84.37 | 84.32 | 84.31 | |
| Overall | 85.61 | 85.07 | 85.01 |
| Dataset | W-P | W-R | W-F1 | |
| IMDB | DLLP | 86.06 | 85.89 | 85.87 |
| LαTV* | 88.40 | 88.36 | 88.36 | |
| Overall | 92.49 | 94.48 | 92.47 | |
| Financial Phrasebank | DLLP | 69.97 | 70.18 | 65.85 |
| LαTV* | 76.36 | 74.94 | 71.48 | |
| Overall | 81.38 | 79.50 | 79.06 |
| Dataset | # Train | # Val | # Test | Average # Words |
| Hyperpartisan | 516 | 64 | 65 | 565 |
| IMDB | 22500 | 2500 | 25000 | 231 |
| Rotten Tomatoes | 7464 | 1068 | 2130 | 21 |
| Medical QP | 2134 | 305 | 609 | 40 |
| Fin. Phrasebank | 3394 | 486 | 966 | 23 |
| Model | Dataset | |||||||||||||
| Hyperpartisan | IMDB | Rotten Tomatoes | ||||||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1.0 | 0.5 | 32 | 10 | 2e-4 | 1.0 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 2.5 | 32 | 10 | 3e-3 | 0.0 | 1.0 | 32 | 10 | 2e-4 | 1.0 |
| BERT | 16 | 20 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 1e-4 | 3.5 | 32 | 10 | 2e-3 | 1e-3 |
| DistilBERT | 16 | 10 | 5e-5 | 1.0 | 1.0 | 32 | 10 | 3e-3 | 0.0 | 4.0 | 32 | 10 | 2e-4 | 1e-6 |
| Model | Dataset | ||||||||
| Medical QP | Financial Phrasebank | ||||||||
| Longformer | B | E | LR | λ | α | B | E | LR | λ |
| 16 | 10 | 3e-5 | 1.0 | 1.0 | 16 | 10 | 2e-4 | 1e-2 | |
| RoBERTa | 16 | 10 | 3e-3 | 1.0 | 5.0 | 16 | 10 | 2e-4 | 1e-2 |
| BERT | 16 | 10 | 3e-5 | 1e-2 | 0.5 | 16 | 10 | 2e-3 | 0.1 |
| DistilBERT | 16 | 10 | 3e-3 | 1.0 | 0.33 | 16 | 10 | 3e-3 | 1e-2 |
| Dataset | Count | Pos-o | Pos-l | Uncl. |
| CivComments | 418 | - | 10 (2.3) | 17 (4.5) |
| SBIC | 20 | 9 (45.0) | 2 (10.0) | 1 (5.0) |
| HateXplain | 20 | 8 (40.0) | 10 (50.0) | 7 (13.0) |
| IHS | 20 | 9 (45.0) | 0 (0.0) | 0 (0.0) |
| CounterContext | 10 | 7 (70.0) | 0 (0.0) | 0 (0.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Definition | [target] states [description]. Protected groups include [groups]. |
| Classification | Is this post a violation of [target]? Please answer Yes, No, or Unclear. |
| Reasoning | Is this post a violation of [target]? Please answer Yes/No/Unclear. If Yes, explain why. |
| Model | Coarse-grained | Fine-grained | Fine-grained + reasoning | |||||||||||
| Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | |
| GPT-3.5 | 30.7 | 46.5 | 52.5 | 27.1 | 0.05 | 45.4 | 30.0 | 96.5 | 45.3 | 31.8 | 19.1 | 99.7 | 32.3 | 0.6 |
| GPT-3.5-16k | 30.9 | 31.1 | 43.3 | 73.9 | 0.6 | 43.4 | 28.2 | 97.8 | 44.2 | 33.7 | 20.7 | 97.1 | 32.6 | 0.6 |
| GPT-4 | 41.4 | 38.1 | 48.0 | 92.5 | 0.04 | - | - | - | - | 52.2 | 43.9 | 67.7 | 55.1 | 0.04 |
| Falcon-7b | 17.4 | 25.1 | 24.6 | 39.7 | 14.0 | 12.1 | 7.0 | 46.1 | 12.1 | 10.0 | 6.0 | 31.6 | 10.2 | 63.4 |
| Alpaca-LoRA-7b | 21.1 | 28.6 | 34.7 | 46.4 | 0.06 | 15.1 | 8.4 | 87.1 | 14.8 | 10.5 | 5.6 | 92.9 | 10.9 | 21.7 |
| WizardLM-13b | 9.2 | 28.5 | 25.9 | 11.3 | 28.5 | 20.2 | 11.8 | 78.1 | 15.2 | 15.2 | 8.4 | 87.1 | 14.8 | 75.6 |
| Llama2-13b | 8.5 | 26.8 | 24.7 | 16.3 | 24.2 | 13.5 | 7.5 | 74.0 | 13.8 | 20.8 | 12.8 | 59.5 | 21.1 | 54.2 |
| RoBERTa-L (LM-BFF) | 5.4 | 35.8 | 33.7 | 8.5 | - | 8.8 | 4.8 | 62.7 | 9.1 | - | - | - | - | - |
| RoBERTa-L (T) | 1.1 | 33.3 | 0.6 | 1.7 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| RoBERTa-L (ST) | 32.0 | 32.2 | 33.3 | 90.1 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| Random | 22.6 | 30.8 | 33.3 | 46.3 | - | 10.6 | 6.0 | 51.0 | 10.7 | - | - | - | - | - |
| Model | Coarse-grained | Reasoning | ||
| Ma-f1 ↑ | Mi-f1 ↑ | HR ↓ | J ↑ | |
| GPT-3.5 | 31.5 | 77.0 | 0.10 | 37.2 |
| GPT-3.5-16k | 31.3 | 76.6 | 0.04 | 36.4 |
| GPT-4 | 41.4 | 92.5 | 0.05 | 38.4 |
| Sample | Did he kill your sex life? Women are objects so you can just buy another one anyways. |
| Source | Response |
| GPT-3.5 | Yes, this post is a violation of TOS_red. It promotes hate based on gender and objectifies women, which goes against the rule on hate and promoting attacks of hate. |
| GPT-4 | Yes, this post violates several of the above policies. 1. CC_318 and CC_319: This post can be seen as promoting hatred and disrespect towards women, which is a protected group under these policies. It devalues women by referring to them as objects that can be bought, which can be seen as advocating for their mistreatment... |
| WizardLM | Yes, this post is a violation of CC_318 as it promotes the idea that women are objects to be used and discarded, which is an act that falls under the definition of genocide. |
| Annotator | ‘Wilful promotion of hatred’, ‘degrades women’, ‘The view of women as objects rises to the level of ardent and extreme feelings constituting hatred required to uphold the constitutionality of a prohibition of expression in human rights legislation’, ‘Hate speech and discriminatory against identifiable group’ |
| Forum | Count | Pos-o | Pos-l | Uncl. |
| CivilComments | 418 | - | 10 (2.4) | 17 (4.1) |
| 39 | 14 (35.9) | 3 (7.7) | 5 (12.8) | |
| 16 | 11 (68.8) | 0 (0.0) | 2 (12.5) | |
| Gab | 13 | 8 (61.5) | 9 (69.2) | 3 (23.1) |
| Stormfront | 2 | 0 | 0 | 1 (50.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Legal source | Criminal Code, R.S.C. 1985, c C-46, s. 318 |
| Prompt | CC_318 is defined as Every person who advocates or promotes genocide is guilty of an indictable offence and liable to imprisonment for a term of not more than five years. In this section, genocide means any of the following acts committed with intent to destroy in whole or in part any identifiable group, namely, killing members of the group; or deliberately inflicting on the group conditions of life calculated to bring about its physical destruction. An intent to directly prompt or provoke another person to commit genocide is enough to establish the mens rea of the offence.\nProtected/identifiable groups include colour, race, religion, national or ethnic origin, age, sex, sexual orientation, gender identity or expression, mental or physical disability\nPost: The Indians need to purchase their own land and building. Canadians do not allow 6 year olds to have the temper tantrums that Indians inflict on Canadians.\nQ: Is this post a violation of CC_318? Please answer yes or no. A: <mask> |
| Criminal Code s.318: Advocating Genocide | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Criminal Code s.319: Public Incitement of Hatred | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Alberta | Age, Religion/Religious beliefs, Source of income, Family status, Sex/Gender identity or expression, Marital status, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, British Columbia | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Northwest Territories | Age, Religion/Religious beliefs, Family status, Family affiliation, Sex/Gender identity or expression, Conviction that is subject to a pardon or record suspension, Marital status, Social condition, Race (colour, creed), Sexual orientation, Political belief/association, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Saskatchewan | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Receipt of public assistance, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service, Meta | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Sex/Gender identity or expression, Serious disease, Race (colour, creed), Sexual orientation, Physical or mental disability, Family affiliation |
| Terms of Service, Reddit | Religion/Religious beliefs, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service,TikTok | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Immigration status, Serious disease, Family affiliation |
| Terms of Service,Twitter | Age, National or ethnic origin (nationality, ethnicity, ancestry), Religion/Religious beliefs, Sex/Gender identity or expression, Serious disease, Sexual orientation, Race (colour, creed), Physical or mental disability, Family affiliation |
| Terms of Service,Youtube | Age, Religion/Religious beliefs, Family affiliation, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Veteran status, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Race (colour, creed) | actual and perceived race, colour, coloured, black, creed, racially, races, race or perceived race |
| National or ethnic origin (nation-ality, ethnicity, ancestry) | ancestry, national, ethnic origin, ethnicity, place of origin, national origin, national or ethnic origin, nationality, indigenous identity |
| Political belief/association | political association, political belief |
| Sex/Gender identity or expres-sion | gender, woman, men, transgender, boy, man, gender identity, gender identity and expression, gender iden-tity or expression, sex, children, whore, sexist, baby, sex/gender |
| Religion/Religious beliefs | religion, cross, belief, christianity, judaism, faith, crosses, religions, nuns, religious affiliation, reli-gion/religious, religious beliefs |
| Sexual orientation | sexual orientation |
| Social condition | social condition |
| Immigration status | immigration status, immigration |
| Source of income | source of income |
| Age | age, young, ages, old |
| Physical or mental disability | physical or mental disability, mental or physical dis-ability, physical disability, mental disability, disabil-ity, impairment, disabilities, pregnancy or disability, adhd |
| Family affiliation | family affiliation, caste |
| Conviction that is subject to a par-\ndon or record suspension | conviction that is subject to a pardon or record sus-pension |
| Receipt of public assistance | receipt of public assistance |
| Serious disease | serious disease, aids |
| Family status | family status |
| Pregnancy | pregnancy |
| Victims of a major violent event \nand their families/kin | victims of a major violent event and their kin, victims of a major event and their families |
| Veteran status | veteran status |
| Marital status | marital status |
| Dataset | Count | Pos-o | Pos-l | Uncl. |
| CivComments | 418 | - | 10 (2.3) | 17 (4.5) |
| SBIC | 20 | 9 (45.0) | 2 (10.0) | 1 (5.0) |
| HateXplain | 20 | 8 (40.0) | 10 (50.0) | 7 (13.0) |
| IHS | 20 | 9 (45.0) | 0 (0.0) | 0 (0.0) |
| CounterContext | 10 | 7 (70.0) | 0 (0.0) | 0 (0.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Definition | [target] states [description]. Protected groups include [groups]. |
| Classification | Is this post a violation of [target]? Please answer Yes, No, or Unclear. |
| Reasoning | Is this post a violation of [target]? Please answer Yes/No/Unclear. If Yes, explain why. |
| Model | Coarse-grained | Fine-grained | Fine-grained + reasoning | |||||||||||
| Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | |
| GPT-3.5 | 30.7 | 46.5 | 52.5 | 27.1 | 0.05 | 45.4 | 30.0 | 96.5 | 45.3 | 31.8 | 19.1 | 99.7 | 32.3 | 0.6 |
| GPT-3.5-16k | 30.9 | 31.1 | 43.3 | 73.9 | 0.6 | 43.4 | 28.2 | 97.8 | 44.2 | 33.7 | 20.7 | 97.1 | 32.6 | 0.6 |
| GPT-4 | 41.4 | 38.1 | 48.0 | 92.5 | 0.04 | - | - | - | - | 52.2 | 43.9 | 67.7 | 55.1 | 0.04 |
| Falcon-7b | 17.4 | 25.1 | 24.6 | 39.7 | 14.0 | 12.1 | 7.0 | 46.1 | 12.1 | 10.0 | 6.0 | 31.6 | 10.2 | 63.4 |
| Alpaca-LoRA-7b | 21.1 | 28.6 | 34.7 | 46.4 | 0.06 | 15.1 | 8.4 | 87.1 | 14.8 | 10.5 | 5.6 | 92.9 | 10.9 | 21.7 |
| WizardLM-13b | 9.2 | 28.5 | 25.9 | 11.3 | 28.5 | 20.2 | 11.8 | 78.1 | 15.2 | 15.2 | 8.4 | 87.1 | 14.8 | 75.6 |
| Llama2-13b | 8.5 | 26.8 | 24.7 | 16.3 | 24.2 | 13.5 | 7.5 | 74.0 | 13.8 | 20.8 | 12.8 | 59.5 | 21.1 | 54.2 |
| RoBERTa-L (LM-BFF) | 5.4 | 35.8 | 33.7 | 8.5 | - | 8.8 | 4.8 | 62.7 | 9.1 | - | - | - | - | - |
| RoBERTa-L (T) | 1.1 | 33.3 | 0.6 | 1.7 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| RoBERTa-L (ST) | 32.0 | 32.2 | 33.3 | 90.1 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| Random | 22.6 | 30.8 | 33.3 | 46.3 | - | 10.6 | 6.0 | 51.0 | 10.7 | - | - | - | - | - |
| Model | Coarse-grained | Reasoning | ||
| Ma-f1 ↑ | Mi-f1 ↑ | HR ↓ | J ↑ | |
| GPT-3.5 | 31.5 | 77.0 | 0.10 | 37.2 |
| GPT-3.5-16k | 31.3 | 76.6 | 0.04 | 36.4 |
| GPT-4 | 41.4 | 92.5 | 0.05 | 38.4 |
| Sample | Did he kill your sex life? Women are objects so you can just buy another one anyways. |
| Source | Response |
| GPT-3.5 | Yes, this post is a violation of TOS_red. It promotes hate based on gender and objectifies women, which goes against the rule on hate and promoting attacks of hate. |
| GPT-4 | Yes, this post violates several of the above policies. 1. CC_318 and CC_319: This post can be seen as promoting hatred and disrespect towards women, which is a protected group under these policies. It devalues women by referring to them as objects that can be bought, which can be seen as advocating for their mistreatment... |
| WizardLM | Yes, this post is a violation of CC_318 as it promotes the idea that women are objects to be used and discarded, which is an act that falls under the definition of genocide. |
| Annotator | ‘Wilful promotion of hatred’, ‘degrades women’, ‘The view of women as objects rises to the level of ardent and extreme feelings constituting hatred required to uphold the constitutionality of a prohibition of expression in human rights legislation’, ‘Hate speech and discriminatory against identifiable group’ |
| Forum | Count | Pos-o | Pos-l | Uncl. |
| CivilComments | 418 | - | 10 (2.4) | 17 (4.1) |
| 39 | 14 (35.9) | 3 (7.7) | 5 (12.8) | |
| 16 | 11 (68.8) | 0 (0.0) | 2 (12.5) | |
| Gab | 13 | 8 (61.5) | 9 (69.2) | 3 (23.1) |
| Stormfront | 2 | 0 | 0 | 1 (50.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Legal source | Criminal Code, R.S.C. 1985, c C-46, s. 318 |
| Prompt | CC_318 is defined as Every person who advocates or promotes genocide is guilty of an indictable offence and liable to imprisonment for a term of not more than five years. In this section, genocide means any of the following acts committed with intent to destroy in whole or in part any identifiable group, namely, killing members of the group; or deliberately inflicting on the group conditions of life calculated to bring about its physical destruction. An intent to directly prompt or provoke another person to commit genocide is enough to establish the mens rea of the offence.\nProtected/identifiable groups include colour, race, religion, national or ethnic origin, age, sex, sexual orientation, gender identity or expression, mental or physical disability\nPost: The Indians need to purchase their own land and building. Canadians do not allow 6 year olds to have the temper tantrums that Indians inflict on Canadians.\nQ: Is this post a violation of CC_318? Please answer yes or no. A: <mask> |
| Criminal Code s.318: Advocating Genocide | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Criminal Code s.319: Public Incitement of Hatred | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Alberta | Age, Religion/Religious beliefs, Source of income, Family status, Sex/Gender identity or expression, Marital status, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, British Columbia | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Northwest Territories | Age, Religion/Religious beliefs, Family status, Family affiliation, Sex/Gender identity or expression, Conviction that is subject to a pardon or record suspension, Marital status, Social condition, Race (colour, creed), Sexual orientation, Political belief/association, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Saskatchewan | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Receipt of public assistance, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service, Meta | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Sex/Gender identity or expression, Serious disease, Race (colour, creed), Sexual orientation, Physical or mental disability, Family affiliation |
| Terms of Service, Reddit | Religion/Religious beliefs, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service,TikTok | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Immigration status, Serious disease, Family affiliation |
| Terms of Service,Twitter | Age, National or ethnic origin (nationality, ethnicity, ancestry), Religion/Religious beliefs, Sex/Gender identity or expression, Serious disease, Sexual orientation, Race (colour, creed), Physical or mental disability, Family affiliation |
| Terms of Service,Youtube | Age, Religion/Religious beliefs, Family affiliation, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Veteran status, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Race (colour, creed) | actual and perceived race, colour, coloured, black, creed, racially, races, race or perceived race |
| National or ethnic origin (nation-ality, ethnicity, ancestry) | ancestry, national, ethnic origin, ethnicity, place of origin, national origin, national or ethnic origin, nationality, indigenous identity |
| Political belief/association | political association, political belief |
| Sex/Gender identity or expres-sion | gender, woman, men, transgender, boy, man, gender identity, gender identity and expression, gender iden-tity or expression, sex, children, whore, sexist, baby, sex/gender |
| Religion/Religious beliefs | religion, cross, belief, christianity, judaism, faith, crosses, religions, nuns, religious affiliation, reli-gion/religious, religious beliefs |
| Sexual orientation | sexual orientation |
| Social condition | social condition |
| Immigration status | immigration status, immigration |
| Source of income | source of income |
| Age | age, young, ages, old |
| Physical or mental disability | physical or mental disability, mental or physical dis-ability, physical disability, mental disability, disabil-ity, impairment, disabilities, pregnancy or disability, adhd |
| Family affiliation | family affiliation, caste |
| Conviction that is subject to a par-\ndon or record suspension | conviction that is subject to a pardon or record sus-pension |
| Receipt of public assistance | receipt of public assistance |
| Serious disease | serious disease, aids |
| Family status | family status |
| Pregnancy | pregnancy |
| Victims of a major violent event \nand their families/kin | victims of a major violent event and their kin, victims of a major event and their families |
| Veteran status | veteran status |
| Marital status | marital status |
| Dataset | Count | Pos-o | Pos-l | Uncl. |
| CivComments | 418 | - | 10 (2.3) | 17 (4.5) |
| SBIC | 20 | 9 (45.0) | 2 (10.0) | 1 (5.0) |
| HateXplain | 20 | 8 (40.0) | 10 (50.0) | 7 (13.0) |
| IHS | 20 | 9 (45.0) | 0 (0.0) | 0 (0.0) |
| CounterContext | 10 | 7 (70.0) | 0 (0.0) | 0 (0.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Definition | [target] states [description]. Protected groups include [groups]. |
| Classification | Is this post a violation of [target]? Please answer Yes, No, or Unclear. |
| Reasoning | Is this post a violation of [target]? Please answer Yes/No/Unclear. If Yes, explain why. |
| Model | Coarse-grained | Fine-grained | Fine-grained + reasoning | |||||||||||
| Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | |
| GPT-3.5 | 30.7 | 46.5 | 52.5 | 27.1 | 0.05 | 45.4 | 30.0 | 96.5 | 45.3 | 31.8 | 19.1 | 99.7 | 32.3 | 0.6 |
| GPT-3.5-16k | 30.9 | 31.1 | 43.3 | 73.9 | 0.6 | 43.4 | 28.2 | 97.8 | 44.2 | 33.7 | 20.7 | 97.1 | 32.6 | 0.6 |
| GPT-4 | 41.4 | 38.1 | 48.0 | 92.5 | 0.04 | - | - | - | - | 52.2 | 43.9 | 67.7 | 55.1 | 0.04 |
| Falcon-7b | 17.4 | 25.1 | 24.6 | 39.7 | 14.0 | 12.1 | 7.0 | 46.1 | 12.1 | 10.0 | 6.0 | 31.6 | 10.2 | 63.4 |
| Alpaca-LoRA-7b | 21.1 | 28.6 | 34.7 | 46.4 | 0.06 | 15.1 | 8.4 | 87.1 | 14.8 | 10.5 | 5.6 | 92.9 | 10.9 | 21.7 |
| WizardLM-13b | 9.2 | 28.5 | 25.9 | 11.3 | 28.5 | 20.2 | 11.8 | 78.1 | 15.2 | 15.2 | 8.4 | 87.1 | 14.8 | 75.6 |
| Llama2-13b | 8.5 | 26.8 | 24.7 | 16.3 | 24.2 | 13.5 | 7.5 | 74.0 | 13.8 | 20.8 | 12.8 | 59.5 | 21.1 | 54.2 |
| RoBERTa-L (LM-BFF) | 5.4 | 35.8 | 33.7 | 8.5 | - | 8.8 | 4.8 | 62.7 | 9.1 | - | - | - | - | - |
| RoBERTa-L (T) | 1.1 | 33.3 | 0.6 | 1.7 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| RoBERTa-L (ST) | 32.0 | 32.2 | 33.3 | 90.1 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| Random | 22.6 | 30.8 | 33.3 | 46.3 | - | 10.6 | 6.0 | 51.0 | 10.7 | - | - | - | - | - |
| Model | Coarse-grained | Reasoning | ||
| Ma-f1 ↑ | Mi-f1 ↑ | HR ↓ | J ↑ | |
| GPT-3.5 | 31.5 | 77.0 | 0.10 | 37.2 |
| GPT-3.5-16k | 31.3 | 76.6 | 0.04 | 36.4 |
| GPT-4 | 41.4 | 92.5 | 0.05 | 38.4 |
| Sample | Did he kill your sex life? Women are objects so you can just buy another one anyways. |
| Source | Response |
| GPT-3.5 | Yes, this post is a violation of TOS_red. It promotes hate based on gender and objectifies women, which goes against the rule on hate and promoting attacks of hate. |
| GPT-4 | Yes, this post violates several of the above policies. 1. CC_318 and CC_319: This post can be seen as promoting hatred and disrespect towards women, which is a protected group under these policies. It devalues women by referring to them as objects that can be bought, which can be seen as advocating for their mistreatment... |
| WizardLM | Yes, this post is a violation of CC_318 as it promotes the idea that women are objects to be used and discarded, which is an act that falls under the definition of genocide. |
| Annotator | ‘Wilful promotion of hatred’, ‘degrades women’, ‘The view of women as objects rises to the level of ardent and extreme feelings constituting hatred required to uphold the constitutionality of a prohibition of expression in human rights legislation’, ‘Hate speech and discriminatory against identifiable group’ |
| Forum | Count | Pos-o | Pos-l | Uncl. |
| CivilComments | 418 | - | 10 (2.4) | 17 (4.1) |
| 39 | 14 (35.9) | 3 (7.7) | 5 (12.8) | |
| 16 | 11 (68.8) | 0 (0.0) | 2 (12.5) | |
| Gab | 13 | 8 (61.5) | 9 (69.2) | 3 (23.1) |
| Stormfront | 2 | 0 | 0 | 1 (50.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Legal source | Criminal Code, R.S.C. 1985, c C-46, s. 318 |
| Prompt | CC_318 is defined as Every person who advocates or promotes genocide is guilty of an indictable offence and liable to imprisonment for a term of not more than five years. In this section, genocide means any of the following acts committed with intent to destroy in whole or in part any identifiable group, namely, killing members of the group; or deliberately inflicting on the group conditions of life calculated to bring about its physical destruction. An intent to directly prompt or provoke another person to commit genocide is enough to establish the mens rea of the offence. +Protected/identifiable groups include colour, race, religion, national or ethnic origin, age, sex, sexual orientation, gender identity or expression, mental or physical disability +Post: The Indians need to purchase their own land and building. Canadians do not allow 6 year olds to have the temper tantrums that Indians inflict on Canadians. +Q: Is this post a violation of CC_318? Please answer yes or no. A: <mask> |
| Criminal Code s.318: Advocating Genocide | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Criminal Code s.319: Public Incitement of Hatred | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Alberta | Age, Religion/Religious beliefs, Source of income, Family status, Sex/Gender identity or expression, Marital status, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, British Columbia | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Northwest Territories | Age, Religion/Religious beliefs, Family status, Family affiliation, Sex/Gender identity or expression, Conviction that is subject to a pardon or record suspension, Marital status, Social condition, Race (colour, creed), Sexual orientation, Political belief/association, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Saskatchewan | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Receipt of public assistance, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service, Meta | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Sex/Gender identity or expression, Serious disease, Race (colour, creed), Sexual orientation, Physical or mental disability, Family affiliation |
| Terms of Service, Reddit | Religion/Religious beliefs, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service,TikTok | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Immigration status, Serious disease, Family affiliation |
| Terms of Service,Twitter | Age, National or ethnic origin (nationality, ethnicity, ancestry), Religion/Religious beliefs, Sex/Gender identity or expression, Serious disease, Sexual orientation, Race (colour, creed), Physical or mental disability, Family affiliation |
| Terms of Service,Youtube | Age, Religion/Religious beliefs, Family affiliation, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Veteran status, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Race (colour, creed) | actual and perceived race, colour, coloured, black, creed, racially, races, race or perceived race |
| National or ethnic origin (nation-ality, ethnicity, ancestry) | ancestry, national, ethnic origin, ethnicity, place of origin, national origin, national or ethnic origin, nationality, indigenous identity |
| Political belief/association | political association, political belief |
| Sex/Gender identity or expres-sion | gender, woman, men, transgender, boy, man, gender identity, gender identity and expression, gender iden-tity or expression, sex, children, whore, sexist, baby, sex/gender |
| Religion/Religious beliefs | religion, cross, belief, christianity, judaism, faith, crosses, religions, nuns, religious affiliation, reli-gion/religious, religious beliefs |
| Sexual orientation | sexual orientation |
| Social condition | social condition |
| Immigration status | immigration status, immigration |
| Source of income | source of income |
| Age | age, young, ages, old |
| Physical or mental disability | physical or mental disability, mental or physical dis-ability, physical disability, mental disability, disabil-ity, impairment, disabilities, pregnancy or disability, adhd |
| Family affiliation | family affiliation, caste |
| Conviction that is subject to a par- +don or record suspension | conviction that is subject to a pardon or record sus-pension |
| Receipt of public assistance | receipt of public assistance |
| Serious disease | serious disease, aids |
| Family status | family status |
| Pregnancy | pregnancy |
| Victims of a major violent event +and their families/kin | victims of a major violent event and their kin, victims of a major event and their families |
| Veteran status | veteran status |
| Marital status | marital status |
| Dataset | Count | Pos-o | Pos-l | Uncl. |
| CivComments | 418 | - | 10 (2.3) | 17 (4.5) |
| SBIC | 20 | 9 (45.0) | 2 (10.0) | 1 (5.0) |
| HateXplain | 20 | 8 (40.0) | 10 (50.0) | 7 (13.0) |
| IHS | 20 | 9 (45.0) | 0 (0.0) | 0 (0.0) |
| CounterContext | 10 | 7 (70.0) | 0 (0.0) | 0 (0.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Definition | [target] states [description]. Protected groups include [groups]. |
| Classification | Is this post a violation of [target]? Please answer Yes, No, or Unclear. |
| Reasoning | Is this post a violation of [target]? Please answer Yes/No/Unclear. If Yes, explain why. |
| Model | Coarse-grained | Fine-grained | Fine-grained + reasoning | |||||||||||
| Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | Ma-f1↑ | Ma-P↑ | Ma-R↑ | Mi-f1↑ | HR↓ | |
| GPT-3.5 | 30.7 | 46.5 | 52.5 | 27.1 | 0.05 | 45.4 | 30.0 | 96.5 | 45.3 | 31.8 | 19.1 | 99.7 | 32.3 | 0.6 |
| GPT-3.5-16k | 30.9 | 31.1 | 43.3 | 73.9 | 0.6 | 43.4 | 28.2 | 97.8 | 44.2 | 33.7 | 20.7 | 97.1 | 32.6 | 0.6 |
| GPT-4 | 41.4 | 38.1 | 48.0 | 92.5 | 0.04 | - | - | - | - | 52.2 | 43.9 | 67.7 | 55.1 | 0.04 |
| Falcon-7b | 17.4 | 25.1 | 24.6 | 39.7 | 14.0 | 12.1 | 7.0 | 46.1 | 12.1 | 10.0 | 6.0 | 31.6 | 10.2 | 63.4 |
| Alpaca-LoRA-7b | 21.1 | 28.6 | 34.7 | 46.4 | 0.06 | 15.1 | 8.4 | 87.1 | 14.8 | 10.5 | 5.6 | 92.9 | 10.9 | 21.7 |
| WizardLM-13b | 9.2 | 28.5 | 25.9 | 11.3 | 28.5 | 20.2 | 11.8 | 78.1 | 15.2 | 15.2 | 8.4 | 87.1 | 14.8 | 75.6 |
| Llama2-13b | 8.5 | 26.8 | 24.7 | 16.3 | 24.2 | 13.5 | 7.5 | 74.0 | 13.8 | 20.8 | 12.8 | 59.5 | 21.1 | 54.2 |
| RoBERTa-L (LM-BFF) | 5.4 | 35.8 | 33.7 | 8.5 | - | 8.8 | 4.8 | 62.7 | 9.1 | - | - | - | - | - |
| RoBERTa-L (T) | 1.1 | 33.3 | 0.6 | 1.7 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| RoBERTa-L (ST) | 32.0 | 32.2 | 33.3 | 90.1 | - | 0 | 0 | 0 | 0 | - | - | - | - | - |
| Random | 22.6 | 30.8 | 33.3 | 46.3 | - | 10.6 | 6.0 | 51.0 | 10.7 | - | - | - | - | - |
| Model | Coarse-grained | Reasoning | ||
| Ma-f1 ↑ | Mi-f1 ↑ | HR ↓ | J ↑ | |
| GPT-3.5 | 31.5 | 77.0 | 0.10 | 37.2 |
| GPT-3.5-16k | 31.3 | 76.6 | 0.04 | 36.4 |
| GPT-4 | 41.4 | 92.5 | 0.05 | 38.4 |
| Sample | Did he kill your sex life? Women are objects so you can just buy another one anyways. |
| Source | Response |
| GPT-3.5 | Yes, this post is a violation of TOS_red. It promotes hate based on gender and objectifies women, which goes against the rule on hate and promoting attacks of hate. |
| GPT-4 | Yes, this post violates several of the above policies. 1. CC_318 and CC_319: This post can be seen as promoting hatred and disrespect towards women, which is a protected group under these policies. It devalues women by referring to them as objects that can be bought, which can be seen as advocating for their mistreatment... |
| WizardLM | Yes, this post is a violation of CC_318 as it promotes the idea that women are objects to be used and discarded, which is an act that falls under the definition of genocide. |
| Annotator | ‘Wilful promotion of hatred’, ‘degrades women’, ‘The view of women as objects rises to the level of ardent and extreme feelings constituting hatred required to uphold the constitutionality of a prohibition of expression in human rights legislation’, ‘Hate speech and discriminatory against identifiable group’ |
| Forum | Count | Pos-o | Pos-l | Uncl. |
| CivilComments | 418 | - | 10 (2.4) | 17 (4.1) |
| 39 | 14 (35.9) | 3 (7.7) | 5 (12.8) | |
| 16 | 11 (68.8) | 0 (0.0) | 2 (12.5) | |
| Gab | 13 | 8 (61.5) | 9 (69.2) | 3 (23.1) |
| Stormfront | 2 | 0 | 0 | 1 (50.0) |
| Total | 488 | 39 (7.9) | 22 (4.5) | 28 (5.7) |
| with Augment. | 704 | - | 165 (23.4) | 38 (5.4) |
| Legal source | Criminal Code, R.S.C. 1985, c C-46, s. 318 |
| Prompt | CC_318 is defined as Every person who advocates or promotes genocide is guilty of an indictable offence and liable to imprisonment for a term of not more than five years. In this section, genocide means any of the following acts committed with intent to destroy in whole or in part any identifiable group, namely, killing members of the group; or deliberately inflicting on the group conditions of life calculated to bring about its physical destruction. An intent to directly prompt or provoke another person to commit genocide is enough to establish the mens rea of the offence.\nProtected/identifiable groups include colour, race, religion, national or ethnic origin, age, sex, sexual orientation, gender identity or expression, mental or physical disability\nPost: The Indians need to purchase their own land and building. Canadians do not allow 6 year olds to have the temper tantrums that Indians inflict on Canadians.\nQ: Is this post a violation of CC_318? Please answer yes or no. A: <mask> |
| Criminal Code s.318: Advocating Genocide | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Criminal Code s.319: Public Incitement of Hatred | Age, Religion/Religious beliefs, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Alberta | Age, Religion/Religious beliefs, Source of income, Family status, Sex/Gender identity or expression, Marital status, Race (colour, creed), Sexual orientation, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, British Columbia | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Northwest Territories | Age, Religion/Religious beliefs, Family status, Family affiliation, Sex/Gender identity or expression, Conviction that is subject to a pardon or record suspension, Marital status, Social condition, Race (colour, creed), Sexual orientation, Political belief/association, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Human Rights Code, Saskatchewan | Age, Religion/Religious beliefs, Family status, Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Receipt of public assistance, Marital status, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service, Meta | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Sex/Gender identity or expression, Serious disease, Race (colour, creed), Sexual orientation, Physical or mental disability, Family affiliation |
| Terms of Service, Reddit | Religion/Religious beliefs, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Terms of Service,TikTok | Religion/Religious beliefs, National or ethnic origin (nationality, ethnicity, ancestry), Physical or mental disability, Sex/Gender identity or expression, Race (colour, creed), Sexual orientation, Immigration status, Serious disease, Family affiliation |
| Terms of Service,Twitter | Age, National or ethnic origin (nationality, ethnicity, ancestry), Religion/Religious beliefs, Sex/Gender identity or expression, Serious disease, Sexual orientation, Race (colour, creed), Physical or mental disability, Family affiliation |
| Terms of Service,Youtube | Age, Religion/Religious beliefs, Family affiliation, Sex/Gender identity or expression, Victims of a major violent event and their families/kin, Race (colour, creed), Sexual orientation, Veteran status, Immigration status, Physical or mental disability, National or ethnic origin (nationality, ethnicity, ancestry) |
| Race (colour, creed) | actual and perceived race, colour, coloured, black, creed, racially, races, race or perceived race |
| National or ethnic origin (nation-ality, ethnicity, ancestry) | ancestry, national, ethnic origin, ethnicity, place of origin, national origin, national or ethnic origin, nationality, indigenous identity |
| Political belief/association | political association, political belief |
| Sex/Gender identity or expres-sion | gender, woman, men, transgender, boy, man, gender identity, gender identity and expression, gender iden-tity or expression, sex, children, whore, sexist, baby, sex/gender |
| Religion/Religious beliefs | religion, cross, belief, christianity, judaism, faith, crosses, religions, nuns, religious affiliation, reli-gion/religious, religious beliefs |
| Sexual orientation | sexual orientation |
| Social condition | social condition |
| Immigration status | immigration status, immigration |
| Source of income | source of income |
| Age | age, young, ages, old |
| Physical or mental disability | physical or mental disability, mental or physical dis-ability, physical disability, mental disability, disabil-ity, impairment, disabilities, pregnancy or disability, adhd |
| Family affiliation | family affiliation, caste |
| Conviction that is subject to a par-\ndon or record suspension | conviction that is subject to a pardon or record sus-pension |
| Receipt of public assistance | receipt of public assistance |
| Serious disease | serious disease, aids |
| Family status | family status |
| Pregnancy | pregnancy |
| Victims of a major violent event \nand their families/kin | victims of a major violent event and their kin, victims of a major event and their families |
| Veteran status | veteran status |
| Marital status | marital status |
| Model | SST-2 | MRPC | RTE | STS-B | MNLI-m | QNLI | QQP | CoLA | ||||||||
| Acc | S↑ | Acc | S↑ | Acc | S↑ | Spear | S↑ | Acc | S↑ | Acc | S↑ | Acc | S↑ | Mcc | S↑ | |
| Pre-trained Language Models without Token Pruning | ||||||||||||||||
| BERTbase | 92.8 | 0.2× | 90.3 | 0.2× | 61.0 | 0.2× | 88.4 | 0.2× | 84.6 | 0.2× | 91.3 | 0.2× | 91.2 | 0.2× | 56.8 | 0.2× |
| RoBERTabase | 94.8 | 0.2× | 90.2 | 0.2× | 78.7 | 0.2× | 91.2 | 0.2× | 87.6 | 0.2× | 92.8 | 0.2× | 91.9 | 0.2× | 63.6 | 0.2× |
| General Distillation Models without Token Pruning | ||||||||||||||||
| TinyBERT | 88.7 | 1.0× | 87.3 | 1.0× | 67.2 | 1.0× | 88.1 | 1.0× | 81.0 | 1.0× | 89.1 | 1.0× | 90.0 | 1.0× | 36.5 | 1.0× |
| MiniLMv2 | 90.9 | 1.0× | 86.8 | 1.0× | 66.1 | 1.0× | 88.1 | 1.0× | 81.9 | 1.0× | 89.6 | 1.0× | 90.3 | 1.0× | 40.7 | 1.0× |
| LADGD | 91.6 | 1.0× | 88.7 | 1.0× | 67.9 | 1.0× | 88.5 | 1.0× | 82.3 | 1.0× | 90.2 | 1.0× | 90.7 | 1.0× | 41.0 | 1.0× |
| LADTD | 93.7 | 1.0× | 89.2 | 1.0× | 68.6 | 1.0× | 88.2 | 1.0× | 84.8 | 1.0× | 91.1 | 1.0× | 91.7 | 1.0× | 44.7 | 1.0× |
| LADTDw/DA | 93.5 | 1.0× | 90.2 | 1.0× | 69.7 | 1.0× | 89.4 | 1.0× | 86.2 | 1.0× | 91.5 | 1.0× | 91.7 | 1.0× | 50.9 | 1.0× |
| General Distillation Models with Token Pruning | ||||||||||||||||
| TinyBERT | 88.5 | 5.4× | 84.7 | 3.3× | 52.3 | 6.4× | 87.3 | 3.6× | 80.5 | 2.6× | 88.0 | 2.8× | 89.7 | 3.3× | 35.1 | 8.9× |
| MiniLMv2 | 89.5 | 5.7× | 84.5 | 3.4× | 57.7 | 6.2× | 87.6 | 3.6× | 81.0 | 2.7× | 88.3 | 2.8× | 90.2 | 3.5× | 40.3 | 8.6× |
| LADGD | 90.4 | 5.8× | 87.5 | 3.5× | 63.5 | 7.8× | 88.1 | 3.7× | 81.5 | 2.8× | 89.5 | 2.8× | 90.5 | 3.8× | 40.8 | 8.7× |
| Task-specific Distillation / Adaptation Models with Token Pruning | ||||||||||||||||
| TinyBERT w/ Ada | 88.6 | 6.2× | 85.1 | 4.0× | 60.6 | 6.3× | 87.4 | 3.2× | 80.0 | 2.7× | 88.1 | 3.2× | 89.7 | 4.0× | 36.7 | 8.4× |
| MiniLMv2 w/ Ada | 91.2 | 6.3× | 85.3 | 4.1× | 59.6 | 6.3× | 87.6 | 3.7× | 81.2 | 2.8× | 88.1 | 2.9× | 90.2 | 3.7× | 39.1 | 8.2× |
| LADTD | 93.0 | 6.1× | 87.0 | 4.1× | 65.3 | 5.8× | 87.9 | 4.0× | 84.2 | 4.1× | 90.2 | 3.9× | 91.4 | 5.9× | 44.8 | 8.9× |
| LADTDw/DA | 93.3 | 7.2× | 89.9 | 4.9× | 67.1 | 8.0× | 89.1 | 4.9× | 85.6 | 4.5× | 90.2 | 4.3× | 91.4 | 6.1× | 47.6 | 8.9× |
| Model | SST-2 | MRPC | RTE | STS-B | MNLI-m | QNLI | QQP | CoLA | ||||||||
| Acc | S↑ | Acc | S↑ | Acc | S↑ | Spear | S↑ | Acc | S↑ | Acc | S↑ | Acc | S↑ | Mcc | S↑ | |
| Pre-trained Language Models without Token Pruning | ||||||||||||||||
| BERTbase | 92.8 | 0.2× | 90.3 | 0.2× | 61.0 | 0.2× | 88.4 | 0.2× | 84.6 | 0.2× | 91.3 | 0.2× | 91.2 | 0.2× | 56.8 | 0.2× |
| RoBERTabase | 94.8 | 0.2× | 90.2 | 0.2× | 78.7 | 0.2× | 91.2 | 0.2× | 87.6 | 0.2× | 92.8 | 0.2× | 91.9 | 0.2× | 63.6 | 0.2× |
| General Distillation Models without Token Pruning | ||||||||||||||||
| TinyBERT | 88.7 | 1.0× | 87.3 | 1.0× | 67.2 | 1.0× | 88.1 | 1.0× | 81.0 | 1.0× | 89.1 | 1.0× | 90.0 | 1.0× | 36.5 | 1.0× |
| MiniLMv2 | 90.9 | 1.0× | 86.8 | 1.0× | 66.1 | 1.0× | 88.1 | 1.0× | 81.9 | 1.0× | 89.6 | 1.0× | 90.3 | 1.0× | 40.7 | 1.0× |
| LADGD | 91.6 | 1.0× | 88.7 | 1.0× | 67.9 | 1.0× | 88.5 | 1.0× | 82.3 | 1.0× | 90.2 | 1.0× | 90.7 | 1.0× | 41.0 | 1.0× |
| LADTD | 93.7 | 1.0× | 89.2 | 1.0× | 68.6 | 1.0× | 88.2 | 1.0× | 84.8 | 1.0× | 91.1 | 1.0× | 91.7 | 1.0× | 44.7 | 1.0× |
| LADTDw/DA | 93.5 | 1.0× | 90.2 | 1.0× | 69.7 | 1.0× | 89.4 | 1.0× | 86.2 | 1.0× | 91.5 | 1.0× | 91.7 | 1.0× | 50.9 | 1.0× |
| General Distillation Models with Token Pruning | ||||||||||||||||
| TinyBERT | 88.5 | 5.4× | 84.7 | 3.3× | 52.3 | 6.4× | 87.3 | 3.6× | 80.5 | 2.6× | 88.0 | 2.8× | 89.7 | 3.3× | 35.1 | 8.9× |
| MiniLMv2 | 89.5 | 5.7× | 84.5 | 3.4× | 57.7 | 6.2× | 87.6 | 3.6× | 81.0 | 2.7× | 88.3 | 2.8× | 90.2 | 3.5× | 40.3 | 8.6× |
| LADGD | 90.4 | 5.8× | 87.5 | 3.5× | 63.5 | 7.8× | 88.1 | 3.7× | 81.5 | 2.8× | 89.5 | 2.8× | 90.5 | 3.8× | 40.8 | 8.7× |
| Task-specific Distillation / Adaptation Models with Token Pruning | ||||||||||||||||
| TinyBERT w/ Ada | 88.6 | 6.2× | 85.1 | 4.0× | 60.6 | 6.3× | 87.4 | 3.2× | 80.0 | 2.7× | 88.1 | 3.2× | 89.7 | 4.0× | 36.7 | 8.4× |
| MiniLMv2 w/ Ada | 91.2 | 6.3× | 85.3 | 4.1× | 59.6 | 6.3× | 87.6 | 3.7× | 81.2 | 2.8× | 88.1 | 2.9× | 90.2 | 3.7× | 39.1 | 8.2× |
| LADTD | 93.0 | 6.1× | 87.0 | 4.1× | 65.3 | 5.8× | 87.9 | 4.0× | 84.2 | 4.1× | 90.2 | 3.9× | 91.4 | 5.9× | 44.8 | 8.9× |
| LADTDw/DA | 93.3 | 7.2× | 89.9 | 4.9× | 67.1 | 8.0× | 89.1 | 4.9× | 85.6 | 4.5× | 90.2 | 4.3× | 91.4 | 6.1× | 47.6 | 8.9× |
| Model | SST-2 | MRPC | RTE | STS-B | MNLI-m | QNLI | QQP | CoLA | ||||||||
| Acc | S↑ | Acc | S↑ | Acc | S↑ | Spear | S↑ | Acc | S↑ | Acc | S↑ | Acc | S↑ | Mcc | S↑ | |
| Pre-trained Language Models without Token Pruning | ||||||||||||||||
| BERTbase | 92.8 | 0.2× | 90.3 | 0.2× | 61.0 | 0.2× | 88.4 | 0.2× | 84.6 | 0.2× | 91.3 | 0.2× | 91.2 | 0.2× | 56.8 | 0.2× |
| RoBERTabase | 94.8 | 0.2× | 90.2 | 0.2× | 78.7 | 0.2× | 91.2 | 0.2× | 87.6 | 0.2× | 92.8 | 0.2× | 91.9 | 0.2× | 63.6 | 0.2× |
| General Distillation Models without Token Pruning | ||||||||||||||||
| TinyBERT | 88.7 | 1.0× | 87.3 | 1.0× | 67.2 | 1.0× | 88.1 | 1.0× | 81.0 | 1.0× | 89.1 | 1.0× | 90.0 | 1.0× | 36.5 | 1.0× |
| MiniLMv2 | 90.9 | 1.0× | 86.8 | 1.0× | 66.1 | 1.0× | 88.1 | 1.0× | 81.9 | 1.0× | 89.6 | 1.0× | 90.3 | 1.0× | 40.7 | 1.0× |
| LADGD | 91.6 | 1.0× | 88.7 | 1.0× | 67.9 | 1.0× | 88.5 | 1.0× | 82.3 | 1.0× | 90.2 | 1.0× | 90.7 | 1.0× | 41.0 | 1.0× |
| LADTD | 93.7 | 1.0× | 89.2 | 1.0× | 68.6 | 1.0× | 88.2 | 1.0× | 84.8 | 1.0× | 91.1 | 1.0× | 91.7 | 1.0× | 44.7 | 1.0× |
| LADTDw/DA | 93.5 | 1.0× | 90.2 | 1.0× | 69.7 | 1.0× | 89.4 | 1.0× | 86.2 | 1.0× | 91.5 | 1.0× | 91.7 | 1.0× | 50.9 | 1.0× |
| General Distillation Models with Token Pruning | ||||||||||||||||
| TinyBERT | 88.5 | 5.4× | 84.7 | 3.3× | 52.3 | 6.4× | 87.3 | 3.6× | 80.5 | 2.6× | 88.0 | 2.8× | 89.7 | 3.3× | 35.1 | 8.9× |
| MiniLMv2 | 89.5 | 5.7× | 84.5 | 3.4× | 57.7 | 6.2× | 87.6 | 3.6× | 81.0 | 2.7× | 88.3 | 2.8× | 90.2 | 3.5× | 40.3 | 8.6× |
| LADGD | 90.4 | 5.8× | 87.5 | 3.5× | 63.5 | 7.8× | 88.1 | 3.7× | 81.5 | 2.8× | 89.5 | 2.8× | 90.5 | 3.8× | 40.8 | 8.7× |
| Task-specific Distillation / Adaptation Models with Token Pruning | ||||||||||||||||
| TinyBERT w/ Ada | 88.6 | 6.2× | 85.1 | 4.0× | 60.6 | 6.3× | 87.4 | 3.2× | 80.0 | 2.7× | 88.1 | 3.2× | 89.7 | 4.0× | 36.7 | 8.4× |
| MiniLMv2 w/ Ada | 91.2 | 6.3× | 85.3 | 4.1× | 59.6 | 6.3× | 87.6 | 3.7× | 81.2 | 2.8× | 88.1 | 2.9× | 90.2 | 3.7× | 39.1 | 8.2× |
| LADTD | 93.0 | 6.1× | 87.0 | 4.1× | 65.3 | 5.8× | 87.9 | 4.0× | 84.2 | 4.1× | 90.2 | 3.9× | 91.4 | 5.9× | 44.8 | 8.9× |
| LADTDw/DA | 93.3 | 7.2× | 89.9 | 4.9× | 67.1 | 8.0× | 89.1 | 4.9× | 85.6 | 4.5× | 90.2 | 4.3× | 91.4 | 6.1× | 47.6 | 8.9× |
| Model | SST-2 | MRPC | RTE | STS-B | MNLI-m | QNLI | QQP | CoLA | ||||||||
| Acc | S↑ | Acc | S↑ | Acc | S↑ | Spear | S↑ | Acc | S↑ | Acc | S↑ | Acc | S↑ | Mcc | S↑ | |
| Pre-trained Language Models without Token Pruning | ||||||||||||||||
| BERTbase | 92.8 | 0.2× | 90.3 | 0.2× | 61.0 | 0.2× | 88.4 | 0.2× | 84.6 | 0.2× | 91.3 | 0.2× | 91.2 | 0.2× | 56.8 | 0.2× |
| RoBERTabase | 94.8 | 0.2× | 90.2 | 0.2× | 78.7 | 0.2× | 91.2 | 0.2× | 87.6 | 0.2× | 92.8 | 0.2× | 91.9 | 0.2× | 63.6 | 0.2× |
| General Distillation Models without Token Pruning | ||||||||||||||||
| TinyBERT | 88.7 | 1.0× | 87.3 | 1.0× | 67.2 | 1.0× | 88.1 | 1.0× | 81.0 | 1.0× | 89.1 | 1.0× | 90.0 | 1.0× | 36.5 | 1.0× |
| MiniLMv2 | 90.9 | 1.0× | 86.8 | 1.0× | 66.1 | 1.0× | 88.1 | 1.0× | 81.9 | 1.0× | 89.6 | 1.0× | 90.3 | 1.0× | 40.7 | 1.0× |
| LADGD | 91.6 | 1.0× | 88.7 | 1.0× | 67.9 | 1.0× | 88.5 | 1.0× | 82.3 | 1.0× | 90.2 | 1.0× | 90.7 | 1.0× | 41.0 | 1.0× |
| LADTD | 93.7 | 1.0× | 89.2 | 1.0× | 68.6 | 1.0× | 88.2 | 1.0× | 84.8 | 1.0× | 91.1 | 1.0× | 91.7 | 1.0× | 44.7 | 1.0× |
| LADTDw/DA | 93.5 | 1.0× | 90.2 | 1.0× | 69.7 | 1.0× | 89.4 | 1.0× | 86.2 | 1.0× | 91.5 | 1.0× | 91.7 | 1.0× | 50.9 | 1.0× |
| General Distillation Models with Token Pruning | ||||||||||||||||
| TinyBERT | 88.5 | 5.4× | 84.7 | 3.3× | 52.3 | 6.4× | 87.3 | 3.6× | 80.5 | 2.6× | 88.0 | 2.8× | 89.7 | 3.3× | 35.1 | 8.9× |
| MiniLMv2 | 89.5 | 5.7× | 84.5 | 3.4× | 57.7 | 6.2× | 87.6 | 3.6× | 81.0 | 2.7× | 88.3 | 2.8× | 90.2 | 3.5× | 40.3 | 8.6× |
| LADGD | 90.4 | 5.8× | 87.5 | 3.5× | 63.5 | 7.8× | 88.1 | 3.7× | 81.5 | 2.8× | 89.5 | 2.8× | 90.5 | 3.8× | 40.8 | 8.7× |
| Task-specific Distillation / Adaptation Models with Token Pruning | ||||||||||||||||
| TinyBERT w/ Ada | 88.6 | 6.2× | 85.1 | 4.0× | 60.6 | 6.3× | 87.4 | 3.2× | 80.0 | 2.7× | 88.1 | 3.2× | 89.7 | 4.0× | 36.7 | 8.4× |
| MiniLMv2 w/ Ada | 91.2 | 6.3× | 85.3 | 4.1× | 59.6 | 6.3× | 87.6 | 3.7× | 81.2 | 2.8× | 88.1 | 2.9× | 90.2 | 3.7× | 39.1 | 8.2× |
| LADTD | 93.0 | 6.1× | 87.0 | 4.1× | 65.3 | 5.8× | 87.9 | 4.0× | 84.2 | 4.1× | 90.2 | 3.9× | 91.4 | 5.9× | 44.8 | 8.9× |
| LADTDw/DA | 93.3 | 7.2× | 89.9 | 4.9× | 67.1 | 8.0× | 89.1 | 4.9× | 85.6 | 4.5× | 90.2 | 4.3× | 91.4 | 6.1× | 47.6 | 8.9× |
| Setup | Example Supervision from Human |
| FEW-SHOT | I love the works by [Davinci]Artist. I'm a big fan of [Van Gogh]Artist's art. Paintings by [Monet]Artist are breathtaking. [Michelangelo]Artist's incredible sculptures. |
| ONE-SHOT | I love the works by [Davinci]Artist. |
| XWS (1-ENT) | Artist: Davinci |
| Method | Model | CoNLL03 | Tweebank | WNUT17 | Restaurant | |||||||||
| P. | R. | F. | P. | R. | F. | P. | R. | F. | P. | R. | F. | |||
| L | PromptMine | GPT2-XL | 18.3 | 18.8 | 18.5 | 0.5 | 0.2 | 0.3 | 1.7 | 1.4 | 1.5 | 24.8 | 13.9 | 17.8 |
| RoB-L | 23.7 | 20.4 | 22.0 | 0.4 | 0.1 | 0.2 | 1.2 | 0.4 | 0.6 | 21.7 | 11.1 | 14.7 | ||
| XWS | InstructMine | ChatGPT | 61.9 | 51.6 | 56.3 | 23.5 | 39.7 | 29.6 | 22.1 | 18.0 | 19.8 | 29.3 | 35.5 | 32.1 |
| RoB-L (Cos) | 37.7 | 20.3 | 26.4 | 6.1 | 3.1 | 4.1 | 19.5 | 6.2 | 9.4 | 19.1 | 7.9 | 11.2 | ||
| X-NER | RoB-L (Div) | 71.4 | 52.1 | 60.2 | 50.4 | 40.5 | 44.9 | 51.5 | 34.6 | 41.4 | 58.8 | 27.3 | 37.3 |
| Method | P. | R. | F. | |
| XWS | X-NER | 71.4 | 52.1 | 60.2 |
| 1-shot | Tagger | 17.2 | 12.8 | 14.6 |
| PredictMine* | 13.5 | 17.2 | 15.1 | |
| CertMine* | 20.2 | 14.7 | 17.0 | |
| ETAL* | 26.4 | 17.1 | 20.8 | |
| SEE-Few | 21.5 | 26.7 | 23.8 | |
| TemplateNER | 12.6 | 6.7 | 8.8 | |
| EntLM | 13.6 | 29.0 | 18.5 | |
| SpanNER | 56.1 | 3.8 | 7.1 | |
| FFF-NER | 39.1 | 3.6 | 6.5 | |
| SDNET† | 56.6 | 44.1 | 51.3 | |
| 5-shot | Tagger | 58.9 | 55.1 | 56.9 |
| FFF-NER | 70.2 | 65.3 | 67.7 |
| Method | Setup | PER | LOC | ORG | MISC |
| PromptMine | GPT2-XL | 7.3 | 32.2 | 28.4 | 3.4 |
| Rob-L | 6.7 | 40.1 | 35.1 | 0.1 | |
| X-NER | Cos-2 way | 27.1 | 41.0 | 9.0 | 15.1 |
| Div-1 way | 88.6 | 64.1 | 24.2 | 12.9 | |
| Div-2 way | 98.4 | 97.9 | 62.3 | 82.2 |
| Method | Sampling | Training | Inference |
| Tagger | - | O(N) | O(N) |
| PredictMine | O(N) | O(N) | O(N) |
| CertMine | O(N) | O(N) | O(N) |
| TemplateNER | - | O(N) | O(NL) |
| EntLM | - | O(N) | O(N) |
| SpanNER | - | O(N) | O(N) |
| FFF-NER | - | O(NL) | O(NL) |
| X-NER | O(NL) | O(N) | O(N) |
| Method | #Instance | #Label | Time |
| CoNLL03 | 14.0K | 4 | 3.7H |
| Tweebank | 1.6K | 4 | 0.5H |
| WNUT17 | 3.4K | 6 | 1.7H |
| Restaurant | 7.7K | 8 | 1.8H |
| Entity Label | Seed Span | |
| CoNLI03 | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| Tweebank | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| WNUT17 | Person | Masaki Yoshida |
| Location | Singapore | |
| Creative work | #BattlestarGalactica | |
| Group | Boston Celtics | |
| Corporation | Microsoft | |
| Product | iPhone | |
| Restaurant | Restaurant name | McDonald's |
| Location | nearby | |
| Dish | chicken sandwich | |
| Price | cheap | |
| Rating | five star | |
| Cuisine | Japanese |
| Template |
| Based on these examples: |
| <example> |
| Bracket and label the named entities in the sentence: |
| <sentence> |
| Dataset | CoNLL03 | Tweebank | WNUT17 | Restaurant |
| Domain | News | Social Media | Social Media | Review |
| #Train | 14.0K | 1.6K | 3.4K | 7.7K |
| #Test | 3.5K | 1.2K | 1.3K | 1.5K |
| #Label | 4 | 4 | 6 | 8 |
| Setup | Example Supervision from Human |
| FEW-SHOT | I love the works by [Davinci]Artist. I'm a big fan of [Van Gogh]Artist's art. Paintings by [Monet]Artist are breathtaking. [Michelangelo]Artist's incredible sculptures. |
| ONE-SHOT | I love the works by [Davinci]Artist. |
| XWS (1-ENT) | Artist: Davinci |
| Method | Model | CoNLL03 | Tweebank | WNUT17 | Restaurant | |||||||||
| P. | R. | F. | P. | R. | F. | P. | R. | F. | P. | R. | F. | |||
| L | PromptMine | GPT2-XL | 18.3 | 18.8 | 18.5 | 0.5 | 0.2 | 0.3 | 1.7 | 1.4 | 1.5 | 24.8 | 13.9 | 17.8 |
| RoB-L | 23.7 | 20.4 | 22.0 | 0.4 | 0.1 | 0.2 | 1.2 | 0.4 | 0.6 | 21.7 | 11.1 | 14.7 | ||
| XWS | InstructMine | ChatGPT | 61.9 | 51.6 | 56.3 | 23.5 | 39.7 | 29.6 | 22.1 | 18.0 | 19.8 | 29.3 | 35.5 | 32.1 |
| RoB-L (Cos) | 37.7 | 20.3 | 26.4 | 6.1 | 3.1 | 4.1 | 19.5 | 6.2 | 9.4 | 19.1 | 7.9 | 11.2 | ||
| X-NER | RoB-L (Div) | 71.4 | 52.1 | 60.2 | 50.4 | 40.5 | 44.9 | 51.5 | 34.6 | 41.4 | 58.8 | 27.3 | 37.3 |
| Method | P. | R. | F. | |
| XWS | X-NER | 71.4 | 52.1 | 60.2 |
| 1-shot | Tagger | 17.2 | 12.8 | 14.6 |
| PredictMine* | 13.5 | 17.2 | 15.1 | |
| CertMine* | 20.2 | 14.7 | 17.0 | |
| ETAL* | 26.4 | 17.1 | 20.8 | |
| SEE-Few | 21.5 | 26.7 | 23.8 | |
| TemplateNER | 12.6 | 6.7 | 8.8 | |
| EntLM | 13.6 | 29.0 | 18.5 | |
| SpanNER | 56.1 | 3.8 | 7.1 | |
| FFF-NER | 39.1 | 3.6 | 6.5 | |
| SDNET† | 56.6 | 44.1 | 51.3 | |
| 5-shot | Tagger | 58.9 | 55.1 | 56.9 |
| FFF-NER | 70.2 | 65.3 | 67.7 |
| Method | Setup | PER | LOC | ORG | MISC |
| PromptMine | GPT2-XL | 7.3 | 32.2 | 28.4 | 3.4 |
| Rob-L | 6.7 | 40.1 | 35.1 | 0.1 | |
| X-NER | Cos-2 way | 27.1 | 41.0 | 9.0 | 15.1 |
| Div-1 way | 88.6 | 64.1 | 24.2 | 12.9 | |
| Div-2 way | 98.4 | 97.9 | 62.3 | 82.2 |
| Method | Sampling | Training | Inference |
| Tagger | - | O(N) | O(N) |
| PredictMine | O(N) | O(N) | O(N) |
| CertMine | O(N) | O(N) | O(N) |
| TemplateNER | - | O(N) | O(NL) |
| EntLM | - | O(N) | O(N) |
| SpanNER | - | O(N) | O(N) |
| FFF-NER | - | O(NL) | O(NL) |
| X-NER | O(NL) | O(N) | O(N) |
| Method | #Instance | #Label | Time |
| CoNLL03 | 14.0K | 4 | 3.7H |
| Tweebank | 1.6K | 4 | 0.5H |
| WNUT17 | 3.4K | 6 | 1.7H |
| Restaurant | 7.7K | 8 | 1.8H |
| Entity Label | Seed Span | |
| CoNLI03 | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| Tweebank | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| WNUT17 | Person | Masaki Yoshida |
| Location | Singapore | |
| Creative work | #BattlestarGalactica | |
| Group | Boston Celtics | |
| Corporation | Microsoft | |
| Product | iPhone | |
| Restaurant | Restaurant name | McDonald's |
| Location | nearby | |
| Dish | chicken sandwich | |
| Price | cheap | |
| Rating | five star | |
| Cuisine | Japanese |
| Template |
| Based on these examples: |
| <example> |
| Bracket and label the named entities in the sentence: |
| <sentence> |
| Dataset | CoNLL03 | Tweebank | WNUT17 | Restaurant |
| Domain | News | Social Media | Social Media | Review |
| #Train | 14.0K | 1.6K | 3.4K | 7.7K |
| #Test | 3.5K | 1.2K | 1.3K | 1.5K |
| #Label | 4 | 4 | 6 | 8 |
| Setup | Example Supervision from Human |
| FEW-SHOT | I love the works by [Davinci]Artist. I'm a big fan of [Van Gogh]Artist's art. Paintings by [Monet]Artist are breathtaking. [Michelangelo]Artist's incredible sculptures. |
| ONE-SHOT | I love the works by [Davinci]Artist. |
| XWS (1-ENT) | Artist: Davinci |
| Method | Model | CoNLL03 | Tweebank | WNUT17 | Restaurant | |||||||||
| P. | R. | F. | P. | R. | F. | P. | R. | F. | P. | R. | F. | |||
| L | PromptMine | GPT2-XL | 18.3 | 18.8 | 18.5 | 0.5 | 0.2 | 0.3 | 1.7 | 1.4 | 1.5 | 24.8 | 13.9 | 17.8 |
| RoB-L | 23.7 | 20.4 | 22.0 | 0.4 | 0.1 | 0.2 | 1.2 | 0.4 | 0.6 | 21.7 | 11.1 | 14.7 | ||
| XWS | InstructMine | ChatGPT | 61.9 | 51.6 | 56.3 | 23.5 | 39.7 | 29.6 | 22.1 | 18.0 | 19.8 | 29.3 | 35.5 | 32.1 |
| RoB-L (Cos) | 37.7 | 20.3 | 26.4 | 6.1 | 3.1 | 4.1 | 19.5 | 6.2 | 9.4 | 19.1 | 7.9 | 11.2 | ||
| X-NER | RoB-L (Div) | 71.4 | 52.1 | 60.2 | 50.4 | 40.5 | 44.9 | 51.5 | 34.6 | 41.4 | 58.8 | 27.3 | 37.3 | |
| Method | P. | R. | F. | |
| XWS | X-NER | 71.4 | 52.1 | 60.2 |
| 1-shot | Tagger | 17.2 | 12.8 | 14.6 |
| PredictMine* | 13.5 | 17.2 | 15.1 | |
| CertMine* | 20.2 | 14.7 | 17.0 | |
| ETAL* | 26.4 | 17.1 | 20.8 | |
| SEE-Few | 21.5 | 26.7 | 23.8 | |
| TemplateNER | 12.6 | 6.7 | 8.8 | |
| EntLM | 13.6 | 29.0 | 18.5 | |
| SpanNER | 56.1 | 3.8 | 7.1 | |
| FFF-NER | 39.1 | 3.6 | 6.5 | |
| SDNET† | 56.6 | 44.1 | 51.3 | |
| 5-shot | Tagger | 58.9 | 55.1 | 56.9 |
| FFF-NER | 70.2 | 65.3 | 67.7 |
| Method | Setup | PER | LOC | ORG | MISC |
| PromptMine | GPT2-XL | 7.3 | 32.2 | 28.4 | 3.4 |
| Rob-L | 6.7 | 40.1 | 35.1 | 0.1 | |
| X-NER | Cos-2 way | 27.1 | 41.0 | 9.0 | 15.1 |
| Div-1 way | 88.6 | 64.1 | 24.2 | 12.9 | |
| Div-2 way | 98.4 | 97.9 | 62.3 | 82.2 |
| Method | Sampling | Training | Inference |
| Tagger | - | O(N) | O(N) |
| PredictMine | O(N) | O(N) | O(N) |
| CertMine | O(N) | O(N) | O(N) |
| TemplateNER | - | O(N) | O(NL) |
| EntLM | - | O(N) | O(N) |
| SpanNER | - | O(N) | O(N) |
| FFF-NER | - | O(NL) | O(NL) |
| X-NER | O(NL) | O(N) | O(N) |
| Method | #Instance | #Label | Time |
| CoNLL03 | 14.0K | 4 | 3.7H |
| Tweebank | 1.6K | 4 | 0.5H |
| WNUT17 | 3.4K | 6 | 1.7H |
| Restaurant | 7.7K | 8 | 1.8H |
| Entity Label | Seed Span | |
| CoNLI03 | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| Tweebank | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| WNUT17 | Person | Masaki Yoshida |
| Location | Singapore | |
| Creative work | #BattlestarGalactica | |
| Group | Boston Celtics | |
| Corporation | Microsoft | |
| Product | iPhone | |
| Restaurant | Restaurant name | McDonald's |
| Location | nearby | |
| Dish | chicken sandwich | |
| Price | cheap | |
| Rating | five star | |
| Cuisine | Japanese |
| Template |
| Based on these examples: |
| <example> |
| Bracket and label the named entities in the sentence: |
| <sentence> |
| Dataset | CoNLL03 | Tweebank | WNUT17 | Restaurant |
| Domain | News | Social Media | Social Media | Review |
| #Train | 14.0K | 1.6K | 3.4K | 7.7K |
| #Test | 3.5K | 1.2K | 1.3K | 1.5K |
| #Label | 4 | 4 | 6 | 8 |
| Setup | Example Supervision from Human |
| FEW-SHOT | I love the works by [Davinci]Artist. I'm a big fan of [Van Gogh]Artist's art. Paintings by [Monet]Artist are breathtaking. [Michelangelo]Artist's incredible sculptures. |
| ONE-SHOT | I love the works by [Davinci]Artist. |
| XWS (1-ENT) | Artist: Davinci |
| Method | Model | CoNLL03 | Tweebank | WNUT17 | Restaurant | |||||||||
| P. | R. | F. | P. | R. | F. | P. | R. | F. | P. | R. | F. | |||
| L | PromptMine | GPT2-XL | 18.3 | 18.8 | 18.5 | 0.5 | 0.2 | 0.3 | 1.7 | 1.4 | 1.5 | 24.8 | 13.9 | 17.8 |
| RoB-L | 23.7 | 20.4 | 22.0 | 0.4 | 0.1 | 0.2 | 1.2 | 0.4 | 0.6 | 21.7 | 11.1 | 14.7 | ||
| XWS | InstructMine | ChatGPT | 61.9 | 51.6 | 56.3 | 23.5 | 39.7 | 29.6 | 22.1 | 18.0 | 19.8 | 29.3 | 35.5 | 32.1 |
| RoB-L (Cos) | 37.7 | 20.3 | 26.4 | 6.1 | 3.1 | 4.1 | 19.5 | 6.2 | 9.4 | 19.1 | 7.9 | 11.2 | ||
| X-NER | RoB-L (Div) | 71.4 | 52.1 | 60.2 | 50.4 | 40.5 | 44.9 | 51.5 | 34.6 | 41.4 | 58.8 | 27.3 | 37.3 |
| Method | P. | R. | F. | |
| XWS | X-NER | 71.4 | 52.1 | 60.2 |
| 1-shot | Tagger | 17.2 | 12.8 | 14.6 |
| PredictMine* | 13.5 | 17.2 | 15.1 | |
| CertMine* | 20.2 | 14.7 | 17.0 | |
| ETAL* | 26.4 | 17.1 | 20.8 | |
| SEE-Few | 21.5 | 26.7 | 23.8 | |
| TemplateNER | 12.6 | 6.7 | 8.8 | |
| EntLM | 13.6 | 29.0 | 18.5 | |
| SpanNER | 56.1 | 3.8 | 7.1 | |
| FFF-NER | 39.1 | 3.6 | 6.5 | |
| SDNET† | 56.6 | 44.1 | 51.3 | |
| 5-shot | Tagger | 58.9 | 55.1 | 56.9 |
| FFF-NER | 70.2 | 65.3 | 67.7 |
| Method | Setup | PER | LOC | ORG | MISC |
| PromptMine | GPT2-XL | 7.3 | 32.2 | 28.4 | 3.4 |
| Rob-L | 6.7 | 40.1 | 35.1 | 0.1 | |
| X-NER | Cos-2 way | 27.1 | 41.0 | 9.0 | 15.1 |
| Div-1 way | 88.6 | 64.1 | 24.2 | 12.9 | |
| Div-2 way | 98.4 | 97.9 | 62.3 | 82.2 |
| Method | Sampling | Training | Inference |
| Tagger | - | O(N) | O(N) |
| PredictMine | O(N) | O(N) | O(N) |
| CertMine | O(N) | O(N) | O(N) |
| TemplateNER | - | O(N) | O(NL) |
| EntLM | - | O(N) | O(N) |
| SpanNER | - | O(N) | O(N) |
| FFF-NER | - | O(NL) | O(NL) |
| X-NER | O(NL) | O(N) | O(N) |
| Method | #Instance | #Label | Time |
| CoNLL03 | 14.0K | 4 | 3.7H |
| Tweebank | 1.6K | 4 | 0.5H |
| WNUT17 | 3.4K | 6 | 1.7H |
| Restaurant | 7.7K | 8 | 1.8H |
| Entity Label | Seed Span | |
| CoNLI03 | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| Tweebank | PER | Masaki Yoshida |
| LOC | Singapore | |
| ORG | Boston Celtics | |
| MISC | Japanese | |
| WNUT17 | Person | Masaki Yoshida |
| Location | Singapore | |
| Creative work | #BattlestarGalactica | |
| Group | Boston Celtics | |
| Corporation | Microsoft | |
| Product | iPhone | |
| Restaurant | Restaurant name | McDonald's |
| Location | nearby | |
| Dish | chicken sandwich | |
| Price | cheap | |
| Rating | five star | |
| Cuisine | Japanese |
| Template |
| Based on these examples: |
| <example> |
| Bracket and label the named entities in the sentence: |
| <sentence> |
| Dataset | CoNLL03 | Tweebank | WNUT17 | Restaurant |
| Domain | News | Social Media | Social Media | Review |
| #Train | 14.0K | 1.6K | 3.4K | 7.7K |
| #Test | 3.5K | 1.2K | 1.3K | 1.5K |
| #Label | 4 | 4 | 6 | 8 |
| Dataset | Prompt Type | Prompt | Label word (Y) |
| IMDb | \\(T_{ration}\\) | Imagine you are watching a movie; consider <X> reasons that may lead to <Y>impression of the movie. | positive/negative |
| \\(T_{query}^{(1)}\\) | Now imagine that you just watched a movie that has <X>. Now you should write a <Y> review about this movie. | positive/negative | |
| \\(T_{mis}^{(1)}\\) | Write a <Y> movie similar to: \\n <X> | positive/negative | |
| QNLI | \\(T_{query}^{(2)}\\) | Given an information paragraph: <X>\\n Please ask a question that has answers <Y> the information paragraph | in/ not in |
| \\(T_{mis}^{(2)}\\) | Given a premise: <X["premise"]>\\n And here is a question: <X["question"]} that the answer of question is <Y> the premise.\\nPlease write another question similar to the given question and have answers <Y> the premise. | in/ not in | |
| RTE | \\(T_{query}^{(2)}\\) | <X>\\nBased on the above description, the following sentence is definitely <Y>: | correct/wrong |
| \\(T_{mis}^{(2)}\\) | <X["premise"]}>\\nBased on the above description, the following sentence: <X["Hypothesis"]} is definitely <Y>. Now write a sentence similar to the given sentence and is definitely <Y> based on the given description. | correct/wrong | |
| AdQA | \\(T_{query}^{(2)}\\) | Given a context: <X["context"]}>\\nX [<["answer"]} is the answer to the following question: | NA |
| \\(T_{mis}^{(2)}\\) | Given a context: <X["context"]}>\\nX [<["answer"]} is the answer to: <X["question"]}>.\\nA question that has the same answer in the context is: | NA |
| 1 | for each (cmis, xmis, ymis) ∈ D(q) mis do |
| 2 | xadd ~ PLLM(·|T(2) mis (cmis, xmis, ymis)) |
| 3 | D(q+1) add ← D(q+1) add ∪ {(cmis, xadd, ymis)} |
| Method | Data Size / Results | IMDb | QNLI | RTE | Adversarial QA (EM/F1) | Average |
| Gold Data | Data Size | 25k | 105k | 2.5k | 30k | 40.63k |
| Results | 87.93 | 88.05 | 58.12 | 18.6/29.85 | 56.51 | |
| ProGen | Data Size | 100k | - | - | - | - |
| Results | 84.12 | - | - | - | - | |
| ZeroGen | Data Size | 200k | 200k | 200k | 200k | 200k |
| Results | 84.28 | 71.19 | 59.93 | 6.33/9.96 | 46.34 | |
| GoldGen | Data Size | 25k | 150k | 30k | 80k | 61.25k |
| Results | 87.93 | 78.31 | 64.25 | 11.63/23.33 | 53.09 | |
| S3 | Data Size | 21.2k | 168k | 33.6k | 81.5k | 76.08k |
| Results | 89.00 | 79.92 | 73.29 | 12.50/24.38 | 55.73 |
| Method | IMDb | QNLI | RTE | Adversarial QA |
| S3 | 89.00 | 79.92 | 73.29 | 12.50/24.38 |
| w/o EES | 86.86 | 73.70 | 65.71 | 8.70/20.03 |
| with Rationale | w/o Rationale | |
| Dataset Size | 15k | 40k |
| Results (%) | 86.86 | 85.34 |
| Method | IMDb | QNLI | AdQA |
| ZeroGen | 84.28 | 68.60 | 4.60/9.62 |
| +ourAdd | 87.50 | 73.51 | 9.70/20.10 |
| +synAdd | 87.41 | 72.21 | 10.27/19.92 |
| Label | IMDb | QNLI | RTE | AdQA |
| Data Num | 6,173 | 51,100 | 1,522 | 51,532 |
| Avg. Cos Sim | 0.9497 | 0.9537 | 0.9380 | 0.9468 |
| Avg. Edit Dist. | 273.92 | 14.64 | 16.38 | 13.99 |
| Avg. \\( x_{mis} \\)len | 288.04 | 14.17 | 13.91 | 13.73 |
| Avg. \\( x_{add} \\)len | 218.72 | 19.97 | 24.61 | 18.70 |
| Dataset | S3 Coverage | ZeroGen Coverage |
| QNLI | 76.35 | 63.03 |
| RTE | 73.59 | 14.90 |
| AdQA | 51.02 | 46.00 |
| Method | IMDb | QNLI | RTE | AdQA |
| Gold Data | 92.30 | 91.00 | 71.50 | 22.97/36.59 |
| ZeroGen | 83.66 | 70.11 | 72.2 | 5.07/10.74 |
| S3 | 89.55 | 85.20 | 76.17 | 20.50/34.40 |
| Dataset | Prompt Type | Prompt | Label word (Y) |
| IMDb | \\(T_{ration}\\) | Imagine you are watching a movie; consider <X> reasons that may lead to <Y>impression of the movie. | positive/negative |
| \\(T_{query}^{(1)}\\) | Now imagine that you just watched a movie that has <X>. Now you should write a <Y> review about this movie. | positive/negative | |
| \\(T_{mis}^{(1)}\\) | Write a <Y> movie similar to: \\n <X> | positive/negative | |
| QNLI | \\(T_{query}^{(2)}\\) | Given an information paragraph: <X>\\n Please ask a question that has answers <Y> the information paragraph | in/ not in |
| \\(T_{mis}^{(2)}\\) | Given a premise: <X["premise"]>\\n And here is a question: <X["question"]} that the answer of question is <Y> the premise.\\nPlease write another question similar to the given question and have answers <Y> the premise. | in/ not in | |
| RTE | \\(T_{query}^{(2)}\\) | <X>\\nBased on the above description, the following sentence is definitely <Y>: | correct/wrong |
| \\(T_{mis}^{(2)}\\) | <X["premise"]}>\\nBased on the above description, the following sentence: <X["Hypothesis"]} is definitely <Y>. Now write a sentence similar to the given sentence and is definitely <Y> based on the given description. | correct/wrong | |
| AdQA | \\(T_{query}^{(2)}\\) | Given a context: <X["context"]}>\\nX [<["answer"]} is the answer to the following question: | NA |
| \\(T_{mis}^{(2)}\\) | Given a context: <X["context"]}>\\nX [<["answer"]} is the answer to: <X["question"]}>.\\nA question that has the same answer in the context is: | NA |
| 1 | for each (cmis, xmis, ymis) ∈ D(q) mis do |
| 2 | xadd ~ PLLM(·|T(2) mis (cmis, xmis, ymis)) |
| 3 | D(q+1) add ← D(q+1) add ∪ {(cmis, xadd, ymis)} |
| Method | Data Size / Results | IMDb | QNLI | RTE | Adversarial QA (EM/F1) | Average |
| Gold Data | Data Size | 25k | 105k | 2.5k | 30k | 40.63k |
| Results | 87.93 | 88.05 | 58.12 | 18.6/29.85 | 56.51 | |
| ProGen | Data Size | 100k | - | - | - | - |
| Results | 84.12 | - | - | - | - | |
| ZeroGen | Data Size | 200k | 200k | 200k | 200k | 200k |
| Results | 84.28 | 71.19 | 59.93 | 6.33/9.96 | 46.34 | |
| GoldGen | Data Size | 25k | 150k | 30k | 80k | 61.25k |
| Results | 87.93 | 78.31 | 64.25 | 11.63/23.33 | 53.09 | |
| S3 | Data Size | 21.2k | 168k | 33.6k | 81.5k | 76.08k |
| Results | 89.00 | 79.92 | 73.29 | 12.50/24.38 | 55.73 |
| Method | IMDb | QNLI | RTE | Adversarial QA |
| S3 | 89.00 | 79.92 | 73.29 | 12.50/24.38 |
| w/o EES | 86.86 | 73.70 | 65.71 | 8.70/20.03 |
| with Rationale | w/o Rationale | |
| Dataset Size | 15k | 40k |
| Results (%) | 86.86 | 85.34 |
| Method | IMDb | QNLI | AdQA |
| ZeroGen | 84.28 | 68.60 | 4.60/9.62 |
| +ourAdd | 87.50 | 73.51 | 9.70/20.10 |
| +synAdd | 87.41 | 72.21 | 10.27/19.92 |
| Label | IMDb | QNLI | RTE | AdQA |
| Data Num | 6,173 | 51,100 | 1,522 | 51,532 |
| Avg. Cos Sim | 0.9497 | 0.9537 | 0.9380 | 0.9468 |
| Avg. Edit Dist. | 273.92 | 14.64 | 16.38 | 13.99 |
| Avg. \\( x_{mis} \\)len | 288.04 | 14.17 | 13.91 | 13.73 |
| Avg. \\( x_{add} \\)len | 218.72 | 19.97 | 24.61 | 18.70 |
| Dataset | S3 Coverage | ZeroGen Coverage |
| QNLI | 76.35 | 63.03 |
| RTE | 73.59 | 14.90 |
| AdQA | 51.02 | 46.00 |
| Method | IMDb | QNLI | RTE | AdQA |
| Gold Data | 92.30 | 91.00 | 71.50 | 22.97/36.59 |
| ZeroGen | 83.66 | 70.11 | 72.2 | 5.07/10.74 |
| S3 | 89.55 | 85.20 | 76.17 | 20.50/34.40 |
| Dataset | Prompt Type | Prompt | Label word (Y) |
| IMDb | \(T_{ration}\) | Imagine you are watching a movie; consider <X> reasons that may lead to <Y>impression of the movie. | positive/negative |
| \(T_{query}^{(1)}\) | Now imagine that you just watched a movie that has <X>. Now you should write a <Y> review about this movie. | positive/negative | |
| \(T_{mis}^{(1)}\) | Write a <Y> movie similar to: \n <X> | positive/negative | |
| QNLI | \(T_{query}^{(2)}\) | Given an information paragraph: <X>\n Please ask a question that has answers <Y> the information paragraph | in/ not in |
| \(T_{mis}^{(2)}\) | Given a premise: <X["premise"]>\n And here is a question: <X["question"]} that the answer of question is <Y> the premise.\nPlease write another question similar to the given question and have answers <Y> the premise. | in/ not in | |
| RTE | \(T_{query}^{(2)}\) | <X>\nBased on the above description, the following sentence is definitely <Y>: | correct/wrong |
| \(T_{mis}^{(2)}\) | <X["premise"]}>\nBased on the above description, the following sentence: <X["Hypothesis"]} is definitely <Y>. Now write a sentence similar to the given sentence and is definitely <Y> based on the given description. | correct/wrong | |
| AdQA | \(T_{query}^{(2)}\) | Given a context: <X["context"]}>\nX [<["answer"]} is the answer to the following question: | NA |
| \(T_{mis}^{(2)}\) | Given a context: <X["context"]}>\nX [<["answer"]} is the answer to: <X["question"]}>.\nA question that has the same answer in the context is: | NA |
| 1 | for each (cmis, xmis, ymis) ∈ D(q) mis do |
| 2 | xadd ~ PLLM(·|T(2) mis (cmis, xmis, ymis)) |
| 3 | D(q+1) add ← D(q+1) add ∪ {(cmis, xadd, ymis)} |
| Method | Data Size / Results | IMDb | QNLI | RTE | Adversarial QA (EM/F1) | Average |
| Gold Data | Data Size | 25k | 105k | 2.5k | 30k | 40.63k |
| Results | 87.93 | 88.05 | 58.12 | 18.6/29.85 | 56.51 | |
| ProGen | Data Size | 100k | - | - | - | - |
| Results | 84.12 | - | - | - | - | |
| ZeroGen | Data Size | 200k | 200k | 200k | 200k | 200k |
| Results | 84.28 | 71.19 | 59.93 | 6.33/9.96 | 46.34 | |
| GoldGen | Data Size | 25k | 150k | 30k | 80k | 61.25k |
| Results | 87.93 | 78.31 | 64.25 | 11.63/23.33 | 53.09 | |
| S3 | Data Size | 21.2k | 168k | 33.6k | 81.5k | 76.08k |
| Results | 89.00 | 79.92 | 73.29 | 12.50/24.38 | 55.73 |
| Method | IMDb | QNLI | RTE | Adversarial QA |
| S3 | 89.00 | 79.92 | 73.29 | 12.50/24.38 |
| w/o EES | 86.86 | 73.70 | 65.71 | 8.70/20.03 |
| with Rationale | w/o Rationale | |
| Dataset Size | 15k | 40k |
| Results (%) | 86.86 | 85.34 |
| Method | IMDb | QNLI | AdQA |
| ZeroGen | 84.28 | 68.60 | 4.60/9.62 |
| +ourAdd | 87.50 | 73.51 | 9.70/20.10 |
| +synAdd | 87.41 | 72.21 | 10.27/19.92 |
| Label | IMDb | QNLI | RTE | AdQA |
| Data Num | 6,173 | 51,100 | 1,522 | 51,532 |
| Avg. Cos Sim | 0.9497 | 0.9537 | 0.9380 | 0.9468 |
| Avg. Edit Dist. | 273.92 | 14.64 | 16.38 | 13.99 |
| Avg. \( x_{mis} \)len | 288.04 | 14.17 | 13.91 | 13.73 |
| Avg. \( x_{add} \)len | 218.72 | 19.97 | 24.61 | 18.70 |
| Dataset | S3 Coverage | ZeroGen Coverage |
| QNLI | 76.35 | 63.03 |
| RTE | 73.59 | 14.90 |
| AdQA | 51.02 | 46.00 |
| Method | IMDb | QNLI | RTE | AdQA |
| Gold Data | 92.30 | 91.00 | 71.50 | 22.97/36.59 |
| ZeroGen | 83.66 | 70.11 | 72.2 | 5.07/10.74 |
| S3 | 89.55 | 85.20 | 76.17 | 20.50/34.40 |
| Dataset | Prompt Type | Prompt | Label word (Y) |
| IMDb | \\(T_{ration}\\) | Imagine you are watching a movie; consider <X> reasons that may lead to <Y>impression of the movie. | positive/negative |
| \\(T_{query}^{(1)}\\) | Now imagine that you just watched a movie that has <X>. Now you should write a <Y> review about this movie. | positive/negative | |
| \\(T_{mis}^{(1)}\\) | Write a <Y> movie similar to: \\n <X> | positive/negative | |
| QNLI | \\(T_{query}^{(2)}\\) | Given an information paragraph: <X>\\n Please ask a question that has answers <Y> the information paragraph | in/ not in |
| \\(T_{mis}^{(2)}\\) | Given a premise: <X["premise"]>\\n And here is a question: <X["question"]} that the answer of question is <Y> the premise.\\nPlease write another question similar to the given question and have answers <Y> the premise. | in/ not in | |
| RTE | \\(T_{query}^{(2)}\\) | <X>\\nBased on the above description, the following sentence is definitely <Y>: | correct/wrong |
| \\(T_{mis}^{(2)}\\) | <X["premise"]}>\\nBased on the above description, the following sentence: <X["Hypothesis"]} is definitely <Y>. Now write a sentence similar to the given sentence and is definitely <Y> based on the given description. | correct/wrong | |
| AdQA | \\(T_{query}^{(2)}\\) | Given a context: <X["context"]}>\\nX [<["answer"]} is the answer to the following question: | NA |
| \\(T_{mis}^{(2)}\\) | Given a context: <X["context"]}>\\nX [<["answer"]} is the answer to: <X["question"]}>.\\nA question that has the same answer in the context is: | NA |
| 1 | for each (cmis, xmis, ymis) ∈ D(q) mis do |
| 2 | xadd ~ PLLM(·|T(2) mis (cmis, xmis, ymis)) |
| 3 | D(q+1) add ← D(q+1) add ∪ {(cmis, xadd, ymis)} |
| Method | Data Size / Results | IMDb | QNLI | RTE | Adversarial QA (EM/F1) | Average |
| Gold Data | Data Size | 25k | 105k | 2.5k | 30k | 40.63k |
| Results | 87.93 | 88.05 | 58.12 | 18.6/29.85 | 56.51 | |
| ProGen | Data Size | 100k | - | - | - | - |
| Results | 84.12 | - | - | - | - | |
| ZeroGen | Data Size | 200k | 200k | 200k | 200k | 200k |
| Results | 84.28 | 71.19 | 59.93 | 6.33/9.96 | 46.34 | |
| GoldGen | Data Size | 25k | 150k | 30k | 80k | 61.25k |
| Results | 87.93 | 78.31 | 64.25 | 11.63/23.33 | 53.09 | |
| S3 | Data Size | 21.2k | 168k | 33.6k | 81.5k | 76.08k |
| Results | 89.00 | 79.92 | 73.29 | 12.50/24.38 | 55.73 |
| Method | IMDb | QNLI | RTE | Adversarial QA |
| S3 | 89.00 | 79.92 | 73.29 | 12.50/24.38 |
| w/o EES | 86.86 | 73.70 | 65.71 | 8.70/20.03 |
| with Rationale | w/o Rationale | |
| Dataset Size | 15k | 40k |
| Results (%) | 86.86 | 85.34 |
| Method | IMDb | QNLI | AdQA |
| ZeroGen | 84.28 | 68.60 | 4.60/9.62 |
| +ourAdd | 87.50 | 73.51 | 9.70/20.10 |
| +synAdd | 87.41 | 72.21 | 10.27/19.92 |
| Label | IMDb | QNLI | RTE | AdQA |
| Data Num | 6,173 | 51,100 | 1,522 | 51,532 |
| Avg. Cos Sim | 0.9497 | 0.9537 | 0.9380 | 0.9468 |
| Avg. Edit Dist. | 273.92 | 14.64 | 16.38 | 13.99 |
| Avg. \\( x_{mis} \\)len | 288.04 | 14.17 | 13.91 | 13.73 |
| Avg. \\( x_{add} \\)len | 218.72 | 19.97 | 24.61 | 18.70 |
| Dataset | S3 Coverage | ZeroGen Coverage |
| QNLI | 76.35 | 63.03 |
| RTE | 73.59 | 14.90 |
| AdQA | 51.02 | 46.00 |
| Method | IMDb | QNLI | RTE | AdQA |
| Gold Data | 92.30 | 91.00 | 71.50 | 22.97/36.59 |
| ZeroGen | 83.66 | 70.11 | 72.2 | 5.07/10.74 |
| S3 | 89.55 | 85.20 | 76.17 | 20.50/34.40 |
| #N | #R | #Images | #Comments | |
| Weibo-19 | 877 | 590 | 1,467 | 4,534 |
| Weibo-17 | 4,749 | 4,779 | 9,528 | 0 |
| 6,026 | 7,898 | 514 | 0 | |
| Pheme | 1,428 | 590 | 2,018 | 21,564 |
| Weibo-19 | Weibo-17 | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 80.96±2.26 | 80.19±2.37 | 79.86±2.46 | 79.87±2.40 | 82.83±2.64 | 82.76±2.42 | 82.68±2.58 | 82.71±2.43 |
| MVAE | 71.67±0.89 | 70.52±0.95 | 70.21±1.01 | 70.34±0.98 | 81.23±1.25 | 84.54±1.75 | 74.68±1.32 | 79.94±1.58 |
| SAFE | 84.95±0.85 | 84.98±0.82 | 84.95±0.91 | 84.96±0.86 | 75.56±1.43 | 76.54±1.35 | 74.95±1.51 | 74.80±1.46 |
| EBGCN | 83.14±2.01 | 85.46±2.12 | 81.76±1.54 | 81.45±1.74 | - | - | - | - |
| GLAN | 82.44±2.02 | 82.45±2.26 | 80.86±1.71 | 81.26±1.93 | - | - | - | - |
| MFAN | 88.95±1.43 | 88.91±1.60 | 88.13±1.68 | 88.33±1.53 | 88.23±1.83 | 88.56±1.91 | 88.07±1.98 | 88.17±1.73 |
| ChatGPT | 29.83±0 | 28.27±0 | 28.95±0 | 28.52±0 | 0±0 | 0±0 | 0±0 | 0±0 |
| CLKD-IMRD | 93.36±1.12 | 93.38±1.23 | 92.91±1.28 | 93.07±1.20 | 91.25±1.01 | 91.29±1.32 | 91.34±1.33 | 91.14±1.71 |
| Pheme | ||||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 71.52±1.96 | 77.95±2.14 | 70.65±3.23 | 72.79±2.69 | 77.13±0.96 | 71.39±1.07 | 70.07±2.19 | 70.44±1.69 |
| MVAE | 79.91±0.96 | 84.88±1.23 | 71.48±1.10 | 71.57±1.25 | 77.62±0.64 | 73.49±0.81 | 72.25±0.90 | 72.77±0.81 |
| SAFE | 76.49±1.24 | 82.18±1.62 | 79.50±1.01 | 79.68±0.98 | 81.49±0.84 | 79.88±1.22 | 79.50±0.81 | 79.68±0.70 |
| EBGCN | - | - | - | - | 82.99±0.65 | 81.31±0.73 | 79.29±0.71 | 79.82±0.64 |
| GLAN | - | - | - | - | 83.32±1.64 | 81.25±2.06 | 77.13±3.26 | 78.51±2.68 |
| MFAN | 86.08±1.12 | 85.03±1.67 | 83.61±1.93 | 84.36±1.64 | 88.73±0.83 | 87.07±1.41 | 85.61±1.65 | 86.16±1.04 |
| ChatGPT | 0±0 | 0±0 | 0±0 | 0±0 | 34.29±0 | 24.26±0 | 26.94±0 | 25.53±0 |
| CLKD-IMRD | 88.01±0.91 | 88.15±1.03 | 84.64±0.65 | 86.53±0.97 | 89.09±0.47 | 87.46±0.96 | 85.86±0.16 | 86.58±0.41 |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| Teacher | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| Student-1 initial | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| Student-1 after KD | 94.24(↑1.02) | 94.67(↑1.34) | 93.40(↑0.85) | 93.93(↑1.03) | 88.83(↑0.52) | 87.39(↑0.67) | 85.11(↑0.62) | 86.13(↑0.65) |
| Student-2 initial | 91.53 | 90.99 | 91.67 | 91.27 | 86.49 | 84.63 | 81.91 | 83.08 |
| Student-2 after KD | 93.22(↑1.69) | 93.67(↑2.68) | 92.28(↑0.61) | 92.85(↑1.58) | 88.31(↑1.82) | 86.23(↑1.60) | 85.26(↑3.35) | 85.72(↑2.64) |
| Student-3 | 92.20 | 92.33 | 91.42 | 91.82 | 87.01 | 84.56 | 83.86 | 84.18 |
| Student-3 after KD | 93.22(↑1.02) | 93.19(↑0.86) | 92.69(↑1.27) | 91.92(↑0.10) | 88.31(↑1.30) | 86.91(↑2.35) | 84.23(↑0.37) | 85.40(↑1.22) |
| Student-4 | 90.85 | 91.85 | 89.33 | 90.23 | 86.23 | 83.65 | 82.76 | 83.18 |
| Student-4 after KD | 92.20(↑1.35) | 92.33(↑0.48) | 91.42(↑2.09) | 91.81(↑1.58) | 87.79(↑1.56) | 86.65(↑3.00) | 83.08(↑0.32) | 84.57(↑1.39) |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| CLKD-IMRD | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| w/o text | 81.02 | 80.36 | 80.00 | 80.17 | 77.14 | 72.76 | 68.56 | 69.89 |
| w/o image | 92.20 | 92.49 | 91.29 | 91.79 | 88.05 | 86.30 | 84.30 | 85.20 |
| w/o comment | 92.54 | 93.36 | 91.30 | 92.08 | 86.75 | 84.43 | 83.12 | 83.73 |
| w/o contrastive learning | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| w/o projection gradient descent | 92.54 | 92.17 | 92.39 | 92.27 | 87.27 | 84.60 | 84.78 | 84.79 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| 0% comments | 92.54 | 93.36 | 91.30 | 92.08 |
| only the first comment | 94.24 | 94.07 | 93.95 | 94.01 |
| 20% comments | 93.56 | 93.61 | 92.97 | 93.26 |
| 50 % comments | 92.88 | 92.37 | 93.08 | 92.67 |
| 80 % comments | 92.88 | 92.90 | 92.27 | 92.55 |
| all comments | 93.22 | 93.33 | 92.55 | 92.90 |
| Pheme | ||||
| 0% comments | 86.75 | 84.43 | 83.12 | 83.73 |
| only the first comment | 89.61 | 88.49 | 85.92 | 87.06 |
| 20% comments | 87.27 | 84.71 | 82.53 | 84.62 |
| 50 % comments | 87.01 | 85.14 | 82.79 | 83.82 |
| 80 % comments | 88.02 | 87.04 | 83.14 | 84.74 |
| all comments | 88.02 | 85.50 | 85.50 | 85.50 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| Zero co-att | 92.54 | 92.04 | 92.66 | 92.31 |
| One co-att | 93.22 | 93.06 | 92.82 | 92.94 |
| Two co-att | 93.90 | 93.77 | 93.53 | 93.64 |
| Three co-att | 94.24 | 94.07 | 93.95 | 94.01 |
| Pheme | ||||
| Zero co-att | 86.49 | 83.31 | 85.79 | 84.33 |
| One co-att | 88.57 | 87.58 | 84.15 | 85.60 |
| Two co-att | 88.83 | 87.80 | 84.59 | 85.96 |
| Three co-att | 89.61 | 88.49 | 85.92 | 87.06 |
| #N | #R | #Images | #Comments | |
| Weibo-19 | 877 | 590 | 1,467 | 4,534 |
| Weibo-17 | 4,749 | 4,779 | 9,528 | 0 |
| 6,026 | 7,898 | 514 | 0 | |
| Pheme | 1,428 | 590 | 2,018 | 21,564 |
| Weibo-19 | Weibo-17 | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 80.96±2.26 | 80.19±2.37 | 79.86±2.46 | 79.87±2.40 | 82.83±2.64 | 82.76±2.42 | 82.68±2.58 | 82.71±2.43 |
| MVAE | 71.67±0.89 | 70.52±0.95 | 70.21±1.01 | 70.34±0.98 | 81.23±1.25 | 84.54±1.75 | 74.68±1.32 | 79.94±1.58 |
| SAFE | 84.95±0.85 | 84.98±0.82 | 84.95±0.91 | 84.96±0.86 | 75.56±1.43 | 76.54±1.35 | 74.95±1.51 | 74.80±1.46 |
| EBGCN | 83.14±2.01 | 85.46±2.12 | 81.76±1.54 | 81.45±1.74 | - | - | - | - |
| GLAN | 82.44±2.02 | 82.45±2.26 | 80.86±1.71 | 81.26±1.93 | - | - | - | - |
| MFAN | 88.95±1.43 | 88.91±1.60 | 88.13±1.68 | 88.33±1.53 | 88.23±1.83 | 88.56±1.91 | 88.07±1.98 | 88.17±1.73 |
| ChatGPT | 29.83±0 | 28.27±0 | 28.95±0 | 28.52±0 | 0±0 | 0±0 | 0±0 | 0±0 |
| CLKD-IMRD | 93.36±1.12 | 93.38±1.23 | 92.91±1.28 | 93.07±1.20 | 91.25±1.01 | 91.29±1.32 | 91.34±1.33 | 91.14±1.71 |
| Pheme | ||||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 71.52±1.96 | 77.95±2.14 | 70.65±3.23 | 72.79±2.69 | 77.13±0.96 | 71.39±1.07 | 70.07±2.19 | 70.44±1.69 |
| MVAE | 79.91±0.96 | 84.88±1.23 | 71.48±1.10 | 71.57±1.25 | 77.62±0.64 | 73.49±0.81 | 72.25±0.90 | 72.77±0.81 |
| SAFE | 76.49±1.24 | 82.18±1.62 | 79.50±1.01 | 79.68±0.98 | 81.49±0.84 | 79.88±1.22 | 79.50±0.81 | 79.68±0.70 |
| EBGCN | - | - | - | - | 82.99±0.65 | 81.31±0.73 | 79.29±0.71 | 79.82±0.64 |
| GLAN | - | - | - | - | 83.32±1.64 | 81.25±2.06 | 77.13±3.26 | 78.51±2.68 |
| MFAN | 86.08±1.12 | 85.03±1.67 | 83.61±1.93 | 84.36±1.64 | 88.73±0.83 | 87.07±1.41 | 85.61±1.65 | 86.16±1.04 |
| ChatGPT | 0±0 | 0±0 | 0±0 | 0±0 | 34.29±0 | 24.26±0 | 26.94±0 | 25.53±0 |
| CLKD-IMRD | 88.01±0.91 | 88.15±1.03 | 84.64±0.65 | 86.53±0.97 | 89.09±0.47 | 87.46±0.96 | 85.86±0.16 | 86.58±0.41 |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| Teacher | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| Student-1 initial | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| Student-1 after KD | 94.24(↑1.02) | 94.67(↑1.34) | 93.40(↑0.85) | 93.93(↑1.03) | 88.83(↑0.52) | 87.39(↑0.67) | 85.11(↑0.62) | 86.13(↑0.65) |
| Student-2 initial | 91.53 | 90.99 | 91.67 | 91.27 | 86.49 | 84.63 | 81.91 | 83.08 |
| Student-2 after KD | 93.22(↑1.69) | 93.67(↑2.68) | 92.28(↑0.61) | 92.85(↑1.58) | 88.31(↑1.82) | 86.23(↑1.60) | 85.26(↑3.35) | 85.72(↑2.64) |
| Student-3 | 92.20 | 92.33 | 91.42 | 91.82 | 87.01 | 84.56 | 83.86 | 84.18 |
| Student-3 after KD | 93.22(↑1.02) | 93.19(↑0.86) | 92.69(↑1.27) | 91.92(↑0.10) | 88.31(↑1.30) | 86.91(↑2.35) | 84.23(↑0.37) | 85.40(↑1.22) |
| Student-4 | 90.85 | 91.85 | 89.33 | 90.23 | 86.23 | 83.65 | 82.76 | 83.18 |
| Student-4 after KD | 92.20(↑1.35) | 92.33(↑0.48) | 91.42(↑2.09) | 91.81(↑1.58) | 87.79(↑1.56) | 86.65(↑3.00) | 83.08(↑0.32) | 84.57(↑1.39) |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| CLKD-IMRD | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| w/o text | 81.02 | 80.36 | 80.00 | 80.17 | 77.14 | 72.76 | 68.56 | 69.89 |
| w/o image | 92.20 | 92.49 | 91.29 | 91.79 | 88.05 | 86.30 | 84.30 | 85.20 |
| w/o comment | 92.54 | 93.36 | 91.30 | 92.08 | 86.75 | 84.43 | 83.12 | 83.73 |
| w/o contrastive learning | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| w/o projection gradient descent | 92.54 | 92.17 | 92.39 | 92.27 | 87.27 | 84.60 | 84.78 | 84.79 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| 0% comments | 92.54 | 93.36 | 91.30 | 92.08 |
| only the first comment | 94.24 | 94.07 | 93.95 | 94.01 |
| 20% comments | 93.56 | 93.61 | 92.97 | 93.26 |
| 50 % comments | 92.88 | 92.37 | 93.08 | 92.67 |
| 80 % comments | 92.88 | 92.90 | 92.27 | 92.55 |
| all comments | 93.22 | 93.33 | 92.55 | 92.90 |
| Pheme | ||||
| 0% comments | 86.75 | 84.43 | 83.12 | 83.73 |
| only the first comment | 89.61 | 88.49 | 85.92 | 87.06 |
| 20% comments | 87.27 | 84.71 | 82.53 | 84.62 |
| 50 % comments | 87.01 | 85.14 | 82.79 | 83.82 |
| 80 % comments | 88.02 | 87.04 | 83.14 | 84.74 |
| all comments | 88.02 | 85.50 | 85.50 | 85.50 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| Zero co-att | 92.54 | 92.04 | 92.66 | 92.31 |
| One co-att | 93.22 | 93.06 | 92.82 | 92.94 |
| Two co-att | 93.90 | 93.77 | 93.53 | 93.64 |
| Three co-att | 94.24 | 94.07 | 93.95 | 94.01 |
| Pheme | ||||
| Zero co-att | 86.49 | 83.31 | 85.79 | 84.33 |
| One co-att | 88.57 | 87.58 | 84.15 | 85.60 |
| Two co-att | 88.83 | 87.80 | 84.59 | 85.96 |
| Three co-att | 89.61 | 88.49 | 85.92 | 87.06 |
| #N | #R | #Images | #Comments | |
| Weibo-19 | 877 | 590 | 1,467 | 4,534 |
| Weibo-17 | 4,749 | 4,779 | 9,528 | 0 |
| 6,026 | 7,898 | 514 | 0 | |
| Pheme | 1,428 | 590 | 2,018 | 21,564 |
| Weibo-19 | Weibo-17 | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 80.96±2.26 | 80.19±2.37 | 79.86±2.46 | 79.87±2.40 | 82.83±2.64 | 82.76±2.42 | 82.68±2.58 | 82.71±2.43 |
| MVAE | 71.67±0.89 | 70.52±0.95 | 70.21±1.01 | 70.34±0.98 | 81.23±1.25 | 84.54±1.75 | 74.68±1.32 | 79.94±1.58 |
| SAFE | 84.95±0.85 | 84.98±0.82 | 84.95±0.91 | 84.96±0.86 | 75.56±1.43 | 76.54±1.35 | 74.95±1.51 | 74.80±1.46 |
| EBGCN | 83.14±2.01 | 85.46±2.12 | 81.76±1.54 | 81.45±1.74 | - | - | - | - |
| GLAN | 82.44±2.02 | 82.45±2.26 | 80.86±1.71 | 81.26±1.93 | - | - | - | - |
| MFAN | 88.95±1.43 | 88.91±1.60 | 88.13±1.68 | 88.33±1.53 | 88.23±1.83 | 88.56±1.91 | 88.07±1.98 | 88.17±1.73 |
| ChatGPT | 29.83±0 | 28.27±0 | 28.95±0 | 28.52±0 | 0±0 | 0±0 | 0±0 | 0±0 |
| CLKD-IMRD | 93.36±1.12 | 93.38±1.23 | 92.91±1.28 | 93.07±1.20 | 91.25±1.01 | 91.29±1.32 | 91.34±1.33 | 91.14±1.71 |
| Pheme | ||||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 71.52±1.96 | 77.95±2.14 | 70.65±3.23 | 72.79±2.69 | 77.13±0.96 | 71.39±1.07 | 70.07±2.19 | 70.44±1.69 |
| MVAE | 79.91±0.96 | 84.88±1.23 | 71.48±1.10 | 71.57±1.25 | 77.62±0.64 | 73.49±0.81 | 72.25±0.90 | 72.77±0.81 |
| SAFE | 76.49±1.24 | 82.18±1.62 | 79.50±1.01 | 79.68±0.98 | 81.49±0.84 | 79.88±1.22 | 79.50±0.81 | 79.68±0.70 |
| EBGCN | - | - | - | - | 82.99±0.65 | 81.31±0.73 | 79.29±0.71 | 79.82±0.64 |
| GLAN | - | - | - | - | 83.32±1.64 | 81.25±2.06 | 77.13±3.26 | 78.51±2.68 |
| MFAN | 86.08±1.12 | 85.03±1.67 | 83.61±1.93 | 84.36±1.64 | 88.73±0.83 | 87.07±1.41 | 85.61±1.65 | 86.16±1.04 |
| ChatGPT | 0±0 | 0±0 | 0±0 | 0±0 | 34.29±0 | 24.26±0 | 26.94±0 | 25.53±0 |
| CLKD-IMRD | 88.01±0.91 | 88.15±1.03 | 84.64±0.65 | 86.53±0.97 | 89.09±0.47 | 87.46±0.96 | 85.86±0.16 | 86.58±0.41 |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| Teacher | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| Student-1 initial | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| Student-1 after KD | 94.24(↑1.02) | 94.67(↑1.34) | 93.40(↑0.85) | 93.93(↑1.03) | 88.83(↑0.52) | 87.39(↑0.67) | 85.11(↑0.62) | 86.13(↑0.65) |
| Student-2 initial | 91.53 | 90.99 | 91.67 | 91.27 | 86.49 | 84.63 | 81.91 | 83.08 |
| Student-2 after KD | 93.22(↑1.69) | 93.67(↑2.68) | 92.28(↑0.61) | 92.85(↑1.58) | 88.31(↑1.82) | 86.23(↑1.60) | 85.26(↑3.35) | 85.72(↑2.64) |
| Student-3 | 92.20 | 92.33 | 91.42 | 91.82 | 87.01 | 84.56 | 83.86 | 84.18 |
| Student-3 after KD | 93.22(↑1.02) | 93.19(↑0.86) | 92.69(↑1.27) | 91.92(↑0.10) | 88.31(↑1.30) | 86.91(↑2.35) | 84.23(↑0.37) | 85.40(↑1.22) |
| Student-4 | 90.85 | 91.85 | 89.33 | 90.23 | 86.23 | 83.65 | 82.76 | 83.18 |
| Student-4 after KD | 92.20(↑1.35) | 92.33(↑0.48) | 91.42(↑2.09) | 91.81(↑1.58) | 87.79(↑1.56) | 86.65(↑3.00) | 83.08(↑0.32) | 84.57(↑1.39) |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| CLKD-IMRD | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| w/o text | 81.02 | 80.36 | 80.00 | 80.17 | 77.14 | 72.76 | 68.56 | 69.89 |
| w/o image | 92.20 | 92.49 | 91.29 | 91.79 | 88.05 | 86.30 | 84.30 | 85.20 |
| w/o comment | 92.54 | 93.36 | 91.30 | 92.08 | 86.75 | 84.43 | 83.12 | 83.73 |
| w/o contrastive learning | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| w/o projection gradient descent | 92.54 | 92.17 | 92.39 | 92.27 | 87.27 | 84.60 | 84.78 | 84.79 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| 0% comments | 92.54 | 93.36 | 91.30 | 92.08 |
| only the first comment | 94.24 | 94.07 | 93.95 | 94.01 |
| 20% comments | 93.56 | 93.61 | 92.97 | 93.26 |
| 50 % comments | 92.88 | 92.37 | 93.08 | 92.67 |
| 80 % comments | 92.88 | 92.90 | 92.27 | 92.55 |
| all comments | 93.22 | 93.33 | 92.55 | 92.90 |
| Pheme | ||||
| 0% comments | 86.75 | 84.43 | 83.12 | 83.73 |
| only the first comment | 89.61 | 88.49 | 85.92 | 87.06 |
| 20% comments | 87.27 | 84.71 | 82.53 | 84.62 |
| 50 % comments | 87.01 | 85.14 | 82.79 | 83.82 |
| 80 % comments | 88.02 | 87.04 | 83.14 | 84.74 |
| all comments | 88.02 | 85.50 | 85.50 | 85.50 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| Zero co-att | 92.54 | 92.04 | 92.66 | 92.31 |
| One co-att | 93.22 | 93.06 | 92.82 | 92.94 |
| Two co-att | 93.90 | 93.77 | 93.53 | 93.64 |
| Three co-att | 94.24 | 94.07 | 93.95 | 94.01 |
| Pheme | ||||
| Zero co-att | 86.49 | 83.31 | 85.79 | 84.33 |
| One co-att | 88.57 | 87.58 | 84.15 | 85.60 |
| Two co-att | 88.83 | 87.80 | 84.59 | 85.96 |
| Three co-att | 89.61 | 88.49 | 85.92 | 87.06 |
| #N | #R | #Images | #Comments | |
| Weibo-19 | 877 | 590 | 1,467 | 4,534 |
| Weibo-17 | 4,749 | 4,779 | 9,528 | 0 |
| 6,026 | 7,898 | 514 | 0 | |
| Pheme | 1,428 | 590 | 2,018 | 21,564 |
| Weibo-19 | Weibo-17 | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 80.96±2.26 | 80.19±2.37 | 79.86±2.46 | 79.87±2.40 | 82.83±2.64 | 82.76±2.42 | 82.68±2.58 | 82.71±2.43 |
| MVAE | 71.67±0.89 | 70.52±0.95 | 70.21±1.01 | 70.34±0.98 | 81.23±1.25 | 84.54±1.75 | 74.68±1.32 | 79.94±1.58 |
| SAFE | 84.95±0.85 | 84.98±0.82 | 84.95±0.91 | 84.96±0.86 | 75.56±1.43 | 76.54±1.35 | 74.95±1.51 | 74.80±1.46 |
| EBGCN | 83.14±2.01 | 85.46±2.12 | 81.76±1.54 | 81.45±1.74 | - | - | - | - |
| GLAN | 82.44±2.02 | 82.45±2.26 | 80.86±1.71 | 81.26±1.93 | - | - | - | - |
| MFAN | 88.95±1.43 | 88.91±1.60 | 88.13±1.68 | 88.33±1.53 | 88.23±1.83 | 88.56±1.91 | 88.07±1.98 | 88.17±1.73 |
| ChatGPT | 29.83±0 | 28.27±0 | 28.95±0 | 28.52±0 | 0±0 | 0±0 | 0±0 | 0±0 |
| CLKD-IMRD | 93.36±1.12 | 93.38±1.23 | 92.91±1.28 | 93.07±1.20 | 91.25±1.01 | 91.29±1.32 | 91.34±1.33 | 91.14±1.71 |
| Pheme | ||||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| EANN | 71.52±1.96 | 77.95±2.14 | 70.65±3.23 | 72.79±2.69 | 77.13±0.96 | 71.39±1.07 | 70.07±2.19 | 70.44±1.69 |
| MVAE | 79.91±0.96 | 84.88±1.23 | 71.48±1.10 | 71.57±1.25 | 77.62±0.64 | 73.49±0.81 | 72.25±0.90 | 72.77±0.81 |
| SAFE | 76.49±1.24 | 82.18±1.62 | 79.50±1.01 | 79.68±0.98 | 81.49±0.84 | 79.88±1.22 | 79.50±0.81 | 79.68±0.70 |
| EBGCN | - | - | - | - | 82.99±0.65 | 81.31±0.73 | 79.29±0.71 | 79.82±0.64 |
| GLAN | - | - | - | - | 83.32±1.64 | 81.25±2.06 | 77.13±3.26 | 78.51±2.68 |
| MFAN | 86.08±1.12 | 85.03±1.67 | 83.61±1.93 | 84.36±1.64 | 88.73±0.83 | 87.07±1.41 | 85.61±1.65 | 86.16±1.04 |
| ChatGPT | 0±0 | 0±0 | 0±0 | 0±0 | 34.29±0 | 24.26±0 | 26.94±0 | 25.53±0 |
| CLKD-IMRD | 88.01±0.91 | 88.15±1.03 | 84.64±0.65 | 86.53±0.97 | 89.09±0.47 | 87.46±0.96 | 85.86±0.16 | 86.58±0.41 |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| Teacher | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| Student-1 initial | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| Student-1 after KD | 94.24(↑1.02) | 94.67(↑1.34) | 93.40(↑0.85) | 93.93(↑1.03) | 88.83(↑0.52) | 87.39(↑0.67) | 85.11(↑0.62) | 86.13(↑0.65) |
| Student-2 initial | 91.53 | 90.99 | 91.67 | 91.27 | 86.49 | 84.63 | 81.91 | 83.08 |
| Student-2 after KD | 93.22(↑1.69) | 93.67(↑2.68) | 92.28(↑0.61) | 92.85(↑1.58) | 88.31(↑1.82) | 86.23(↑1.60) | 85.26(↑3.35) | 85.72(↑2.64) |
| Student-3 | 92.20 | 92.33 | 91.42 | 91.82 | 87.01 | 84.56 | 83.86 | 84.18 |
| Student-3 after KD | 93.22(↑1.02) | 93.19(↑0.86) | 92.69(↑1.27) | 91.92(↑0.10) | 88.31(↑1.30) | 86.91(↑2.35) | 84.23(↑0.37) | 85.40(↑1.22) |
| Student-4 | 90.85 | 91.85 | 89.33 | 90.23 | 86.23 | 83.65 | 82.76 | 83.18 |
| Student-4 after KD | 92.20(↑1.35) | 92.33(↑0.48) | 91.42(↑2.09) | 91.81(↑1.58) | 87.79(↑1.56) | 86.65(↑3.00) | 83.08(↑0.32) | 84.57(↑1.39) |
| Weibo-19 | Pheme | |||||||
| Accuracy | Precision | Recall | F1-Score | Accuracy | Precision | Recall | F1-Score | |
| CLKD-IMRD | 94.24 | 94.07 | 93.95 | 94.01 | 89.61 | 88.49 | 85.92 | 87.06 |
| w/o text | 81.02 | 80.36 | 80.00 | 80.17 | 77.14 | 72.76 | 68.56 | 69.89 |
| w/o image | 92.20 | 92.49 | 91.29 | 91.79 | 88.05 | 86.30 | 84.30 | 85.20 |
| w/o comment | 92.54 | 93.36 | 91.30 | 92.08 | 86.75 | 84.43 | 83.12 | 83.73 |
| w/o contrastive learning | 93.22 | 93.33 | 92.55 | 92.90 | 88.31 | 86.72 | 84.49 | 85.48 |
| w/o projection gradient descent | 92.54 | 92.17 | 92.39 | 92.27 | 87.27 | 84.60 | 84.78 | 84.79 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| 0% comments | 92.54 | 93.36 | 91.30 | 92.08 |
| only the first comment | 94.24 | 94.07 | 93.95 | 94.01 |
| 20% comments | 93.56 | 93.61 | 92.97 | 93.26 |
| 50 % comments | 92.88 | 92.37 | 93.08 | 92.67 |
| 80 % comments | 92.88 | 92.90 | 92.27 | 92.55 |
| all comments | 93.22 | 93.33 | 92.55 | 92.90 |
| Pheme | ||||
| 0% comments | 86.75 | 84.43 | 83.12 | 83.73 |
| only the first comment | 89.61 | 88.49 | 85.92 | 87.06 |
| 20% comments | 87.27 | 84.71 | 82.53 | 84.62 |
| 50 % comments | 87.01 | 85.14 | 82.79 | 83.82 |
| 80 % comments | 88.02 | 87.04 | 83.14 | 84.74 |
| all comments | 88.02 | 85.50 | 85.50 | 85.50 |
| Accuracy | Precision | Recall | F1-Score | |
| Weibo-19 | ||||
| Zero co-att | 92.54 | 92.04 | 92.66 | 92.31 |
| One co-att | 93.22 | 93.06 | 92.82 | 92.94 |
| Two co-att | 93.90 | 93.77 | 93.53 | 93.64 |
| Three co-att | 94.24 | 94.07 | 93.95 | 94.01 |
| Pheme | ||||
| Zero co-att | 86.49 | 83.31 | 85.79 | 84.33 |
| One co-att | 88.57 | 87.58 | 84.15 | 85.60 |
| Two co-att | 88.83 | 87.80 | 84.59 | 85.96 |
| Three co-att | 89.61 | 88.49 | 85.92 | 87.06 |
| Research Question | Measurement | Datasets |
| Nature of the Post-Edited Translation | TER(T', Zero-Shot) vs TER(T', T) | WMT-22 |
| General Quality Improvements | COMET*(S, T) vs COMET*(S, T') | WMT-22 |
| Edits On Human Annotated Error Spans | Edit Efficacy over Error Spans (E3S) | WMT-20, 21, 22 |
| Trustworthiness of the Proposed Edits | Edit Realization Rate (ERR) | WMT-22 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 92.0 | 70.3 |
| Without CoT | 84.6 | 94.5 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 42.9 | 22.0 |
| Without CoT | 38.1 | 34.9 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 81.38 | 39.96 | 85.04 | 56.60 |
| MS Translator | 81.04 | 38.64 | 84.68 | 55.28 |
| MS Translator + GPT-4 | 81.66 | 42.15 | 85.41 | 58.21 |
| MS Translator + GPT-4-CoT | 81.40 | 41.05 | 85.28 | 57.84 |
| MS Translator + GPT-4-SCoT | 81.39 | 41.40 | 85.18 | 57.45 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.51 | 41.36 | 85.26 | 57.53 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 77.66 | 23.98 | 81.02 | 45.21 |
| MS Translator | 77.58 | 23.97 | 80.35 | 40.40 |
| MS Translator + GPT-4 | 79.75 | 31.84 | 82.79 | 53.42 |
| MS Translator + GPT-4-CoT | 79.02 | 28.96 | 82.20 | 50.77 |
| MS Translator + GPT-4-SCoT | 78.94 | 28.80 | 82.09 | 49.65 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 79.29 | 30.13 | 82.49 | 51.78 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 83.56 | 44.67 | 87.21 | 62.35 |
| MS Translator | 83.35 | 43.48 | 86.78 | 62.06 |
| MS Translator + GPT-4 | 83.69 | 44.50 | 87.37 | 62.85 |
| MS Translator + GPT-4-CoT | 83.32 | 43.96 | 87.13 | 62.62 |
| MS Translator + GPT-4-SCoT | 83.12 | 44.17 | 86.90 | 61.94 |
| MS Translator + GPT-3.5-CoT | 81.36 | 43.12 | 84.55 | 50.52 |
| GPT-4-Zero-Shot | 82.95 | 44.69 | 86.80 | 60.85 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 82.04 | 32.11 | 86.69 | 61.04 |
| MS Translator | 81.39 | 31.46 | 86.11 | 59.43 |
| MS Translator + GPT-4 | 82.68 | 34.47 | 87.53 | 63.21 |
| MS Translator + GPT-4-CoT | 81.60 | 32.01 | 86.43 | 59.97 |
| MS Translator + GPT-4-SCoT | 81.81 | 32.56 | 86.56 | 60.20 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.73 | 32.61 | 86.51 | 58.66 |
| System | Initial-QE | PE-QE | E3S |
| PROMT | 74.58 | 78.90 | 57.86 % |
| M2M100 | 73.06 | 79.64 | 70.30 % |
| QUARTZ | 79.36 | 80.38 | 63.63 % |
| JDExplore | 79.43 | 79.61 | 52.32 % |
| Average | 76.61 | 79.63 | 61.03 % |
| System | Initial-QE | PE-QE | E3S |
| AISP-SJTU | 71.87 | 75.66 | 71.62 % |
| Lan-Bridge | 75.52 | 75.82 | 64.14 % |
| LanguageX | 72.80 | 75.80 | 69.34 % |
| M2M100 | 68.24 | 76.49 | 80.72 % |
| Average | 72.11 | 75.94 | 71.46 % |
| System | Initial-QE | PE-QE | E3S |
| eTranslation | 73.52 | 78.61 | 51.42 % |
| HuaweiTSC | 75.30 | 79.22 | 53.35 % |
| M2M100 | 74.80 | 79.93 | 55.70 % |
| PROMT | 75.82 | 78.96 | 48.82 % |
| QUARTZ | 77.87 | 81.02 | 67.81 % |
| JDExplore | 78.91 | 78.45 | 41.18 % |
| Average | 76.04 | 79.37 | 53.04 % |
| Illustration: Edit Realization Rate Difference between GPT-Turbo and GPT-4 |
| Source: Stephanie Lay said she found the piles of letters from the insurance company stashed in her mailbox between Thursday and Monday. \nTranslation: Stephanie Lay sagte, sie habe die Briefstapel der Versicherung zwischen Donnerstag und Montag in ihrem Briefkasten versteckt gefunden. |
| gpt-3.5-turbo |
| Proposed Improvements: \n1. The word "stashed" is not adequately translated in the German text. \n2. The word "mailbox" is not translated correctly in the German text. \n3. The word "piles" is not translated correctly in the German text. \n4. The word "found" is not translated adequately in the German text. \n5. The word "between" is not translated correctly in the German text. \nImproved Translation: \nStephanie Lay sagte, sie haben zwischen Donnerstag und Montag im Briefkasten Stapel von Briefen der Versicherung gefunden. |
| GPT-4 |
| Proposed Improvements: \n1. Replace "Briefstapel der Versicherung" with "Stapel von Briefen der Versicherungsgesellschaft" to better convey the meaning of "pile of letters from the insurance company". \n2. Remove "versteckt" as it is not supported in the input and is not necessary for the meaning of the sentence. \nImproved Translation: \nStephanie Lay sagte, sie haben die Stapel von Briefen der Versicherungsgesellschaft zwischen Donnerstag und Montag in ihrem Briefkasten gefunden. |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.59 | 53.01 % |
| OPPO | 79.25 | 82.32 | 56.11 % |
| eTranslation | 78.82 | 82.72 | 57.04 % |
| Tencent | 80.03 | 82.29 | 52.84 % |
| Huoshan | 78.49 | 82.26 | 56.33 % |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.68 | 39.13 % |
| Facebook-AI | 82.88 | 82.73 | 38.16 % |
| HuaweiTSC | 80.98 | 82.70 | 64.65 % |
| UEdin | 80.82 | 81.34 | 46.67 % |
| eTranslation | 80.04 | 81.60 | 74.05 % |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.50 | 69.86 |
| OPPO | 79.25 | 81.59 | 73.49 |
| eTranslation | 78.82 | 82.03 | 73.45 |
| Tencent | 80.03 | 82.49 | 71.48 |
| Huoshan | 78.49 | 82.34 | 71.72 |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.60 | 56.52 |
| Facebook-AI | 82.88 | 82.45 | 60.53 |
| HuaweiTSC | 80.98 | 82.64 | 70.71 |
| UEdin | 80.82 | 81.99 | 74.17 |
| eTranslation | 80.04 | 81.60 | 74.05 |
| Language | System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| En-Ha | MS Translator | 57.18 | 2.68 | 72.51 | -13.04 |
| En-Ha | MS Translator + GPT-4-CoT | 59.02 | 3.03 | 74.10 | -4.27 |
| Ha-En | MS Translator | 68.45 | 15.62 | 73.26 | 13.23 |
| Ha-En | MS Translator + GPT-4-CoT | 73.19 | 22.05 | 77.83 | 32.76 |
| En-Ja | MS Translator | 85.27 | 36.80 | 87.95 | 57.97 |
| En-Ja | MS Translator + GPT-4-CoT | 85.39 | 38.73 | 89.04 | 62.41 |
| Ja-En | MS Translator | 80.08 | 22.20 | 81.53 | 36.14 |
| Ja-En | MS Translator + GPT-4-CoT | 80.92 | 25.00 | 82.93 | 42.96 |
| Cs-En | MS Translator | 82.18 | 32.90 | 87.44 | 72.02 |
| Cs-En | MS Translator + GPT-4-CoT | 82.65 | 34.76 | 87.56 | 72.28 |
| En-Cs | MS Translator | 84.16 | 59.02 | 90.63 | 94.06 |
| En-Cs | MS Translator + GPT-4-CoT | 84.27 | 58.95 | 90.62 | 93.02 |
| En-Ru | MS Translator | 82.85 | 45.97 | 87.44 | 67.37 |
| En-Ru | MS Translator + GPT-4-CoT | 82.94 | 47.75 | 88.05 | 68.56 |
| Ru-En | MS Translator | 80.72 | 31.11 | 85.16 | 62.24 |
| Ru-En | MS Translator + GPT-4-CoT | 81.28 | 32.79 | 85.66 | 63.94 |
| En-Is | MS Translator | 80.20 | 34.50 | 84.25 | 65.75 |
| En-Is | MS Translator + GPT-4-CoT | 79.27 | 31.96 | 83.47 | 62.45 |
| Is-En | MS Translator | 80.33 | 31.96 | 85.91 | 65.98 |
| Is-En | MS Translator + GPT-4-CoT | 81.11 | 31.96 | 83.47 | 62.45 |
| En-Uk | MS Translator | 81.85 | 47.73 | 86.13 | 61.09 |
| En-Uk | MS Translator + GPT-4-CoT | 81.96 | 48.89 | 86.96 | 63.08 |
| Uk-En | MS Translator | 79.72 | 25.28 | 83.47 | 52.37 |
| Uk-En | MS Translator + GPT-4-CoT | 81.18 | 28.38 | 85.42 | 60.27 |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Wir wollen jene mit Tischplatten Durchmesser von #NUMBER# \nTranslation: We wanted the ones with table tops diameter of #NUMBER# |
| Proposed Improvements: \n1. Change "table tops diameter" to "tablet top diameters" \n2. Add "a" before "diameter" \nImproved Translation: \nWe wanted the ones with tabletop diameters of #NUMBER# |
| Table 18: Edit Hallucination Example on a WMT-22 De-En Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. \nTranslation: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. |
| Proposed Improvements: \n1. Change "will notify" to "informs" to better convey the present tense of "teilt...mit" in the German input. \n2. Add "with" after "informs" to better convey the meaning of "teilt...mit" in the German input. \nImproved Translation: \nThe dealer informs the customer of the cost of the part before the repair. |
| Table 19: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: schicken Sie miritte irgendeine Transaktionsnummer.. oder sonst was, was damit bereits \nTranslation: please send me any transaction number.. or anything else that could help |
| Proposed Improvements: \n1. Replace the double period with a single period after "transaction number. \n2. Add a comma after "or anything else". \nImproved Translation: \nPlease send me any transaction number or anything else that could help. |
| Table 20: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Spiegel hat einen Kratzer und am Rahmen ist ebenfalle ein Fehler, der mit einem gelben Füssel fucht wurde kaschiert zu werden. \nTranslation: The mirror has a scratch and there is also an error on the frame that was attempted to be covered with a yellow filler. |
| Proposed Improvements: \n1. Replace "error" with "blemish" to better convey the meaning of "Fehler" in this context. \n2. Rearrange the sentence structure to improve fluency. \nImproved Translation: \nThe mirror has a scratch, and there is also a blemish on the frame that was attempted to be concealed with a yellow filler. |
| Research Question | Measurement | Datasets |
| Nature of the Post-Edited Translation | TER(T', Zero-Shot) vs TER(T', T) | WMT-22 |
| General Quality Improvements | COMET*(S, T) vs COMET*(S, T') | WMT-22 |
| Edits On Human Annotated Error Spans | Edit Efficacy over Error Spans (E3S) | WMT-20, 21, 22 |
| Trustworthiness of the Proposed Edits | Edit Realization Rate (ERR) | WMT-22 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 92.0 | 70.3 |
| Without CoT | 84.6 | 94.5 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 42.9 | 22.0 |
| Without CoT | 38.1 | 34.9 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 81.38 | 39.96 | 85.04 | 56.60 |
| MS Translator | 81.04 | 38.64 | 84.68 | 55.28 |
| MS Translator + GPT-4 | 81.66 | 42.15 | 85.41 | 58.21 |
| MS Translator + GPT-4-CoT | 81.40 | 41.05 | 85.28 | 57.84 |
| MS Translator + GPT-4-SCoT | 81.39 | 41.40 | 85.18 | 57.45 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.51 | 41.36 | 85.26 | 57.53 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 77.66 | 23.98 | 81.02 | 45.21 |
| MS Translator | 77.58 | 23.97 | 80.35 | 40.40 |
| MS Translator + GPT-4 | 79.75 | 31.84 | 82.79 | 53.42 |
| MS Translator + GPT-4-CoT | 79.02 | 28.96 | 82.20 | 50.77 |
| MS Translator + GPT-4-SCoT | 78.94 | 28.80 | 82.09 | 49.65 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 79.29 | 30.13 | 82.49 | 51.78 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 83.56 | 44.67 | 87.21 | 62.35 |
| MS Translator | 83.35 | 43.48 | 86.78 | 62.06 |
| MS Translator + GPT-4 | 83.69 | 44.50 | 87.37 | 62.85 |
| MS Translator + GPT-4-CoT | 83.32 | 43.96 | 87.13 | 62.62 |
| MS Translator + GPT-4-SCoT | 83.12 | 44.17 | 86.90 | 61.94 |
| MS Translator + GPT-3.5-CoT | 81.36 | 43.12 | 84.55 | 50.52 |
| GPT-4-Zero-Shot | 82.95 | 44.69 | 86.80 | 60.85 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 82.04 | 32.11 | 86.69 | 61.04 |
| MS Translator | 81.39 | 31.46 | 86.11 | 59.43 |
| MS Translator + GPT-4 | 82.68 | 34.47 | 87.53 | 63.21 |
| MS Translator + GPT-4-CoT | 81.60 | 32.01 | 86.43 | 59.97 |
| MS Translator + GPT-4-SCoT | 81.81 | 32.56 | 86.56 | 60.20 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.73 | 32.61 | 86.51 | 58.66 |
| System | Initial-QE | PE-QE | E3S |
| PROMT | 74.58 | 78.90 | 57.86 % |
| M2M100 | 73.06 | 79.64 | 70.30 % |
| QUARTZ | 79.36 | 80.38 | 63.63 % |
| JDExplore | 79.43 | 79.61 | 52.32 % |
| Average | 76.61 | 79.63 | 61.03 % |
| System | Initial-QE | PE-QE | E3S |
| AISP-SJTU | 71.87 | 75.66 | 71.62 % |
| Lan-Bridge | 75.52 | 75.82 | 64.14 % |
| LanguageX | 72.80 | 75.80 | 69.34 % |
| M2M100 | 68.24 | 76.49 | 80.72 % |
| Average | 72.11 | 75.94 | 71.46 % |
| System | Initial-QE | PE-QE | E3S |
| eTranslation | 73.52 | 78.61 | 51.42 % |
| HuaweiTSC | 75.30 | 79.22 | 53.35 % |
| M2M100 | 74.80 | 79.93 | 55.70 % |
| PROMT | 75.82 | 78.96 | 48.82 % |
| QUARTZ | 77.87 | 81.02 | 67.81 % |
| JDExplore | 78.91 | 78.45 | 41.18 % |
| Average | 76.04 | 79.37 | 53.04 % |
| Illustration: Edit Realization Rate Difference between GPT-Turbo and GPT-4 |
| Source: Stephanie Lay said she found the piles of letters from the insurance company stashed in her mailbox between Thursday and Monday. \nTranslation: Stephanie Lay sagte, sie habe die Briefstapel der Versicherung zwischen Donnerstag und Montag in ihrem Briefkasten versteckt gefunden. |
| gpt-3.5-turbo |
| Proposed Improvements: \n1. The word "stashed" is not adequately translated in the German text. \n2. The word "mailbox" is not translated correctly in the German text. \n3. The word "piles" is not translated correctly in the German text. \n4. The word "found" is not translated adequately in the German text. \n5. The word "between" is not translated correctly in the German text. \nImproved Translation: \nStephanie Lay sagte, sie haben zwischen Donnerstag und Montag im Briefkasten Stapel von Briefen der Versicherung gefunden. |
| GPT-4 |
| Proposed Improvements: \n1. Replace "Briefstapel der Versicherung" with "Stapel von Briefen der Versicherungsgesellschaft" to better convey the meaning of "pile of letters from the insurance company". \n2. Remove "versteckt" as it is not supported in the input and is not necessary for the meaning of the sentence. \nImproved Translation: \nStephanie Lay sagte, sie haben die Stapel von Briefen der Versicherungsgesellschaft zwischen Donnerstag und Montag in ihrem Briefkasten gefunden. |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.59 | 53.01 % |
| OPPO | 79.25 | 82.32 | 56.11 % |
| eTranslation | 78.82 | 82.72 | 57.04 % |
| Tencent | 80.03 | 82.29 | 52.84 % |
| Huoshan | 78.49 | 82.26 | 56.33 % |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.68 | 39.13 % |
| Facebook-AI | 82.88 | 82.73 | 38.16 % |
| HuaweiTSC | 80.98 | 82.70 | 64.65 % |
| UEdin | 80.82 | 81.34 | 46.67 % |
| eTranslation | 80.04 | 81.60 | 74.05 % |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.50 | 69.86 |
| OPPO | 79.25 | 81.59 | 73.49 |
| eTranslation | 78.82 | 82.03 | 73.45 |
| Tencent | 80.03 | 82.49 | 71.48 |
| Huoshan | 78.49 | 82.34 | 71.72 |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.60 | 56.52 |
| Facebook-AI | 82.88 | 82.45 | 60.53 |
| HuaweiTSC | 80.98 | 82.64 | 70.71 |
| UEdin | 80.82 | 81.99 | 74.17 |
| eTranslation | 80.04 | 81.60 | 74.05 |
| Language | System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| En-Ha | MS Translator | 57.18 | 2.68 | 72.51 | -13.04 |
| En-Ha | MS Translator + GPT-4-CoT | 59.02 | 3.03 | 74.10 | -4.27 |
| Ha-En | MS Translator | 68.45 | 15.62 | 73.26 | 13.23 |
| Ha-En | MS Translator + GPT-4-CoT | 73.19 | 22.05 | 77.83 | 32.76 |
| En-Ja | MS Translator | 85.27 | 36.80 | 87.95 | 57.97 |
| En-Ja | MS Translator + GPT-4-CoT | 85.39 | 38.73 | 89.04 | 62.41 |
| Ja-En | MS Translator | 80.08 | 22.20 | 81.53 | 36.14 |
| Ja-En | MS Translator + GPT-4-CoT | 80.92 | 25.00 | 82.93 | 42.96 |
| Cs-En | MS Translator | 82.18 | 32.90 | 87.44 | 72.02 |
| Cs-En | MS Translator + GPT-4-CoT | 82.65 | 34.76 | 87.56 | 72.28 |
| En-Cs | MS Translator | 84.16 | 59.02 | 90.63 | 94.06 |
| En-Cs | MS Translator + GPT-4-CoT | 84.27 | 58.95 | 90.62 | 93.02 |
| En-Ru | MS Translator | 82.85 | 45.97 | 87.44 | 67.37 |
| En-Ru | MS Translator + GPT-4-CoT | 82.94 | 47.75 | 88.05 | 68.56 |
| Ru-En | MS Translator | 80.72 | 31.11 | 85.16 | 62.24 |
| Ru-En | MS Translator + GPT-4-CoT | 81.28 | 32.79 | 85.66 | 63.94 |
| En-Is | MS Translator | 80.20 | 34.50 | 84.25 | 65.75 |
| En-Is | MS Translator + GPT-4-CoT | 79.27 | 31.96 | 83.47 | 62.45 |
| Is-En | MS Translator | 80.33 | 31.96 | 85.91 | 65.98 |
| Is-En | MS Translator + GPT-4-CoT | 81.11 | 31.96 | 83.47 | 62.45 |
| En-Uk | MS Translator | 81.85 | 47.73 | 86.13 | 61.09 |
| En-Uk | MS Translator + GPT-4-CoT | 81.96 | 48.89 | 86.96 | 63.08 |
| Uk-En | MS Translator | 79.72 | 25.28 | 83.47 | 52.37 |
| Uk-En | MS Translator + GPT-4-CoT | 81.18 | 28.38 | 85.42 | 60.27 |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Wir wollen jene mit Tischplatten Durchmesser von #NUMBER# \nTranslation: We wanted the ones with table tops diameter of #NUMBER# |
| Proposed Improvements: \n1. Change "table tops diameter" to "tablet top diameters" \n2. Add "a" before "diameter" \nImproved Translation: \nWe wanted the ones with tabletop diameters of #NUMBER# |
| Table 18: Edit Hallucination Example on a WMT-22 De-En Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. \nTranslation: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. |
| Proposed Improvements: \n1. Change "will notify" to "informs" to better convey the present tense of "teilt...mit" in the German input. \n2. Add "with" after "informs" to better convey the meaning of "teilt...mit" in the German input. \nImproved Translation: \nThe dealer informs the customer of the cost of the part before the repair. |
| Table 19: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: schicken Sie miritte irgendeine Transaktionsnummer.. oder sonst was, was damit bereits \nTranslation: please send me any transaction number.. or anything else that could help |
| Proposed Improvements: \n1. Replace the double period with a single period after "transaction number. \n2. Add a comma after "or anything else". \nImproved Translation: \nPlease send me any transaction number or anything else that could help. |
| Table 20: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Spiegel hat einen Kratzer und am Rahmen ist ebenfalle ein Fehler, der mit einem gelben Füssel fucht wurde kaschiert zu werden. \nTranslation: The mirror has a scratch and there is also an error on the frame that was attempted to be covered with a yellow filler. |
| Proposed Improvements: \n1. Replace "error" with "blemish" to better convey the meaning of "Fehler" in this context. \n2. Rearrange the sentence structure to improve fluency. \nImproved Translation: \nThe mirror has a scratch, and there is also a blemish on the frame that was attempted to be concealed with a yellow filler. |
| Research Question | Measurement | Datasets |
| Nature of the Post-Edited Translation | TER(T', Zero-Shot) vs TER(T', T) | WMT-22 |
| General Quality Improvements | COMET*(S, T) vs COMET*(S, T') | WMT-22 |
| Edits On Human Annotated Error Spans | Edit Efficacy over Error Spans (E3S) | WMT-20, 21, 22 |
| Trustworthiness of the Proposed Edits | Edit Realization Rate (ERR) | WMT-22 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 92.0 | 70.3 |
| Without CoT | 84.6 | 94.5 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 42.9 | 22.0 |
| Without CoT | 38.1 | 34.9 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 81.38 | 39.96 | 85.04 | 56.60 |
| MS Translator | 81.04 | 38.64 | 84.68 | 55.28 |
| MS Translator + GPT-4 | 81.66 | 42.15 | 85.41 | 58.21 |
| MS Translator + GPT-4-CoT | 81.40 | 41.05 | 85.28 | 57.84 |
| MS Translator + GPT-4-SCoT | 81.39 | 41.40 | 85.18 | 57.45 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.51 | 41.36 | 85.26 | 57.53 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 77.66 | 23.98 | 81.02 | 45.21 |
| MS Translator | 77.58 | 23.97 | 80.35 | 40.40 |
| MS Translator + GPT-4 | 79.75 | 31.84 | 82.79 | 53.42 |
| MS Translator + GPT-4-CoT | 79.02 | 28.96 | 82.20 | 50.77 |
| MS Translator + GPT-4-SCoT | 78.94 | 28.80 | 82.09 | 49.65 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 79.29 | 30.13 | 82.49 | 51.78 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 83.56 | 44.67 | 87.21 | 62.35 |
| MS Translator | 83.35 | 43.48 | 86.78 | 62.06 |
| MS Translator + GPT-4 | 83.69 | 44.50 | 87.37 | 62.85 |
| MS Translator + GPT-4-CoT | 83.32 | 43.96 | 87.13 | 62.62 |
| MS Translator + GPT-4-SCoT | 83.12 | 44.17 | 86.90 | 61.94 |
| MS Translator + GPT-3.5-CoT | 81.36 | 43.12 | 84.55 | 50.52 |
| GPT-4-Zero-Shot | 82.95 | 44.69 | 86.80 | 60.85 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 82.04 | 32.11 | 86.69 | 61.04 |
| MS Translator | 81.39 | 31.46 | 86.11 | 59.43 |
| MS Translator + GPT-4 | 82.68 | 34.47 | 87.53 | 63.21 |
| MS Translator + GPT-4-CoT | 81.60 | 32.01 | 86.43 | 59.97 |
| MS Translator + GPT-4-SCoT | 81.81 | 32.56 | 86.56 | 60.20 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.73 | 32.61 | 86.51 | 58.66 |
| System | Initial-QE | PE-QE | E3S |
| PROMT | 74.58 | 78.90 | 57.86 % |
| M2M100 | 73.06 | 79.64 | 70.30 % |
| QUARTZ | 79.36 | 80.38 | 63.63 % |
| JDExplore | 79.43 | 79.61 | 52.32 % |
| Average | 76.61 | 79.63 | 61.03 % |
| System | Initial-QE | PE-QE | E3S |
| AISP-SJTU | 71.87 | 75.66 | 71.62 % |
| Lan-Bridge | 75.52 | 75.82 | 64.14 % |
| LanguageX | 72.80 | 75.80 | 69.34 % |
| M2M100 | 68.24 | 76.49 | 80.72 % |
| Average | 72.11 | 75.94 | 71.46 % |
| System | Initial-QE | PE-QE | E3S |
| eTranslation | 73.52 | 78.61 | 51.42 % |
| HuaweiTSC | 75.30 | 79.22 | 53.35 % |
| M2M100 | 74.80 | 79.93 | 55.70 % |
| PROMT | 75.82 | 78.96 | 48.82 % |
| QUARTZ | 77.87 | 81.02 | 67.81 % |
| JDExplore | 78.91 | 78.45 | 41.18 % |
| Average | 76.04 | 79.37 | 53.04 % |
| Illustration: Edit Realization Rate Difference between GPT-Turbo and GPT-4 |
| Source: Stephanie Lay said she found the piles of letters from the insurance company stashed in her mailbox between Thursday and Monday. +Translation: Stephanie Lay sagte, sie habe die Briefstapel der Versicherung zwischen Donnerstag und Montag in ihrem Briefkasten versteckt gefunden. |
| gpt-3.5-turbo |
| Proposed Improvements: +1. The word "stashed" is not adequately translated in the German text. +2. The word "mailbox" is not translated correctly in the German text. +3. The word "piles" is not translated correctly in the German text. +4. The word "found" is not translated adequately in the German text. +5. The word "between" is not translated correctly in the German text. +Improved Translation: +Stephanie Lay sagte, sie haben zwischen Donnerstag und Montag im Briefkasten Stapel von Briefen der Versicherung gefunden. |
| GPT-4 |
| Proposed Improvements: +1. Replace "Briefstapel der Versicherung" with "Stapel von Briefen der Versicherungsgesellschaft" to better convey the meaning of "pile of letters from the insurance company". +2. Remove "versteckt" as it is not supported in the input and is not necessary for the meaning of the sentence. +Improved Translation: +Stephanie Lay sagte, sie haben die Stapel von Briefen der Versicherungsgesellschaft zwischen Donnerstag und Montag in ihrem Briefkasten gefunden. |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.59 | 53.01 % |
| OPPO | 79.25 | 82.32 | 56.11 % |
| eTranslation | 78.82 | 82.72 | 57.04 % |
| Tencent | 80.03 | 82.29 | 52.84 % |
| Huoshan | 78.49 | 82.26 | 56.33 % |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.68 | 39.13 % |
| Facebook-AI | 82.88 | 82.73 | 38.16 % |
| HuaweiTSC | 80.98 | 82.70 | 64.65 % |
| UEdin | 80.82 | 81.34 | 46.67 % |
| eTranslation | 80.04 | 81.60 | 74.05 % |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.50 | 69.86 |
| OPPO | 79.25 | 81.59 | 73.49 |
| eTranslation | 78.82 | 82.03 | 73.45 |
| Tencent | 80.03 | 82.49 | 71.48 |
| Huoshan | 78.49 | 82.34 | 71.72 |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.60 | 56.52 |
| Facebook-AI | 82.88 | 82.45 | 60.53 |
| HuaweiTSC | 80.98 | 82.64 | 70.71 |
| UEdin | 80.82 | 81.99 | 74.17 |
| eTranslation | 80.04 | 81.60 | 74.05 |
| Language | System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| En-Ha | MS Translator | 57.18 | 2.68 | 72.51 | -13.04 |
| En-Ha | MS Translator + GPT-4-CoT | 59.02 | 3.03 | 74.10 | -4.27 |
| Ha-En | MS Translator | 68.45 | 15.62 | 73.26 | 13.23 |
| Ha-En | MS Translator + GPT-4-CoT | 73.19 | 22.05 | 77.83 | 32.76 |
| En-Ja | MS Translator | 85.27 | 36.80 | 87.95 | 57.97 |
| En-Ja | MS Translator + GPT-4-CoT | 85.39 | 38.73 | 89.04 | 62.41 |
| Ja-En | MS Translator | 80.08 | 22.20 | 81.53 | 36.14 |
| Ja-En | MS Translator + GPT-4-CoT | 80.92 | 25.00 | 82.93 | 42.96 |
| Cs-En | MS Translator | 82.18 | 32.90 | 87.44 | 72.02 |
| Cs-En | MS Translator + GPT-4-CoT | 82.65 | 34.76 | 87.56 | 72.28 |
| En-Cs | MS Translator | 84.16 | 59.02 | 90.63 | 94.06 |
| En-Cs | MS Translator + GPT-4-CoT | 84.27 | 58.95 | 90.62 | 93.02 |
| En-Ru | MS Translator | 82.85 | 45.97 | 87.44 | 67.37 |
| En-Ru | MS Translator + GPT-4-CoT | 82.94 | 47.75 | 88.05 | 68.56 |
| Ru-En | MS Translator | 80.72 | 31.11 | 85.16 | 62.24 |
| Ru-En | MS Translator + GPT-4-CoT | 81.28 | 32.79 | 85.66 | 63.94 |
| En-Is | MS Translator | 80.20 | 34.50 | 84.25 | 65.75 |
| En-Is | MS Translator + GPT-4-CoT | 79.27 | 31.96 | 83.47 | 62.45 |
| Is-En | MS Translator | 80.33 | 31.96 | 85.91 | 65.98 |
| Is-En | MS Translator + GPT-4-CoT | 81.11 | 31.96 | 83.47 | 62.45 |
| En-Uk | MS Translator | 81.85 | 47.73 | 86.13 | 61.09 |
| En-Uk | MS Translator + GPT-4-CoT | 81.96 | 48.89 | 86.96 | 63.08 |
| Uk-En | MS Translator | 79.72 | 25.28 | 83.47 | 52.37 |
| Uk-En | MS Translator + GPT-4-CoT | 81.18 | 28.38 | 85.42 | 60.27 |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Wir wollen jene mit Tischplatten Durchmesser von #NUMBER# +Translation: We wanted the ones with table tops diameter of #NUMBER# |
| Proposed Improvements: +1. Change "table tops diameter" to "tablet top diameters" +2. Add "a" before "diameter" +Improved Translation: +We wanted the ones with tabletop diameters of #NUMBER# |
| Table 18: Edit Hallucination Example on a WMT-22 De-En Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. +Translation: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. |
| Proposed Improvements: +1. Change "will notify" to "informs" to better convey the present tense of "teilt...mit" in the German input. +2. Add "with" after "informs" to better convey the meaning of "teilt...mit" in the German input. +Improved Translation: +The dealer informs the customer of the cost of the part before the repair. |
| Table 19: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: schicken Sie miritte irgendeine Transaktionsnummer.. oder sonst was, was damit bereits +Translation: please send me any transaction number.. or anything else that could help |
| Proposed Improvements: +1. Replace the double period with a single period after "transaction number. +2. Add a comma after "or anything else". +Improved Translation: +Please send me any transaction number or anything else that could help. |
| Table 20: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Spiegel hat einen Kratzer und am Rahmen ist ebenfalle ein Fehler, der mit einem gelben Füssel fucht wurde kaschiert zu werden. +Translation: The mirror has a scratch and there is also an error on the frame that was attempted to be covered with a yellow filler. |
| Proposed Improvements: +1. Replace "error" with "blemish" to better convey the meaning of "Fehler" in this context. +2. Rearrange the sentence structure to improve fluency. +Improved Translation: +The mirror has a scratch, and there is also a blemish on the frame that was attempted to be concealed with a yellow filler. |
| Research Question | Measurement | Datasets |
| Nature of the Post-Edited Translation | TER(T', Zero-Shot) vs TER(T', T) | WMT-22 |
| General Quality Improvements | COMET*(S, T) vs COMET*(S, T') | WMT-22 |
| Edits On Human Annotated Error Spans | Edit Efficacy over Error Spans (E3S) | WMT-20, 21, 22 |
| Trustworthiness of the Proposed Edits | Edit Realization Rate (ERR) | WMT-22 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 92.0 | 70.3 |
| Without CoT | 84.6 | 94.5 |
| PE Setting | TER (T', Z) | TER (T', T) |
| With CoT | 42.9 | 22.0 |
| Without CoT | 38.1 | 34.9 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 81.38 | 39.96 | 85.04 | 56.60 |
| MS Translator | 81.04 | 38.64 | 84.68 | 55.28 |
| MS Translator + GPT-4 | 81.66 | 42.15 | 85.41 | 58.21 |
| MS Translator + GPT-4-CoT | 81.40 | 41.05 | 85.28 | 57.84 |
| MS Translator + GPT-4-SCoT | 81.39 | 41.40 | 85.18 | 57.45 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.51 | 41.36 | 85.26 | 57.53 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 77.66 | 23.98 | 81.02 | 45.21 |
| MS Translator | 77.58 | 23.97 | 80.35 | 40.40 |
| MS Translator + GPT-4 | 79.75 | 31.84 | 82.79 | 53.42 |
| MS Translator + GPT-4-CoT | 79.02 | 28.96 | 82.20 | 50.77 |
| MS Translator + GPT-4-SCoT | 78.94 | 28.80 | 82.09 | 49.65 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 79.29 | 30.13 | 82.49 | 51.78 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 83.56 | 44.67 | 87.21 | 62.35 |
| MS Translator | 83.35 | 43.48 | 86.78 | 62.06 |
| MS Translator + GPT-4 | 83.69 | 44.50 | 87.37 | 62.85 |
| MS Translator + GPT-4-CoT | 83.32 | 43.96 | 87.13 | 62.62 |
| MS Translator + GPT-4-SCoT | 83.12 | 44.17 | 86.90 | 61.94 |
| MS Translator + GPT-3.5-CoT | 81.36 | 43.12 | 84.55 | 50.52 |
| GPT-4-Zero-Shot | 82.95 | 44.69 | 86.80 | 60.85 |
| System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| WMT-Best | 82.04 | 32.11 | 86.69 | 61.04 |
| MS Translator | 81.39 | 31.46 | 86.11 | 59.43 |
| MS Translator + GPT-4 | 82.68 | 34.47 | 87.53 | 63.21 |
| MS Translator + GPT-4-CoT | 81.60 | 32.01 | 86.43 | 59.97 |
| MS Translator + GPT-4-SCoT | 81.81 | 32.56 | 86.56 | 60.20 |
| MS Translator + GPT-3.5-CoT | 79.32 | 41.56 | 82.71 | 44.82 |
| GPT-4-Zero-Shot | 81.73 | 32.61 | 86.51 | 58.66 |
| System | Initial-QE | PE-QE | E3S |
| PROMT | 74.58 | 78.90 | 57.86 % |
| M2M100 | 73.06 | 79.64 | 70.30 % |
| QUARTZ | 79.36 | 80.38 | 63.63 % |
| JDExplore | 79.43 | 79.61 | 52.32 % |
| Average | 76.61 | 79.63 | 61.03 % |
| System | Initial-QE | PE-QE | E3S |
| AISP-SJTU | 71.87 | 75.66 | 71.62 % |
| Lan-Bridge | 75.52 | 75.82 | 64.14 % |
| LanguageX | 72.80 | 75.80 | 69.34 % |
| M2M100 | 68.24 | 76.49 | 80.72 % |
| Average | 72.11 | 75.94 | 71.46 % |
| System | Initial-QE | PE-QE | E3S |
| eTranslation | 73.52 | 78.61 | 51.42 % |
| HuaweiTSC | 75.30 | 79.22 | 53.35 % |
| M2M100 | 74.80 | 79.93 | 55.70 % |
| PROMT | 75.82 | 78.96 | 48.82 % |
| QUARTZ | 77.87 | 81.02 | 67.81 % |
| JDExplore | 78.91 | 78.45 | 41.18 % |
| Average | 76.04 | 79.37 | 53.04 % |
| Illustration: Edit Realization Rate Difference between GPT-Turbo and GPT-4 |
| Source: Stephanie Lay said she found the piles of letters from the insurance company stashed in her mailbox between Thursday and Monday. \nTranslation: Stephanie Lay sagte, sie habe die Briefstapel der Versicherung zwischen Donnerstag und Montag in ihrem Briefkasten versteckt gefunden. |
| gpt-3.5-turbo |
| Proposed Improvements: \n1. The word "stashed" is not adequately translated in the German text. \n2. The word "mailbox" is not translated correctly in the German text. \n3. The word "piles" is not translated correctly in the German text. \n4. The word "found" is not translated adequately in the German text. \n5. The word "between" is not translated correctly in the German text. \nImproved Translation: \nStephanie Lay sagte, sie haben zwischen Donnerstag und Montag im Briefkasten Stapel von Briefen der Versicherung gefunden. |
| GPT-4 |
| Proposed Improvements: \n1. Replace "Briefstapel der Versicherung" with "Stapel von Briefen der Versicherungsgesellschaft" to better convey the meaning of "pile of letters from the insurance company". \n2. Remove "versteckt" as it is not supported in the input and is not necessary for the meaning of the sentence. \nImproved Translation: \nStephanie Lay sagte, sie haben die Stapel von Briefen der Versicherungsgesellschaft zwischen Donnerstag und Montag in ihrem Briefkasten gefunden. |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.59 | 53.01 % |
| OPPO | 79.25 | 82.32 | 56.11 % |
| eTranslation | 78.82 | 82.72 | 57.04 % |
| Tencent | 80.03 | 82.29 | 52.84 % |
| Huoshan | 78.49 | 82.26 | 56.33 % |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.68 | 39.13 % |
| Facebook-AI | 82.88 | 82.73 | 38.16 % |
| HuaweiTSC | 80.98 | 82.70 | 64.65 % |
| UEdin | 80.82 | 81.34 | 46.67 % |
| eTranslation | 80.04 | 81.60 | 74.05 % |
| System | Initial-QE | PE-QE | E3S |
| WMT-20 | |||
| Tohoku | 80.93 | 82.50 | 69.86 |
| OPPO | 79.25 | 81.59 | 73.49 |
| eTranslation | 78.82 | 82.03 | 73.45 |
| Tencent | 80.03 | 82.49 | 71.48 |
| Huoshan | 78.49 | 82.34 | 71.72 |
| WMT-21 | |||
| VolcTrans-GLAT | 80.22 | 82.60 | 56.52 |
| Facebook-AI | 82.88 | 82.45 | 60.53 |
| HuaweiTSC | 80.98 | 82.64 | 70.71 |
| UEdin | 80.82 | 81.99 | 74.17 |
| eTranslation | 80.04 | 81.60 | 74.05 |
| Language | System | COMET-KIWI | COMET-QE | COMET-22 | COMET-20 |
| En-Ha | MS Translator | 57.18 | 2.68 | 72.51 | -13.04 |
| En-Ha | MS Translator + GPT-4-CoT | 59.02 | 3.03 | 74.10 | -4.27 |
| Ha-En | MS Translator | 68.45 | 15.62 | 73.26 | 13.23 |
| Ha-En | MS Translator + GPT-4-CoT | 73.19 | 22.05 | 77.83 | 32.76 |
| En-Ja | MS Translator | 85.27 | 36.80 | 87.95 | 57.97 |
| En-Ja | MS Translator + GPT-4-CoT | 85.39 | 38.73 | 89.04 | 62.41 |
| Ja-En | MS Translator | 80.08 | 22.20 | 81.53 | 36.14 |
| Ja-En | MS Translator + GPT-4-CoT | 80.92 | 25.00 | 82.93 | 42.96 |
| Cs-En | MS Translator | 82.18 | 32.90 | 87.44 | 72.02 |
| Cs-En | MS Translator + GPT-4-CoT | 82.65 | 34.76 | 87.56 | 72.28 |
| En-Cs | MS Translator | 84.16 | 59.02 | 90.63 | 94.06 |
| En-Cs | MS Translator + GPT-4-CoT | 84.27 | 58.95 | 90.62 | 93.02 |
| En-Ru | MS Translator | 82.85 | 45.97 | 87.44 | 67.37 |
| En-Ru | MS Translator + GPT-4-CoT | 82.94 | 47.75 | 88.05 | 68.56 |
| Ru-En | MS Translator | 80.72 | 31.11 | 85.16 | 62.24 |
| Ru-En | MS Translator + GPT-4-CoT | 81.28 | 32.79 | 85.66 | 63.94 |
| En-Is | MS Translator | 80.20 | 34.50 | 84.25 | 65.75 |
| En-Is | MS Translator + GPT-4-CoT | 79.27 | 31.96 | 83.47 | 62.45 |
| Is-En | MS Translator | 80.33 | 31.96 | 85.91 | 65.98 |
| Is-En | MS Translator + GPT-4-CoT | 81.11 | 31.96 | 83.47 | 62.45 |
| En-Uk | MS Translator | 81.85 | 47.73 | 86.13 | 61.09 |
| En-Uk | MS Translator + GPT-4-CoT | 81.96 | 48.89 | 86.96 | 63.08 |
| Uk-En | MS Translator | 79.72 | 25.28 | 83.47 | 52.37 |
| Uk-En | MS Translator + GPT-4-CoT | 81.18 | 28.38 | 85.42 | 60.27 |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Wir wollen jene mit Tischplatten Durchmesser von #NUMBER# \nTranslation: We wanted the ones with table tops diameter of #NUMBER# |
| Proposed Improvements: \n1. Change "table tops diameter" to "tablet top diameters" \n2. Add "a" before "diameter" \nImproved Translation: \nWe wanted the ones with tabletop diameters of #NUMBER# |
| Table 18: Edit Hallucination Example on a WMT-22 De-En Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. \nTranslation: Der Handlert teilt dem Kunden die Kosten des Teils vor der Reparatur mit. |
| Proposed Improvements: \n1. Change "will notify" to "informs" to better convey the present tense of "teilt...mit" in the German input. \n2. Add "with" after "informs" to better convey the meaning of "teilt...mit" in the German input. \nImproved Translation: \nThe dealer informs the customer of the cost of the part before the repair. |
| Table 19: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: schicken Sie miritte irgendeine Transaktionsnummer.. oder sonst was, was damit bereits \nTranslation: please send me any transaction number.. or anything else that could help |
| Proposed Improvements: \n1. Replace the double period with a single period after "transaction number. \n2. Add a comma after "or anything else". \nImproved Translation: \nPlease send me any transaction number or anything else that could help. |
| Table 20: Edit Hallucination Example on a WMT-22 En-De Sample: GPT-4 produces an unrealized edit. |
| Example of Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting) |
| Source: Der Spiegel hat einen Kratzer und am Rahmen ist ebenfalle ein Fehler, der mit einem gelben Füssel fucht wurde kaschiert zu werden. \nTranslation: The mirror has a scratch and there is also an error on the frame that was attempted to be covered with a yellow filler. |
| Proposed Improvements: \n1. Replace "error" with "blemish" to better convey the meaning of "Fehler" in this context. \n2. Rearrange the sentence structure to improve fluency. \nImproved Translation: \nThe mirror has a scratch, and there is also a blemish on the frame that was attempted to be concealed with a yellow filler. |
| Tasks | Non-Adaptation Baselines | Adaptation Ratio 0.5 | Adaptation Ratio 1.0 | |||
| ExEnt | ExEnt MV | ExEnt-FT | TALC | ExEnt-FT | TALC | |
| banknote-authentication | 46.9 | 48.4 | 45.1(0.5) | 49.5(0.0) | 45.2(0.2) | 49.7(0.3) |
| tic-tac-toe-endgame | 32.8 | 32.3 | 32.3(0.0) | 41.1(0.0) | 32.3(0.0) | 41.1(0.0) |
| car-evaluation | 10.7 | 17.6 | 4.6(2.3) | 14.1(3.4) | 3.8(0.4) | 16.5(0.0) |
| contraceptive-choice | 42.7 | 43.7 | 43.4(0.0) | 44.0(0.7) | 43.4(0.0) | 43.8(0.3) |
| indian-liver-patient | 48.7 | 40.0 | 54.5(7.8) | 44.3(2.8) | 62.0(4.6) | 47.8(0.0) |
| travel-insurance | 31.9 | 33.9 | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) |
| Average | 35.6 | 36.0 | 35.7 | 37.9 | 36.8 | 38.9 |
| Tasks | Non-Adaptation Baselines | Adaptation Methods | ||
| ExEnt-A | ExEnt MV-A | ExEnt-FT-A | TALC-A | |
| banknote-authentication | 8.0 | 36.7 | 27.1(2.5) | 53.9(0.1) |
| tic-tac-toe-endgame | 2.6 | 30.2 | 32.3(0.1) | 32.4(0.4) |
| car-evaluation | 2.3 | 2.6 | 14.8(1.8) | 13.3(0.8) |
| contraceptive-choice | 22.7 | 32.8 | 19.3(2.7) | 31.6(1.3) |
| indian-liver-patient | 36.5 | 34.7 | 30.1(0.2) | 40.2(0.9) |
| travel-insurance | 15.3 | 24.6 | 15.2(2.6) | 23.4(0.0) |
| Average | 14.6 | 26.9 | 23.1 | 32.5 |
| Tasks | T0(3B) | OPT(2.7B) | FLAN-T5-XXL | ||||||
| Concat | MV | TALC | Concat | MV | TALC | Concat | MV | TALC | |
| banknote-authentication | 44.7 | 44.7 | 44.7(0.0) | 44.7 | 45.1 | 44.7(0.0) | 54.2 | 31.3 | 43.6(0.0) |
| tic-tac-toe-endgame | 67.7 | 67.7 | 67.7(0.0) | 32.3 | 49.0 | 64.9(1.3) | 32.3 | 40.6 | 56.8(0.0) |
| car-evaluation | 3.8 | 71.7 | 71.7(0.0) | 2.3 | 3.8 | 3.8(0.0) | 3.8 | 26.6 | 26.6(0.0) |
| contraceptive-choice | 23.4 | 23.9 | 23.4(0.0) | 27.5 | 33.2 | 33.2(0.0) | 38.3 | 45.4 | 45.4(0.0) |
| indian-liver-patient | 67.8 | 33.0 | 48.7(0.0) | 32.2 | 32.2 | 32.2(0.0) | 66.1 | 48.7 | 56.5(0.4) |
| travel-insurance | 38.9 | 65.8 | 65.8(0.0) | 42.5 | 65.8 | 65.8(0.0) | 71.9 | 71.9 | 72.6(0.4) |
| Average | 41.1 | 51.1 | 53.7 | 30.2 | 38.2 | 40.8 | 44.4 | 44.1 | 50.3 |
| Models | Prompt |
| T0FLAN-T5 | Explanations: <explanation_1>. <explanation_2>. . . . <explanation_n> Note Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... From the explanations, is the note fake or original? Answer: |
| OPT | The following is a classification task that uses the following explanations. Based on the explanation classify the subsequent sample: \nExplanations: \n- <explanation_1> \n- <explanation_2> \n: \n- <explanation_n> \nNote Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... \nFrom the explanations, is the note fake or original? \nAnswer: |
| TASK | LEARNING RATE | EPOCH |
| banknote-authentication | 4.38e-08 | 500 |
| tic-tac-toe-endgame | 7.31e-08 | 40 |
| car-evaluation | 2.48e-04 | 500 |
| contraceptive-method-choice | 5.53e-03 | 100 |
| indian-liver-patient | 7.91e-03 | 50 |
| travel-insurance | 7.31e-08 | 40 |
| Dataset | Quantifier | No Quantifier | Conjunction | No Conjunction |
| banknote | 0.18 | 0.14 | - | 0.14 |
| authentication | ||||
| tic-tac-toe | 0.20 | 0.12 | 0.29 | 0.16 |
| endgame | ||||
| car evaluation | 0.01 | 0.18 | 0.15 | 0.18 |
| contraceptive choice | 0.15 | 0.21 | 0.14 | 0.22 |
| indian-liver patient | - | 0.51 | 0.33 | 0.57 |
| travel insurance | 0.18 | 0.17 | 0.16 | 0.18 |
| Average | 0.14 | 0.22 | 0.21 | 0.24 |
| Dataset | k=4 | k=8 | k=16 | k=32 |
| banknote authentication | 47.5(2.7) | 51.2(1.5) | 54.9(5.2) | 53.6(4.1) |
| tic-tac-toe endgame | 55.6(8.1) | 59.1(6.5) | 63.7(2.8) | 64.2(1.8) |
| car evaluation | 61.7(6.2) | 52.6(1.6) | 68.1(1.4) | 69.4(2.3) |
| contraceptive choice | 34.4(4.3) | 32.9(4.4) | 36.8(0.2) | 40.0(2.3) |
| indian-liver patient | 66.9(2.6) | 67.4(1.3) | 63.8(3.4) | 67.5(0.8) |
| travel insurance | 53.9(10.3) | 57.0(7.1) | 53.9(1.7) | 62.5(3.7) |
| Average | 53.3 | 53.4 | 56.9 | 59.5 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If the variance of the note is a negative number, it's more likely to be an original note. | 56.7% | 25.05 |
| 1 | If the kurtosis of the bill is a negative number, it's more likely to be a fake note. | 41.1% | 31.82 |
| 2 | A fake banknote has a variance level over 4.0. | 46.2% | 177.85 |
| 3 | Notes that have a negative Variance level will also have a negative Entropy level. | 45.1% | 39.02 |
| 4 | Notes that have a positive Skewness level will have a negative Kurtosis level. | 45.8% | 61.06 |
| 5 | Variance above 4 leads to the results of FAKE. | 43.3% | 160.22 |
| 6 | Below 3.80 skewness leads to the original. | 56.0% | 548.93 |
| 7 | Variance above 1.00 leads to the FAKE. | 53.1% | 105.15 |
| 8 | Below the 3.75 Sketwness leads to the ORIGINAL. | 54.2% | 410.40 |
| 9 | Entropy is low value so it is fake. | 49.1% | 160.25 |
| 10 | Kurtosis is high value so it is original. | 54.9% | 346.58 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The game is usually not over yet when there are at least two blank squares. | 34.4% | 56.82 |
| 1 | The game is usually over when the players have taken all the middle squares. | 42.7% | 51.84 |
| 2 | The top middle square results in x winning. | 33.3% | 800.38 |
| 3 | The bottom middle square results in x winning. | 32.3% | 1132.84 |
| 4 | When the middle-right-square is left blank, x is less likely to win. | 32.3% | 94.01 |
| 5 | Whoever marks the top-left-square is unlikely to win. | 38.5% | 78.83 |
| 6 | The player who moves second is much more likely to lose if they place in the middle-left-square. | 37.0% | 45.58 |
| 7 | The middle-right-square is the square that is most likely to go unused during a game. | 35.9% | 36.20 |
| 8 | Top left square X indicates the Negative group. | 61.5% | 969.94 |
| 9 | Without b categories in middle middle square comes under the Positive group. | 32.3% | 2158.20 |
| 10 | An O in both the top-left and bottom-right is likely to be positive. | 39.6% | 24.43 |
| 11 | A blank in the middle-right will likely lead to negative. | 65.1% | 139.76 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If safety is high, then the car will not be unacceptable. | 22.3% | 56.95 |
| 1 | If maintenance cost is medium, then the car will not be unacceptable. | 42.8% | 121.90 |
| 2 | A capacity for 4 or more persons makes the vehicle acceptable for resale. | 9.0% | 115.36 |
| 3 | High safety ratings generally make vehicles acceptable for resale. | 9.8% | 158.10 |
| 4 | A low buying cost is a good indicator of a vehicle being acceptable for resale. | 8.4% | 55.96 |
| 5 | Most people having the passenger capacity of 4 or more have good acceptability for car resale. | 18.2% | 117.22 |
| 6 | Cars with low buying and maintenance cost are highly acceptable for resale | 5.5% | 436.58 |
| 7 | Cars with higher safety and capacity are highly acceptable for resale. | 12.4% | 156.38 |
| 8 | Cars with higher safety and medium luggage boot size are highly acceptable for resale. | 4.0% | 323.92 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | A husband's education has a high chance of a long-term contraceptive being used. | 44.7% | 85.88 |
| 1 | A wife's education level usually determines if a long term contraceptive was used. | 46.8% | 166.37 |
| 2 | Women with low education are more likely to use short-term contraception. | 39.3% | 12.18 |
| 3 | Women with high education age 40 and under are more likely to use long-term contraception. | 43.4% | 22.01 |
| 4 | Women with two or more children used short and long term methods. | 42.7% | 47.27 |
| 5 | The least educated women either used short-term method or didn't use any contraceptive method. | 24.1% | 72.05 |
| 6 | If the wife's education is high, then the contraceptive method used is long-term. | 41.0% | 44.04 |
| 7 | If the wife's education is not high, then the contraceptive method used is no-use or short-term. | 27.5% | 41.12 |
| 8 | If the wife's education is not high, then the contraceptive method used is not long-term. | 23.4% | 31.23 |
| 9 | If the wife's education is not high, then the contraceptive method is short-term. | 38.6% | 37.03 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The SGPT High percentage so the liver patient was yes. | 49.6% | 4184.78 |
| 1 | The SGPT Low percentage so the liver patient was no. | 52.2% | 4891.02 |
| 2 | Patients over the age of forty are liver patients. | 42.6% | 54.92 |
| 3 | Patients who has SGOT range greater than forty are liver pa-tients. | 42.6% | 373.03 |
| 4 | Decreased SGPT Value ensures no liver patient. | 52.2% | 4404.52 |
| 5 | Age group above 40 ensures liver patient. | 37.4% | 4994.60 |
| 6 | Some people have more age and have the SGOT and they are liver patient. | 54.8% | 272.55 |
| 7 | Some people have minimum age and they are liver patient. They are somewhat accurate. | 50.4% | 226.11 |
| No. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | Frequent flyers with an annual income over 1 million usually take travel insurance. | 33.2% | 104.36 |
| 1 | People with an annual income over 1 million and under 30 years old usually take travel insurance. | 30.4% | 53.40 |
| 2 | Travelers who are 29 and older take traveler insurance. | 34.4% | 105.36 |
| 3 | Travelers who have not traveled abroad before are more likely to take traveler insurance. | 34.2% | 22.74 |
| 4 | Frequent flyer travelers with an annual income above 1 million frequently take travel insurance. | 31.2% | 121.70 |
| 5 | Travelers older than 25 years old and with an income below 1 million do not usually take travel insurance. | 42.5% | 27.45 |
| 6 | Most passengers who are not frequent fliers do not use travel insurance. | 65.8% 33.28 | |
| 7 | Most passengers who have not traveled abroad do not use travel insurance. | 77.1% | 42.87 |
| 8 | Most passenger with an income higher than 100,000 use travel insurance. | 27.9% | 131.69 |
| 9 | People who have never traveled abroad before are more likely to have taken travel insurance. | 34.2% | 15.95 |
| 10 | People with an annual income below 1,000,000 are less likely to have traveled abroad than those with annual incomes above 1,000,000. | 23.1% | 8.36 |
| 11 | People with an annual income above 1,000,000 are more likely to have taken travel insurance. | 39.9% | 26.63 |
| 12 | More frequent flyers have taken travel insurance. | 34.9% | 334.32 |
| 13 | More people that travel abroad have taken travel insurance. | 34.2% | 114.91 |
| 14 | People who are non-frequent flyers and are college graduates are less likely to get travel insurance. | 27.4% | 22.28 |
| 15 | People who make a million or more and are frequent fliers are more likely to get travel insurance. | 32.4% | 22.73 |
| 16 | Most people who didn't travel abroad before had taken Travel Insurance. | 34.2% | 84.50 |
| 17 | Most college graduate have taken Travel Insurance. | 64.8% | 780.98 |
| 18 | Those who are college graduates and in their 20s are somewhat likely to purchase travel insurance. | 22.9% | 28.62 |
| 19 | Those who have never travelled abroad and are not frequent flyers often do not purchase travel insurance. | 67.1% | 23.85 |
| 20 | Travelled Abroad Before "No" categories indicates the "No" Travel Insurance Taken. | 24.4% | 490.45 |
| 21 | College Graduate "Yes" categories leads to the "Yes" Travel Insurance Taken. | 30.9% | 678.44 |
| 22 | Frequent Flyer "No" indicates the "No" Travel Insurance Taken. | 53.8% | 194.31 |
| 23 | Annual income categories above 1049999 indicates the "Yes" Travel Insurance Taken. | 43.5% | 2291.89 |
| 24 | Most college graduates that make more than 1000000 annually have taken travel insurance. | 37.2% | 231.38 |
| 25 | Most college graduates from 26 to 34 have taken travel insurance. | 43.2% | 185.20 |
| 26 | About have of college graduates that have not travelled abroad before have taken travel insurance. | 34.2% | 186.52 |
| 27 | Frequently flyer "No" categories indicates the "No" travel insurance taken group. | 65.6% | 586.77 |
| 28 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 37.4% | 769.28 |
| 29 | Frequent flyer "No" categories indicate the "No" travel insurance taken group. | 29.6% | 375.25 |
| 30 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 42.5% | 695.34 |
| Tasks | Non-Adaptation Baselines | Adaptation Ratio 0.5 | Adaptation Ratio 1.0 | |||
| ExEnt | ExEnt MV | ExEnt-FT | TALC | ExEnt-FT | TALC | |
| banknote-authentication | 46.9 | 48.4 | 45.1(0.5) | 49.5(0.0) | 45.2(0.2) | 49.7(0.3) |
| tic-tac-toe-endgame | 32.8 | 32.3 | 32.3(0.0) | 41.1(0.0) | 32.3(0.0) | 41.1(0.0) |
| car-evaluation | 10.7 | 17.6 | 4.6(2.3) | 14.1(3.4) | 3.8(0.4) | 16.5(0.0) |
| contraceptive-choice | 42.7 | 43.7 | 43.4(0.0) | 44.0(0.7) | 43.4(0.0) | 43.8(0.3) |
| indian-liver-patient | 48.7 | 40.0 | 54.5(7.8) | 44.3(2.8) | 62.0(4.6) | 47.8(0.0) |
| travel-insurance | 31.9 | 33.9 | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) |
| Average | 35.6 | 36.0 | 35.7 | 37.9 | 36.8 | 38.9 |
| Tasks | Non-Adaptation Baselines | Adaptation Methods | ||
| ExEnt-A | ExEnt MV-A | ExEnt-FT-A | TALC-A | |
| banknote-authentication | 8.0 | 36.7 | 27.1(2.5) | 53.9(0.1) |
| tic-tac-toe-endgame | 2.6 | 30.2 | 32.3(0.1) | 32.4(0.4) |
| car-evaluation | 2.3 | 2.6 | 14.8(1.8) | 13.3(0.8) |
| contraceptive-choice | 22.7 | 32.8 | 19.3(2.7) | 31.6(1.3) |
| indian-liver-patient | 36.5 | 34.7 | 30.1(0.2) | 40.2(0.9) |
| travel-insurance | 15.3 | 24.6 | 15.2(2.6) | 23.4(0.0) |
| Average | 14.6 | 26.9 | 23.1 | 32.5 |
| Tasks | T0(3B) | OPT(2.7B) | FLAN-T5-XXL | ||||||
| Concat | MV | TALC | Concat | MV | TALC | Concat | MV | TALC | |
| banknote-authentication | 44.7 | 44.7 | 44.7(0.0) | 44.7 | 45.1 | 44.7(0.0) | 54.2 | 31.3 | 43.6(0.0) |
| tic-tac-toe-endgame | 67.7 | 67.7 | 67.7(0.0) | 32.3 | 49.0 | 64.9(1.3) | 32.3 | 40.6 | 56.8(0.0) |
| car-evaluation | 3.8 | 71.7 | 71.7(0.0) | 2.3 | 3.8 | 3.8(0.0) | 3.8 | 26.6 | 26.6(0.0) |
| contraceptive-choice | 23.4 | 23.9 | 23.4(0.0) | 27.5 | 33.2 | 33.2(0.0) | 38.3 | 45.4 | 45.4(0.0) |
| indian-liver-patient | 67.8 | 33.0 | 48.7(0.0) | 32.2 | 32.2 | 32.2(0.0) | 66.1 | 48.7 | 56.5(0.4) |
| travel-insurance | 38.9 | 65.8 | 65.8(0.0) | 42.5 | 65.8 | 65.8(0.0) | 71.9 | 71.9 | 72.6(0.4) |
| Average | 41.1 | 51.1 | 53.7 | 30.2 | 38.2 | 40.8 | 44.4 | 44.1 | 50.3 |
| Models | Prompt |
| T0FLAN-T5 | Explanations: <explanation_1>. <explanation_2>. . . . <explanation_n> Note Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... From the explanations, is the note fake or original? Answer: |
| OPT | The following is a classification task that uses the following explanations. Based on the explanation classify the subsequent sample: \nExplanations: \n- <explanation_1> \n- <explanation_2> \n: \n- <explanation_n> \nNote Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... \nFrom the explanations, is the note fake or original? \nAnswer: |
| TASK | LEARNING RATE | EPOCH |
| banknote-authentication | 4.38e-08 | 500 |
| tic-tac-toe-endgame | 7.31e-08 | 40 |
| car-evaluation | 2.48e-04 | 500 |
| contraceptive-method-choice | 5.53e-03 | 100 |
| indian-liver-patient | 7.91e-03 | 50 |
| travel-insurance | 7.31e-08 | 40 |
| Dataset | Quantifier | No Quantifier | Conjunction | No Conjunction |
| banknote | 0.18 | 0.14 | - | 0.14 |
| authentication | ||||
| tic-tac-toe | 0.20 | 0.12 | 0.29 | 0.16 |
| endgame | ||||
| car evaluation | 0.01 | 0.18 | 0.15 | 0.18 |
| contraceptive choice | 0.15 | 0.21 | 0.14 | 0.22 |
| indian-liver patient | - | 0.51 | 0.33 | 0.57 |
| travel insurance | 0.18 | 0.17 | 0.16 | 0.18 |
| Average | 0.14 | 0.22 | 0.21 | 0.24 |
| Dataset | k=4 | k=8 | k=16 | k=32 |
| banknote authentication | 47.5(2.7) | 51.2(1.5) | 54.9(5.2) | 53.6(4.1) |
| tic-tac-toe endgame | 55.6(8.1) | 59.1(6.5) | 63.7(2.8) | 64.2(1.8) |
| car evaluation | 61.7(6.2) | 52.6(1.6) | 68.1(1.4) | 69.4(2.3) |
| contraceptive choice | 34.4(4.3) | 32.9(4.4) | 36.8(0.2) | 40.0(2.3) |
| indian-liver patient | 66.9(2.6) | 67.4(1.3) | 63.8(3.4) | 67.5(0.8) |
| travel insurance | 53.9(10.3) | 57.0(7.1) | 53.9(1.7) | 62.5(3.7) |
| Average | 53.3 | 53.4 | 56.9 | 59.5 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If the variance of the note is a negative number, it's more likely to be an original note. | 56.7% | 25.05 |
| 1 | If the kurtosis of the bill is a negative number, it's more likely to be a fake note. | 41.1% | 31.82 |
| 2 | A fake banknote has a variance level over 4.0. | 46.2% | 177.85 |
| 3 | Notes that have a negative Variance level will also have a negative Entropy level. | 45.1% | 39.02 |
| 4 | Notes that have a positive Skewness level will have a negative Kurtosis level. | 45.8% | 61.06 |
| 5 | Variance above 4 leads to the results of FAKE. | 43.3% | 160.22 |
| 6 | Below 3.80 skewness leads to the original. | 56.0% | 548.93 |
| 7 | Variance above 1.00 leads to the FAKE. | 53.1% | 105.15 |
| 8 | Below the 3.75 Sketwness leads to the ORIGINAL. | 54.2% | 410.40 |
| 9 | Entropy is low value so it is fake. | 49.1% | 160.25 |
| 10 | Kurtosis is high value so it is original. | 54.9% | 346.58 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The game is usually not over yet when there are at least two blank squares. | 34.4% | 56.82 |
| 1 | The game is usually over when the players have taken all the middle squares. | 42.7% | 51.84 |
| 2 | The top middle square results in x winning. | 33.3% | 800.38 |
| 3 | The bottom middle square results in x winning. | 32.3% | 1132.84 |
| 4 | When the middle-right-square is left blank, x is less likely to win. | 32.3% | 94.01 |
| 5 | Whoever marks the top-left-square is unlikely to win. | 38.5% | 78.83 |
| 6 | The player who moves second is much more likely to lose if they place in the middle-left-square. | 37.0% | 45.58 |
| 7 | The middle-right-square is the square that is most likely to go unused during a game. | 35.9% | 36.20 |
| 8 | Top left square X indicates the Negative group. | 61.5% | 969.94 |
| 9 | Without b categories in middle middle square comes under the Positive group. | 32.3% | 2158.20 |
| 10 | An O in both the top-left and bottom-right is likely to be positive. | 39.6% | 24.43 |
| 11 | A blank in the middle-right will likely lead to negative. | 65.1% | 139.76 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If safety is high, then the car will not be unacceptable. | 22.3% | 56.95 |
| 1 | If maintenance cost is medium, then the car will not be unacceptable. | 42.8% | 121.90 |
| 2 | A capacity for 4 or more persons makes the vehicle acceptable for resale. | 9.0% | 115.36 |
| 3 | High safety ratings generally make vehicles acceptable for resale. | 9.8% | 158.10 |
| 4 | A low buying cost is a good indicator of a vehicle being acceptable for resale. | 8.4% | 55.96 |
| 5 | Most people having the passenger capacity of 4 or more have good acceptability for car resale. | 18.2% | 117.22 |
| 6 | Cars with low buying and maintenance cost are highly acceptable for resale | 5.5% | 436.58 |
| 7 | Cars with higher safety and capacity are highly acceptable for resale. | 12.4% | 156.38 |
| 8 | Cars with higher safety and medium luggage boot size are highly acceptable for resale. | 4.0% | 323.92 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | A husband's education has a high chance of a long-term contraceptive being used. | 44.7% | 85.88 |
| 1 | A wife's education level usually determines if a long term contraceptive was used. | 46.8% | 166.37 |
| 2 | Women with low education are more likely to use short-term contraception. | 39.3% | 12.18 |
| 3 | Women with high education age 40 and under are more likely to use long-term contraception. | 43.4% | 22.01 |
| 4 | Women with two or more children used short and long term methods. | 42.7% | 47.27 |
| 5 | The least educated women either used short-term method or didn't use any contraceptive method. | 24.1% | 72.05 |
| 6 | If the wife's education is high, then the contraceptive method used is long-term. | 41.0% | 44.04 |
| 7 | If the wife's education is not high, then the contraceptive method used is no-use or short-term. | 27.5% | 41.12 |
| 8 | If the wife's education is not high, then the contraceptive method used is not long-term. | 23.4% | 31.23 |
| 9 | If the wife's education is not high, then the contraceptive method is short-term. | 38.6% | 37.03 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The SGPT High percentage so the liver patient was yes. | 49.6% | 4184.78 |
| 1 | The SGPT Low percentage so the liver patient was no. | 52.2% | 4891.02 |
| 2 | Patients over the age of forty are liver patients. | 42.6% | 54.92 |
| 3 | Patients who has SGOT range greater than forty are liver pa-tients. | 42.6% | 373.03 |
| 4 | Decreased SGPT Value ensures no liver patient. | 52.2% | 4404.52 |
| 5 | Age group above 40 ensures liver patient. | 37.4% | 4994.60 |
| 6 | Some people have more age and have the SGOT and they are liver patient. | 54.8% | 272.55 |
| 7 | Some people have minimum age and they are liver patient. They are somewhat accurate. | 50.4% | 226.11 |
| No. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | Frequent flyers with an annual income over 1 million usually take travel insurance. | 33.2% | 104.36 |
| 1 | People with an annual income over 1 million and under 30 years old usually take travel insurance. | 30.4% | 53.40 |
| 2 | Travelers who are 29 and older take traveler insurance. | 34.4% | 105.36 |
| 3 | Travelers who have not traveled abroad before are more likely to take traveler insurance. | 34.2% | 22.74 |
| 4 | Frequent flyer travelers with an annual income above 1 million frequently take travel insurance. | 31.2% | 121.70 |
| 5 | Travelers older than 25 years old and with an income below 1 million do not usually take travel insurance. | 42.5% | 27.45 |
| 6 | Most passengers who are not frequent fliers do not use travel insurance. | 65.8% 33.28 | |
| 7 | Most passengers who have not traveled abroad do not use travel insurance. | 77.1% | 42.87 |
| 8 | Most passenger with an income higher than 100,000 use travel insurance. | 27.9% | 131.69 |
| 9 | People who have never traveled abroad before are more likely to have taken travel insurance. | 34.2% | 15.95 |
| 10 | People with an annual income below 1,000,000 are less likely to have traveled abroad than those with annual incomes above 1,000,000. | 23.1% | 8.36 |
| 11 | People with an annual income above 1,000,000 are more likely to have taken travel insurance. | 39.9% | 26.63 |
| 12 | More frequent flyers have taken travel insurance. | 34.9% | 334.32 |
| 13 | More people that travel abroad have taken travel insurance. | 34.2% | 114.91 |
| 14 | People who are non-frequent flyers and are college graduates are less likely to get travel insurance. | 27.4% | 22.28 |
| 15 | People who make a million or more and are frequent fliers are more likely to get travel insurance. | 32.4% | 22.73 |
| 16 | Most people who didn't travel abroad before had taken Travel Insurance. | 34.2% | 84.50 |
| 17 | Most college graduate have taken Travel Insurance. | 64.8% | 780.98 |
| 18 | Those who are college graduates and in their 20s are somewhat likely to purchase travel insurance. | 22.9% | 28.62 |
| 19 | Those who have never travelled abroad and are not frequent flyers often do not purchase travel insurance. | 67.1% | 23.85 |
| 20 | Travelled Abroad Before "No" categories indicates the "No" Travel Insurance Taken. | 24.4% | 490.45 |
| 21 | College Graduate "Yes" categories leads to the "Yes" Travel Insurance Taken. | 30.9% | 678.44 |
| 22 | Frequent Flyer "No" indicates the "No" Travel Insurance Taken. | 53.8% | 194.31 |
| 23 | Annual income categories above 1049999 indicates the "Yes" Travel Insurance Taken. | 43.5% | 2291.89 |
| 24 | Most college graduates that make more than 1000000 annually have taken travel insurance. | 37.2% | 231.38 |
| 25 | Most college graduates from 26 to 34 have taken travel insurance. | 43.2% | 185.20 |
| 26 | About have of college graduates that have not travelled abroad before have taken travel insurance. | 34.2% | 186.52 |
| 27 | Frequently flyer "No" categories indicates the "No" travel insurance taken group. | 65.6% | 586.77 |
| 28 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 37.4% | 769.28 |
| 29 | Frequent flyer "No" categories indicate the "No" travel insurance taken group. | 29.6% | 375.25 |
| 30 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 42.5% | 695.34 |
| Tasks | Non-Adaptation Baselines | Adaptation Ratio 0.5 | Adaptation Ratio 1.0 | |||
| ExEnt | ExEnt MV | ExEnt-FT | TALC | ExEnt-FT | TALC | |
| banknote-authentication | 46.9 | 48.4 | 45.1(0.5) | 49.5(0.0) | 45.2(0.2) | 49.7(0.3) |
| tic-tac-toe-endgame | 32.8 | 32.3 | 32.3(0.0) | 41.1(0.0) | 32.3(0.0) | 41.1(0.0) |
| car-evaluation | 10.7 | 17.6 | 4.6(2.3) | 14.1(3.4) | 3.8(0.4) | 16.5(0.0) |
| contraceptive-choice | 42.7 | 43.7 | 43.4(0.0) | 44.0(0.7) | 43.4(0.0) | 43.8(0.3) |
| indian-liver-patient | 48.7 | 40.0 | 54.5(7.8) | 44.3(2.8) | 62.0(4.6) | 47.8(0.0) |
| travel-insurance | 31.9 | 33.9 | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) |
| Average | 35.6 | 36.0 | 35.7 | 37.9 | 36.8 | 38.9 |
| Tasks | Non-Adaptation Baselines | Adaptation Methods | ||
| ExEnt-A | ExEnt MV-A | ExEnt-FT-A | TALC-A | |
| banknote-authentication | 8.0 | 36.7 | 27.1(2.5) | 53.9(0.1) |
| tic-tac-toe-endgame | 2.6 | 30.2 | 32.3(0.1) | 32.4(0.4) |
| car-evaluation | 2.3 | 2.6 | 14.8(1.8) | 13.3(0.8) |
| contraceptive-choice | 22.7 | 32.8 | 19.3(2.7) | 31.6(1.3) |
| indian-liver-patient | 36.5 | 34.7 | 30.1(0.2) | 40.2(0.9) |
| travel-insurance | 15.3 | 24.6 | 15.2(2.6) | 23.4(0.0) |
| Average | 14.6 | 26.9 | 23.1 | 32.5 |
| Tasks | T0(3B) | OPT(2.7B) | FLAN-T5-XXL | ||||||
| Concat | MV | TALC | Concat | MV | TALC | Concat | MV | TALC | |
| banknote-authentication | 44.7 | 44.7 | 44.7(0.0) | 44.7 | 45.1 | 44.7(0.0) | 54.2 | 31.3 | 43.6(0.0) |
| tic-tac-toe-endgame | 67.7 | 67.7 | 67.7(0.0) | 32.3 | 49.0 | 64.9(1.3) | 32.3 | 40.6 | 56.8(0.0) |
| car-evaluation | 3.8 | 71.7 | 71.7(0.0) | 2.3 | 3.8 | 3.8(0.0) | 3.8 | 26.6 | 26.6(0.0) |
| contraceptive-choice | 23.4 | 23.9 | 23.4(0.0) | 27.5 | 33.2 | 33.2(0.0) | 38.3 | 45.4 | 45.4(0.0) |
| indian-liver-patient | 67.8 | 33.0 | 48.7(0.0) | 32.2 | 32.2 | 32.2(0.0) | 66.1 | 48.7 | 56.5(0.4) |
| travel-insurance | 38.9 | 65.8 | 65.8(0.0) | 42.5 | 65.8 | 65.8(0.0) | 71.9 | 71.9 | 72.6(0.4) |
| Average | 41.1 | 51.1 | 53.7 | 30.2 | 38.2 | 40.8 | 44.4 | 44.1 | 50.3 |
| Models | Prompt |
| T0FLAN-T5 | Explanations: <explanation_1>. <explanation_2>. . . . <explanation_n> Note Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... From the explanations, is the note fake or original? Answer: |
| OPT | The following is a classification task that uses the following explanations. Based on the explanation classify the subsequent sample: +Explanations: +- <explanation_1> +- <explanation_2> +: +- <explanation_n> +Note Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... +From the explanations, is the note fake or original? +Answer: |
| TASK | LEARNING RATE | EPOCH |
| banknote-authentication | 4.38e-08 | 500 |
| tic-tac-toe-endgame | 7.31e-08 | 40 |
| car-evaluation | 2.48e-04 | 500 |
| contraceptive-method-choice | 5.53e-03 | 100 |
| indian-liver-patient | 7.91e-03 | 50 |
| travel-insurance | 7.31e-08 | 40 |
| Dataset | Quantifier | No Quantifier | Conjunction | No Conjunction |
| banknote | 0.18 | 0.14 | - | 0.14 |
| authentication | ||||
| tic-tac-toe | 0.20 | 0.12 | 0.29 | 0.16 |
| endgame | ||||
| car evaluation | 0.01 | 0.18 | 0.15 | 0.18 |
| contraceptive choice | 0.15 | 0.21 | 0.14 | 0.22 |
| indian-liver patient | - | 0.51 | 0.33 | 0.57 |
| travel insurance | 0.18 | 0.17 | 0.16 | 0.18 |
| Average | 0.14 | 0.22 | 0.21 | 0.24 |
| Dataset | k=4 | k=8 | k=16 | k=32 |
| banknote authentication | 47.5(2.7) | 51.2(1.5) | 54.9(5.2) | 53.6(4.1) |
| tic-tac-toe endgame | 55.6(8.1) | 59.1(6.5) | 63.7(2.8) | 64.2(1.8) |
| car evaluation | 61.7(6.2) | 52.6(1.6) | 68.1(1.4) | 69.4(2.3) |
| contraceptive choice | 34.4(4.3) | 32.9(4.4) | 36.8(0.2) | 40.0(2.3) |
| indian-liver patient | 66.9(2.6) | 67.4(1.3) | 63.8(3.4) | 67.5(0.8) |
| travel insurance | 53.9(10.3) | 57.0(7.1) | 53.9(1.7) | 62.5(3.7) |
| Average | 53.3 | 53.4 | 56.9 | 59.5 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If the variance of the note is a negative number, it's more likely to be an original note. | 56.7% | 25.05 |
| 1 | If the kurtosis of the bill is a negative number, it's more likely to be a fake note. | 41.1% | 31.82 |
| 2 | A fake banknote has a variance level over 4.0. | 46.2% | 177.85 |
| 3 | Notes that have a negative Variance level will also have a negative Entropy level. | 45.1% | 39.02 |
| 4 | Notes that have a positive Skewness level will have a negative Kurtosis level. | 45.8% | 61.06 |
| 5 | Variance above 4 leads to the results of FAKE. | 43.3% | 160.22 |
| 6 | Below 3.80 skewness leads to the original. | 56.0% | 548.93 |
| 7 | Variance above 1.00 leads to the FAKE. | 53.1% | 105.15 |
| 8 | Below the 3.75 Sketwness leads to the ORIGINAL. | 54.2% | 410.40 |
| 9 | Entropy is low value so it is fake. | 49.1% | 160.25 |
| 10 | Kurtosis is high value so it is original. | 54.9% | 346.58 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The game is usually not over yet when there are at least two blank squares. | 34.4% | 56.82 |
| 1 | The game is usually over when the players have taken all the middle squares. | 42.7% | 51.84 |
| 2 | The top middle square results in x winning. | 33.3% | 800.38 |
| 3 | The bottom middle square results in x winning. | 32.3% | 1132.84 |
| 4 | When the middle-right-square is left blank, x is less likely to win. | 32.3% | 94.01 |
| 5 | Whoever marks the top-left-square is unlikely to win. | 38.5% | 78.83 |
| 6 | The player who moves second is much more likely to lose if they place in the middle-left-square. | 37.0% | 45.58 |
| 7 | The middle-right-square is the square that is most likely to go unused during a game. | 35.9% | 36.20 |
| 8 | Top left square X indicates the Negative group. | 61.5% | 969.94 |
| 9 | Without b categories in middle middle square comes under the Positive group. | 32.3% | 2158.20 |
| 10 | An O in both the top-left and bottom-right is likely to be positive. | 39.6% | 24.43 |
| 11 | A blank in the middle-right will likely lead to negative. | 65.1% | 139.76 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If safety is high, then the car will not be unacceptable. | 22.3% | 56.95 |
| 1 | If maintenance cost is medium, then the car will not be unacceptable. | 42.8% | 121.90 |
| 2 | A capacity for 4 or more persons makes the vehicle acceptable for resale. | 9.0% | 115.36 |
| 3 | High safety ratings generally make vehicles acceptable for resale. | 9.8% | 158.10 |
| 4 | A low buying cost is a good indicator of a vehicle being acceptable for resale. | 8.4% | 55.96 |
| 5 | Most people having the passenger capacity of 4 or more have good acceptability for car resale. | 18.2% | 117.22 |
| 6 | Cars with low buying and maintenance cost are highly acceptable for resale | 5.5% | 436.58 |
| 7 | Cars with higher safety and capacity are highly acceptable for resale. | 12.4% | 156.38 |
| 8 | Cars with higher safety and medium luggage boot size are highly acceptable for resale. | 4.0% | 323.92 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | A husband's education has a high chance of a long-term contraceptive being used. | 44.7% | 85.88 |
| 1 | A wife's education level usually determines if a long term contraceptive was used. | 46.8% | 166.37 |
| 2 | Women with low education are more likely to use short-term contraception. | 39.3% | 12.18 |
| 3 | Women with high education age 40 and under are more likely to use long-term contraception. | 43.4% | 22.01 |
| 4 | Women with two or more children used short and long term methods. | 42.7% | 47.27 |
| 5 | The least educated women either used short-term method or didn't use any contraceptive method. | 24.1% | 72.05 |
| 6 | If the wife's education is high, then the contraceptive method used is long-term. | 41.0% | 44.04 |
| 7 | If the wife's education is not high, then the contraceptive method used is no-use or short-term. | 27.5% | 41.12 |
| 8 | If the wife's education is not high, then the contraceptive method used is not long-term. | 23.4% | 31.23 |
| 9 | If the wife's education is not high, then the contraceptive method is short-term. | 38.6% | 37.03 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The SGPT High percentage so the liver patient was yes. | 49.6% | 4184.78 |
| 1 | The SGPT Low percentage so the liver patient was no. | 52.2% | 4891.02 |
| 2 | Patients over the age of forty are liver patients. | 42.6% | 54.92 |
| 3 | Patients who has SGOT range greater than forty are liver pa-tients. | 42.6% | 373.03 |
| 4 | Decreased SGPT Value ensures no liver patient. | 52.2% | 4404.52 |
| 5 | Age group above 40 ensures liver patient. | 37.4% | 4994.60 |
| 6 | Some people have more age and have the SGOT and they are liver patient. | 54.8% | 272.55 |
| 7 | Some people have minimum age and they are liver patient. They are somewhat accurate. | 50.4% | 226.11 |
| No. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | Frequent flyers with an annual income over 1 million usually take travel insurance. | 33.2% | 104.36 |
| 1 | People with an annual income over 1 million and under 30 years old usually take travel insurance. | 30.4% | 53.40 |
| 2 | Travelers who are 29 and older take traveler insurance. | 34.4% | 105.36 |
| 3 | Travelers who have not traveled abroad before are more likely to take traveler insurance. | 34.2% | 22.74 |
| 4 | Frequent flyer travelers with an annual income above 1 million frequently take travel insurance. | 31.2% | 121.70 |
| 5 | Travelers older than 25 years old and with an income below 1 million do not usually take travel insurance. | 42.5% | 27.45 |
| 6 | Most passengers who are not frequent fliers do not use travel insurance. | 65.8% 33.28 | |
| 7 | Most passengers who have not traveled abroad do not use travel insurance. | 77.1% | 42.87 |
| 8 | Most passenger with an income higher than 100,000 use travel insurance. | 27.9% | 131.69 |
| 9 | People who have never traveled abroad before are more likely to have taken travel insurance. | 34.2% | 15.95 |
| 10 | People with an annual income below 1,000,000 are less likely to have traveled abroad than those with annual incomes above 1,000,000. | 23.1% | 8.36 |
| 11 | People with an annual income above 1,000,000 are more likely to have taken travel insurance. | 39.9% | 26.63 |
| 12 | More frequent flyers have taken travel insurance. | 34.9% | 334.32 |
| 13 | More people that travel abroad have taken travel insurance. | 34.2% | 114.91 |
| 14 | People who are non-frequent flyers and are college graduates are less likely to get travel insurance. | 27.4% | 22.28 |
| 15 | People who make a million or more and are frequent fliers are more likely to get travel insurance. | 32.4% | 22.73 |
| 16 | Most people who didn't travel abroad before had taken Travel Insurance. | 34.2% | 84.50 |
| 17 | Most college graduate have taken Travel Insurance. | 64.8% | 780.98 |
| 18 | Those who are college graduates and in their 20s are somewhat likely to purchase travel insurance. | 22.9% | 28.62 |
| 19 | Those who have never travelled abroad and are not frequent flyers often do not purchase travel insurance. | 67.1% | 23.85 |
| 20 | Travelled Abroad Before "No" categories indicates the "No" Travel Insurance Taken. | 24.4% | 490.45 |
| 21 | College Graduate "Yes" categories leads to the "Yes" Travel Insurance Taken. | 30.9% | 678.44 |
| 22 | Frequent Flyer "No" indicates the "No" Travel Insurance Taken. | 53.8% | 194.31 |
| 23 | Annual income categories above 1049999 indicates the "Yes" Travel Insurance Taken. | 43.5% | 2291.89 |
| 24 | Most college graduates that make more than 1000000 annually have taken travel insurance. | 37.2% | 231.38 |
| 25 | Most college graduates from 26 to 34 have taken travel insurance. | 43.2% | 185.20 |
| 26 | About have of college graduates that have not travelled abroad before have taken travel insurance. | 34.2% | 186.52 |
| 27 | Frequently flyer "No" categories indicates the "No" travel insurance taken group. | 65.6% | 586.77 |
| 28 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 37.4% | 769.28 |
| 29 | Frequent flyer "No" categories indicate the "No" travel insurance taken group. | 29.6% | 375.25 |
| 30 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 42.5% | 695.34 |
| Tasks | Non-Adaptation Baselines | Adaptation Ratio 0.5 | Adaptation Ratio 1.0 | |||
| ExEnt | ExEnt MV | ExEnt-FT | TALC | ExEnt-FT | TALC | |
| banknote-authentication | 46.9 | 48.4 | 45.1(0.5) | 49.5(0.0) | 45.2(0.2) | 49.7(0.3) |
| tic-tac-toe-endgame | 32.8 | 32.3 | 32.3(0.0) | 41.1(0.0) | 32.3(0.0) | 41.1(0.0) |
| car-evaluation | 10.7 | 17.6 | 4.6(2.3) | 14.1(3.4) | 3.8(0.4) | 16.5(0.0) |
| contraceptive-choice | 42.7 | 43.7 | 43.4(0.0) | 44.0(0.7) | 43.4(0.0) | 43.8(0.3) |
| indian-liver-patient | 48.7 | 40.0 | 54.5(7.8) | 44.3(2.8) | 62.0(4.6) | 47.8(0.0) |
| travel-insurance | 31.9 | 33.9 | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) | 34.2(0.0) |
| Average | 35.6 | 36.0 | 35.7 | 37.9 | 36.8 | 38.9 |
| Tasks | Non-Adaptation Baselines | Adaptation Methods | ||
| ExEnt-A | ExEnt MV-A | ExEnt-FT-A | TALC-A | |
| banknote-authentication | 8.0 | 36.7 | 27.1(2.5) | 53.9(0.1) |
| tic-tac-toe-endgame | 2.6 | 30.2 | 32.3(0.1) | 32.4(0.4) |
| car-evaluation | 2.3 | 2.6 | 14.8(1.8) | 13.3(0.8) |
| contraceptive-choice | 22.7 | 32.8 | 19.3(2.7) | 31.6(1.3) |
| indian-liver-patient | 36.5 | 34.7 | 30.1(0.2) | 40.2(0.9) |
| travel-insurance | 15.3 | 24.6 | 15.2(2.6) | 23.4(0.0) |
| Average | 14.6 | 26.9 | 23.1 | 32.5 |
| Tasks | T0(3B) | OPT(2.7B) | FLAN-T5-XXL | ||||||
| Concat | MV | TALC | Concat | MV | TALC | Concat | MV | TALC | |
| banknote-authentication | 44.7 | 44.7 | 44.7(0.0) | 44.7 | 45.1 | 44.7(0.0) | 54.2 | 31.3 | 43.6(0.0) |
| tic-tac-toe-endgame | 67.7 | 67.7 | 67.7(0.0) | 32.3 | 49.0 | 64.9(1.3) | 32.3 | 40.6 | 56.8(0.0) |
| car-evaluation | 3.8 | 71.7 | 71.7(0.0) | 2.3 | 3.8 | 3.8(0.0) | 3.8 | 26.6 | 26.6(0.0) |
| contraceptive-choice | 23.4 | 23.9 | 23.4(0.0) | 27.5 | 33.2 | 33.2(0.0) | 38.3 | 45.4 | 45.4(0.0) |
| indian-liver-patient | 67.8 | 33.0 | 48.7(0.0) | 32.2 | 32.2 | 32.2(0.0) | 66.1 | 48.7 | 56.5(0.4) |
| travel-insurance | 38.9 | 65.8 | 65.8(0.0) | 42.5 | 65.8 | 65.8(0.0) | 71.9 | 71.9 | 72.6(0.4) |
| Average | 41.1 | 51.1 | 53.7 | 30.2 | 38.2 | 40.8 | 44.4 | 44.1 | 50.3 |
| Models | Prompt |
| T0FLAN-T5 | Explanations: <explanation_1>. <explanation_2>. . . . <explanation_n> Note Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... From the explanations, is the note fake or original? Answer: |
| OPT | The following is a classification task that uses the following explanations. Based on the explanation classify the subsequent sample: \nExplanations: \n- <explanation_1> \n- <explanation_2> \n: \n- <explanation_n> \nNote Details: <feat_1> equal to <feat_1_value>. <feat_2> equal to <feat_2_value> ... \nFrom the explanations, is the note fake or original? \nAnswer: |
| TASK | LEARNING RATE | EPOCH |
| banknote-authentication | 4.38e-08 | 500 |
| tic-tac-toe-endgame | 7.31e-08 | 40 |
| car-evaluation | 2.48e-04 | 500 |
| contraceptive-method-choice | 5.53e-03 | 100 |
| indian-liver-patient | 7.91e-03 | 50 |
| travel-insurance | 7.31e-08 | 40 |
| Dataset | Quantifier | No Quantifier | Conjunction | No Conjunction |
| banknote | 0.18 | 0.14 | - | 0.14 |
| authentication | ||||
| tic-tac-toe | 0.20 | 0.12 | 0.29 | 0.16 |
| endgame | ||||
| car evaluation | 0.01 | 0.18 | 0.15 | 0.18 |
| contraceptive choice | 0.15 | 0.21 | 0.14 | 0.22 |
| indian-liver patient | - | 0.51 | 0.33 | 0.57 |
| travel insurance | 0.18 | 0.17 | 0.16 | 0.18 |
| Average | 0.14 | 0.22 | 0.21 | 0.24 |
| Dataset | k=4 | k=8 | k=16 | k=32 |
| banknote authentication | 47.5(2.7) | 51.2(1.5) | 54.9(5.2) | 53.6(4.1) |
| tic-tac-toe endgame | 55.6(8.1) | 59.1(6.5) | 63.7(2.8) | 64.2(1.8) |
| car evaluation | 61.7(6.2) | 52.6(1.6) | 68.1(1.4) | 69.4(2.3) |
| contraceptive choice | 34.4(4.3) | 32.9(4.4) | 36.8(0.2) | 40.0(2.3) |
| indian-liver patient | 66.9(2.6) | 67.4(1.3) | 63.8(3.4) | 67.5(0.8) |
| travel insurance | 53.9(10.3) | 57.0(7.1) | 53.9(1.7) | 62.5(3.7) |
| Average | 53.3 | 53.4 | 56.9 | 59.5 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If the variance of the note is a negative number, it's more likely to be an original note. | 56.7% | 25.05 |
| 1 | If the kurtosis of the bill is a negative number, it's more likely to be a fake note. | 41.1% | 31.82 |
| 2 | A fake banknote has a variance level over 4.0. | 46.2% | 177.85 |
| 3 | Notes that have a negative Variance level will also have a negative Entropy level. | 45.1% | 39.02 |
| 4 | Notes that have a positive Skewness level will have a negative Kurtosis level. | 45.8% | 61.06 |
| 5 | Variance above 4 leads to the results of FAKE. | 43.3% | 160.22 |
| 6 | Below 3.80 skewness leads to the original. | 56.0% | 548.93 |
| 7 | Variance above 1.00 leads to the FAKE. | 53.1% | 105.15 |
| 8 | Below the 3.75 Sketwness leads to the ORIGINAL. | 54.2% | 410.40 |
| 9 | Entropy is low value so it is fake. | 49.1% | 160.25 |
| 10 | Kurtosis is high value so it is original. | 54.9% | 346.58 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The game is usually not over yet when there are at least two blank squares. | 34.4% | 56.82 |
| 1 | The game is usually over when the players have taken all the middle squares. | 42.7% | 51.84 |
| 2 | The top middle square results in x winning. | 33.3% | 800.38 |
| 3 | The bottom middle square results in x winning. | 32.3% | 1132.84 |
| 4 | When the middle-right-square is left blank, x is less likely to win. | 32.3% | 94.01 |
| 5 | Whoever marks the top-left-square is unlikely to win. | 38.5% | 78.83 |
| 6 | The player who moves second is much more likely to lose if they place in the middle-left-square. | 37.0% | 45.58 |
| 7 | The middle-right-square is the square that is most likely to go unused during a game. | 35.9% | 36.20 |
| 8 | Top left square X indicates the Negative group. | 61.5% | 969.94 |
| 9 | Without b categories in middle middle square comes under the Positive group. | 32.3% | 2158.20 |
| 10 | An O in both the top-left and bottom-right is likely to be positive. | 39.6% | 24.43 |
| 11 | A blank in the middle-right will likely lead to negative. | 65.1% | 139.76 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | If safety is high, then the car will not be unacceptable. | 22.3% | 56.95 |
| 1 | If maintenance cost is medium, then the car will not be unacceptable. | 42.8% | 121.90 |
| 2 | A capacity for 4 or more persons makes the vehicle acceptable for resale. | 9.0% | 115.36 |
| 3 | High safety ratings generally make vehicles acceptable for resale. | 9.8% | 158.10 |
| 4 | A low buying cost is a good indicator of a vehicle being acceptable for resale. | 8.4% | 55.96 |
| 5 | Most people having the passenger capacity of 4 or more have good acceptability for car resale. | 18.2% | 117.22 |
| 6 | Cars with low buying and maintenance cost are highly acceptable for resale | 5.5% | 436.58 |
| 7 | Cars with higher safety and capacity are highly acceptable for resale. | 12.4% | 156.38 |
| 8 | Cars with higher safety and medium luggage boot size are highly acceptable for resale. | 4.0% | 323.92 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | A husband's education has a high chance of a long-term contraceptive being used. | 44.7% | 85.88 |
| 1 | A wife's education level usually determines if a long term contraceptive was used. | 46.8% | 166.37 |
| 2 | Women with low education are more likely to use short-term contraception. | 39.3% | 12.18 |
| 3 | Women with high education age 40 and under are more likely to use long-term contraception. | 43.4% | 22.01 |
| 4 | Women with two or more children used short and long term methods. | 42.7% | 47.27 |
| 5 | The least educated women either used short-term method or didn't use any contraceptive method. | 24.1% | 72.05 |
| 6 | If the wife's education is high, then the contraceptive method used is long-term. | 41.0% | 44.04 |
| 7 | If the wife's education is not high, then the contraceptive method used is no-use or short-term. | 27.5% | 41.12 |
| 8 | If the wife's education is not high, then the contraceptive method used is not long-term. | 23.4% | 31.23 |
| 9 | If the wife's education is not high, then the contraceptive method is short-term. | 38.6% | 37.03 |
| NO. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | The SGPT High percentage so the liver patient was yes. | 49.6% | 4184.78 |
| 1 | The SGPT Low percentage so the liver patient was no. | 52.2% | 4891.02 |
| 2 | Patients over the age of forty are liver patients. | 42.6% | 54.92 |
| 3 | Patients who has SGOT range greater than forty are liver pa-tients. | 42.6% | 373.03 |
| 4 | Decreased SGPT Value ensures no liver patient. | 52.2% | 4404.52 |
| 5 | Age group above 40 ensures liver patient. | 37.4% | 4994.60 |
| 6 | Some people have more age and have the SGOT and they are liver patient. | 54.8% | 272.55 |
| 7 | Some people have minimum age and they are liver patient. They are somewhat accurate. | 50.4% | 226.11 |
| No. | EXPLANATION | ACCURACY | PERPLEXITY |
| 0 | Frequent flyers with an annual income over 1 million usually take travel insurance. | 33.2% | 104.36 |
| 1 | People with an annual income over 1 million and under 30 years old usually take travel insurance. | 30.4% | 53.40 |
| 2 | Travelers who are 29 and older take traveler insurance. | 34.4% | 105.36 |
| 3 | Travelers who have not traveled abroad before are more likely to take traveler insurance. | 34.2% | 22.74 |
| 4 | Frequent flyer travelers with an annual income above 1 million frequently take travel insurance. | 31.2% | 121.70 |
| 5 | Travelers older than 25 years old and with an income below 1 million do not usually take travel insurance. | 42.5% | 27.45 |
| 6 | Most passengers who are not frequent fliers do not use travel insurance. | 65.8% 33.28 | |
| 7 | Most passengers who have not traveled abroad do not use travel insurance. | 77.1% | 42.87 |
| 8 | Most passenger with an income higher than 100,000 use travel insurance. | 27.9% | 131.69 |
| 9 | People who have never traveled abroad before are more likely to have taken travel insurance. | 34.2% | 15.95 |
| 10 | People with an annual income below 1,000,000 are less likely to have traveled abroad than those with annual incomes above 1,000,000. | 23.1% | 8.36 |
| 11 | People with an annual income above 1,000,000 are more likely to have taken travel insurance. | 39.9% | 26.63 |
| 12 | More frequent flyers have taken travel insurance. | 34.9% | 334.32 |
| 13 | More people that travel abroad have taken travel insurance. | 34.2% | 114.91 |
| 14 | People who are non-frequent flyers and are college graduates are less likely to get travel insurance. | 27.4% | 22.28 |
| 15 | People who make a million or more and are frequent fliers are more likely to get travel insurance. | 32.4% | 22.73 |
| 16 | Most people who didn't travel abroad before had taken Travel Insurance. | 34.2% | 84.50 |
| 17 | Most college graduate have taken Travel Insurance. | 64.8% | 780.98 |
| 18 | Those who are college graduates and in their 20s are somewhat likely to purchase travel insurance. | 22.9% | 28.62 |
| 19 | Those who have never travelled abroad and are not frequent flyers often do not purchase travel insurance. | 67.1% | 23.85 |
| 20 | Travelled Abroad Before "No" categories indicates the "No" Travel Insurance Taken. | 24.4% | 490.45 |
| 21 | College Graduate "Yes" categories leads to the "Yes" Travel Insurance Taken. | 30.9% | 678.44 |
| 22 | Frequent Flyer "No" indicates the "No" Travel Insurance Taken. | 53.8% | 194.31 |
| 23 | Annual income categories above 1049999 indicates the "Yes" Travel Insurance Taken. | 43.5% | 2291.89 |
| 24 | Most college graduates that make more than 1000000 annually have taken travel insurance. | 37.2% | 231.38 |
| 25 | Most college graduates from 26 to 34 have taken travel insurance. | 43.2% | 185.20 |
| 26 | About have of college graduates that have not travelled abroad before have taken travel insurance. | 34.2% | 186.52 |
| 27 | Frequently flyer "No" categories indicates the "No" travel insurance taken group. | 65.6% | 586.77 |
| 28 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 37.4% | 769.28 |
| 29 | Frequent flyer "No" categories indicate the "No" travel insurance taken group. | 29.6% | 375.25 |
| 30 | Annual income above 1049999 indicate the "Yes" travel insurance taken group. | 42.5% | 695.34 |
| EM | F1 | Precision | Recall | EM | F1 | Precision | Recall | |
| Methods | HotpotQA | 2WikiMultiHopQA | ||||||
| Base | 0.59 | 0.733 | 0.745 | 0.779 | 0.43 | 0.532 | 0.501 | 0.724 |
| G-Full prompt | 0.62 | 0.763 | 0.777 | 0.813 | 0.46 | 0.565 | 0.534 | 0.757 |
| SG-Multi prompt | 0.64 | 0.785 | 0.791 | 0.849 | 0.49 | 0.601 | 0.573 | 0.772 |
| SG-One prompt | 0.61 | 0.749 | 0.741 | 0.823 | 0.47 | 0.579 | 0.563 | 0.748 |
| F1 | Precision | Recall | F1 | Precision | Recall | |
| Methods | CoT Setting | Few-shot Setting | ||||
| Base | 0.716 | 0.729 | 0.830 | 0.6280 | 0.5970 | 0.7821 |
| G-Full prompt | 0.714 | 0.732 | 0.805 | 0.6065 | 0.5910 | 0.7802 |
| SG-Multi prompt | 0.728 | 0.731 | 0.872 | 0.5925 | 0.5596 | 0.8205 |
| EM | F1 | Prec | Recall | |
| ρ | 0.51 | 0.77 | 0.76 | 0.88 |
| τ | 0.51 | 0.69 | 0.68 | 0.87 |
| method | Human | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Base Prompt | 0.81 | 0.680 | 0.520 | 0.614 |
| G-Full prompt | 0.87 | 0.687 | 0.536 | 0.624 |
| SG-Multi prompt | 0.88 | 0.689 | 0.543 | 0.628 |
| EM | F1 | Precision | Recall | EM | F1 | Precision | Recall | |
| Methods | HotpotQA | 2WikiMultiHopQA | ||||||
| Base | 0.59 | 0.733 | 0.745 | 0.779 | 0.43 | 0.532 | 0.501 | 0.724 |
| G-Full prompt | 0.62 | 0.763 | 0.777 | 0.813 | 0.46 | 0.565 | 0.534 | 0.757 |
| SG-Multi prompt | 0.64 | 0.785 | 0.791 | 0.849 | 0.49 | 0.601 | 0.573 | 0.772 |
| SG-One prompt | 0.61 | 0.749 | 0.741 | 0.823 | 0.47 | 0.579 | 0.563 | 0.748 |
| F1 | Precision | Recall | F1 | Precision | Recall | |
| Methods | CoT Setting | Few-shot Setting | ||||
| Base | 0.716 | 0.729 | 0.830 | 0.6280 | 0.5970 | 0.7821 |
| G-Full prompt | 0.714 | 0.732 | 0.805 | 0.6065 | 0.5910 | 0.7802 |
| SG-Multi prompt | 0.728 | 0.731 | 0.872 | 0.5925 | 0.5596 | 0.8205 |
| EM | F1 | Prec | Recall | |
| ρ | 0.51 | 0.77 | 0.76 | 0.88 |
| τ | 0.51 | 0.69 | 0.68 | 0.87 |
| method | Human | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Base Prompt | 0.81 | 0.680 | 0.520 | 0.614 |
| G-Full prompt | 0.87 | 0.687 | 0.536 | 0.624 |
| SG-Multi prompt | 0.88 | 0.689 | 0.543 | 0.628 |
| EM | F1 | Precision | Recall | EM | F1 | Precision | Recall | |
| Methods | HotpotQA | 2WikiMultiHopQA | ||||||
| Base | 0.59 | 0.733 | 0.745 | 0.779 | 0.43 | 0.532 | 0.501 | 0.724 |
| G-Full prompt | 0.62 | 0.763 | 0.777 | 0.813 | 0.46 | 0.565 | 0.534 | 0.757 |
| SG-Multi prompt | 0.64 | 0.785 | 0.791 | 0.849 | 0.49 | 0.601 | 0.573 | 0.772 |
| SG-One prompt | 0.61 | 0.749 | 0.741 | 0.823 | 0.47 | 0.579 | 0.563 | 0.748 |
| F1 | Precision | Recall | F1 | Precision | Recall | |
| Methods | CoT Setting | Few-shot Setting | ||||
| Base | 0.716 | 0.729 | 0.830 | 0.6280 | 0.5970 | 0.7821 |
| G-Full prompt | 0.714 | 0.732 | 0.805 | 0.6065 | 0.5910 | 0.7802 |
| SG-Multi prompt | 0.728 | 0.731 | 0.872 | 0.5925 | 0.5596 | 0.8205 |
| EM | F1 | Prec | Recall | |
| ρ | 0.51 | 0.77 | 0.76 | 0.88 |
| τ | 0.51 | 0.69 | 0.68 | 0.87 |
| method | Human | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Base Prompt | 0.81 | 0.680 | 0.520 | 0.614 |
| G-Full prompt | 0.87 | 0.687 | 0.536 | 0.624 |
| SG-Multi prompt | 0.88 | 0.689 | 0.543 | 0.628 |
| EM | F1 | Precision | Recall | EM | F1 | Precision | Recall | |
| Methods | HotpotQA | 2WikiMultiHopQA | ||||||
| Base | 0.59 | 0.733 | 0.745 | 0.779 | 0.43 | 0.532 | 0.501 | 0.724 |
| G-Full prompt | 0.62 | 0.763 | 0.777 | 0.813 | 0.46 | 0.565 | 0.534 | 0.757 |
| SG-Multi prompt | 0.64 | 0.785 | 0.791 | 0.849 | 0.49 | 0.601 | 0.573 | 0.772 |
| SG-One prompt | 0.61 | 0.749 | 0.741 | 0.823 | 0.47 | 0.579 | 0.563 | 0.748 |
| F1 | Precision | Recall | F1 | Precision | Recall | |
| Methods | CoT Setting | Few-shot Setting | ||||
| Base | 0.716 | 0.729 | 0.830 | 0.6280 | 0.5970 | 0.7821 |
| G-Full prompt | 0.714 | 0.732 | 0.805 | 0.6065 | 0.5910 | 0.7802 |
| SG-Multi prompt | 0.728 | 0.731 | 0.872 | 0.5925 | 0.5596 | 0.8205 |
| EM | F1 | Prec | Recall | |
| ρ | 0.51 | 0.77 | 0.76 | 0.88 |
| τ | 0.51 | 0.69 | 0.68 | 0.87 |
| method | Human | ROUGE-1 | ROUGE-2 | ROUGE-L |
| Base Prompt | 0.81 | 0.680 | 0.520 | 0.614 |
| G-Full prompt | 0.87 | 0.687 | 0.536 | 0.624 |
| SG-Multi prompt | 0.88 | 0.689 | 0.543 | 0.628 |
| No. | Speaker | Utterance |
| 1 | User | Im looking for an Italian restaurant . I also need it to be near the center of town . |
| 2 | Agent | There are 9 Italian restaurants in the center of town . Do you have a particular price range in mind? |
| 3 | User | I need a cheap restaurant . |
| 4 | Agent | There is Pizza Hut City Centre, Ask , and Zizzi Cambridge, which would you like to book? |
| 5 | User | Can you book a table for 7 at Ask ? We're looking at 14:00 on Saturday. |
| 6 | Agent | Sorry, there are no tables available at Ask at that time. Would you like to book another day or time, or perhaps another restaurant ? |
| 7 | User | Can you try 13:00, then? |
| 8 | Agent | 13:00 it is. Your table at Ask will be held for 15 minutes, and your reference number is L2KZYL26. Can I help you with anything else? |
| 9 | User | I'm also looking for a hotel . |
| 10 | Agent | We have 33 different hotels . Did you have a price range or star rating in mind? |
| 11 | User | Something in the cheap price range including free parking. |
| 12 | Agent | What area would you prefer to stay in? This will help narrow down your options. |
| 13 | User | I would like to book a hotel close to our restaurant here in the centre , it will also need to be cheap, and have free parking. |
| 14 | Agent | There are two hotels available. Would you like me to book one for you? |
| 15 | User | Yes please. The our group of 7 will all be staying there for 3 nights starting Saturday. |
| 16 | Agent | I've booked you a 3 day stay at the Alexander Bed and Breakfast for 7 people. Do you need anything else? |
| 17 | User | Could I have the reference number for that reservation please? |
| 18 | Agent | Yes, your reference number is 1JBO1R08. |
| 19 | User | Thank you so much for all of your help. That is all I needed for now. Have a nice day . |
| 20 | Agent | Thank you for choosing us to assist you on your travels. Have a good day . |
| Model | Utterance (Case Study 1) |
| USERAGENT | Yes, please book it for 2 people on Saturday for 4 nights. Will you give me the confirmation number when that is complete?Sure, booking was successful. Your confirmation number is 7MG032V9. |
| AUGPTMINTLMHDSA MARCO | Booking was successful. Reference number is: 00000016.I have booked you at the Express by Holiday Inn Cambridge for 4 nights starting Saturday. Your reference number is D77JCGJ2. Is there anything else I can help?Booking was successful. Reference number is 7MG032V9. Is there anything else I can help you with?I have booked you at the [hotel_name]. Your reference number is 7MG032V9. Is there anything else I can help you with? |
| Model | Utterance (Case Study 2) |
| USERAGENT | I need a cab too please, arriving by my booked time , car type and contact number would be helpful too please.You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| AUGPTMINTLHDSA MARCO | I have booked a blue Skoda for you. The contact number is 85534112262.I have booked you a BMW. The contact number is 07338019809. Where will you be departing from?Where will you be departing from and where are you going?I would be happy to book a taxi for you . where are you departing from and what is your destination? |
| Domain | Taxi | Hotel | Attraction | Train | Restaurant |
| Train | 1655 | 3387 | 2718 | 3117 | 3817 |
| Valid | 207 | 416 | 401 | 484 | 438 |
| Test | 195 | 394 | 396 | 495 | 437 |
| Human | AUGPT | MINTL | HDSA | MARCO |
| 0.798 | 0.676† | 0.558† | 0.654† | 0.719† |
| Dialogue History Length = 1 | Dialogue History Length = 2 | Full Dialogue History Length | |||||||
| Recall | Precision | F1 Score | Recall | Precision | F1 Score | Recall | Precision | F1 Score | |
| *End-to-End Approach | |||||||||
| Baseline | 30.7 | 58.3 | 40.3 | 40.5 | 49.7 | 44.7 | 70.0 | 49.3 | 57.8 |
| + user samples | 33.9(▲10.4%) | 52.0(▼10.8%) | 41.0(▲1.7%) | 40.8(▲0.7%) | 47.9(▼3.6%) | 44.1(▼1.3%) | 72.8(▲4%) | 45.9(▼6.9%) | 56.3(▼2.6%) |
| + agent acts | 28.2(▼8.1%) | 65.5(▲12.3%) | 39.4(▼2.2%) | 38.7(▼4.4%) | 54.6(▲9.9%) | 45.3(▲1.3%) | 77.7(▲11%) | 39.9(▼19.1%) | 52.7(▼8.8%) |
| *Pipeline Approach | |||||||||
| Baseline | 34.9 | 39.4 | 37.0 | 40.8 | 33.3 | 36.7 | 53.8 | 21.4 | 30.7 |
| + user samples | 35.5(▲1.7%) | 40.6(▲3.0%) | 37.9(▲2.4%) | 41.5(▲1.7%) | 35.1(▲5.4%) | 38.1(▲3.8%) | 53.5(▲0.6%) | 22.7(▲0.6%) | 31.8(▲0.6%) |
| + agent acts | 36.9(▲5.7%) | 21.1(▼46.4%) | 26.9(▼27.3%) | 44.6(▲9.3%) | 12.2(▼63.4%) | 19.1(▼48.0%) | 55.7(▲4.1%) | 7.8(▼53.6%) | 13.7(▼55.4%) |
| Domains | Recall | Precision | F1 Score |
| Taxi | 63.53 | 44.17 | 52.11 |
| Hotel | 68.79 | 42.16 | 52.28 |
| Attraction | 62.56 | 49.80 | 55.01 |
| Train | 64.68 | 45.85 | 53.66 |
| Restaurant | 63.77 | 48.11 | 54.84 |
| All | 70.0 | 49.3 | 57.8 |
| IERu | IERa | ERRu | ERRa | |
| Mean. | 75.02 | 24.98 | 5.29 | 5.76 |
| Std. | 21.39 | 21.39 | 2.59 | 2.94 |
| Mean. | Std. | Min. | Max. | Median. | Mode. | |
| Frequency | 2.62 | 0.98 | 2 | 17 | 2 | 2 |
| Size | 1.47 | 0.88 | 1 | 17 | 1 | 1 |
| Span | 5.02 | 4.05 | 2 | 40 | 4 | 2 |
| Priming | 1.11 | 0.34 | 1 | 6 | 1 | 1 |
| Distance | 2.55 | 2.90 | 1 | 35 | 1 | 1 |
| Density | 0.72 | 0.30 | 0.06 | 2.00 | 0.75 | 1.00 |
| Hyperparameter | Assignment |
| number of steps | 100 epochs or 50K steps |
| batch size | 24 |
| maximum learning rate | 5e-5 |
| learning rate optimizer | AdamW |
| max sequence length | 512 |
| Adam epsilon | 1e-6 |
| Adam beta weights | 0.9, 0.98 |
| learning rate scheduler | Warmup linear |
| Weight decay | 0.01 |
| Warmup proportion | 0.06 |
| learning rate decay | linear |
| fusion layers | 2 |
| Model | Utterance |
| USER | I need a cab too please, arriving by my booked time, car type and contact number would be helpful too please. |
| Agent | You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| USER | From the Cambridge Punter, please. |
| AGENT | A black Tesla is booked, with contact number 07774975314. |
| Generated responses for the first agent response | |
| AUGPT | I have booked a blue skoda for you. The contact number is 85534112262. |
| MINTL | I have booked you a BMW. The contact number is 07338019809. Where will you be departing from? |
| HDSA | Where will you be departing from and where are you going? |
| MARCO | I would be happy to book a taxi for you. where are you departing from and what is your destination? |
| Generated responses for the second agent response | |
| AUGPT | I have booked you a blackHonda and the contact number is 91788856987. Is there anything else I can help you with? |
| MINTL | I have booked your taxi. Be expecting a BMW. If you need to reach them, please call 07338019809. |
| HDSA | I have booked you a [taxi_type]. the contact number is [taxi_phone]. is there anything else i can help you with? |
| MARCO | I have booked a [taxi_type] for you. the contact number is [taxi_phone]. is there anything else i can help you with? |
| Model | Utterance |
| USER | Yes, I would like to book a stay for 1 person, starting Monday and for 5 nights. |
| AGENT | Your room has been booked at the Bridge Guest House for 5 nights starting Monday. Your reference code is 7VM9FIPC. Is there anything else I can do to help you? |
| Generated responses for the first agent response | |
| AUGPT | I have made those reservations and your reference number is 00000011. |
| MINTL | I have booked your room. Your reference number is QJTIXK8C. |
| HDSA | Booking was successful. Reference number is 7vm9fipc. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is 7vm9fipc. Is there anything else i can help you with? |
| Model | Utterance |
| USER | I need it for 8 people , and for 4 nights starting Saturday. |
| AGENT | Great, I have booked your stay for 8 people for 4 nights . Your reference number is PB41WUOT. May I help with anything else? |
| Generated responses for the first agent response | |
| AUGPT | Booking was successful. Reference number is: 00000003. |
| MINTL | I have booked your room at Allenbell for saturday for 4 nights for 8 people . Your reference number is 6KNJSQXY. Is there anything else i can help you with? |
| HDSA | Booking was successful. Reference number is pb41wuot. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is pb41wuot. Is there anything else i can help you with? |
| No. | Speaker | Utterance |
| 1 | User | Im looking for an Italian restaurant . I also need it to be near the center of town . |
| 2 | Agent | There are 9 Italian restaurants in the center of town . Do you have a particular price range in mind? |
| 3 | User | I need a cheap restaurant . |
| 4 | Agent | There is Pizza Hut City Centre, Ask , and Zizzi Cambridge, which would you like to book? |
| 5 | User | Can you book a table for 7 at Ask ? We're looking at 14:00 on Saturday. |
| 6 | Agent | Sorry, there are no tables available at Ask at that time. Would you like to book another day or time, or perhaps another restaurant ? |
| 7 | User | Can you try 13:00, then? |
| 8 | Agent | 13:00 it is. Your table at Ask will be held for 15 minutes, and your reference number is L2KZYL26. Can I help you with anything else? |
| 9 | User | I'm also looking for a hotel . |
| 10 | Agent | We have 33 different hotels . Did you have a price range or star rating in mind? |
| 11 | User | Something in the cheap price range including free parking. |
| 12 | Agent | What area would you prefer to stay in? This will help narrow down your options. |
| 13 | User | I would like to book a hotel close to our restaurant here in the centre , it will also need to be cheap, and have free parking. |
| 14 | Agent | There are two hotels available. Would you like me to book one for you? |
| 15 | User | Yes please. The our group of 7 will all be staying there for 3 nights starting Saturday. |
| 16 | Agent | I've booked you a 3 day stay at the Alexander Bed and Breakfast for 7 people. Do you need anything else? |
| 17 | User | Could I have the reference number for that reservation please? |
| 18 | Agent | Yes, your reference number is 1JBO1R08. |
| 19 | User | Thank you so much for all of your help. That is all I needed for now. Have a nice day . |
| 20 | Agent | Thank you for choosing us to assist you on your travels. Have a good day . |
| Model | Utterance (Case Study 1) |
| USERAGENT | Yes, please book it for 2 people on Saturday for 4 nights. Will you give me the confirmation number when that is complete?Sure, booking was successful. Your confirmation number is 7MG032V9. |
| AUGPTMINTLMHDSA MARCO | Booking was successful. Reference number is: 00000016.I have booked you at the Express by Holiday Inn Cambridge for 4 nights starting Saturday. Your reference number is D77JCGJ2. Is there anything else I can help?Booking was successful. Reference number is 7MG032V9. Is there anything else I can help you with?I have booked you at the [hotel_name]. Your reference number is 7MG032V9. Is there anything else I can help you with? |
| Model | Utterance (Case Study 2) |
| USERAGENT | I need a cab too please, arriving by my booked time , car type and contact number would be helpful too please.You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| AUGPTMINTLHDSA MARCO | I have booked a blue Skoda for you. The contact number is 85534112262.I have booked you a BMW. The contact number is 07338019809. Where will you be departing from?Where will you be departing from and where are you going?I would be happy to book a taxi for you . where are you departing from and what is your destination? |
| Domain | Taxi | Hotel | Attraction | Train | Restaurant |
| Train | 1655 | 3387 | 2718 | 3117 | 3817 |
| Valid | 207 | 416 | 401 | 484 | 438 |
| Test | 195 | 394 | 396 | 495 | 437 |
| Human | AUGPT | MINTL | HDSA | MARCO |
| 0.798 | 0.676† | 0.558† | 0.654† | 0.719† |
| Dialogue History Length = 1 | Dialogue History Length = 2 | Full Dialogue History Length | |||||||
| Recall | Precision | F1 Score | Recall | Precision | F1 Score | Recall | Precision | F1 Score | |
| *End-to-End Approach | |||||||||
| Baseline | 30.7 | 58.3 | 40.3 | 40.5 | 49.7 | 44.7 | 70.0 | 49.3 | 57.8 |
| + user samples | 33.9(▲10.4%) | 52.0(▼10.8%) | 41.0(▲1.7%) | 40.8(▲0.7%) | 47.9(▼3.6%) | 44.1(▼1.3%) | 72.8(▲4%) | 45.9(▼6.9%) | 56.3(▼2.6%) |
| + agent acts | 28.2(▼8.1%) | 65.5(▲12.3%) | 39.4(▼2.2%) | 38.7(▼4.4%) | 54.6(▲9.9%) | 45.3(▲1.3%) | 77.7(▲11%) | 39.9(▼19.1%) | 52.7(▼8.8%) |
| *Pipeline Approach | |||||||||
| Baseline | 34.9 | 39.4 | 37.0 | 40.8 | 33.3 | 36.7 | 53.8 | 21.4 | 30.7 |
| + user samples | 35.5(▲1.7%) | 40.6(▲3.0%) | 37.9(▲2.4%) | 41.5(▲1.7%) | 35.1(▲5.4%) | 38.1(▲3.8%) | 53.5(▲0.6%) | 22.7(▲0.6%) | 31.8(▲0.6%) |
| + agent acts | 36.9(▲5.7%) | 21.1(▼46.4%) | 26.9(▼27.3%) | 44.6(▲9.3%) | 12.2(▼63.4%) | 19.1(▼48.0%) | 55.7(▲4.1%) | 7.8(▼53.6%) | 13.7(▼55.4%) |
| Domains | Recall | Precision | F1 Score |
| Taxi | 63.53 | 44.17 | 52.11 |
| Hotel | 68.79 | 42.16 | 52.28 |
| Attraction | 62.56 | 49.80 | 55.01 |
| Train | 64.68 | 45.85 | 53.66 |
| Restaurant | 63.77 | 48.11 | 54.84 |
| All | 70.0 | 49.3 | 57.8 |
| IERu | IERa | ERRu | ERRa | |
| Mean. | 75.02 | 24.98 | 5.29 | 5.76 |
| Std. | 21.39 | 21.39 | 2.59 | 2.94 |
| Mean. | Std. | Min. | Max. | Median. | Mode. | |
| Frequency | 2.62 | 0.98 | 2 | 17 | 2 | 2 |
| Size | 1.47 | 0.88 | 1 | 17 | 1 | 1 |
| Span | 5.02 | 4.05 | 2 | 40 | 4 | 2 |
| Priming | 1.11 | 0.34 | 1 | 6 | 1 | 1 |
| Distance | 2.55 | 2.90 | 1 | 35 | 1 | 1 |
| Density | 0.72 | 0.30 | 0.06 | 2.00 | 0.75 | 1.00 |
| Hyperparameter | Assignment |
| number of steps | 100 epochs or 50K steps |
| batch size | 24 |
| maximum learning rate | 5e-5 |
| learning rate optimizer | AdamW |
| max sequence length | 512 |
| Adam epsilon | 1e-6 |
| Adam beta weights | 0.9, 0.98 |
| learning rate scheduler | Warmup linear |
| Weight decay | 0.01 |
| Warmup proportion | 0.06 |
| learning rate decay | linear |
| fusion layers | 2 |
| Model | Utterance |
| USER | I need a cab too please, arriving by my booked time, car type and contact number would be helpful too please. |
| Agent | You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| USER | From the Cambridge Punter, please. |
| AGENT | A black Tesla is booked, with contact number 07774975314. |
| Generated responses for the first agent response | |
| AUGPT | I have booked a blue skoda for you. The contact number is 85534112262. |
| MINTL | I have booked you a BMW. The contact number is 07338019809. Where will you be departing from? |
| HDSA | Where will you be departing from and where are you going? |
| MARCO | I would be happy to book a taxi for you. where are you departing from and what is your destination? |
| Generated responses for the second agent response | |
| AUGPT | I have booked you a blackHonda and the contact number is 91788856987. Is there anything else I can help you with? |
| MINTL | I have booked your taxi. Be expecting a BMW. If you need to reach them, please call 07338019809. |
| HDSA | I have booked you a [taxi_type]. the contact number is [taxi_phone]. is there anything else i can help you with? |
| MARCO | I have booked a [taxi_type] for you. the contact number is [taxi_phone]. is there anything else i can help you with? |
| Model | Utterance |
| USER | Yes, I would like to book a stay for 1 person, starting Monday and for 5 nights. |
| AGENT | Your room has been booked at the Bridge Guest House for 5 nights starting Monday. Your reference code is 7VM9FIPC. Is there anything else I can do to help you? |
| Generated responses for the first agent response | |
| AUGPT | I have made those reservations and your reference number is 00000011. |
| MINTL | I have booked your room. Your reference number is QJTIXK8C. |
| HDSA | Booking was successful. Reference number is 7vm9fipc. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is 7vm9fipc. Is there anything else i can help you with? |
| Model | Utterance |
| USER | I need it for 8 people , and for 4 nights starting Saturday. |
| AGENT | Great, I have booked your stay for 8 people for 4 nights . Your reference number is PB41WUOT. May I help with anything else? |
| Generated responses for the first agent response | |
| AUGPT | Booking was successful. Reference number is: 00000003. |
| MINTL | I have booked your room at Allenbell for saturday for 4 nights for 8 people . Your reference number is 6KNJSQXY. Is there anything else i can help you with? |
| HDSA | Booking was successful. Reference number is pb41wuot. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is pb41wuot. Is there anything else i can help you with? |
| No. | Speaker | Utterance |
| 1 | User | Im looking for an Italian restaurant . I also need it to be near the center of town . |
| 2 | Agent | There are 9 Italian restaurants in the center of town . Do you have a particular price range in mind? |
| 3 | User | I need a cheap restaurant . |
| 4 | Agent | There is Pizza Hut City Centre, Ask , and Zizzi Cambridge, which would you like to book? |
| 5 | User | Can you book a table for 7 at Ask ? We're looking at 14:00 on Saturday. |
| 6 | Agent | Sorry, there are no tables available at Ask at that time. Would you like to book another day or time, or perhaps another restaurant ? |
| 7 | User | Can you try 13:00, then? |
| 8 | Agent | 13:00 it is. Your table at Ask will be held for 15 minutes, and your reference number is L2KZYL26. Can I help you with anything else? |
| 9 | User | I'm also looking for a hotel . |
| 10 | Agent | We have 33 different hotels . Did you have a price range or star rating in mind? |
| 11 | User | Something in the cheap price range including free parking. |
| 12 | Agent | What area would you prefer to stay in? This will help narrow down your options. |
| 13 | User | I would like to book a hotel close to our restaurant here in the centre , it will also need to be cheap, and have free parking. |
| 14 | Agent | There are two hotels available. Would you like me to book one for you? |
| 15 | User | Yes please. The our group of 7 will all be staying there for 3 nights starting Saturday. |
| 16 | Agent | I've booked you a 3 day stay at the Alexander Bed and Breakfast for 7 people. Do you need anything else? |
| 17 | User | Could I have the reference number for that reservation please? |
| 18 | Agent | Yes, your reference number is 1JBO1R08. |
| 19 | User | Thank you so much for all of your help. That is all I needed for now. Have a nice day . |
| 20 | Agent | Thank you for choosing us to assist you on your travels. Have a good day . |
| Model | Utterance (Case Study 1) |
| USERAGENT | Yes, please book it for 2 people on Saturday for 4 nights. Will you give me the confirmation number when that is complete?Sure, booking was successful. Your confirmation number is 7MG032V9. |
| AUGPTMINTLMHDSA MARCO | Booking was successful. Reference number is: 00000016.I have booked you at the Express by Holiday Inn Cambridge for 4 nights starting Saturday. Your reference number is D77JCGJ2. Is there anything else I can help?Booking was successful. Reference number is 7MG032V9. Is there anything else I can help you with?I have booked you at the [hotel_name]. Your reference number is 7MG032V9. Is there anything else I can help you with? |
| Model | Utterance (Case Study 2) |
| USERAGENT | I need a cab too please, arriving by my booked time , car type and contact number would be helpful too please.You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| AUGPTMINTLHDSA MARCO | I have booked a blue Skoda for you. The contact number is 85534112262.I have booked you a BMW. The contact number is 07338019809. Where will you be departing from?Where will you be departing from and where are you going?I would be happy to book a taxi for you . where are you departing from and what is your destination? |
| Domain | Taxi | Hotel | Attraction | Train | Restaurant |
| Train | 1655 | 3387 | 2718 | 3117 | 3817 |
| Valid | 207 | 416 | 401 | 484 | 438 |
| Test | 195 | 394 | 396 | 495 | 437 |
| Human | AUGPT | MINTL | HDSA | MARCO |
| 0.798 | 0.676† | 0.558† | 0.654† | 0.719† |
| Dialogue History Length = 1 | Dialogue History Length = 2 | Full Dialogue History Length | |||||||
| Recall | Precision | F1 Score | Recall | Precision | F1 Score | Recall | Precision | F1 Score | |
| *End-to-End Approach | |||||||||
| Baseline | 30.7 | 58.3 | 40.3 | 40.5 | 49.7 | 44.7 | 70.0 | 49.3 | 57.8 |
| + user samples | 33.9(▲10.4%) | 52.0(▼10.8%) | 41.0(▲1.7%) | 40.8(▲0.7%) | 47.9(▼3.6%) | 44.1(▼1.3%) | 72.8(▲4%) | 45.9(▼6.9%) | 56.3(▼2.6%) |
| + agent acts | 28.2(▼8.1%) | 65.5(▲12.3%) | 39.4(▼2.2%) | 38.7(▼4.4%) | 54.6(▲9.9%) | 45.3(▲1.3%) | 77.7(▲11%) | 39.9(▼19.1%) | 52.7(▼8.8%) |
| *Pipeline Approach | |||||||||
| Baseline | 34.9 | 39.4 | 37.0 | 40.8 | 33.3 | 36.7 | 53.8 | 21.4 | 30.7 |
| + user samples | 35.5(▲1.7%) | 40.6(▲3.0%) | 37.9(▲2.4%) | 41.5(▲1.7%) | 35.1(▲5.4%) | 38.1(▲3.8%) | 53.5(▲0.6%) | 22.7(▲0.6%) | 31.8(▲0.6%) |
| + agent acts | 36.9(▲5.7%) | 21.1(▼46.4%) | 26.9(▼27.3%) | 44.6(▲9.3%) | 12.2(▼63.4%) | 19.1(▼48.0%) | 55.7(▲4.1%) | 7.8(▼53.6%) | 13.7(▼55.4%) |
| Domains | Recall | Precision | F1 Score |
| Taxi | 63.53 | 44.17 | 52.11 |
| Hotel | 68.79 | 42.16 | 52.28 |
| Attraction | 62.56 | 49.80 | 55.01 |
| Train | 64.68 | 45.85 | 53.66 |
| Restaurant | 63.77 | 48.11 | 54.84 |
| All | 70.0 | 49.3 | 57.8 |
| IERu | IERa | ERRu | ERRa | |
| Mean. | 75.02 | 24.98 | 5.29 | 5.76 |
| Std. | 21.39 | 21.39 | 2.59 | 2.94 |
| Mean. | Std. | Min. | Max. | Median. | Mode. | |
| Frequency | 2.62 | 0.98 | 2 | 17 | 2 | 2 |
| Size | 1.47 | 0.88 | 1 | 17 | 1 | 1 |
| Span | 5.02 | 4.05 | 2 | 40 | 4 | 2 |
| Priming | 1.11 | 0.34 | 1 | 6 | 1 | 1 |
| Distance | 2.55 | 2.90 | 1 | 35 | 1 | 1 |
| Density | 0.72 | 0.30 | 0.06 | 2.00 | 0.75 | 1.00 |
| Hyperparameter | Assignment |
| number of steps | 100 epochs or 50K steps |
| batch size | 24 |
| maximum learning rate | 5e-5 |
| learning rate optimizer | AdamW |
| max sequence length | 512 |
| Adam epsilon | 1e-6 |
| Adam beta weights | 0.9, 0.98 |
| learning rate scheduler | Warmup linear |
| Weight decay | 0.01 |
| Warmup proportion | 0.06 |
| learning rate decay | linear |
| fusion layers | 2 |
| Model | Utterance |
| USER | I need a cab too please, arriving by my booked time, car type and contact number would be helpful too please. |
| Agent | You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| USER | From the Cambridge Punter, please. |
| AGENT | A black Tesla is booked, with contact number 07774975314. |
| Generated responses for the first agent response | |
| AUGPT | I have booked a blue skoda for you. The contact number is 85534112262. |
| MINTL | I have booked you a BMW. The contact number is 07338019809. Where will you be departing from? |
| HDSA | Where will you be departing from and where are you going? |
| MARCO | I would be happy to book a taxi for you. where are you departing from and what is your destination? |
| Generated responses for the second agent response | |
| AUGPT | I have booked you a blackHonda and the contact number is 91788856987. Is there anything else I can help you with? |
| MINTL | I have booked your taxi. Be expecting a BMW. If you need to reach them, please call 07338019809. |
| HDSA | I have booked you a [taxi_type]. the contact number is [taxi_phone]. is there anything else i can help you with? |
| MARCO | I have booked a [taxi_type] for you. the contact number is [taxi_phone]. is there anything else i can help you with? |
| Model | Utterance |
| USER | Yes, I would like to book a stay for 1 person, starting Monday and for 5 nights. |
| AGENT | Your room has been booked at the Bridge Guest House for 5 nights starting Monday. Your reference code is 7VM9FIPC. Is there anything else I can do to help you? |
| Generated responses for the first agent response | |
| AUGPT | I have made those reservations and your reference number is 00000011. |
| MINTL | I have booked your room. Your reference number is QJTIXK8C. |
| HDSA | Booking was successful. Reference number is 7vm9fipc. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is 7vm9fipc. Is there anything else i can help you with? |
| Model | Utterance |
| USER | I need it for 8 people , and for 4 nights starting Saturday. |
| AGENT | Great, I have booked your stay for 8 people for 4 nights . Your reference number is PB41WUOT. May I help with anything else? |
| Generated responses for the first agent response | |
| AUGPT | Booking was successful. Reference number is: 00000003. |
| MINTL | I have booked your room at Allenbell for saturday for 4 nights for 8 people . Your reference number is 6KNJSQXY. Is there anything else i can help you with? |
| HDSA | Booking was successful. Reference number is pb41wuot. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is pb41wuot. Is there anything else i can help you with? |
| No. | Speaker | Utterance |
| 1 | User | Im looking for an Italian restaurant . I also need it to be near the center of town . |
| 2 | Agent | There are 9 Italian restaurants in the center of town . Do you have a particular price range in mind? |
| 3 | User | I need a cheap restaurant . |
| 4 | Agent | There is Pizza Hut City Centre, Ask , and Zizzi Cambridge, which would you like to book? |
| 5 | User | Can you book a table for 7 at Ask ? We're looking at 14:00 on Saturday. |
| 6 | Agent | Sorry, there are no tables available at Ask at that time. Would you like to book another day or time, or perhaps another restaurant ? |
| 7 | User | Can you try 13:00, then? |
| 8 | Agent | 13:00 it is. Your table at Ask will be held for 15 minutes, and your reference number is L2KZYL26. Can I help you with anything else? |
| 9 | User | I'm also looking for a hotel . |
| 10 | Agent | We have 33 different hotels . Did you have a price range or star rating in mind? |
| 11 | User | Something in the cheap price range including free parking. |
| 12 | Agent | What area would you prefer to stay in? This will help narrow down your options. |
| 13 | User | I would like to book a hotel close to our restaurant here in the centre , it will also need to be cheap, and have free parking. |
| 14 | Agent | There are two hotels available. Would you like me to book one for you? |
| 15 | User | Yes please. The our group of 7 will all be staying there for 3 nights starting Saturday. |
| 16 | Agent | I've booked you a 3 day stay at the Alexander Bed and Breakfast for 7 people. Do you need anything else? |
| 17 | User | Could I have the reference number for that reservation please? |
| 18 | Agent | Yes, your reference number is 1JBO1R08. |
| 19 | User | Thank you so much for all of your help. That is all I needed for now. Have a nice day . |
| 20 | Agent | Thank you for choosing us to assist you on your travels. Have a good day . |
| Model | Utterance (Case Study 1) |
| USERAGENT | Yes, please book it for 2 people on Saturday for 4 nights. Will you give me the confirmation number when that is complete?Sure, booking was successful. Your confirmation number is 7MG032V9. |
| AUGPTMINTLMHDSA MARCO | Booking was successful. Reference number is: 00000016.I have booked you at the Express by Holiday Inn Cambridge for 4 nights starting Saturday. Your reference number is D77JCGJ2. Is there anything else I can help?Booking was successful. Reference number is 7MG032V9. Is there anything else I can help you with?I have booked you at the [hotel_name]. Your reference number is 7MG032V9. Is there anything else I can help you with? |
| Model | Utterance (Case Study 2) |
| USERAGENT | I need a cab too please, arriving by my booked time , car type and contact number would be helpful too please.You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| AUGPTMINTLHDSA MARCO | I have booked a blue Skoda for you. The contact number is 85534112262.I have booked you a BMW. The contact number is 07338019809. Where will you be departing from?Where will you be departing from and where are you going?I would be happy to book a taxi for you . where are you departing from and what is your destination? |
| Domain | Taxi | Hotel | Attraction | Train | Restaurant |
| Train | 1655 | 3387 | 2718 | 3117 | 3817 |
| Valid | 207 | 416 | 401 | 484 | 438 |
| Test | 195 | 394 | 396 | 495 | 437 |
| Human | AUGPT | MINTL | HDSA | MARCO |
| 0.798 | 0.676† | 0.558† | 0.654† | 0.719† |
| Dialogue History Length = 1 | Dialogue History Length = 2 | Full Dialogue History Length | |||||||
| Recall | Precision | F1 Score | Recall | Precision | F1 Score | Recall | Precision | F1 Score | |
| *End-to-End Approach | |||||||||
| Baseline | 30.7 | 58.3 | 40.3 | 40.5 | 49.7 | 44.7 | 70.0 | 49.3 | 57.8 |
| + user samples | 33.9(▲10.4%) | 52.0(▼10.8%) | 41.0(▲1.7%) | 40.8(▲0.7%) | 47.9(▼3.6%) | 44.1(▼1.3%) | 72.8(▲4%) | 45.9(▼6.9%) | 56.3(▼2.6%) |
| + agent acts | 28.2(▼8.1%) | 65.5(▲12.3%) | 39.4(▼2.2%) | 38.7(▼4.4%) | 54.6(▲9.9%) | 45.3(▲1.3%) | 77.7(▲11%) | 39.9(▼19.1%) | 52.7(▼8.8%) |
| *Pipeline Approach | |||||||||
| Baseline | 34.9 | 39.4 | 37.0 | 40.8 | 33.3 | 36.7 | 53.8 | 21.4 | 30.7 |
| + user samples | 35.5(▲1.7%) | 40.6(▲3.0%) | 37.9(▲2.4%) | 41.5(▲1.7%) | 35.1(▲5.4%) | 38.1(▲3.8%) | 53.5(▲0.6%) | 22.7(▲0.6%) | 31.8(▲0.6%) |
| + agent acts | 36.9(▲5.7%) | 21.1(▼46.4%) | 26.9(▼27.3%) | 44.6(▲9.3%) | 12.2(▼63.4%) | 19.1(▼48.0%) | 55.7(▲4.1%) | 7.8(▼53.6%) | 13.7(▼55.4%) |
| Domains | Recall | Precision | F1 Score |
| Taxi | 63.53 | 44.17 | 52.11 |
| Hotel | 68.79 | 42.16 | 52.28 |
| Attraction | 62.56 | 49.80 | 55.01 |
| Train | 64.68 | 45.85 | 53.66 |
| Restaurant | 63.77 | 48.11 | 54.84 |
| All | 70.0 | 49.3 | 57.8 |
| IERu | IERa | ERRu | ERRa | |
| Mean. | 75.02 | 24.98 | 5.29 | 5.76 |
| Std. | 21.39 | 21.39 | 2.59 | 2.94 |
| Mean. | Std. | Min. | Max. | Median. | Mode. | |
| Frequency | 2.62 | 0.98 | 2 | 17 | 2 | 2 |
| Size | 1.47 | 0.88 | 1 | 17 | 1 | 1 |
| Span | 5.02 | 4.05 | 2 | 40 | 4 | 2 |
| Priming | 1.11 | 0.34 | 1 | 6 | 1 | 1 |
| Distance | 2.55 | 2.90 | 1 | 35 | 1 | 1 |
| Density | 0.72 | 0.30 | 0.06 | 2.00 | 0.75 | 1.00 |
| Hyperparameter | Assignment |
| number of steps | 100 epochs or 50K steps |
| batch size | 24 |
| maximum learning rate | 5e-5 |
| learning rate optimizer | AdamW |
| max sequence length | 512 |
| Adam epsilon | 1e-6 |
| Adam beta weights | 0.9, 0.98 |
| learning rate scheduler | Warmup linear |
| Weight decay | 0.01 |
| Warmup proportion | 0.06 |
| learning rate decay | linear |
| fusion layers | 2 |
| Model | Utterance |
| USER | I need a cab too please, arriving by my booked time, car type and contact number would be helpful too please. |
| Agent | You need a cab to get you to the restaurant before the booked time and where do you need cab to pick you up at? |
| USER | From the Cambridge Punter, please. |
| AGENT | A black Tesla is booked, with contact number 07774975314. |
| Generated responses for the first agent response | |
| AUGPT | I have booked a blue skoda for you. The contact number is 85534112262. |
| MINTL | I have booked you a BMW. The contact number is 07338019809. Where will you be departing from? |
| HDSA | Where will you be departing from and where are you going? |
| MARCO | I would be happy to book a taxi for you. where are you departing from and what is your destination? |
| Generated responses for the second agent response | |
| AUGPT | I have booked you a blackHonda and the contact number is 91788856987. Is there anything else I can help you with? |
| MINTL | I have booked your taxi. Be expecting a BMW. If you need to reach them, please call 07338019809. |
| HDSA | I have booked you a [taxi_type]. the contact number is [taxi_phone]. is there anything else i can help you with? |
| MARCO | I have booked a [taxi_type] for you. the contact number is [taxi_phone]. is there anything else i can help you with? |
| Model | Utterance |
| USER | Yes, I would like to book a stay for 1 person, starting Monday and for 5 nights. |
| AGENT | Your room has been booked at the Bridge Guest House for 5 nights starting Monday. Your reference code is 7VM9FIPC. Is there anything else I can do to help you? |
| Generated responses for the first agent response | |
| AUGPT | I have made those reservations and your reference number is 00000011. |
| MINTL | I have booked your room. Your reference number is QJTIXK8C. |
| HDSA | Booking was successful. Reference number is 7vm9fipc. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is 7vm9fipc. Is there anything else i can help you with? |
| Model | Utterance |
| USER | I need it for 8 people , and for 4 nights starting Saturday. |
| AGENT | Great, I have booked your stay for 8 people for 4 nights . Your reference number is PB41WUOT. May I help with anything else? |
| Generated responses for the first agent response | |
| AUGPT | Booking was successful. Reference number is: 00000003. |
| MINTL | I have booked your room at Allenbell for saturday for 4 nights for 8 people . Your reference number is 6KNJSQXY. Is there anything else i can help you with? |
| HDSA | Booking was successful. Reference number is pb41wuot. Is there anything else i can help you with? |
| MARCO | I was able to book you at the [hotel_name] for [value_count] nights. Your reference number is pb41wuot. Is there anything else i can help you with? |
| XSUM | CNN | WikiHow | SamSum | |
| BRIO-2/10/1 | 24.85 | 23.39 | 18.60 | 27.31 |
| BRIO-5/25/2 | 24.90 | 23.50 | 18.75 | 27.64 |
| BRIO-10/100/3 | 24.94 | 23.54 | 19.12 | 27.64 |
| BRIO-25/250/5 | 24.99 | 23.79 | 19.16 | 27.72 |
| BRIO-all | 25.31 | 23.73 | 19.34 | 27.90 |
| 25 (FT) | all (FT) | BRIO-25 | BRIO-all | |
| Consistency | ||||
| XSUM | 17% | 13% | 19% | 7% |
| CNN | 7% | 2% | 9% | 8% |
| Relevance | ||||
| XSUM | 34% | 24% | 36% | 13% |
| CNN | 30% | 29% | 34% | 28% |
| XSUM | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-40 | 40-45 | 45-55 | 55-70 | >70 |
| #Samples | 690 | 826 | 961 | 1048 | 1159 | 1216 | 1029 | 863 | 888 | 1152 | 681 | 819 |
| CNN | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | >35 |
| #Samples | 789 | 1587 | 2118 | 2039 | 1726 | 1350 | 823 | 1058 |
| WikiHow | 0-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-45 | 45-55 | >55 |
| #Samples | 640 | 445 | 601 | 702 | 628 | 1120 | 817 | 1043 |
| SamSum | 0-5 | 5-15 | >15 |
| #Samples | 286 | 246 | 287 |
| Threshold | 2 | 5 | 10 | 25 | all |
| XSUM | 44K | 76K | 102K | 136K | 204K |
| CNN | 44K | 86K | 124K | 177K | 287K |
| Threshold | 10 | 25 | 100 | 250 | all |
| WikiHow | 36K | 57K | 99K | 130K | 168K |
| Threshold | 1 | 2 | 3 | 5 | all |
| SamSum | 5.7K | 8.3K | 9.6K | 11.1K | 14.7K |
| 2 | 5 | 10 | 25 | 100 | 250 | |
| C4 | 249K | 510K | 823K | 1.48M | 3.26M | 5.06M |
| NR | 39K | 93K | 155K | 273K | 511K | - |
| CNN | XSUM | WikiHow | SamSum | |
| Random | 20.14 | 19.17 | 16.72 | 26.71 |
| \\( \\theta_{4G}^{nov} \\) | 19.61 | 18.33 | 15.98 | 26.58 |
| \\( \\theta_{sim} \\) | 20.35 | 19.35 | 16.91 | 26.82 |
| θ4G | Correct ↑ | Skipped | Incorrect ↓ |
| 2 | 35.0% | 41.0% | 24.0% |
| 5 | 35.67% | 35.67% | 28.67% |
| 10 | 40.33% | 30.67% | 29.0% |
| 25 | 45.0% | 26.33% | 28.66% |
| all | 34.33% | 29.66% | 36.0% |
| XSUM | CNN | WikiHow | SamSum | |
| BRIO-2/10/1 | 24.85 | 23.39 | 18.60 | 27.31 |
| BRIO-5/25/2 | 24.90 | 23.50 | 18.75 | 27.64 |
| BRIO-10/100/3 | 24.94 | 23.54 | 19.12 | 27.64 |
| BRIO-25/250/5 | 24.99 | 23.79 | 19.16 | 27.72 |
| BRIO-all | 25.31 | 23.73 | 19.34 | 27.90 |
| 25 (FT) | all (FT) | BRIO-25 | BRIO-all | |
| Consistency | ||||
| XSUM | 17% | 13% | 19% | 7% |
| CNN | 7% | 2% | 9% | 8% |
| Relevance | ||||
| XSUM | 34% | 24% | 36% | 13% |
| CNN | 30% | 29% | 34% | 28% |
| XSUM | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-40 | 40-45 | 45-55 | 55-70 | >70 |
| #Samples | 690 | 826 | 961 | 1048 | 1159 | 1216 | 1029 | 863 | 888 | 1152 | 681 | 819 |
| CNN | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | >35 |
| #Samples | 789 | 1587 | 2118 | 2039 | 1726 | 1350 | 823 | 1058 |
| WikiHow | 0-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-45 | 45-55 | >55 |
| #Samples | 640 | 445 | 601 | 702 | 628 | 1120 | 817 | 1043 |
| SamSum | 0-5 | 5-15 | >15 |
| #Samples | 286 | 246 | 287 |
| Threshold | 2 | 5 | 10 | 25 | all |
| XSUM | 44K | 76K | 102K | 136K | 204K |
| CNN | 44K | 86K | 124K | 177K | 287K |
| Threshold | 10 | 25 | 100 | 250 | all |
| WikiHow | 36K | 57K | 99K | 130K | 168K |
| Threshold | 1 | 2 | 3 | 5 | all |
| SamSum | 5.7K | 8.3K | 9.6K | 11.1K | 14.7K |
| 2 | 5 | 10 | 25 | 100 | 250 | |
| C4 | 249K | 510K | 823K | 1.48M | 3.26M | 5.06M |
| NR | 39K | 93K | 155K | 273K | 511K | - |
| CNN | XSUM | WikiHow | SamSum | |
| Random | 20.14 | 19.17 | 16.72 | 26.71 |
| \\( \\theta_{4G}^{nov} \\) | 19.61 | 18.33 | 15.98 | 26.58 |
| \\( \\theta_{sim} \\) | 20.35 | 19.35 | 16.91 | 26.82 |
| θ4G | Correct ↑ | Skipped | Incorrect ↓ |
| 2 | 35.0% | 41.0% | 24.0% |
| 5 | 35.67% | 35.67% | 28.67% |
| 10 | 40.33% | 30.67% | 29.0% |
| 25 | 45.0% | 26.33% | 28.66% |
| all | 34.33% | 29.66% | 36.0% |
| XSUM | CNN | WikiHow | SamSum | |
| BRIO-2/10/1 | 24.85 | 23.39 | 18.60 | 27.31 |
| BRIO-5/25/2 | 24.90 | 23.50 | 18.75 | 27.64 |
| BRIO-10/100/3 | 24.94 | 23.54 | 19.12 | 27.64 |
| BRIO-25/250/5 | 24.99 | 23.79 | 19.16 | 27.72 |
| BRIO-all | 25.31 | 23.73 | 19.34 | 27.90 |
| 25 (FT) | all (FT) | BRIO-25 | BRIO-all | |
| Consistency | ||||
| XSUM | 17% | 13% | 19% | 7% |
| CNN | 7% | 2% | 9% | 8% |
| Relevance | ||||
| XSUM | 34% | 24% | 36% | 13% |
| CNN | 30% | 29% | 34% | 28% |
| XSUM | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-40 | 40-45 | 45-55 | 55-70 | >70 |
| #Samples | 690 | 826 | 961 | 1048 | 1159 | 1216 | 1029 | 863 | 888 | 1152 | 681 | 819 |
| CNN | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | >35 |
| #Samples | 789 | 1587 | 2118 | 2039 | 1726 | 1350 | 823 | 1058 |
| WikiHow | 0-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-45 | 45-55 | >55 |
| #Samples | 640 | 445 | 601 | 702 | 628 | 1120 | 817 | 1043 |
| SamSum | 0-5 | 5-15 | >15 |
| #Samples | 286 | 246 | 287 |
| Threshold | 2 | 5 | 10 | 25 | all |
| XSUM | 44K | 76K | 102K | 136K | 204K |
| CNN | 44K | 86K | 124K | 177K | 287K |
| Threshold | 10 | 25 | 100 | 250 | all |
| WikiHow | 36K | 57K | 99K | 130K | 168K |
| Threshold | 1 | 2 | 3 | 5 | all |
| SamSum | 5.7K | 8.3K | 9.6K | 11.1K | 14.7K |
| 2 | 5 | 10 | 25 | 100 | 250 | |
| C4 | 249K | 510K | 823K | 1.48M | 3.26M | 5.06M |
| NR | 39K | 93K | 155K | 273K | 511K | - |
| CNN | XSUM | WikiHow | SamSum | |
| Random | 20.14 | 19.17 | 16.72 | 26.71 |
| \( \theta_{4G}^{nov} \) | 19.61 | 18.33 | 15.98 | 26.58 |
| \( \theta_{sim} \) | 20.35 | 19.35 | 16.91 | 26.82 |
| θ4G | Correct ↑ | Skipped | Incorrect ↓ |
| 2 | 35.0% | 41.0% | 24.0% |
| 5 | 35.67% | 35.67% | 28.67% |
| 10 | 40.33% | 30.67% | 29.0% |
| 25 | 45.0% | 26.33% | 28.66% |
| all | 34.33% | 29.66% | 36.0% |
| XSUM | CNN | WikiHow | SamSum | |
| BRIO-2/10/1 | 24.85 | 23.39 | 18.60 | 27.31 |
| BRIO-5/25/2 | 24.90 | 23.50 | 18.75 | 27.64 |
| BRIO-10/100/3 | 24.94 | 23.54 | 19.12 | 27.64 |
| BRIO-25/250/5 | 24.99 | 23.79 | 19.16 | 27.72 |
| BRIO-all | 25.31 | 23.73 | 19.34 | 27.90 |
| 25 (FT) | all (FT) | BRIO-25 | BRIO-all | |
| Consistency | ||||
| XSUM | 17% | 13% | 19% | 7% |
| CNN | 7% | 2% | 9% | 8% |
| Relevance | ||||
| XSUM | 34% | 24% | 36% | 13% |
| CNN | 30% | 29% | 34% | 28% |
| XSUM | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-40 | 40-45 | 45-55 | 55-70 | >70 |
| #Samples | 690 | 826 | 961 | 1048 | 1159 | 1216 | 1029 | 863 | 888 | 1152 | 681 | 819 |
| CNN | 0-5 | 5-10 | 10-15 | 15-20 | 20-25 | 25-30 | 30-35 | >35 |
| #Samples | 789 | 1587 | 2118 | 2039 | 1726 | 1350 | 823 | 1058 |
| WikiHow | 0-15 | 15-20 | 20-25 | 25-30 | 30-35 | 35-45 | 45-55 | >55 |
| #Samples | 640 | 445 | 601 | 702 | 628 | 1120 | 817 | 1043 |
| SamSum | 0-5 | 5-15 | >15 |
| #Samples | 286 | 246 | 287 |
| Threshold | 2 | 5 | 10 | 25 | all |
| XSUM | 44K | 76K | 102K | 136K | 204K |
| CNN | 44K | 86K | 124K | 177K | 287K |
| Threshold | 10 | 25 | 100 | 250 | all |
| WikiHow | 36K | 57K | 99K | 130K | 168K |
| Threshold | 1 | 2 | 3 | 5 | all |
| SamSum | 5.7K | 8.3K | 9.6K | 11.1K | 14.7K |
| 2 | 5 | 10 | 25 | 100 | 250 | |
| C4 | 249K | 510K | 823K | 1.48M | 3.26M | 5.06M |
| NR | 39K | 93K | 155K | 273K | 511K | - |
| CNN | XSUM | WikiHow | SamSum | |
| Random | 20.14 | 19.17 | 16.72 | 26.71 |
| \\( \\theta_{4G}^{nov} \\) | 19.61 | 18.33 | 15.98 | 26.58 |
| \\( \\theta_{sim} \\) | 20.35 | 19.35 | 16.91 | 26.82 |
| θ4G | Correct ↑ | Skipped | Incorrect ↓ |
| 2 | 35.0% | 41.0% | 24.0% |
| 5 | 35.67% | 35.67% | 28.67% |
| 10 | 40.33% | 30.67% | 29.0% |
| 25 | 45.0% | 26.33% | 28.66% |
| all | 34.33% | 29.66% | 36.0% |
| Licon-A | Licon-I | Licon-T | |
| Train concepts | 90 | 280 | 681 |
| Test concepts | 18 | 56 | 227 |
| Avg.# Q/Concept | 36.05 | 39.68 | 4 |
| Classification | 28.05 | 30.22 | - |
| Attribute | 2 | 4.34 | - |
| Relationship | 6 | 7.34 | - |
| Vocabulary | 207 | 713 | 9448 |
| Avg.# tokens/Q | 11.79 | 9.75 | 49.83 |
| Avg.# tokens/Des | 22.24 | 13.15 | 13.1 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| RoBERTa.C | 61.84 | 61.84 | 65.99 | 62.82 | 65.45 | 67.11 | 64.18 | 63.19 | 67.94 |
| RoBERTa.D | 60.53 | 60.53 | 72.11 | 55.13 | 68.18 | 61.84 | 63.05 | 61.83 | 67.99 |
| RoBERTa-MNLI | 67.11 | 67.11 | 78.23 | 64.10 | 70.50 | 67.76 | 69.14 | 68.98 | 70.02 |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| 1-T | 1-T1 | 1-T2 | 2-T | 2-T1 | 2-T2 | Avg.#T | Avg.#T1 | Avg.#T2 | |
| CNN-LSTM | 45.83 | 48.37 | 32.77 | 46.52 | 50.47 | 34.31 | 46.18 | 49.42 | 33.54 |
| CLIP-RoBERTa.D | 53.08 | 47.22 | 83.19 | 58.99 | 49.72 | 87.68 | 56.04 | 48.47 | 85.44 |
| ResNet-RoBERTa.D | 53.49 | 48.04 | 81.51 | 60.21 | 51.04 | 88.56 | 56.85 | 49.54 | 74.39 |
| CLIP-RoBERTa.C | 50.89 | 45.42 | 78.99 | 56.99 | 47.06 | 87.68 | 53.94 | 46.24 | 83.33 |
| ResNet-RoBERTa.C | 51.03 | 45.26 | 80.67 | 54.48 | 44.51 | 85.34 | 52.76 | 44.89 | 83.01 |
| ResNet-Enc | 63.61 | 59.31 | 85.71 | 64.31 | 56.17 | 89.44 | 63.96 | 57.74 | 87.58 |
| EN | ZH | |
| RoBERTa.D | 28.71 | 29.21 |
| RoBERTa.C | 32.67 | 30.19 |
| RoBERTa(TE).C | 33.17 | 30.19 |
| EnC | 34.15 | 34.15 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| w/o Box | 71.05 | 71.05 | 83.67 | 70.51 | 71.94 | 74.34 | 73.76 | 74.24 | 72.45 |
| w/o GNN | 76.32 | 73.68 | 82.99 | 69.23 | 76.26 | 74.34 | 75.47 | 76.69 | 71.47 |
| w/o Loss | 77.63 | 76.32 | 84.35 | 73.08 | 76.36 | 76.32 | 77.34 | 78.72 | 72.39 |
| backbone | domestic | teeth | arms |
| no | yes | no | yes |
| backbone | domestic | teeth | arms |
| no | yes | no | no |
| tail | domestic | length | legs |
| no | no | short | 4 |
| Licon-A | Licon-I | Licon-T | |
| Train concepts | 90 | 280 | 681 |
| Test concepts | 18 | 56 | 227 |
| Avg.# Q/Concept | 36.05 | 39.68 | 4 |
| Classification | 28.05 | 30.22 | - |
| Attribute | 2 | 4.34 | - |
| Relationship | 6 | 7.34 | - |
| Vocabulary | 207 | 713 | 9448 |
| Avg.# tokens/Q | 11.79 | 9.75 | 49.83 |
| Avg.# tokens/Des | 22.24 | 13.15 | 13.1 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| RoBERTa.C | 61.84 | 61.84 | 65.99 | 62.82 | 65.45 | 67.11 | 64.18 | 63.19 | 67.94 |
| RoBERTa.D | 60.53 | 60.53 | 72.11 | 55.13 | 68.18 | 61.84 | 63.05 | 61.83 | 67.99 |
| RoBERTa-MNLI | 67.11 | 67.11 | 78.23 | 64.10 | 70.50 | 67.76 | 69.14 | 68.98 | 70.02 |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| 1-T | 1-T1 | 1-T2 | 2-T | 2-T1 | 2-T2 | Avg.#T | Avg.#T1 | Avg.#T2 | |
| CNN-LSTM | 45.83 | 48.37 | 32.77 | 46.52 | 50.47 | 34.31 | 46.18 | 49.42 | 33.54 |
| CLIP-RoBERTa.D | 53.08 | 47.22 | 83.19 | 58.99 | 49.72 | 87.68 | 56.04 | 48.47 | 85.44 |
| ResNet-RoBERTa.D | 53.49 | 48.04 | 81.51 | 60.21 | 51.04 | 88.56 | 56.85 | 49.54 | 74.39 |
| CLIP-RoBERTa.C | 50.89 | 45.42 | 78.99 | 56.99 | 47.06 | 87.68 | 53.94 | 46.24 | 83.33 |
| ResNet-RoBERTa.C | 51.03 | 45.26 | 80.67 | 54.48 | 44.51 | 85.34 | 52.76 | 44.89 | 83.01 |
| ResNet-Enc | 63.61 | 59.31 | 85.71 | 64.31 | 56.17 | 89.44 | 63.96 | 57.74 | 87.58 |
| EN | ZH | |
| RoBERTa.D | 28.71 | 29.21 |
| RoBERTa.C | 32.67 | 30.19 |
| RoBERTa(TE).C | 33.17 | 30.19 |
| EnC | 34.15 | 34.15 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| w/o Box | 71.05 | 71.05 | 83.67 | 70.51 | 71.94 | 74.34 | 73.76 | 74.24 | 72.45 |
| w/o GNN | 76.32 | 73.68 | 82.99 | 69.23 | 76.26 | 74.34 | 75.47 | 76.69 | 71.47 |
| w/o Loss | 77.63 | 76.32 | 84.35 | 73.08 | 76.36 | 76.32 | 77.34 | 78.72 | 72.39 |
| backbone | domestic | teeth | arms |
| no | yes | no | yes |
| backbone | domestic | teeth | arms |
| no | yes | no | no |
| tail | domestic | length | legs |
| no | no | short | 4 |
| Licon-A | Licon-I | Licon-T | |
| Train concepts | 90 | 280 | 681 |
| Test concepts | 18 | 56 | 227 |
| Avg.# Q/Concept | 36.05 | 39.68 | 4 |
| Classification | 28.05 | 30.22 | - |
| Attribute | 2 | 4.34 | - |
| Relationship | 6 | 7.34 | - |
| Vocabulary | 207 | 713 | 9448 |
| Avg.# tokens/Q | 11.79 | 9.75 | 49.83 |
| Avg.# tokens/Des | 22.24 | 13.15 | 13.1 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| RoBERTa.C | 61.84 | 61.84 | 65.99 | 62.82 | 65.45 | 67.11 | 64.18 | 63.19 | 67.94 |
| RoBERTa.D | 60.53 | 60.53 | 72.11 | 55.13 | 68.18 | 61.84 | 63.05 | 61.83 | 67.99 |
| RoBERTa-MNLI | 67.11 | 67.11 | 78.23 | 64.10 | 70.50 | 67.76 | 69.14 | 68.98 | 70.02 |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| 1-T | 1-T1 | 1-T2 | 2-T | 2-T1 | 2-T2 | Avg.#T | Avg.#T1 | Avg.#T2 | |
| CNN-LSTM | 45.83 | 48.37 | 32.77 | 46.52 | 50.47 | 34.31 | 46.18 | 49.42 | 33.54 |
| CLIP-RoBERTa.D | 53.08 | 47.22 | 83.19 | 58.99 | 49.72 | 87.68 | 56.04 | 48.47 | 85.44 |
| ResNet-RoBERTa.D | 53.49 | 48.04 | 81.51 | 60.21 | 51.04 | 88.56 | 56.85 | 49.54 | 74.39 |
| CLIP-RoBERTa.C | 50.89 | 45.42 | 78.99 | 56.99 | 47.06 | 87.68 | 53.94 | 46.24 | 83.33 |
| ResNet-RoBERTa.C | 51.03 | 45.26 | 80.67 | 54.48 | 44.51 | 85.34 | 52.76 | 44.89 | 83.01 |
| ResNet-Enc | 63.61 | 59.31 | 85.71 | 64.31 | 56.17 | 89.44 | 63.96 | 57.74 | 87.58 |
| EN | ZH | |
| RoBERTa.D | 28.71 | 29.21 |
| RoBERTa.C | 32.67 | 30.19 |
| RoBERTa(TE).C | 33.17 | 30.19 |
| EnC | 34.15 | 34.15 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| w/o Box | 71.05 | 71.05 | 83.67 | 70.51 | 71.94 | 74.34 | 73.76 | 74.24 | 72.45 |
| w/o GNN | 76.32 | 73.68 | 82.99 | 69.23 | 76.26 | 74.34 | 75.47 | 76.69 | 71.47 |
| w/o Loss | 77.63 | 76.32 | 84.35 | 73.08 | 76.36 | 76.32 | 77.34 | 78.72 | 72.39 |
| backbone | domestic | teeth | arms |
| no | yes | no | yes |
| backbone | domestic | teeth | arms |
| no | yes | no | no |
| tail | domestic | length | legs |
| no | no | short | 4 |
| Licon-A | Licon-I | Licon-T | |
| Train concepts | 90 | 280 | 681 |
| Test concepts | 18 | 56 | 227 |
| Avg.# Q/Concept | 36.05 | 39.68 | 4 |
| Classification | 28.05 | 30.22 | - |
| Attribute | 2 | 4.34 | - |
| Relationship | 6 | 7.34 | - |
| Vocabulary | 207 | 713 | 9448 |
| Avg.# tokens/Q | 11.79 | 9.75 | 49.83 |
| Avg.# tokens/Des | 22.24 | 13.15 | 13.1 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| RoBERTa.C | 61.84 | 61.84 | 65.99 | 62.82 | 65.45 | 67.11 | 64.18 | 63.19 | 67.94 |
| RoBERTa.D | 60.53 | 60.53 | 72.11 | 55.13 | 68.18 | 61.84 | 63.05 | 61.83 | 67.99 |
| RoBERTa-MNLI | 67.11 | 67.11 | 78.23 | 64.10 | 70.50 | 67.76 | 69.14 | 68.98 | 70.02 |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| 1-T | 1-T1 | 1-T2 | 2-T | 2-T1 | 2-T2 | Avg.#T | Avg.#T1 | Avg.#T2 | |
| CNN-LSTM | 45.83 | 48.37 | 32.77 | 46.52 | 50.47 | 34.31 | 46.18 | 49.42 | 33.54 |
| CLIP-RoBERTa.D | 53.08 | 47.22 | 83.19 | 58.99 | 49.72 | 87.68 | 56.04 | 48.47 | 85.44 |
| ResNet-RoBERTa.D | 53.49 | 48.04 | 81.51 | 60.21 | 51.04 | 88.56 | 56.85 | 49.54 | 74.39 |
| CLIP-RoBERTa.C | 50.89 | 45.42 | 78.99 | 56.99 | 47.06 | 87.68 | 53.94 | 46.24 | 83.33 |
| ResNet-RoBERTa.C | 51.03 | 45.26 | 80.67 | 54.48 | 44.51 | 85.34 | 52.76 | 44.89 | 83.01 |
| ResNet-Enc | 63.61 | 59.31 | 85.71 | 64.31 | 56.17 | 89.44 | 63.96 | 57.74 | 87.58 |
| EN | ZH | |
| RoBERTa.D | 28.71 | 29.21 |
| RoBERTa.C | 32.67 | 30.19 |
| RoBERTa(TE).C | 33.17 | 30.19 |
| EnC | 34.15 | 34.15 |
| 1-T | 2-T | 3-T | 4-T | 5-T | 6-T | Avg.#T | Avg.#T1 | Avg.#T2 | |
| EnC | 78.74 | 77.63 | 85.71 | 75.64 | 77.70 | 77.63 | 78.84 | 80.86 | 72.91 |
| w/o Box | 71.05 | 71.05 | 83.67 | 70.51 | 71.94 | 74.34 | 73.76 | 74.24 | 72.45 |
| w/o GNN | 76.32 | 73.68 | 82.99 | 69.23 | 76.26 | 74.34 | 75.47 | 76.69 | 71.47 |
| w/o Loss | 77.63 | 76.32 | 84.35 | 73.08 | 76.36 | 76.32 | 77.34 | 78.72 | 72.39 |
| backbone | domestic | teeth | arms |
| no | yes | no | yes |
| backbone | domestic | teeth | arms |
| no | yes | no | no |
| tail | domestic | length | legs |
| no | no | short | 4 |
| # Source-Summary Pair (train/val/test) | Avg Source Length (sd) | Avg Summary Length (sd) | |
| WikiHow | 1060732/37932/41182 | 82.32 (60.10) | 8.28 (5.27) |
| XSum | 204045/11332/11334 | 433.05 (355.51) | 23.19 (5.81) |
| SciTLDR-Auth | 1992/618/618 | 1142.09 (469.71) | 20.34 (8.38) |
| SciTLDR-Peer | 1992/834/1349 | 1142.09 (469.71) | 23.17 (7.65) |
| Feature | Regression Coefficient (Prediction Accuracy) | |||
| WikiHow | XSum | SciTLDR-Auth | SciTLDR-Peer | |
| # Morphemes | -3.842 (0.637) | -2.877 (0.595) | -5.290 (0.705) | -7.404 (0.690) |
| Word Length | -0.192 (0.531) | -0.274 (0.512) | -2.452 (0.768) | -2.218 (0.724) |
| # Unique POS | 0.722 (0.858) | -0.170 (0.573) | 0.333 (0.695) | 0.401(0.720) |
| Tree Height | 0.678 (0.810) | -0.152 (0.593) | 0.194 (0.632) | -0.128 (0.573) |
| Avg # Modifiers | 0.034 (0.498) | -0.299 (0.590) | 0.004 (0.520) | 0.046 (0.493) |
| Avg Dependency Length | 2.369 (0.746) | -0.174 (0.518) | 0.822 (0.604) | 0.756 (0.610) |
| Type-Token Ratio | -22.308 (0.772) | -3.938 (0.547) | -4.945 (0.661) | -0.787 (0.525) |
| Word Specificity | -0.155 (0.554) | -0.500 (0.587) | -0.107 (0.501) | -1.448 (0.597) |
| Dataset | Regression Coefficient | Accuracy | ||
| # Morphemes | Tree Height | Type-Token Ratio | ||
| WikiHow | -5.463 | 34.862 | -20.857 | 0.916 |
| XSum | -6.247 | -7.190 | -2.550 | 0.657 |
| SciTLDR-Auth | -5.579 | 1.472 | -3.714 | 0.779 |
| SciTLDR-Peer | -8.344 | -3.213 | -5.766 | 0.795 |
| Feature | Dataset | Mean | Median | Q1 | Q3 |
| # Morphemes | WikiHow | -0.082 | -0.055 | -0.178 | 0.059 |
| XSum | -0.040 | -0.025 | -0.126 | 0.063 | |
| SciTLDR-Auth | -0.111 | -0.069 | -0.188 | 0.016 | |
| SciTLDR-Peer | -0.814 | -0.729 | -1.322 | -0.202 | |
| Word Length | WikiHow | -0.101 | -0.054 | -0.593 | 0.434 |
| XSum | -0.055 | -0.018 | -0.419 | 0.338 | |
| SciTLDR-Auth | -0.087 | -0.070 | -0.158 | 0.013 | |
| SciTLDR-Peer | -0.587 | -0.557 | -1.008 | -0.097 | |
| # Unique POS | WikiHow | 4.867 | 5.167 | 3.000 | 7.000 |
| XSum | -0.880 | -0.800 | -3.000 | 1.000 | |
| SciTLDR-Auth | 2.019 | 2.000 | -0.099 | 4.467 | |
| SciTLDR-Peer | 1.747 | 1.856 | 0.000 | 3.744 | |
| Tree Height | WikiHow | 3.563 | 3.750 | 2.000 | 5.500 |
| XSum | -1.007 | -0.800 | -3.200 | 1.250 | |
| SciTLDR-Auth | 1.123 | 1.232 | -0.822 | 3.462 | |
| SciTLDR-Peer | -0.772 | -0.636 | -2.625 | 1.529 | |
| Avg # Modifiers | WikiHow | 0.034 | 0.000 | 0.000 | 0.000 |
| XSum | -0.268 | -0.050 | -0.875 | 0.125 | |
| SciTLDR-Auth | 0.003 | 0.000 | -0.444 | 0.500 | |
| SciTLDR-Peer | 0.040 | 0.000 | -0.500 | 0.564 | |
| Avg Dependency Length | WikiHow | 0.497 | 0.532 | 0.126 | 0.902 |
| XSum | -0.026 | 0.000 | -0.361 | 0.329 | |
| SciTLDR-Auth | 0.178 | 0.236 | -0.171 | 0.622 | |
| SciTLDR-Peer | 0.161 | 0.257 | -0.198 | 0.626 | |
| Type-Token Ratio | WikiHow | -0.052 | -0.055 | -0.087 | -0.024 |
| XSum | -0.008 | -0.012 | -0.048 | 0.028 | |
| SciTLDR-Auth | -0.030 | -0.044 | -0.085 | 0.011 | |
| SciTLDR-Peer | -0.003 | -0.009 | -0.062 | 0.043 | |
| Word Specificity | WikiHow | -0.058 | -0.039 | -0.456 | 0.366 |
| XSum | -0.077 | -0.065 | -0.334 | 0.209 | |
| SciTLDR-Auth | -0.080 | -0.085 | -0.309 | 0.168 | |
| SciTLDR-Peer | -0.126 | -0.103 | -0.324 | 0.106 |
| # Source-Summary Pair (train/val/test) | Avg Source Length (sd) | Avg Summary Length (sd) | |
| WikiHow | 1060732/37932/41182 | 82.32 (60.10) | 8.28 (5.27) |
| XSum | 204045/11332/11334 | 433.05 (355.51) | 23.19 (5.81) |
| SciTLDR-Auth | 1992/618/618 | 1142.09 (469.71) | 20.34 (8.38) |
| SciTLDR-Peer | 1992/834/1349 | 1142.09 (469.71) | 23.17 (7.65) |
| Feature | Regression Coefficient (Prediction Accuracy) | |||
| WikiHow | XSum | SciTLDR-Auth | SciTLDR-Peer | |
| # Morphemes | -3.842 (0.637) | -2.877 (0.595) | -5.290 (0.705) | -7.404 (0.690) |
| Word Length | -0.192 (0.531) | -0.274 (0.512) | -2.452 (0.768) | -2.218 (0.724) |
| # Unique POS | 0.722 (0.858) | -0.170 (0.573) | 0.333 (0.695) | 0.401(0.720) |
| Tree Height | 0.678 (0.810) | -0.152 (0.593) | 0.194 (0.632) | -0.128 (0.573) |
| Avg # Modifiers | 0.034 (0.498) | -0.299 (0.590) | 0.004 (0.520) | 0.046 (0.493) |
| Avg Dependency Length | 2.369 (0.746) | -0.174 (0.518) | 0.822 (0.604) | 0.756 (0.610) |
| Type-Token Ratio | -22.308 (0.772) | -3.938 (0.547) | -4.945 (0.661) | -0.787 (0.525) |
| Word Specificity | -0.155 (0.554) | -0.500 (0.587) | -0.107 (0.501) | -1.448 (0.597) |
| Dataset | Regression Coefficient | Accuracy | ||
| # Morphemes | Tree Height | Type-Token Ratio | ||
| WikiHow | -5.463 | 34.862 | -20.857 | 0.916 |
| XSum | -6.247 | -7.190 | -2.550 | 0.657 |
| SciTLDR-Auth | -5.579 | 1.472 | -3.714 | 0.779 |
| SciTLDR-Peer | -8.344 | -3.213 | -5.766 | 0.795 |
| Feature | Dataset | Mean | Median | Q1 | Q3 |
| # Morphemes | WikiHow | -0.082 | -0.055 | -0.178 | 0.059 |
| XSum | -0.040 | -0.025 | -0.126 | 0.063 | |
| SciTLDR-Auth | -0.111 | -0.069 | -0.188 | 0.016 | |
| SciTLDR-Peer | -0.814 | -0.729 | -1.322 | -0.202 | |
| Word Length | WikiHow | -0.101 | -0.054 | -0.593 | 0.434 |
| XSum | -0.055 | -0.018 | -0.419 | 0.338 | |
| SciTLDR-Auth | -0.087 | -0.070 | -0.158 | 0.013 | |
| SciTLDR-Peer | -0.587 | -0.557 | -1.008 | -0.097 | |
| # Unique POS | WikiHow | 4.867 | 5.167 | 3.000 | 7.000 |
| XSum | -0.880 | -0.800 | -3.000 | 1.000 | |
| SciTLDR-Auth | 2.019 | 2.000 | -0.099 | 4.467 | |
| SciTLDR-Peer | 1.747 | 1.856 | 0.000 | 3.744 | |
| Tree Height | WikiHow | 3.563 | 3.750 | 2.000 | 5.500 |
| XSum | -1.007 | -0.800 | -3.200 | 1.250 | |
| SciTLDR-Auth | 1.123 | 1.232 | -0.822 | 3.462 | |
| SciTLDR-Peer | -0.772 | -0.636 | -2.625 | 1.529 | |
| Avg # Modifiers | WikiHow | 0.034 | 0.000 | 0.000 | 0.000 |
| XSum | -0.268 | -0.050 | -0.875 | 0.125 | |
| SciTLDR-Auth | 0.003 | 0.000 | -0.444 | 0.500 | |
| SciTLDR-Peer | 0.040 | 0.000 | -0.500 | 0.564 | |
| Avg Dependency Length | WikiHow | 0.497 | 0.532 | 0.126 | 0.902 |
| XSum | -0.026 | 0.000 | -0.361 | 0.329 | |
| SciTLDR-Auth | 0.178 | 0.236 | -0.171 | 0.622 | |
| SciTLDR-Peer | 0.161 | 0.257 | -0.198 | 0.626 | |
| Type-Token Ratio | WikiHow | -0.052 | -0.055 | -0.087 | -0.024 |
| XSum | -0.008 | -0.012 | -0.048 | 0.028 | |
| SciTLDR-Auth | -0.030 | -0.044 | -0.085 | 0.011 | |
| SciTLDR-Peer | -0.003 | -0.009 | -0.062 | 0.043 | |
| Word Specificity | WikiHow | -0.058 | -0.039 | -0.456 | 0.366 |
| XSum | -0.077 | -0.065 | -0.334 | 0.209 | |
| SciTLDR-Auth | -0.080 | -0.085 | -0.309 | 0.168 | |
| SciTLDR-Peer | -0.126 | -0.103 | -0.324 | 0.106 |
| # Source-Summary Pair (train/val/test) | Avg Source Length (sd) | Avg Summary Length (sd) | |
| WikiHow | 1060732/37932/41182 | 82.32 (60.10) | 8.28 (5.27) |
| XSum | 204045/11332/11334 | 433.05 (355.51) | 23.19 (5.81) |
| SciTLDR-Auth | 1992/618/618 | 1142.09 (469.71) | 20.34 (8.38) |
| SciTLDR-Peer | 1992/834/1349 | 1142.09 (469.71) | 23.17 (7.65) |
| Feature | Regression Coefficient (Prediction Accuracy) | |||
| WikiHow | XSum | SciTLDR-Auth | SciTLDR-Peer | |
| # Morphemes | -3.842 (0.637) | -2.877 (0.595) | -5.290 (0.705) | -7.404 (0.690) |
| Word Length | -0.192 (0.531) | -0.274 (0.512) | -2.452 (0.768) | -2.218 (0.724) |
| # Unique POS | 0.722 (0.858) | -0.170 (0.573) | 0.333 (0.695) | 0.401(0.720) |
| Tree Height | 0.678 (0.810) | -0.152 (0.593) | 0.194 (0.632) | -0.128 (0.573) |
| Avg # Modifiers | 0.034 (0.498) | -0.299 (0.590) | 0.004 (0.520) | 0.046 (0.493) |
| Avg Dependency Length | 2.369 (0.746) | -0.174 (0.518) | 0.822 (0.604) | 0.756 (0.610) |
| Type-Token Ratio | -22.308 (0.772) | -3.938 (0.547) | -4.945 (0.661) | -0.787 (0.525) |
| Word Specificity | -0.155 (0.554) | -0.500 (0.587) | -0.107 (0.501) | -1.448 (0.597) |
| Dataset | Regression Coefficient | Accuracy | ||
| # Morphemes | Tree Height | Type-Token Ratio | ||
| WikiHow | -5.463 | 34.862 | -20.857 | 0.916 |
| XSum | -6.247 | -7.190 | -2.550 | 0.657 |
| SciTLDR-Auth | -5.579 | 1.472 | -3.714 | 0.779 |
| SciTLDR-Peer | -8.344 | -3.213 | -5.766 | 0.795 |
| Feature | Dataset | Mean | Median | Q1 | Q3 |
| # Morphemes | WikiHow | -0.082 | -0.055 | -0.178 | 0.059 |
| XSum | -0.040 | -0.025 | -0.126 | 0.063 | |
| SciTLDR-Auth | -0.111 | -0.069 | -0.188 | 0.016 | |
| SciTLDR-Peer | -0.814 | -0.729 | -1.322 | -0.202 | |
| Word Length | WikiHow | -0.101 | -0.054 | -0.593 | 0.434 |
| XSum | -0.055 | -0.018 | -0.419 | 0.338 | |
| SciTLDR-Auth | -0.087 | -0.070 | -0.158 | 0.013 | |
| SciTLDR-Peer | -0.587 | -0.557 | -1.008 | -0.097 | |
| # Unique POS | WikiHow | 4.867 | 5.167 | 3.000 | 7.000 |
| XSum | -0.880 | -0.800 | -3.000 | 1.000 | |
| SciTLDR-Auth | 2.019 | 2.000 | -0.099 | 4.467 | |
| SciTLDR-Peer | 1.747 | 1.856 | 0.000 | 3.744 | |
| Tree Height | WikiHow | 3.563 | 3.750 | 2.000 | 5.500 |
| XSum | -1.007 | -0.800 | -3.200 | 1.250 | |
| SciTLDR-Auth | 1.123 | 1.232 | -0.822 | 3.462 | |
| SciTLDR-Peer | -0.772 | -0.636 | -2.625 | 1.529 | |
| Avg # Modifiers | WikiHow | 0.034 | 0.000 | 0.000 | 0.000 |
| XSum | -0.268 | -0.050 | -0.875 | 0.125 | |
| SciTLDR-Auth | 0.003 | 0.000 | -0.444 | 0.500 | |
| SciTLDR-Peer | 0.040 | 0.000 | -0.500 | 0.564 | |
| Avg Dependency Length | WikiHow | 0.497 | 0.532 | 0.126 | 0.902 |
| XSum | -0.026 | 0.000 | -0.361 | 0.329 | |
| SciTLDR-Auth | 0.178 | 0.236 | -0.171 | 0.622 | |
| SciTLDR-Peer | 0.161 | 0.257 | -0.198 | 0.626 | |
| Type-Token Ratio | WikiHow | -0.052 | -0.055 | -0.087 | -0.024 |
| XSum | -0.008 | -0.012 | -0.048 | 0.028 | |
| SciTLDR-Auth | -0.030 | -0.044 | -0.085 | 0.011 | |
| SciTLDR-Peer | -0.003 | -0.009 | -0.062 | 0.043 | |
| Word Specificity | WikiHow | -0.058 | -0.039 | -0.456 | 0.366 |
| XSum | -0.077 | -0.065 | -0.334 | 0.209 | |
| SciTLDR-Auth | -0.080 | -0.085 | -0.309 | 0.168 | |
| SciTLDR-Peer | -0.126 | -0.103 | -0.324 | 0.106 |
| # Source-Summary Pair (train/val/test) | Avg Source Length (sd) | Avg Summary Length (sd) | |
| WikiHow | 1060732/37932/41182 | 82.32 (60.10) | 8.28 (5.27) |
| XSum | 204045/11332/11334 | 433.05 (355.51) | 23.19 (5.81) |
| SciTLDR-Auth | 1992/618/618 | 1142.09 (469.71) | 20.34 (8.38) |
| SciTLDR-Peer | 1992/834/1349 | 1142.09 (469.71) | 23.17 (7.65) |
| Feature | Regression Coefficient (Prediction Accuracy) | |||
| WikiHow | XSum | SciTLDR-Auth | SciTLDR-Peer | |
| # Morphemes | -3.842 (0.637) | -2.877 (0.595) | -5.290 (0.705) | -7.404 (0.690) |
| Word Length | -0.192 (0.531) | -0.274 (0.512) | -2.452 (0.768) | -2.218 (0.724) |
| # Unique POS | 0.722 (0.858) | -0.170 (0.573) | 0.333 (0.695) | 0.401(0.720) |
| Tree Height | 0.678 (0.810) | -0.152 (0.593) | 0.194 (0.632) | -0.128 (0.573) |
| Avg # Modifiers | 0.034 (0.498) | -0.299 (0.590) | 0.004 (0.520) | 0.046 (0.493) |
| Avg Dependency Length | 2.369 (0.746) | -0.174 (0.518) | 0.822 (0.604) | 0.756 (0.610) |
| Type-Token Ratio | -22.308 (0.772) | -3.938 (0.547) | -4.945 (0.661) | -0.787 (0.525) |
| Word Specificity | -0.155 (0.554) | -0.500 (0.587) | -0.107 (0.501) | -1.448 (0.597) |
| Dataset | Regression Coefficient | Accuracy | ||
| # Morphemes | Tree Height | Type-Token Ratio | ||
| WikiHow | -5.463 | 34.862 | -20.857 | 0.916 |
| XSum | -6.247 | -7.190 | -2.550 | 0.657 |
| SciTLDR-Auth | -5.579 | 1.472 | -3.714 | 0.779 |
| SciTLDR-Peer | -8.344 | -3.213 | -5.766 | 0.795 |
| Feature | Dataset | Mean | Median | Q1 | Q3 |
| # Morphemes | WikiHow | -0.082 | -0.055 | -0.178 | 0.059 |
| XSum | -0.040 | -0.025 | -0.126 | 0.063 | |
| SciTLDR-Auth | -0.111 | -0.069 | -0.188 | 0.016 | |
| SciTLDR-Peer | -0.814 | -0.729 | -1.322 | -0.202 | |
| Word Length | WikiHow | -0.101 | -0.054 | -0.593 | 0.434 |
| XSum | -0.055 | -0.018 | -0.419 | 0.338 | |
| SciTLDR-Auth | -0.087 | -0.070 | -0.158 | 0.013 | |
| SciTLDR-Peer | -0.587 | -0.557 | -1.008 | -0.097 | |
| # Unique POS | WikiHow | 4.867 | 5.167 | 3.000 | 7.000 |
| XSum | -0.880 | -0.800 | -3.000 | 1.000 | |
| SciTLDR-Auth | 2.019 | 2.000 | -0.099 | 4.467 | |
| SciTLDR-Peer | 1.747 | 1.856 | 0.000 | 3.744 | |
| Tree Height | WikiHow | 3.563 | 3.750 | 2.000 | 5.500 |
| XSum | -1.007 | -0.800 | -3.200 | 1.250 | |
| SciTLDR-Auth | 1.123 | 1.232 | -0.822 | 3.462 | |
| SciTLDR-Peer | -0.772 | -0.636 | -2.625 | 1.529 | |
| Avg # Modifiers | WikiHow | 0.034 | 0.000 | 0.000 | 0.000 |
| XSum | -0.268 | -0.050 | -0.875 | 0.125 | |
| SciTLDR-Auth | 0.003 | 0.000 | -0.444 | 0.500 | |
| SciTLDR-Peer | 0.040 | 0.000 | -0.500 | 0.564 | |
| Avg Dependency Length | WikiHow | 0.497 | 0.532 | 0.126 | 0.902 |
| XSum | -0.026 | 0.000 | -0.361 | 0.329 | |
| SciTLDR-Auth | 0.178 | 0.236 | -0.171 | 0.622 | |
| SciTLDR-Peer | 0.161 | 0.257 | -0.198 | 0.626 | |
| Type-Token Ratio | WikiHow | -0.052 | -0.055 | -0.087 | -0.024 |
| XSum | -0.008 | -0.012 | -0.048 | 0.028 | |
| SciTLDR-Auth | -0.030 | -0.044 | -0.085 | 0.011 | |
| SciTLDR-Peer | -0.003 | -0.009 | -0.062 | 0.043 | |
| Word Specificity | WikiHow | -0.058 | -0.039 | -0.456 | 0.366 |
| XSum | -0.077 | -0.065 | -0.334 | 0.209 | |
| SciTLDR-Auth | -0.080 | -0.085 | -0.309 | 0.168 | |
| SciTLDR-Peer | -0.126 | -0.103 | -0.324 | 0.106 |
| Experiment | Sign | Phrase | Efficiency | ||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | ||
| E0 | Moryossef et al. (2020) | — | 0.46 | 1.09 | — | 0.70 | 1.00 | 102K | 0:50:17 |
| E1 | Baseline | 0.56 | 0.66 | 0.91 | 0.59 | 0.80 | 2.50 | 454K | 1:01:50 |
| E2 | E1 + Face | 0.53 | 0.58 | 0.64 | 0.57 | 0.76 | 1.87 | 552K | 1:50:31 |
| E3 | E1 + Optical Flow | 0.58 | 0.62 | 1.12 | 0.60 | 0.82 | 3.19 | 473K | 1:20:17 |
| E4 | E3 + Hand Norm | 0.56 | 0.61 | 1.07 | 0.60 | 0.80 | 3.24 | 516K | 1:30:59 |
| E1s | E1 + Depth=4 | 0.63 | 0.69 | 1.11 | 0.65 | 0.82 | 1.63 | 1.6M | 4:08:48 |
| E2s | E2 + Depth=4 | 0.62 | 0.69 | 1.07 | 0.63 | 0.84 | 2.68 | 1.7M | 3:14:03 |
| E3s | E3 + Depth=4 | 0.60 | 0.63 | 1.13 | 0.64 | 0.80 | 1.53 | 1.7M | 4:08:30 |
| E4s | E4 + Depth=4 | 0.59 | 0.63 | 1.13 | 0.62 | 0.79 | 1.43 | 1.7M | 4:35:29 |
| E1s* | E1s + Tuned Decoding | — | 0.69 | 1.03 | — | 0.85 | 1.02 | — | — |
| E4s* | E4s + Tuned Decoding | — | 0.63 | 1.06 | — | 0.79 | 1.12 | — | — |
| E5 | E4s + Autoregressive | 0.45 | 0.47 | 0.88 | 0.52 | 0.63 | 2.72 | 1.3M | ~3 days |
| Data | Model | ROC-AUC | F1-M |
| LSF | full (theirs) | 0.87 | — |
| body (theirs) | 0.87 | — | |
| E1s (ours, zero-shot) | 0.71 | 0.41 | |
| E4s (ours, zero-shot) | 0.76 | 0.44 | |
| E1s (ours, trained) | 0.87 | 0.49 | |
| E4s (ours, trained) | 0.87 | 0.51 | |
| DGS | E1s (ours) | 0.91 | 0.65 |
| E4s (ours) | 0.90 | 0.62 |
| Experiment | Sign | Phrase | ||||||
| F1 | IoU | % | F1 | IoU | % | |||
| P0 | Moryossef et al. (2020) | test | — | 0.4 | 1.45 | — | 0.65 | 0.82 |
| dev | — | 0.35 | 1.36 | — | 0.6 | 0.77 | ||
| P0.1 | P0 + Holistic 25fps | test | — | 0.39 | 0.86 | — | 0.64 | 0.5 |
| dev | — | 0.32 | 0.81 | — | 0.58 | 0.52 | ||
| P1 | P1 baseline | test | 0.55 | 0.49 | 0.83 | 0.6 | 0.67 | 2.63 |
| dev | 0.56 | 0.43 | 0.75 | 0.58 | 0.62 | 2.61 | ||
| P1.1 | P1 - encoder_bidirectional | test | 0.48 | 0.45 | 0.68 | 0.5 | 0.64 | 2.68 |
| dev | 0.46 | 0.41 | 0.64 | 0.51 | 0.61 | 2.56 | ||
| P1.2.1 | P1 + hidden_dim=512 | test | 0.47 | 0.42 | 0.44 | 0.52 | 0.63 | 1.7 |
| dev | 0.46 | 0.4 | 0.43 | 0.52 | 0.61 | 1.69 | ||
| P1.2.2 | P1 + hidden_dim=1024 | test | 0.48 | 0.45 | 0.42 | 0.58 | 0.65 | 1.53 |
| dev | 0.46 | 0.41 | 0.36 | 0.53 | 0.61 | 1.49 | ||
| P1.3.1 | P1 + encoder_depth=2 | test | 0.55 | 0.48 | 0.76 | 0.58 | 0.67 | 2.56 |
| dev | 0.56 | 0.43 | 0.69 | 0.58 | 0.62 | 2.52 | ||
| P1.3.2 | P1 + encoder_depth=4 | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.67 | 1.41 |
| dev | 0.61 | 0.47 | 0.84 | 0.64 | 0.6 | 1.39 | ||
| P1.4.1 | P1 + hidden_dim=128 + encoder_depth=2 | test | 0.58 | 0.48 | 0.8 | 0.6 | 0.67 | 2.0 |
| dev | 0.55 | 0.43 | 0.75 | 0.54 | 0.62 | 2.03 | ||
| P1.4.2 | P1 + hidden_dim=128 + encoder_depth=4 | test | 0.62 | 0.51 | 0.91 | 0.64 | 0.68 | 2.43 |
| dev | 0.6 | 0.47 | 0.83 | 0.6 | 0.62 | 2.57 | ||
| P1.4.3 | P1 + hidden_dim=128 + encoder_depth=8 | test | 0.59 | 0.52 | 0.91 | 0.63 | 0.68 | 3.04 |
| dev | 0.6 | 0.47 | 0.84 | 0.6 | 0.62 | 3.02 | ||
| P1.5.1 | P1 + hidden_dim=64 + encoder_depth=4 | test | 0.57 | 0.5 | 0.8 | 0.6 | 0.68 | 2.41 |
| dev | 0.58 | 0.45 | 0.75 | 0.59 | 0.62 | 2.39 | ||
| P1.5.2 | P1 + hidden_dim=64 + encoder_depth=8 | test | 0.62 | 0.51 | 0.85 | 0.64 | 0.68 | 2.53 |
| dev | 0.6 | 0.46 | 0.79 | 0.6 | 0.62 | 2.53 | ||
| P2 | P1 + optical_flow | test | 0.58 | 0.5 | 0.95 | 0.63 | 0.68 | 3.17 |
| dev | 0.59 | 0.45 | 0.84 | 0.59 | 0.61 | 3.08 | ||
| P2.1 | P1.3.2 + optical_flow | test | 0.63 | 0.51 | 0.92 | 0.66 | 0.67 | 1.51 |
| dev | 0.62 | 0.46 | 0.81 | 0.62 | 0.6 | 1.53 | ||
| P3 | P1 + hand_normalization | test | 0.55 | 0.48 | 0.77 | 0.58 | 0.67 | 2.79 |
| dev | 0.55 | 0.42 | 0.71 | 0.57 | 0.62 | 2.73 | ||
| P3.1 | P1.3.2 + hand_normalization | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.66 | 1.43 |
| dev | 0.61 | 0.46 | 0.82 | 0.64 | 0.61 | 1.46 | ||
| P4 | P2.1 + P3.1 | test | 0.56 | 0.51 | 0.92 | 0.61 | 0.66 | 1.45 |
| dev | 0.61 | 0.46 | 0.81 | 0.63 | 0.6 | 1.41 | ||
| P4.1 | P4 + encoder_depth=8 | test | 0.6 | 0.51 | 0.95 | 0.62 | 0.67 | 1.08 |
| dev | 0.61 | 0.47 | 0.86 | 0.62 | 0.6 | 1.12 | ||
| P5 | P1.3.2 + reduced_face | test | 0.63 | 0.51 | 0.94 | 0.64 | 0.66 | 1.16 |
| dev | 0.61 | 0.47 | 0.86 | 0.64 | 0.58 | 1.14 | ||
| P5.1 | P1.3.2 + full_face | test | 0.54 | 0.49 | 0.8 | 0.6 | 0.68 | 2.29 |
| dev | 0.57 | 0.45 | 0.7 | 0.59 | 0.62 | 2.29 |
| Experiment | Sign | Phrase | Efficiency | |||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | |||
| E0 | Moryossef et al. (2020) | test | — | 0.46 ± 0.03 | 1.09 ± 0.41 | — | 0.70 ± 0.01 | 1.00 ± 0.06 | 102K | 0:50:17 |
| dev | — | 0.42 ± 0.05 | 1.21 ± 0.59 | — | 0.61 ± 0.06 | 2.47 ± 0.85 | 102K | 0:50:17 | ||
| E1 | Baseline | test | 0.56 ± 0.03 | 0.66 ± 0.01 | 0.91 ± 0.05 | 0.59 ± 0.02 | 0.80 ± 0.03 | 2.50 ± 0.13 | 454K | 1:01:50 |
| dev | 0.55 ± 0.01 | 0.59 ± 0.00 | 1.12 ± 0.11 | 0.56 ± 0.02 | 0.75 ± 0.05 | 2.94 ± 0.08 | 454K | 1:01:50 | ||
| E2 | E1 + Face | test | 0.53 ± 0.05 | 0.58 ± 0.07 | 0.64 ± 0.30 | 0.57 ± 0.02 | 0.76 ± 0.03 | 1.87 ± 0.83 | 552K | 1:50:31 |
| dev | 0.50 ± 0.07 | 0.53 ± 0.11 | 0.90 ± 0.19 | 0.53 ± 0.05 | 0.71 ± 0.07 | 2.43 ± 1.02 | 552K | 1:50:31 | ||
| E3 | E1 + Optical Flow | test | 0.58 ± 0.01 | 0.62 ± 0.00 | 1.12 ± 0.05 | 0.60 ± 0.03 | 0.82 ± 0.03 | 3.19 ± 0.11 | 473K | 1:20:17 |
| dev | 0.58 ± 0.00 | 0.62 ± 0.00 | 1.50 ± 0.19 | 0.59 ± 0.01 | 0.79 ± 0.00 | 3.94 ± 0.14 | 473K | 1:20:17 | ||
| E4 | E3 + Hand Norm | test | 0.56 ± 0.02 | 0.61 ± 0.00 | 1.07 ± 0.05 | 0.60 ± 0.00 | 0.80 ± 0.00 | 3.24 ± 0.17 | 516K | 1:30:59 |
| dev | 0.57 ± 0.01 | 0.61 ± 0.01 | 1.50 ± 0.07 | 0.58 ± 0.00 | 0.79 ± 0.00 | 4.04 ± 0.31 | 516K | 1:30:59 | ||
| E1s | E1 + Depth=4 | test | 0.63 ± 0.01 | 0.69 ± 0.00 | 1.11 ± 0.01 | 0.65 ± 0.02 | 0.82 ± 0.04 | 1.63 ± 0.10 | 1.6M | 4:08:48 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.27 ± 0.01 | 0.63 ± 0.01 | 0.77 ± 0.01 | 2.17 ± 0.18 | 1.6M | 4:08:48 | ||
| E2s | E2 + Depth=4 | test | 0.62 ± 0.02 | 0.69 ± 0.00 | 1.07 ± 0.03 | 0.63 ± 0.01 | 0.84 ± 0.03 | 2.68 ± 0.53 | 1.7M | 3:14:03 |
| dev | 0.60 ± 0.01 | 0.63 ± 0.01 | 1.20 ± 0.12 | 0.59 ± 0.02 | 0.76 ± 0.05 | 3.30 ± 0.62 | 1.7M | 3:14:03 | ||
| E3s | E3 + Depth=4 | test | 0.60 ± 0.01 | 0.63 ± 0.00 | 1.13 ± 0.01 | 0.64 ± 0.03 | 0.80 ± 0.03 | 1.53 ± 0.18 | 1.7M | 4:08:30 |
| dev | 0.62 ± 0.00 | 0.63 ± 0.00 | 1.63 ± 0.05 | 0.63 ± 0.00 | 0.76 ± 0.00 | 2.14 ± 0.09 | 1.7M | 4:08:30 | ||
| E4s | E4 + Depth=4 | test | 0.59 ± 0.00 | 0.63 ± 0.00 | 1.13 ± 0.03 | 0.62 ± 0.00 | 0.79 ± 0.00 | 1.43 ± 0.10 | 1.7M | 4:35:29 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.56 ± 0.04 | 0.63 ± 0.00 | 0.77 ± 0.01 | 1.89 ± 0.07 | 1.7M | 4:35:29 | ||
| E4ba | E4s + Autoregressive | test | 0.45 ± 0.03 | 0.47 ± 0.05 | 0.88 ± 0.08 | 0.52 ± 0.02 | 0.63 ± 0.10 | 2.72 ± 1.33 | 1.3M | 2 days, 21:28:42 |
| dev | 0.40 ± 0.01 | 0.40 ± 0.01 | 2.02 ± 0.73 | 0.47 ± 0.00 | 0.57 ± 0.04 | 4.26 ± 1.26 | 1.3M | 2 days, 21:28:42 |
| Experiment | Sign | Phrase | Efficiency | ||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | ||
| E0 | Moryossef et al. (2020) | — | 0.46 | 1.09 | — | 0.70 | 1.00 | 102K | 0:50:17 |
| E1 | Baseline | 0.56 | 0.66 | 0.91 | 0.59 | 0.80 | 2.50 | 454K | 1:01:50 |
| E2 | E1 + Face | 0.53 | 0.58 | 0.64 | 0.57 | 0.76 | 1.87 | 552K | 1:50:31 |
| E3 | E1 + Optical Flow | 0.58 | 0.62 | 1.12 | 0.60 | 0.82 | 3.19 | 473K | 1:20:17 |
| E4 | E3 + Hand Norm | 0.56 | 0.61 | 1.07 | 0.60 | 0.80 | 3.24 | 516K | 1:30:59 |
| E1s | E1 + Depth=4 | 0.63 | 0.69 | 1.11 | 0.65 | 0.82 | 1.63 | 1.6M | 4:08:48 |
| E2s | E2 + Depth=4 | 0.62 | 0.69 | 1.07 | 0.63 | 0.84 | 2.68 | 1.7M | 3:14:03 |
| E3s | E3 + Depth=4 | 0.60 | 0.63 | 1.13 | 0.64 | 0.80 | 1.53 | 1.7M | 4:08:30 |
| E4s | E4 + Depth=4 | 0.59 | 0.63 | 1.13 | 0.62 | 0.79 | 1.43 | 1.7M | 4:35:29 |
| E1s* | E1s + Tuned Decoding | — | 0.69 | 1.03 | — | 0.85 | 1.02 | — | — |
| E4s* | E4s + Tuned Decoding | — | 0.63 | 1.06 | — | 0.79 | 1.12 | — | — |
| E5 | E4s + Autoregressive | 0.45 | 0.47 | 0.88 | 0.52 | 0.63 | 2.72 | 1.3M | ~3 days |
| Data | Model | ROC-AUC | F1-M |
| LSF | full (theirs) | 0.87 | — |
| body (theirs) | 0.87 | — | |
| E1s (ours, zero-shot) | 0.71 | 0.41 | |
| E4s (ours, zero-shot) | 0.76 | 0.44 | |
| E1s (ours, trained) | 0.87 | 0.49 | |
| E4s (ours, trained) | 0.87 | 0.51 | |
| DGS | E1s (ours) | 0.91 | 0.65 |
| E4s (ours) | 0.90 | 0.62 |
| Experiment | Sign | Phrase | ||||||
| F1 | IoU | % | F1 | IoU | % | |||
| P0 | Moryossef et al. (2020) | test | — | 0.4 | 1.45 | — | 0.65 | 0.82 |
| dev | — | 0.35 | 1.36 | — | 0.6 | 0.77 | ||
| P0.1 | P0 + Holistic 25fps | test | — | 0.39 | 0.86 | — | 0.64 | 0.5 |
| dev | — | 0.32 | 0.81 | — | 0.58 | 0.52 | ||
| P1 | P1 baseline | test | 0.55 | 0.49 | 0.83 | 0.6 | 0.67 | 2.63 |
| dev | 0.56 | 0.43 | 0.75 | 0.58 | 0.62 | 2.61 | ||
| P1.1 | P1 - encoder_bidirectional | test | 0.48 | 0.45 | 0.68 | 0.5 | 0.64 | 2.68 |
| dev | 0.46 | 0.41 | 0.64 | 0.51 | 0.61 | 2.56 | ||
| P1.2.1 | P1 + hidden_dim=512 | test | 0.47 | 0.42 | 0.44 | 0.52 | 0.63 | 1.7 |
| dev | 0.46 | 0.4 | 0.43 | 0.52 | 0.61 | 1.69 | ||
| P1.2.2 | P1 + hidden_dim=1024 | test | 0.48 | 0.45 | 0.42 | 0.58 | 0.65 | 1.53 |
| dev | 0.46 | 0.41 | 0.36 | 0.53 | 0.61 | 1.49 | ||
| P1.3.1 | P1 + encoder_depth=2 | test | 0.55 | 0.48 | 0.76 | 0.58 | 0.67 | 2.56 |
| dev | 0.56 | 0.43 | 0.69 | 0.58 | 0.62 | 2.52 | ||
| P1.3.2 | P1 + encoder_depth=4 | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.67 | 1.41 |
| dev | 0.61 | 0.47 | 0.84 | 0.64 | 0.6 | 1.39 | ||
| P1.4.1 | P1 + hidden_dim=128 + encoder_depth=2 | test | 0.58 | 0.48 | 0.8 | 0.6 | 0.67 | 2.0 |
| dev | 0.55 | 0.43 | 0.75 | 0.54 | 0.62 | 2.03 | ||
| P1.4.2 | P1 + hidden_dim=128 + encoder_depth=4 | test | 0.62 | 0.51 | 0.91 | 0.64 | 0.68 | 2.43 |
| dev | 0.6 | 0.47 | 0.83 | 0.6 | 0.62 | 2.57 | ||
| P1.4.3 | P1 + hidden_dim=128 + encoder_depth=8 | test | 0.59 | 0.52 | 0.91 | 0.63 | 0.68 | 3.04 |
| dev | 0.6 | 0.47 | 0.84 | 0.6 | 0.62 | 3.02 | ||
| P1.5.1 | P1 + hidden_dim=64 + encoder_depth=4 | test | 0.57 | 0.5 | 0.8 | 0.6 | 0.68 | 2.41 |
| dev | 0.58 | 0.45 | 0.75 | 0.59 | 0.62 | 2.39 | ||
| P1.5.2 | P1 + hidden_dim=64 + encoder_depth=8 | test | 0.62 | 0.51 | 0.85 | 0.64 | 0.68 | 2.53 |
| dev | 0.6 | 0.46 | 0.79 | 0.6 | 0.62 | 2.53 | ||
| P2 | P1 + optical_flow | test | 0.58 | 0.5 | 0.95 | 0.63 | 0.68 | 3.17 |
| dev | 0.59 | 0.45 | 0.84 | 0.59 | 0.61 | 3.08 | ||
| P2.1 | P1.3.2 + optical_flow | test | 0.63 | 0.51 | 0.92 | 0.66 | 0.67 | 1.51 |
| dev | 0.62 | 0.46 | 0.81 | 0.62 | 0.6 | 1.53 | ||
| P3 | P1 + hand_normalization | test | 0.55 | 0.48 | 0.77 | 0.58 | 0.67 | 2.79 |
| dev | 0.55 | 0.42 | 0.71 | 0.57 | 0.62 | 2.73 | ||
| P3.1 | P1.3.2 + hand_normalization | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.66 | 1.43 |
| dev | 0.61 | 0.46 | 0.82 | 0.64 | 0.61 | 1.46 | ||
| P4 | P2.1 + P3.1 | test | 0.56 | 0.51 | 0.92 | 0.61 | 0.66 | 1.45 |
| dev | 0.61 | 0.46 | 0.81 | 0.63 | 0.6 | 1.41 | ||
| P4.1 | P4 + encoder_depth=8 | test | 0.6 | 0.51 | 0.95 | 0.62 | 0.67 | 1.08 |
| dev | 0.61 | 0.47 | 0.86 | 0.62 | 0.6 | 1.12 | ||
| P5 | P1.3.2 + reduced_face | test | 0.63 | 0.51 | 0.94 | 0.64 | 0.66 | 1.16 |
| dev | 0.61 | 0.47 | 0.86 | 0.64 | 0.58 | 1.14 | ||
| P5.1 | P1.3.2 + full_face | test | 0.54 | 0.49 | 0.8 | 0.6 | 0.68 | 2.29 |
| dev | 0.57 | 0.45 | 0.7 | 0.59 | 0.62 | 2.29 |
| Experiment | Sign | Phrase | Efficiency | |||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | |||
| E0 | Moryossef et al. (2020) | test | — | 0.46 ± 0.03 | 1.09 ± 0.41 | — | 0.70 ± 0.01 | 1.00 ± 0.06 | 102K | 0:50:17 |
| dev | — | 0.42 ± 0.05 | 1.21 ± 0.59 | — | 0.61 ± 0.06 | 2.47 ± 0.85 | 102K | 0:50:17 | ||
| E1 | Baseline | test | 0.56 ± 0.03 | 0.66 ± 0.01 | 0.91 ± 0.05 | 0.59 ± 0.02 | 0.80 ± 0.03 | 2.50 ± 0.13 | 454K | 1:01:50 |
| dev | 0.55 ± 0.01 | 0.59 ± 0.00 | 1.12 ± 0.11 | 0.56 ± 0.02 | 0.75 ± 0.05 | 2.94 ± 0.08 | 454K | 1:01:50 | ||
| E2 | E1 + Face | test | 0.53 ± 0.05 | 0.58 ± 0.07 | 0.64 ± 0.30 | 0.57 ± 0.02 | 0.76 ± 0.03 | 1.87 ± 0.83 | 552K | 1:50:31 |
| dev | 0.50 ± 0.07 | 0.53 ± 0.11 | 0.90 ± 0.19 | 0.53 ± 0.05 | 0.71 ± 0.07 | 2.43 ± 1.02 | 552K | 1:50:31 | ||
| E3 | E1 + Optical Flow | test | 0.58 ± 0.01 | 0.62 ± 0.00 | 1.12 ± 0.05 | 0.60 ± 0.03 | 0.82 ± 0.03 | 3.19 ± 0.11 | 473K | 1:20:17 |
| dev | 0.58 ± 0.00 | 0.62 ± 0.00 | 1.50 ± 0.19 | 0.59 ± 0.01 | 0.79 ± 0.00 | 3.94 ± 0.14 | 473K | 1:20:17 | ||
| E4 | E3 + Hand Norm | test | 0.56 ± 0.02 | 0.61 ± 0.00 | 1.07 ± 0.05 | 0.60 ± 0.00 | 0.80 ± 0.00 | 3.24 ± 0.17 | 516K | 1:30:59 |
| dev | 0.57 ± 0.01 | 0.61 ± 0.01 | 1.50 ± 0.07 | 0.58 ± 0.00 | 0.79 ± 0.00 | 4.04 ± 0.31 | 516K | 1:30:59 | ||
| E1s | E1 + Depth=4 | test | 0.63 ± 0.01 | 0.69 ± 0.00 | 1.11 ± 0.01 | 0.65 ± 0.02 | 0.82 ± 0.04 | 1.63 ± 0.10 | 1.6M | 4:08:48 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.27 ± 0.01 | 0.63 ± 0.01 | 0.77 ± 0.01 | 2.17 ± 0.18 | 1.6M | 4:08:48 | ||
| E2s | E2 + Depth=4 | test | 0.62 ± 0.02 | 0.69 ± 0.00 | 1.07 ± 0.03 | 0.63 ± 0.01 | 0.84 ± 0.03 | 2.68 ± 0.53 | 1.7M | 3:14:03 |
| dev | 0.60 ± 0.01 | 0.63 ± 0.01 | 1.20 ± 0.12 | 0.59 ± 0.02 | 0.76 ± 0.05 | 3.30 ± 0.62 | 1.7M | 3:14:03 | ||
| E3s | E3 + Depth=4 | test | 0.60 ± 0.01 | 0.63 ± 0.00 | 1.13 ± 0.01 | 0.64 ± 0.03 | 0.80 ± 0.03 | 1.53 ± 0.18 | 1.7M | 4:08:30 |
| dev | 0.62 ± 0.00 | 0.63 ± 0.00 | 1.63 ± 0.05 | 0.63 ± 0.00 | 0.76 ± 0.00 | 2.14 ± 0.09 | 1.7M | 4:08:30 | ||
| E4s | E4 + Depth=4 | test | 0.59 ± 0.00 | 0.63 ± 0.00 | 1.13 ± 0.03 | 0.62 ± 0.00 | 0.79 ± 0.00 | 1.43 ± 0.10 | 1.7M | 4:35:29 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.56 ± 0.04 | 0.63 ± 0.00 | 0.77 ± 0.01 | 1.89 ± 0.07 | 1.7M | 4:35:29 | ||
| E4ba | E4s + Autoregressive | test | 0.45 ± 0.03 | 0.47 ± 0.05 | 0.88 ± 0.08 | 0.52 ± 0.02 | 0.63 ± 0.10 | 2.72 ± 1.33 | 1.3M | 2 days, 21:28:42 |
| dev | 0.40 ± 0.01 | 0.40 ± 0.01 | 2.02 ± 0.73 | 0.47 ± 0.00 | 0.57 ± 0.04 | 4.26 ± 1.26 | 1.3M | 2 days, 21:28:42 |
| Experiment | Sign | Phrase | Efficiency | ||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | ||
| E0 | Moryossef et al. (2020) | — | 0.46 | 1.09 | — | 0.70 | 1.00 | 102K | 0:50:17 |
| E1 | Baseline | 0.56 | 0.66 | 0.91 | 0.59 | 0.80 | 2.50 | 454K | 1:01:50 |
| E2 | E1 + Face | 0.53 | 0.58 | 0.64 | 0.57 | 0.76 | 1.87 | 552K | 1:50:31 |
| E3 | E1 + Optical Flow | 0.58 | 0.62 | 1.12 | 0.60 | 0.82 | 3.19 | 473K | 1:20:17 |
| E4 | E3 + Hand Norm | 0.56 | 0.61 | 1.07 | 0.60 | 0.80 | 3.24 | 516K | 1:30:59 |
| E1s | E1 + Depth=4 | 0.63 | 0.69 | 1.11 | 0.65 | 0.82 | 1.63 | 1.6M | 4:08:48 |
| E2s | E2 + Depth=4 | 0.62 | 0.69 | 1.07 | 0.63 | 0.84 | 2.68 | 1.7M | 3:14:03 |
| E3s | E3 + Depth=4 | 0.60 | 0.63 | 1.13 | 0.64 | 0.80 | 1.53 | 1.7M | 4:08:30 |
| E4s | E4 + Depth=4 | 0.59 | 0.63 | 1.13 | 0.62 | 0.79 | 1.43 | 1.7M | 4:35:29 |
| E1s* | E1s + Tuned Decoding | — | 0.69 | 1.03 | — | 0.85 | 1.02 | — | — |
| E4s* | E4s + Tuned Decoding | — | 0.63 | 1.06 | — | 0.79 | 1.12 | — | — |
| E5 | E4s + Autoregressive | 0.45 | 0.47 | 0.88 | 0.52 | 0.63 | 2.72 | 1.3M | ~3 days |
| Data | Model | ROC-AUC | F1-M |
| LSF | full (theirs) | 0.87 | — |
| body (theirs) | 0.87 | — | |
| E1s (ours, zero-shot) | 0.71 | 0.41 | |
| E4s (ours, zero-shot) | 0.76 | 0.44 | |
| E1s (ours, trained) | 0.87 | 0.49 | |
| E4s (ours, trained) | 0.87 | 0.51 | |
| DGS | E1s (ours) | 0.91 | 0.65 |
| E4s (ours) | 0.90 | 0.62 |
| Experiment | Sign | Phrase | ||||||
| F1 | IoU | % | F1 | IoU | % | |||
| P0 | Moryossef et al. (2020) | test | — | 0.4 | 1.45 | — | 0.65 | 0.82 |
| dev | — | 0.35 | 1.36 | — | 0.6 | 0.77 | ||
| P0.1 | P0 + Holistic 25fps | test | — | 0.39 | 0.86 | — | 0.64 | 0.5 |
| dev | — | 0.32 | 0.81 | — | 0.58 | 0.52 | ||
| P1 | P1 baseline | test | 0.55 | 0.49 | 0.83 | 0.6 | 0.67 | 2.63 |
| dev | 0.56 | 0.43 | 0.75 | 0.58 | 0.62 | 2.61 | ||
| P1.1 | P1 - encoder_bidirectional | test | 0.48 | 0.45 | 0.68 | 0.5 | 0.64 | 2.68 |
| dev | 0.46 | 0.41 | 0.64 | 0.51 | 0.61 | 2.56 | ||
| P1.2.1 | P1 + hidden_dim=512 | test | 0.47 | 0.42 | 0.44 | 0.52 | 0.63 | 1.7 |
| dev | 0.46 | 0.4 | 0.43 | 0.52 | 0.61 | 1.69 | ||
| P1.2.2 | P1 + hidden_dim=1024 | test | 0.48 | 0.45 | 0.42 | 0.58 | 0.65 | 1.53 |
| dev | 0.46 | 0.41 | 0.36 | 0.53 | 0.61 | 1.49 | ||
| P1.3.1 | P1 + encoder_depth=2 | test | 0.55 | 0.48 | 0.76 | 0.58 | 0.67 | 2.56 |
| dev | 0.56 | 0.43 | 0.69 | 0.58 | 0.62 | 2.52 | ||
| P1.3.2 | P1 + encoder_depth=4 | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.67 | 1.41 |
| dev | 0.61 | 0.47 | 0.84 | 0.64 | 0.6 | 1.39 | ||
| P1.4.1 | P1 + hidden_dim=128 + encoder_depth=2 | test | 0.58 | 0.48 | 0.8 | 0.6 | 0.67 | 2.0 |
| dev | 0.55 | 0.43 | 0.75 | 0.54 | 0.62 | 2.03 | ||
| P1.4.2 | P1 + hidden_dim=128 + encoder_depth=4 | test | 0.62 | 0.51 | 0.91 | 0.64 | 0.68 | 2.43 |
| dev | 0.6 | 0.47 | 0.83 | 0.6 | 0.62 | 2.57 | ||
| P1.4.3 | P1 + hidden_dim=128 + encoder_depth=8 | test | 0.59 | 0.52 | 0.91 | 0.63 | 0.68 | 3.04 |
| dev | 0.6 | 0.47 | 0.84 | 0.6 | 0.62 | 3.02 | ||
| P1.5.1 | P1 + hidden_dim=64 + encoder_depth=4 | test | 0.57 | 0.5 | 0.8 | 0.6 | 0.68 | 2.41 |
| dev | 0.58 | 0.45 | 0.75 | 0.59 | 0.62 | 2.39 | ||
| P1.5.2 | P1 + hidden_dim=64 + encoder_depth=8 | test | 0.62 | 0.51 | 0.85 | 0.64 | 0.68 | 2.53 |
| dev | 0.6 | 0.46 | 0.79 | 0.6 | 0.62 | 2.53 | ||
| P2 | P1 + optical_flow | test | 0.58 | 0.5 | 0.95 | 0.63 | 0.68 | 3.17 |
| dev | 0.59 | 0.45 | 0.84 | 0.59 | 0.61 | 3.08 | ||
| P2.1 | P1.3.2 + optical_flow | test | 0.63 | 0.51 | 0.92 | 0.66 | 0.67 | 1.51 |
| dev | 0.62 | 0.46 | 0.81 | 0.62 | 0.6 | 1.53 | ||
| P3 | P1 + hand_normalization | test | 0.55 | 0.48 | 0.77 | 0.58 | 0.67 | 2.79 |
| dev | 0.55 | 0.42 | 0.71 | 0.57 | 0.62 | 2.73 | ||
| P3.1 | P1.3.2 + hand_normalization | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.66 | 1.43 |
| dev | 0.61 | 0.46 | 0.82 | 0.64 | 0.61 | 1.46 | ||
| P4 | P2.1 + P3.1 | test | 0.56 | 0.51 | 0.92 | 0.61 | 0.66 | 1.45 |
| dev | 0.61 | 0.46 | 0.81 | 0.63 | 0.6 | 1.41 | ||
| P4.1 | P4 + encoder_depth=8 | test | 0.6 | 0.51 | 0.95 | 0.62 | 0.67 | 1.08 |
| dev | 0.61 | 0.47 | 0.86 | 0.62 | 0.6 | 1.12 | ||
| P5 | P1.3.2 + reduced_face | test | 0.63 | 0.51 | 0.94 | 0.64 | 0.66 | 1.16 |
| dev | 0.61 | 0.47 | 0.86 | 0.64 | 0.58 | 1.14 | ||
| P5.1 | P1.3.2 + full_face | test | 0.54 | 0.49 | 0.8 | 0.6 | 0.68 | 2.29 |
| dev | 0.57 | 0.45 | 0.7 | 0.59 | 0.62 | 2.29 | ||
| Experiment | Sign | Phrase | Efficiency | |||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | |||
| E0 | Moryossef et al. (2020) | test | — | 0.46 ± 0.03 | 1.09 ± 0.41 | — | 0.70 ± 0.01 | 1.00 ± 0.06 | 102K | 0:50:17 |
| dev | — | 0.42 ± 0.05 | 1.21 ± 0.59 | — | 0.61 ± 0.06 | 2.47 ± 0.85 | 102K | 0:50:17 | ||
| E1 | Baseline | test | 0.56 ± 0.03 | 0.66 ± 0.01 | 0.91 ± 0.05 | 0.59 ± 0.02 | 0.80 ± 0.03 | 2.50 ± 0.13 | 454K | 1:01:50 |
| dev | 0.55 ± 0.01 | 0.59 ± 0.00 | 1.12 ± 0.11 | 0.56 ± 0.02 | 0.75 ± 0.05 | 2.94 ± 0.08 | 454K | 1:01:50 | ||
| E2 | E1 + Face | test | 0.53 ± 0.05 | 0.58 ± 0.07 | 0.64 ± 0.30 | 0.57 ± 0.02 | 0.76 ± 0.03 | 1.87 ± 0.83 | 552K | 1:50:31 |
| dev | 0.50 ± 0.07 | 0.53 ± 0.11 | 0.90 ± 0.19 | 0.53 ± 0.05 | 0.71 ± 0.07 | 2.43 ± 1.02 | 552K | 1:50:31 | ||
| E3 | E1 + Optical Flow | test | 0.58 ± 0.01 | 0.62 ± 0.00 | 1.12 ± 0.05 | 0.60 ± 0.03 | 0.82 ± 0.03 | 3.19 ± 0.11 | 473K | 1:20:17 |
| dev | 0.58 ± 0.00 | 0.62 ± 0.00 | 1.50 ± 0.19 | 0.59 ± 0.01 | 0.79 ± 0.00 | 3.94 ± 0.14 | 473K | 1:20:17 | ||
| E4 | E3 + Hand Norm | test | 0.56 ± 0.02 | 0.61 ± 0.00 | 1.07 ± 0.05 | 0.60 ± 0.00 | 0.80 ± 0.00 | 3.24 ± 0.17 | 516K | 1:30:59 |
| dev | 0.57 ± 0.01 | 0.61 ± 0.01 | 1.50 ± 0.07 | 0.58 ± 0.00 | 0.79 ± 0.00 | 4.04 ± 0.31 | 516K | 1:30:59 | ||
| E1s | E1 + Depth=4 | test | 0.63 ± 0.01 | 0.69 ± 0.00 | 1.11 ± 0.01 | 0.65 ± 0.02 | 0.82 ± 0.04 | 1.63 ± 0.10 | 1.6M | 4:08:48 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.27 ± 0.01 | 0.63 ± 0.01 | 0.77 ± 0.01 | 2.17 ± 0.18 | 1.6M | 4:08:48 | ||
| E2s | E2 + Depth=4 | test | 0.62 ± 0.02 | 0.69 ± 0.00 | 1.07 ± 0.03 | 0.63 ± 0.01 | 0.84 ± 0.03 | 2.68 ± 0.53 | 1.7M | 3:14:03 |
| dev | 0.60 ± 0.01 | 0.63 ± 0.01 | 1.20 ± 0.12 | 0.59 ± 0.02 | 0.76 ± 0.05 | 3.30 ± 0.62 | 1.7M | 3:14:03 | ||
| E3s | E3 + Depth=4 | test | 0.60 ± 0.01 | 0.63 ± 0.00 | 1.13 ± 0.01 | 0.64 ± 0.03 | 0.80 ± 0.03 | 1.53 ± 0.18 | 1.7M | 4:08:30 |
| dev | 0.62 ± 0.00 | 0.63 ± 0.00 | 1.63 ± 0.05 | 0.63 ± 0.00 | 0.76 ± 0.00 | 2.14 ± 0.09 | 1.7M | 4:08:30 | ||
| E4s | E4 + Depth=4 | test | 0.59 ± 0.00 | 0.63 ± 0.00 | 1.13 ± 0.03 | 0.62 ± 0.00 | 0.79 ± 0.00 | 1.43 ± 0.10 | 1.7M | 4:35:29 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.56 ± 0.04 | 0.63 ± 0.00 | 0.77 ± 0.01 | 1.89 ± 0.07 | 1.7M | 4:35:29 | ||
| E4ba | E4s + Autoregressive | test | 0.45 ± 0.03 | 0.47 ± 0.05 | 0.88 ± 0.08 | 0.52 ± 0.02 | 0.63 ± 0.10 | 2.72 ± 1.33 | 1.3M | 2 days, 21:28:42 |
| dev | 0.40 ± 0.01 | 0.40 ± 0.01 | 2.02 ± 0.73 | 0.47 ± 0.00 | 0.57 ± 0.04 | 4.26 ± 1.26 | 1.3M | 2 days, 21:28:42 | ||
| Experiment | Sign | Phrase | Efficiency | ||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | ||
| E0 | Moryossef et al. (2020) | — | 0.46 | 1.09 | — | 0.70 | 1.00 | 102K | 0:50:17 |
| E1 | Baseline | 0.56 | 0.66 | 0.91 | 0.59 | 0.80 | 2.50 | 454K | 1:01:50 |
| E2 | E1 + Face | 0.53 | 0.58 | 0.64 | 0.57 | 0.76 | 1.87 | 552K | 1:50:31 |
| E3 | E1 + Optical Flow | 0.58 | 0.62 | 1.12 | 0.60 | 0.82 | 3.19 | 473K | 1:20:17 |
| E4 | E3 + Hand Norm | 0.56 | 0.61 | 1.07 | 0.60 | 0.80 | 3.24 | 516K | 1:30:59 |
| E1s | E1 + Depth=4 | 0.63 | 0.69 | 1.11 | 0.65 | 0.82 | 1.63 | 1.6M | 4:08:48 |
| E2s | E2 + Depth=4 | 0.62 | 0.69 | 1.07 | 0.63 | 0.84 | 2.68 | 1.7M | 3:14:03 |
| E3s | E3 + Depth=4 | 0.60 | 0.63 | 1.13 | 0.64 | 0.80 | 1.53 | 1.7M | 4:08:30 |
| E4s | E4 + Depth=4 | 0.59 | 0.63 | 1.13 | 0.62 | 0.79 | 1.43 | 1.7M | 4:35:29 |
| E1s* | E1s + Tuned Decoding | — | 0.69 | 1.03 | — | 0.85 | 1.02 | — | — |
| E4s* | E4s + Tuned Decoding | — | 0.63 | 1.06 | — | 0.79 | 1.12 | — | — |
| E5 | E4s + Autoregressive | 0.45 | 0.47 | 0.88 | 0.52 | 0.63 | 2.72 | 1.3M | ~3 days |
| Data | Model | ROC-AUC | F1-M |
| LSF | full (theirs) | 0.87 | — |
| body (theirs) | 0.87 | — | |
| E1s (ours, zero-shot) | 0.71 | 0.41 | |
| E4s (ours, zero-shot) | 0.76 | 0.44 | |
| E1s (ours, trained) | 0.87 | 0.49 | |
| E4s (ours, trained) | 0.87 | 0.51 | |
| DGS | E1s (ours) | 0.91 | 0.65 |
| E4s (ours) | 0.90 | 0.62 |
| Experiment | Sign | Phrase | ||||||
| F1 | IoU | % | F1 | IoU | % | |||
| P0 | Moryossef et al. (2020) | test | — | 0.4 | 1.45 | — | 0.65 | 0.82 |
| dev | — | 0.35 | 1.36 | — | 0.6 | 0.77 | ||
| P0.1 | P0 + Holistic 25fps | test | — | 0.39 | 0.86 | — | 0.64 | 0.5 |
| dev | — | 0.32 | 0.81 | — | 0.58 | 0.52 | ||
| P1 | P1 baseline | test | 0.55 | 0.49 | 0.83 | 0.6 | 0.67 | 2.63 |
| dev | 0.56 | 0.43 | 0.75 | 0.58 | 0.62 | 2.61 | ||
| P1.1 | P1 - encoder_bidirectional | test | 0.48 | 0.45 | 0.68 | 0.5 | 0.64 | 2.68 |
| dev | 0.46 | 0.41 | 0.64 | 0.51 | 0.61 | 2.56 | ||
| P1.2.1 | P1 + hidden_dim=512 | test | 0.47 | 0.42 | 0.44 | 0.52 | 0.63 | 1.7 |
| dev | 0.46 | 0.4 | 0.43 | 0.52 | 0.61 | 1.69 | ||
| P1.2.2 | P1 + hidden_dim=1024 | test | 0.48 | 0.45 | 0.42 | 0.58 | 0.65 | 1.53 |
| dev | 0.46 | 0.41 | 0.36 | 0.53 | 0.61 | 1.49 | ||
| P1.3.1 | P1 + encoder_depth=2 | test | 0.55 | 0.48 | 0.76 | 0.58 | 0.67 | 2.56 |
| dev | 0.56 | 0.43 | 0.69 | 0.58 | 0.62 | 2.52 | ||
| P1.3.2 | P1 + encoder_depth=4 | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.67 | 1.41 |
| dev | 0.61 | 0.47 | 0.84 | 0.64 | 0.6 | 1.39 | ||
| P1.4.1 | P1 + hidden_dim=128 + encoder_depth=2 | test | 0.58 | 0.48 | 0.8 | 0.6 | 0.67 | 2.0 |
| dev | 0.55 | 0.43 | 0.75 | 0.54 | 0.62 | 2.03 | ||
| P1.4.2 | P1 + hidden_dim=128 + encoder_depth=4 | test | 0.62 | 0.51 | 0.91 | 0.64 | 0.68 | 2.43 |
| dev | 0.6 | 0.47 | 0.83 | 0.6 | 0.62 | 2.57 | ||
| P1.4.3 | P1 + hidden_dim=128 + encoder_depth=8 | test | 0.59 | 0.52 | 0.91 | 0.63 | 0.68 | 3.04 |
| dev | 0.6 | 0.47 | 0.84 | 0.6 | 0.62 | 3.02 | ||
| P1.5.1 | P1 + hidden_dim=64 + encoder_depth=4 | test | 0.57 | 0.5 | 0.8 | 0.6 | 0.68 | 2.41 |
| dev | 0.58 | 0.45 | 0.75 | 0.59 | 0.62 | 2.39 | ||
| P1.5.2 | P1 + hidden_dim=64 + encoder_depth=8 | test | 0.62 | 0.51 | 0.85 | 0.64 | 0.68 | 2.53 |
| dev | 0.6 | 0.46 | 0.79 | 0.6 | 0.62 | 2.53 | ||
| P2 | P1 + optical_flow | test | 0.58 | 0.5 | 0.95 | 0.63 | 0.68 | 3.17 |
| dev | 0.59 | 0.45 | 0.84 | 0.59 | 0.61 | 3.08 | ||
| P2.1 | P1.3.2 + optical_flow | test | 0.63 | 0.51 | 0.92 | 0.66 | 0.67 | 1.51 |
| dev | 0.62 | 0.46 | 0.81 | 0.62 | 0.6 | 1.53 | ||
| P3 | P1 + hand_normalization | test | 0.55 | 0.48 | 0.77 | 0.58 | 0.67 | 2.79 |
| dev | 0.55 | 0.42 | 0.71 | 0.57 | 0.62 | 2.73 | ||
| P3.1 | P1.3.2 + hand_normalization | test | 0.63 | 0.51 | 0.91 | 0.66 | 0.66 | 1.43 |
| dev | 0.61 | 0.46 | 0.82 | 0.64 | 0.61 | 1.46 | ||
| P4 | P2.1 + P3.1 | test | 0.56 | 0.51 | 0.92 | 0.61 | 0.66 | 1.45 |
| dev | 0.61 | 0.46 | 0.81 | 0.63 | 0.6 | 1.41 | ||
| P4.1 | P4 + encoder_depth=8 | test | 0.6 | 0.51 | 0.95 | 0.62 | 0.67 | 1.08 |
| dev | 0.61 | 0.47 | 0.86 | 0.62 | 0.6 | 1.12 | ||
| P5 | P1.3.2 + reduced_face | test | 0.63 | 0.51 | 0.94 | 0.64 | 0.66 | 1.16 |
| dev | 0.61 | 0.47 | 0.86 | 0.64 | 0.58 | 1.14 | ||
| P5.1 | P1.3.2 + full_face | test | 0.54 | 0.49 | 0.8 | 0.6 | 0.68 | 2.29 |
| dev | 0.57 | 0.45 | 0.7 | 0.59 | 0.62 | 2.29 |
| Experiment | Sign | Phrase | Efficiency | |||||||
| F1 | IoU | % | F1 | IoU | % | #Params | Time | |||
| E0 | Moryossef et al. (2020) | test | — | 0.46 ± 0.03 | 1.09 ± 0.41 | — | 0.70 ± 0.01 | 1.00 ± 0.06 | 102K | 0:50:17 |
| dev | — | 0.42 ± 0.05 | 1.21 ± 0.59 | — | 0.61 ± 0.06 | 2.47 ± 0.85 | 102K | 0:50:17 | ||
| E1 | Baseline | test | 0.56 ± 0.03 | 0.66 ± 0.01 | 0.91 ± 0.05 | 0.59 ± 0.02 | 0.80 ± 0.03 | 2.50 ± 0.13 | 454K | 1:01:50 |
| dev | 0.55 ± 0.01 | 0.59 ± 0.00 | 1.12 ± 0.11 | 0.56 ± 0.02 | 0.75 ± 0.05 | 2.94 ± 0.08 | 454K | 1:01:50 | ||
| E2 | E1 + Face | test | 0.53 ± 0.05 | 0.58 ± 0.07 | 0.64 ± 0.30 | 0.57 ± 0.02 | 0.76 ± 0.03 | 1.87 ± 0.83 | 552K | 1:50:31 |
| dev | 0.50 ± 0.07 | 0.53 ± 0.11 | 0.90 ± 0.19 | 0.53 ± 0.05 | 0.71 ± 0.07 | 2.43 ± 1.02 | 552K | 1:50:31 | ||
| E3 | E1 + Optical Flow | test | 0.58 ± 0.01 | 0.62 ± 0.00 | 1.12 ± 0.05 | 0.60 ± 0.03 | 0.82 ± 0.03 | 3.19 ± 0.11 | 473K | 1:20:17 |
| dev | 0.58 ± 0.00 | 0.62 ± 0.00 | 1.50 ± 0.19 | 0.59 ± 0.01 | 0.79 ± 0.00 | 3.94 ± 0.14 | 473K | 1:20:17 | ||
| E4 | E3 + Hand Norm | test | 0.56 ± 0.02 | 0.61 ± 0.00 | 1.07 ± 0.05 | 0.60 ± 0.00 | 0.80 ± 0.00 | 3.24 ± 0.17 | 516K | 1:30:59 |
| dev | 0.57 ± 0.01 | 0.61 ± 0.01 | 1.50 ± 0.07 | 0.58 ± 0.00 | 0.79 ± 0.00 | 4.04 ± 0.31 | 516K | 1:30:59 | ||
| E1s | E1 + Depth=4 | test | 0.63 ± 0.01 | 0.69 ± 0.00 | 1.11 ± 0.01 | 0.65 ± 0.02 | 0.82 ± 0.04 | 1.63 ± 0.10 | 1.6M | 4:08:48 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.27 ± 0.01 | 0.63 ± 0.01 | 0.77 ± 0.01 | 2.17 ± 0.18 | 1.6M | 4:08:48 | ||
| E2s | E2 + Depth=4 | test | 0.62 ± 0.02 | 0.69 ± 0.00 | 1.07 ± 0.03 | 0.63 ± 0.01 | 0.84 ± 0.03 | 2.68 ± 0.53 | 1.7M | 3:14:03 |
| dev | 0.60 ± 0.01 | 0.63 ± 0.01 | 1.20 ± 0.12 | 0.59 ± 0.02 | 0.76 ± 0.05 | 3.30 ± 0.62 | 1.7M | 3:14:03 | ||
| E3s | E3 + Depth=4 | test | 0.60 ± 0.01 | 0.63 ± 0.00 | 1.13 ± 0.01 | 0.64 ± 0.03 | 0.80 ± 0.03 | 1.53 ± 0.18 | 1.7M | 4:08:30 |
| dev | 0.62 ± 0.00 | 0.63 ± 0.00 | 1.63 ± 0.05 | 0.63 ± 0.00 | 0.76 ± 0.00 | 2.14 ± 0.09 | 1.7M | 4:08:30 | ||
| E4s | E4 + Depth=4 | test | 0.59 ± 0.00 | 0.63 ± 0.00 | 1.13 ± 0.03 | 0.62 ± 0.00 | 0.79 ± 0.00 | 1.43 ± 0.10 | 1.7M | 4:35:29 |
| dev | 0.61 ± 0.00 | 0.63 ± 0.00 | 1.56 ± 0.04 | 0.63 ± 0.00 | 0.77 ± 0.01 | 1.89 ± 0.07 | 1.7M | 4:35:29 | ||
| E4ba | E4s + Autoregressive | test | 0.45 ± 0.03 | 0.47 ± 0.05 | 0.88 ± 0.08 | 0.52 ± 0.02 | 0.63 ± 0.10 | 2.72 ± 1.33 | 1.3M | 2 days, 21:28:42 |
| dev | 0.40 ± 0.01 | 0.40 ± 0.01 | 2.02 ± 0.73 | 0.47 ± 0.00 | 0.57 ± 0.04 | 4.26 ± 1.26 | 1.3M | 2 days, 21:28:42 |
| Mechanism | Privacy Level | Requires fine-tuning | Generates sanitized doc |
| Madlib (Feyisetan et al., 2020) | Word level Metric-DP | No | Yes |
| Mahanolbis (Xu et al., 2020) | Word level Metric-DP | No | Yes |
| TEM (Carvalho et al., 2021) | Word level Metric-DP | No | Yes |
| Truncated Laplace (Meehan et al., 2022) | Sentence level Pure-DP | No | No |
| Deep Candidate (Meehan et al., 2022) | Sentence level Pure-DP | Yes | No |
| Paraphrase (Mattern et al., 2022b) | Document level Pure-LDP | Yes | Yes |
| DP Prompt (Ours) | Document level Pure-LDP | No | Yes |
| Data | IMDB | Yelp | |||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | |||||||||||||||||||
| clipping | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | |||
| Flan-t5 (3b) | Static Attacker | Yes | 0.74 | 0.67 | 0.56 | 0.45 | 0.38 | 0.26 | 0.21 | 0.13 | 0.07 | 0.05 | 0.69 | 0.62 | 0.58 | 0.54 | 0.48 | 0.21 | 0.17 | 0.12 | 0.09 | 0.07 | |
| No | 0.75 (+0.01) | 0.69 (+0.02) | 0.58 (+0.02) | 0.46 (+0.01) | 0.40 (+0.02) | 0.31 (+0.03) | 0.24 (+0.03) | 0.14 (+0.01) | 0.08 (+0.01) | 0.05 (+0.00) | 0.69 (+0.00) | 0.64 (+0.02) | 0.58 (+0.00) | 0.54 (+0.00) | 0.49 (+0.01) | 0.25 (+0.04) | 0.20 (+0.03) | 0.13 (+0.01) | 0.09 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.72 | 0.66 | 0.58 | 0.52 | 0.51 | 0.34 | 0.25 | 0.16 | 0.13 | 0.11 | 0.67 | 0.59 | 0.54 | 0.52 | 0.49 | 0.25 | 0.20 | 0.14 | 0.11 | 0.10 | ||
| No | 0.73 (+0.01) | 0.67 (+0.01) | 0.58 (+0.00) | 0.53 (+0.01) | 0.52 (+0.01) | 0.38 (+0.04) | 0.28 (+0.03) | 0.19 (+0.03) | 0.13 (+0.00) | 0.11 (+0.00) | 0.67 (0.00) | 0.60 (+0.01) | 0.55 (+0.01) | 0.52 (+0.00) | 0.49 (+0.04) | 0.29 (+0.04) | 0.22 (+0.01) | 0.15 (+0.01) | 0.11 (+0.01) | 0.10 | |||
| Stablelm Tuned (7b) | Static Attacker | Yes | 0.67 | 0.63 | 0.53 | 0.38 | 0.34 | 0.33 | 0.29 | 0.12 | 0.05 | 0.03 | 0.66 | 0.62 | 0.55 | 0.49 | 0.48 | 0.28 | 0.22 | 0.10 | 0.08 | 0.07 | |
| No | 0.68 (+0.01) | 0.66 (+0.03) | 0.59 (+0.06) | 0.45 (+0.07) | 0.36 (+0.02) | 0.37 (+0.04) | 0.31 (+0.02) | 0.18 (+0.06) | 0.07 (+0.02) | 0.05 (+0.02) | 0.68 (+0.02) | 0.65 (+0.03) | 0.59 (+0.04) | 0.50 (+0.01) | 0.49 (+0.01) | 0.26 (-0.02) | 0.21 (-0.03) | 0.13 (+0.03) | 0.08 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.65 | 0.60 | 0.54 | 0.51 | 0.50 | 0.45 | 0.37 | 0.22 | 0.14 | 0.12 | 0.63 | 0.57 | 0.50 | 0.49 | 0.48 | 0.36 | 0.28 | 0.17 | 0.13 | 0.10 | ||
| No | 0.71 (+0.06) | 0.67 (+0.07) | 0.59 (+0.05) | 0.53 (+0.02) | 0.51 (0.01) | 0.53 (+0.08) | 0.46 (+0.09) | 0.30 (+0.02) | 0.17 (+0.03) | 0.14 (+0.02) | 0.66 (+0.03) | 0.64 (+0.07) | 0.55 (+0.05) | 0.51 (+0.02) | 0.49 (0.01) | 0.41 (+0.05) | 0.35 (+0.07) | 0.22 (+0.05) | 0.14 (+0.01) | 0.12 (+0.02) |
| Data | IMDB | Yelp | ||||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | ||||||||||||||||||||
| top_k | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | ||
| Flan-t5 | Static Attacker | all | 0.75 | 0.69 | 0.58 | 0.46 | 0.40 | 0.31 | 0.24 | 0.14 | 0.08 | 0.05 | 0.69 | 0.64 | 0.58 | 0.51 | 0.49 | 0.25 | 0.20 | 0.13 | 0.09 | 0.07 | ||
| 80 | 0.75 (+0.0) | 0.72 (+0.03) | 0.70 (+0.12) | 0.69 (+0.23) | 0.66 (+0.26) | 0.31 (+0.0) | 0.27 (+0.03) | 0.23 (+0.09) | 0.22 (+0.14) | 0.21 (+0.16) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.20 (+0.00) | 0.19 (+0.06) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| 40 | 0.75 (+0.00) | 0.74 (+0.05) | 0.72 (+0.24) | 0.69 (+0.23) | 0.67 (+0.27) | 0.32 (+0.0) | 0.29 (+0.05) | 0.26 (+0.12) | 0.25 (+0.17) | 0.23 (+0.18) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| Adaptive Attacker | all | 0.73 | 0.67 | 0.58 | 0.53 | 0.52 | 0.38 | 0.28 | 0.19 | 0.12 | 0.11 | 0.66 | 0.60 | 0.55 | 0.52 | 0.49 | 0.29 | 0.22 | 0.15 | 0.11 | 0.10 | |||
| 80 | 0.76 (+0.03) | 0.70 (+0.03) | 0.69 (+0.11) | 0.66 (+0.13) | 0.65 (+0.13) | 0.39 (+0.01) | 0.33 (+0.05) | 0.29 (+0.10) | 0.27 (+0.15) | 0.26 (+0.15) | 0.67 (+0.01) | 0.66 (+0.06) | 0.62 (+0.07) | 0.60 (+0.08) | 0.60 (+0.11) | 0.31 (+0.02) | 0.24 (+0.02) | 0.23 (+0.08) | 0.20 (+0.09) | 0.20 (+0.10) | ||||
| 40 | 0.75 (+0.02) | 0.72 (+0.05) | 0.70 (+0.12) | 0.67 (+0.14) | 0.67 (+0.15) | 0.41 (+0.02) | 0.36 (+0.08) | 0.33 (+0.14) | 0.31 (+0.19) | 0.29 (+0.18) | 0.69 (+0.02) | 0.66 (+0.06) | 0.65 (+0.10) | 0.63 (+0.11) | 0.61 (+0.12) | 0.31 (+0.02) | 0.26 (+0.04) | 0.24 (+0.09) | 0.23 (+0.12) | 0.23 (+0.13) | ||||
| Stablem Tuned (7b) | Static Attacker | all | 0.68 | 0.66 | 0.59 | 0.45 | 0.36 | 0.37 | 0.31 | 0.18 | 0.07 | 0.05 | 0.68 | 0.65 | 0.59 | 0.50 | 0.49 | 0.24 | 0.20 | 0.13 | 0.08 | 0.07 | ||
| 80 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.62 (+0.17) | 0.62 (+0.26) | 0.37 (+0.00) | 0.34 (+0.03) | 0.28 (+0.10) | 0.27 (+0.20) | 0.24 (+0.19) | 0.69 (+0.01) | 0.66 (+0.01) | 0.65 (+0.06) | 0.62 (+0.12) | 0.59 (+0.10) | 0.24 (+0.00) | 0.22 (+0.02) | 0.19 (+0.06) | 0.16 (+0.08) | |||||
| 40 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.64 (+0.19) | 0.64 (+0.28) | 0.38 (+0.01) | 0.34 (+0.03) | 0.30 (+0.12) | 0.27 (+0.20) | 0.27 (+0.22) | 0.69 (+0.01) | 0.69 (+0.04) | 0.61 (+0.02) | 0.65 (+0.15) | 0.65 (+0.16) | 0.25 (+0.01) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.11) | |||||
| Adaptive Attacker | all | 0.70 | 0.67 | 0.59 | 0.53 | 0.51 | 0.53 | 0.46 | 0.30 | 0.17 | 0.14 | 0.66 | 0.64 | 0.55 | 0.51 | 0.49 | 0.41 | 0.35 | 0.22 | 0.14 | 0.12 | |||
| 80 | 0.70 (+0.00) | 0.68 (+0.01) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.44 (+0.14) | 0.40 (+0.23) | 0.35 (+0.21) | 0.69 (+0.03) | 0.66 (+0.02) | 0.64 (+0.09) | 0.59 (+0.08) | 0.60 (+0.11) | 0.41 (+0.00) | 0.38 (+0.03) | 0.32 (+0.10) | 0.29 (+0.15) | |||||
| 40 | 0.70 (+0.00) | 0.69 (+0.02) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.45 (+0.15) | 0.42 (+0.25) | 0.39 (+0.25) | 0.69 (+0.03) | 0.67 (+0.03) | 0.64 (+0.09) | 0.60 (+0.09) | 0.62 (+0.13) | 0.42 (+0.01) | 0.37 (+0.02) | 0.34 (+0.12) | 0.32 (+0.18) |
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | |
| Chat GPT | 0.882 | 0.879 | 0.863 | 0.805 | 0.765 | 0.89 | 0.887 | 0.876 | 0.83 | 0.777 |
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | |
| GPT-2 (xl) (fine tuned) | 0.838 | 0.822 | 0.794 | 0.762 | 0.748 | 0.852 | 0.837 | 0.804 | 0.765 | 0.750 |
| T5 (xl) | 0.831 | 0.812 | 0.790 | 0.775 | 0.763 | 0.834 | 0.817 | 0.797 | 0.782 | 0.769 |
| Stablelm-Base (3b) | 0.814 | 0.805 | 0.785 | 0.762 | 0.749 | 0.835 | 0.817 | 0.789 | 0.757 | 0.747 |
| Stablelm-Base (7b) | 0.813 | 0.803 | 0.779 | 0.759 | 0.748 | 0.835 | 0.819 | 0.788 | 0.757 | 0.746 |
| Flan T5 (xl) | 0.843 | 0.823 | 0.792 | 0.762 | 0.750 | 0.849 | 0.830 | 0.801 | 0.769 | 0.753 |
| Stablelm Tuned (3b) | 0.846 | 0.830 | 0.795 | 0.757 | 0.743 | 0.849 | 0.836 | 0.800 | 0.760 | 0.746 |
| Stablelm Tuned (7b) | 0.854 | 0.839 | 0.800 | 0.757 | 0.743 | 0.858 | 0.845 | 0.806 | 0.761 | 0.747 |
| Mechanism | Privacy Level | Requires fine-tuning | Generates sanitized doc |
| Madlib (Feyisetan et al., 2020) | Word level Metric-DP | No | Yes |
| Mahanolbis (Xu et al., 2020) | Word level Metric-DP | No | Yes |
| TEM (Carvalho et al., 2021) | Word level Metric-DP | No | Yes |
| Truncated Laplace (Meehan et al., 2022) | Sentence level Pure-DP | No | No |
| Deep Candidate (Meehan et al., 2022) | Sentence level Pure-DP | Yes | No |
| Paraphrase (Mattern et al., 2022b) | Document level Pure-LDP | Yes | Yes |
| DP Prompt (Ours) | Document level Pure-LDP | No | Yes |
| Data | IMDB | Yelp | |||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | |||||||||||||||||||
| clipping | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | |||
| Flan-t5 (3b) | Static Attacker | Yes | 0.74 | 0.67 | 0.56 | 0.45 | 0.38 | 0.26 | 0.21 | 0.13 | 0.07 | 0.05 | 0.69 | 0.62 | 0.58 | 0.54 | 0.48 | 0.21 | 0.17 | 0.12 | 0.09 | 0.07 | |
| No | 0.75 (+0.01) | 0.69 (+0.02) | 0.58 (+0.02) | 0.46 (+0.01) | 0.40 (+0.02) | 0.31 (+0.03) | 0.24 (+0.03) | 0.14 (+0.01) | 0.08 (+0.01) | 0.05 (+0.00) | 0.69 (+0.00) | 0.64 (+0.02) | 0.58 (+0.00) | 0.54 (+0.00) | 0.49 (+0.01) | 0.25 (+0.04) | 0.20 (+0.03) | 0.13 (+0.01) | 0.09 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.72 | 0.66 | 0.58 | 0.52 | 0.51 | 0.34 | 0.25 | 0.16 | 0.13 | 0.11 | 0.67 | 0.59 | 0.54 | 0.52 | 0.49 | 0.25 | 0.20 | 0.14 | 0.11 | 0.10 | ||
| No | 0.73 (+0.01) | 0.67 (+0.01) | 0.58 (+0.00) | 0.53 (+0.01) | 0.52 (+0.01) | 0.38 (+0.04) | 0.28 (+0.03) | 0.19 (+0.03) | 0.13 (+0.00) | 0.11 (+0.00) | 0.67 (0.00) | 0.60 (+0.01) | 0.55 (+0.01) | 0.52 (+0.00) | 0.49 (+0.04) | 0.29 (+0.04) | 0.22 (+0.01) | 0.15 (+0.01) | 0.11 (+0.01) | 0.10 | |||
| Stablelm Tuned (7b) | Static Attacker | Yes | 0.67 | 0.63 | 0.53 | 0.38 | 0.34 | 0.33 | 0.29 | 0.12 | 0.05 | 0.03 | 0.66 | 0.62 | 0.55 | 0.49 | 0.48 | 0.28 | 0.22 | 0.10 | 0.08 | 0.07 | |
| No | 0.68 (+0.01) | 0.66 (+0.03) | 0.59 (+0.06) | 0.45 (+0.07) | 0.36 (+0.02) | 0.37 (+0.04) | 0.31 (+0.02) | 0.18 (+0.06) | 0.07 (+0.02) | 0.05 (+0.02) | 0.68 (+0.02) | 0.65 (+0.03) | 0.59 (+0.04) | 0.50 (+0.01) | 0.49 (+0.01) | 0.26 (-0.02) | 0.21 (-0.03) | 0.13 (+0.03) | 0.08 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.65 | 0.60 | 0.54 | 0.51 | 0.50 | 0.45 | 0.37 | 0.22 | 0.14 | 0.12 | 0.63 | 0.57 | 0.50 | 0.49 | 0.48 | 0.36 | 0.28 | 0.17 | 0.13 | 0.10 | ||
| No | 0.71 (+0.06) | 0.67 (+0.07) | 0.59 (+0.05) | 0.53 (+0.02) | 0.51 (0.01) | 0.53 (+0.08) | 0.46 (+0.09) | 0.30 (+0.02) | 0.17 (+0.03) | 0.14 (+0.02) | 0.66 (+0.03) | 0.64 (+0.07) | 0.55 (+0.05) | 0.51 (+0.02) | 0.49 (0.01) | 0.41 (+0.05) | 0.35 (+0.07) | 0.22 (+0.05) | 0.14 (+0.01) | 0.12 (+0.02) |
| Data | IMDB | Yelp | ||||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | ||||||||||||||||||||
| top_k | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | ||
| Flan-t5 | Static Attacker | all | 0.75 | 0.69 | 0.58 | 0.46 | 0.40 | 0.31 | 0.24 | 0.14 | 0.08 | 0.05 | 0.69 | 0.64 | 0.58 | 0.51 | 0.49 | 0.25 | 0.20 | 0.13 | 0.09 | 0.07 | ||
| 80 | 0.75 (+0.0) | 0.72 (+0.03) | 0.70 (+0.12) | 0.69 (+0.23) | 0.66 (+0.26) | 0.31 (+0.0) | 0.27 (+0.03) | 0.23 (+0.09) | 0.22 (+0.14) | 0.21 (+0.16) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.20 (+0.00) | 0.19 (+0.06) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| 40 | 0.75 (+0.00) | 0.74 (+0.05) | 0.72 (+0.24) | 0.69 (+0.23) | 0.67 (+0.27) | 0.32 (+0.0) | 0.29 (+0.05) | 0.26 (+0.12) | 0.25 (+0.17) | 0.23 (+0.18) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| Adaptive Attacker | all | 0.73 | 0.67 | 0.58 | 0.53 | 0.52 | 0.38 | 0.28 | 0.19 | 0.12 | 0.11 | 0.66 | 0.60 | 0.55 | 0.52 | 0.49 | 0.29 | 0.22 | 0.15 | 0.11 | 0.10 | |||
| 80 | 0.76 (+0.03) | 0.70 (+0.03) | 0.69 (+0.11) | 0.66 (+0.13) | 0.65 (+0.13) | 0.39 (+0.01) | 0.33 (+0.05) | 0.29 (+0.10) | 0.27 (+0.15) | 0.26 (+0.15) | 0.67 (+0.01) | 0.66 (+0.06) | 0.62 (+0.07) | 0.60 (+0.08) | 0.60 (+0.11) | 0.31 (+0.02) | 0.24 (+0.02) | 0.23 (+0.08) | 0.20 (+0.09) | 0.20 (+0.10) | ||||
| 40 | 0.75 (+0.02) | 0.72 (+0.05) | 0.70 (+0.12) | 0.67 (+0.14) | 0.67 (+0.15) | 0.41 (+0.02) | 0.36 (+0.08) | 0.33 (+0.14) | 0.31 (+0.19) | 0.29 (+0.18) | 0.69 (+0.02) | 0.66 (+0.06) | 0.65 (+0.10) | 0.63 (+0.11) | 0.61 (+0.12) | 0.31 (+0.02) | 0.26 (+0.04) | 0.24 (+0.09) | 0.23 (+0.12) | 0.23 (+0.13) | ||||
| Stablem Tuned (7b) | Static Attacker | all | 0.68 | 0.66 | 0.59 | 0.45 | 0.36 | 0.37 | 0.31 | 0.18 | 0.07 | 0.05 | 0.68 | 0.65 | 0.59 | 0.50 | 0.49 | 0.24 | 0.20 | 0.13 | 0.08 | 0.07 | ||
| 80 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.62 (+0.17) | 0.62 (+0.26) | 0.37 (+0.00) | 0.34 (+0.03) | 0.28 (+0.10) | 0.27 (+0.20) | 0.24 (+0.19) | 0.69 (+0.01) | 0.66 (+0.01) | 0.65 (+0.06) | 0.62 (+0.12) | 0.59 (+0.10) | 0.24 (+0.00) | 0.22 (+0.02) | 0.19 (+0.06) | 0.16 (+0.08) | |||||
| 40 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.64 (+0.19) | 0.64 (+0.28) | 0.38 (+0.01) | 0.34 (+0.03) | 0.30 (+0.12) | 0.27 (+0.20) | 0.27 (+0.22) | 0.69 (+0.01) | 0.69 (+0.04) | 0.61 (+0.02) | 0.65 (+0.15) | 0.65 (+0.16) | 0.25 (+0.01) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.11) | |||||
| Adaptive Attacker | all | 0.70 | 0.67 | 0.59 | 0.53 | 0.51 | 0.53 | 0.46 | 0.30 | 0.17 | 0.14 | 0.66 | 0.64 | 0.55 | 0.51 | 0.49 | 0.41 | 0.35 | 0.22 | 0.14 | 0.12 | |||
| 80 | 0.70 (+0.00) | 0.68 (+0.01) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.44 (+0.14) | 0.40 (+0.23) | 0.35 (+0.21) | 0.69 (+0.03) | 0.66 (+0.02) | 0.64 (+0.09) | 0.59 (+0.08) | 0.60 (+0.11) | 0.41 (+0.00) | 0.38 (+0.03) | 0.32 (+0.10) | 0.29 (+0.15) | |||||
| 40 | 0.70 (+0.00) | 0.69 (+0.02) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.45 (+0.15) | 0.42 (+0.25) | 0.39 (+0.25) | 0.69 (+0.03) | 0.67 (+0.03) | 0.64 (+0.09) | 0.60 (+0.09) | 0.62 (+0.13) | 0.42 (+0.01) | 0.37 (+0.02) | 0.34 (+0.12) | 0.32 (+0.18) |
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | |
| Chat GPT | 0.882 | 0.879 | 0.863 | 0.805 | 0.765 | 0.89 | 0.887 | 0.876 | 0.83 | 0.777 |
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | |
| GPT-2 (xl) (fine tuned) | 0.838 | 0.822 | 0.794 | 0.762 | 0.748 | 0.852 | 0.837 | 0.804 | 0.765 | 0.750 |
| T5 (xl) | 0.831 | 0.812 | 0.790 | 0.775 | 0.763 | 0.834 | 0.817 | 0.797 | 0.782 | 0.769 |
| Stablelm-Base (3b) | 0.814 | 0.805 | 0.785 | 0.762 | 0.749 | 0.835 | 0.817 | 0.789 | 0.757 | 0.747 |
| Stablelm-Base (7b) | 0.813 | 0.803 | 0.779 | 0.759 | 0.748 | 0.835 | 0.819 | 0.788 | 0.757 | 0.746 |
| Flan T5 (xl) | 0.843 | 0.823 | 0.792 | 0.762 | 0.750 | 0.849 | 0.830 | 0.801 | 0.769 | 0.753 |
| Stablelm Tuned (3b) | 0.846 | 0.830 | 0.795 | 0.757 | 0.743 | 0.849 | 0.836 | 0.800 | 0.760 | 0.746 |
| Stablelm Tuned (7b) | 0.854 | 0.839 | 0.800 | 0.757 | 0.743 | 0.858 | 0.845 | 0.806 | 0.761 | 0.747 |
| Mechanism | Privacy Level | Requires fine-tuning | Generates sanitized doc |
| Madlib (Feyisetan et al., 2020) | Word level Metric-DP | No | Yes |
| Mahanolbis (Xu et al., 2020) | Word level Metric-DP | No | Yes |
| TEM (Carvalho et al., 2021) | Word level Metric-DP | No | Yes |
| Truncated Laplace (Meehan et al., 2022) | Sentence level Pure-DP | No | No |
| Deep Candidate (Meehan et al., 2022) | Sentence level Pure-DP | Yes | No |
| Paraphrase (Mattern et al., 2022b) | Document level Pure-LDP | Yes | Yes |
| DP Prompt (Ours) | Document level Pure-LDP | No | Yes |
| Data | IMDB | Yelp | |||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | |||||||||||||||||||
| clipping | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | |||
| Flan-t5 (3b) | Static Attacker | Yes | 0.74 | 0.67 | 0.56 | 0.45 | 0.38 | 0.26 | 0.21 | 0.13 | 0.07 | 0.05 | 0.69 | 0.62 | 0.58 | 0.54 | 0.48 | 0.21 | 0.17 | 0.12 | 0.09 | 0.07 | |
| No | 0.75 (+0.01) | 0.69 (+0.02) | 0.58 (+0.02) | 0.46 (+0.01) | 0.40 (+0.02) | 0.31 (+0.03) | 0.24 (+0.03) | 0.14 (+0.01) | 0.08 (+0.01) | 0.05 (+0.00) | 0.69 (+0.00) | 0.64 (+0.02) | 0.58 (+0.00) | 0.54 (+0.00) | 0.49 (+0.01) | 0.25 (+0.04) | 0.20 (+0.03) | 0.13 (+0.01) | 0.09 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.72 | 0.66 | 0.58 | 0.52 | 0.51 | 0.34 | 0.25 | 0.16 | 0.13 | 0.11 | 0.67 | 0.59 | 0.54 | 0.52 | 0.49 | 0.25 | 0.20 | 0.14 | 0.11 | 0.10 | ||
| No | 0.73 (+0.01) | 0.67 (+0.01) | 0.58 (+0.00) | 0.53 (+0.01) | 0.52 (+0.01) | 0.38 (+0.04) | 0.28 (+0.03) | 0.19 (+0.03) | 0.13 (+0.00) | 0.11 (+0.00) | 0.67 (0.00) | 0.60 (+0.01) | 0.55 (+0.01) | 0.52 (+0.00) | 0.49 (+0.04) | 0.29 (+0.04) | 0.22 (+0.01) | 0.15 (+0.01) | 0.11 (+0.01) | 0.10 | |||
| Stablelm Tuned (7b) | Static Attacker | Yes | 0.67 | 0.63 | 0.53 | 0.38 | 0.34 | 0.33 | 0.29 | 0.12 | 0.05 | 0.03 | 0.66 | 0.62 | 0.55 | 0.49 | 0.48 | 0.28 | 0.22 | 0.10 | 0.08 | 0.07 | |
| No | 0.68 (+0.01) | 0.66 (+0.03) | 0.59 (+0.06) | 0.45 (+0.07) | 0.36 (+0.02) | 0.37 (+0.04) | 0.31 (+0.02) | 0.18 (+0.06) | 0.07 (+0.02) | 0.05 (+0.02) | 0.68 (+0.02) | 0.65 (+0.03) | 0.59 (+0.04) | 0.50 (+0.01) | 0.49 (+0.01) | 0.26 (-0.02) | 0.21 (-0.03) | 0.13 (+0.03) | 0.08 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.65 | 0.60 | 0.54 | 0.51 | 0.50 | 0.45 | 0.37 | 0.22 | 0.14 | 0.12 | 0.63 | 0.57 | 0.50 | 0.49 | 0.48 | 0.36 | 0.28 | 0.17 | 0.13 | 0.10 | ||
| No | 0.71 (+0.06) | 0.67 (+0.07) | 0.59 (+0.05) | 0.53 (+0.02) | 0.51 (0.01) | 0.53 (+0.08) | 0.46 (+0.09) | 0.30 (+0.02) | 0.17 (+0.03) | 0.14 (+0.02) | 0.66 (+0.03) | 0.64 (+0.07) | 0.55 (+0.05) | 0.51 (+0.02) | 0.49 (0.01) | 0.41 (+0.05) | 0.35 (+0.07) | 0.22 (+0.05) | 0.14 (+0.01) | 0.12 (+0.02) | |||
| Data | IMDB | Yelp | ||||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | ||||||||||||||||||||
| top_k | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | ||
| Flan-t5 | Static Attacker | all | 0.75 | 0.69 | 0.58 | 0.46 | 0.40 | 0.31 | 0.24 | 0.14 | 0.08 | 0.05 | 0.69 | 0.64 | 0.58 | 0.51 | 0.49 | 0.25 | 0.20 | 0.13 | 0.09 | 0.07 | ||
| 80 | 0.75 (+0.0) | 0.72 (+0.03) | 0.70 (+0.12) | 0.69 (+0.23) | 0.66 (+0.26) | 0.31 (+0.0) | 0.27 (+0.03) | 0.23 (+0.09) | 0.22 (+0.14) | 0.21 (+0.16) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.20 (+0.00) | 0.19 (+0.06) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| 40 | 0.75 (+0.00) | 0.74 (+0.05) | 0.72 (+0.24) | 0.69 (+0.23) | 0.67 (+0.27) | 0.32 (+0.0) | 0.29 (+0.05) | 0.26 (+0.12) | 0.25 (+0.17) | 0.23 (+0.18) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| Adaptive Attacker | all | 0.73 | 0.67 | 0.58 | 0.53 | 0.52 | 0.38 | 0.28 | 0.19 | 0.12 | 0.11 | 0.66 | 0.60 | 0.55 | 0.52 | 0.49 | 0.29 | 0.22 | 0.15 | 0.11 | 0.10 | |||
| 80 | 0.76 (+0.03) | 0.70 (+0.03) | 0.69 (+0.11) | 0.66 (+0.13) | 0.65 (+0.13) | 0.39 (+0.01) | 0.33 (+0.05) | 0.29 (+0.10) | 0.27 (+0.15) | 0.26 (+0.15) | 0.67 (+0.01) | 0.66 (+0.06) | 0.62 (+0.07) | 0.60 (+0.08) | 0.60 (+0.11) | 0.31 (+0.02) | 0.24 (+0.02) | 0.23 (+0.08) | 0.20 (+0.09) | 0.20 (+0.10) | ||||
| 40 | 0.75 (+0.02) | 0.72 (+0.05) | 0.70 (+0.12) | 0.67 (+0.14) | 0.67 (+0.15) | 0.41 (+0.02) | 0.36 (+0.08) | 0.33 (+0.14) | 0.31 (+0.19) | 0.29 (+0.18) | 0.69 (+0.02) | 0.66 (+0.06) | 0.65 (+0.10) | 0.63 (+0.11) | 0.61 (+0.12) | 0.31 (+0.02) | 0.26 (+0.04) | 0.24 (+0.09) | 0.23 (+0.12) | 0.23 (+0.13) | ||||
| Stablem Tuned (7b) | Static Attacker | all | 0.68 | 0.66 | 0.59 | 0.45 | 0.36 | 0.37 | 0.31 | 0.18 | 0.07 | 0.05 | 0.68 | 0.65 | 0.59 | 0.50 | 0.49 | 0.24 | 0.20 | 0.13 | 0.08 | 0.07 | ||
| 80 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.62 (+0.17) | 0.62 (+0.26) | 0.37 (+0.00) | 0.34 (+0.03) | 0.28 (+0.10) | 0.27 (+0.20) | 0.24 (+0.19) | 0.69 (+0.01) | 0.66 (+0.01) | 0.65 (+0.06) | 0.62 (+0.12) | 0.59 (+0.10) | 0.24 (+0.00) | 0.22 (+0.02) | 0.19 (+0.06) | 0.16 (+0.08) | |||||
| 40 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.64 (+0.19) | 0.64 (+0.28) | 0.38 (+0.01) | 0.34 (+0.03) | 0.30 (+0.12) | 0.27 (+0.20) | 0.27 (+0.22) | 0.69 (+0.01) | 0.69 (+0.04) | 0.61 (+0.02) | 0.65 (+0.15) | 0.65 (+0.16) | 0.25 (+0.01) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.11) | |||||
| Adaptive Attacker | all | 0.70 | 0.67 | 0.59 | 0.53 | 0.51 | 0.53 | 0.46 | 0.30 | 0.17 | 0.14 | 0.66 | 0.64 | 0.55 | 0.51 | 0.49 | 0.41 | 0.35 | 0.22 | 0.14 | 0.12 | |||
| 80 | 0.70 (+0.00) | 0.68 (+0.01) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.44 (+0.14) | 0.40 (+0.23) | 0.35 (+0.21) | 0.69 (+0.03) | 0.66 (+0.02) | 0.64 (+0.09) | 0.59 (+0.08) | 0.60 (+0.11) | 0.41 (+0.00) | 0.38 (+0.03) | 0.32 (+0.10) | 0.29 (+0.15) | |||||
| 40 | 0.70 (+0.00) | 0.69 (+0.02) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.45 (+0.15) | 0.42 (+0.25) | 0.39 (+0.25) | 0.69 (+0.03) | 0.67 (+0.03) | 0.64 (+0.09) | 0.60 (+0.09) | 0.62 (+0.13) | 0.42 (+0.01) | 0.37 (+0.02) | 0.34 (+0.12) | 0.32 (+0.18) | |||||
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | |
| Chat GPT | 0.882 | 0.879 | 0.863 | 0.805 | 0.765 | 0.89 | 0.887 | 0.876 | 0.83 | 0.777 |
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | |
| GPT-2 (xl) (fine tuned) | 0.838 | 0.822 | 0.794 | 0.762 | 0.748 | 0.852 | 0.837 | 0.804 | 0.765 | 0.750 |
| T5 (xl) | 0.831 | 0.812 | 0.790 | 0.775 | 0.763 | 0.834 | 0.817 | 0.797 | 0.782 | 0.769 |
| Stablelm-Base (3b) | 0.814 | 0.805 | 0.785 | 0.762 | 0.749 | 0.835 | 0.817 | 0.789 | 0.757 | 0.747 |
| Stablelm-Base (7b) | 0.813 | 0.803 | 0.779 | 0.759 | 0.748 | 0.835 | 0.819 | 0.788 | 0.757 | 0.746 |
| Flan T5 (xl) | 0.843 | 0.823 | 0.792 | 0.762 | 0.750 | 0.849 | 0.830 | 0.801 | 0.769 | 0.753 |
| Stablelm Tuned (3b) | 0.846 | 0.830 | 0.795 | 0.757 | 0.743 | 0.849 | 0.836 | 0.800 | 0.760 | 0.746 |
| Stablelm Tuned (7b) | 0.854 | 0.839 | 0.800 | 0.757 | 0.743 | 0.858 | 0.845 | 0.806 | 0.761 | 0.747 |
| Mechanism | Privacy Level | Requires fine-tuning | Generates sanitized doc |
| Madlib (Feyisetan et al., 2020) | Word level Metric-DP | No | Yes |
| Mahanolbis (Xu et al., 2020) | Word level Metric-DP | No | Yes |
| TEM (Carvalho et al., 2021) | Word level Metric-DP | No | Yes |
| Truncated Laplace (Meehan et al., 2022) | Sentence level Pure-DP | No | No |
| Deep Candidate (Meehan et al., 2022) | Sentence level Pure-DP | Yes | No |
| Paraphrase (Mattern et al., 2022b) | Document level Pure-LDP | Yes | Yes |
| DP Prompt (Ours) | Document level Pure-LDP | No | Yes |
| Data | IMDB | Yelp | |||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | |||||||||||||||||||
| clipping | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | |||
| Flan-t5 (3b) | Static Attacker | Yes | 0.74 | 0.67 | 0.56 | 0.45 | 0.38 | 0.26 | 0.21 | 0.13 | 0.07 | 0.05 | 0.69 | 0.62 | 0.58 | 0.54 | 0.48 | 0.21 | 0.17 | 0.12 | 0.09 | 0.07 | |
| No | 0.75 (+0.01) | 0.69 (+0.02) | 0.58 (+0.02) | 0.46 (+0.01) | 0.40 (+0.02) | 0.31 (+0.03) | 0.24 (+0.03) | 0.14 (+0.01) | 0.08 (+0.01) | 0.05 (+0.00) | 0.69 (+0.00) | 0.64 (+0.02) | 0.58 (+0.00) | 0.54 (+0.00) | 0.49 (+0.01) | 0.25 (+0.04) | 0.20 (+0.03) | 0.13 (+0.01) | 0.09 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.72 | 0.66 | 0.58 | 0.52 | 0.51 | 0.34 | 0.25 | 0.16 | 0.13 | 0.11 | 0.67 | 0.59 | 0.54 | 0.52 | 0.49 | 0.25 | 0.20 | 0.14 | 0.11 | 0.10 | ||
| No | 0.73 (+0.01) | 0.67 (+0.01) | 0.58 (+0.00) | 0.53 (+0.01) | 0.52 (+0.01) | 0.38 (+0.04) | 0.28 (+0.03) | 0.19 (+0.03) | 0.13 (+0.00) | 0.11 (+0.00) | 0.67 (0.00) | 0.60 (+0.01) | 0.55 (+0.01) | 0.52 (+0.00) | 0.49 (+0.04) | 0.29 (+0.04) | 0.22 (+0.01) | 0.15 (+0.01) | 0.11 (+0.01) | 0.10 | |||
| Stablelm Tuned (7b) | Static Attacker | Yes | 0.67 | 0.63 | 0.53 | 0.38 | 0.34 | 0.33 | 0.29 | 0.12 | 0.05 | 0.03 | 0.66 | 0.62 | 0.55 | 0.49 | 0.48 | 0.28 | 0.22 | 0.10 | 0.08 | 0.07 | |
| No | 0.68 (+0.01) | 0.66 (+0.03) | 0.59 (+0.06) | 0.45 (+0.07) | 0.36 (+0.02) | 0.37 (+0.04) | 0.31 (+0.02) | 0.18 (+0.06) | 0.07 (+0.02) | 0.05 (+0.02) | 0.68 (+0.02) | 0.65 (+0.03) | 0.59 (+0.04) | 0.50 (+0.01) | 0.49 (+0.01) | 0.26 (-0.02) | 0.21 (-0.03) | 0.13 (+0.03) | 0.08 (+0.00) | 0.07 (+0.00) | |||
| Adaptive Attacker | Yes | 0.65 | 0.60 | 0.54 | 0.51 | 0.50 | 0.45 | 0.37 | 0.22 | 0.14 | 0.12 | 0.63 | 0.57 | 0.50 | 0.49 | 0.48 | 0.36 | 0.28 | 0.17 | 0.13 | 0.10 | ||
| No | 0.71 (+0.06) | 0.67 (+0.07) | 0.59 (+0.05) | 0.53 (+0.02) | 0.51 (0.01) | 0.53 (+0.08) | 0.46 (+0.09) | 0.30 (+0.02) | 0.17 (+0.03) | 0.14 (+0.02) | 0.66 (+0.03) | 0.64 (+0.07) | 0.55 (+0.05) | 0.51 (+0.02) | 0.49 (0.01) | 0.41 (+0.05) | 0.35 (+0.07) | 0.22 (+0.05) | 0.14 (+0.01) | 0.12 (+0.02) |
| Data | IMDB | Yelp | ||||||||||||||||||||||
| Metric | Sentiment F1 score | Author Identification F1 Score | Sentiment F1 score | Author Identification F1 Score | ||||||||||||||||||||
| top_k | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | 1.25 | 1.5 | 1.75 | 0.75 | 1.0 | ||
| Flan-t5 | Static Attacker | all | 0.75 | 0.69 | 0.58 | 0.46 | 0.40 | 0.31 | 0.24 | 0.14 | 0.08 | 0.05 | 0.69 | 0.64 | 0.58 | 0.51 | 0.49 | 0.25 | 0.20 | 0.13 | 0.09 | 0.07 | ||
| 80 | 0.75 (+0.0) | 0.72 (+0.03) | 0.70 (+0.12) | 0.69 (+0.23) | 0.66 (+0.26) | 0.31 (+0.0) | 0.27 (+0.03) | 0.23 (+0.09) | 0.22 (+0.14) | 0.21 (+0.16) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.20 (+0.00) | 0.19 (+0.06) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| 40 | 0.75 (+0.00) | 0.74 (+0.05) | 0.72 (+0.24) | 0.69 (+0.23) | 0.67 (+0.27) | 0.32 (+0.0) | 0.29 (+0.05) | 0.26 (+0.12) | 0.25 (+0.17) | 0.23 (+0.18) | 0.71 (+0.02) | 0.69 (+0.05) | 0.68 (+0.10) | 0.64 (+0.13) | 0.64 (+0.15) | 0.25 (+0.00) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.10) | 0.17 (+0.10) | ||||
| Adaptive Attacker | all | 0.73 | 0.67 | 0.58 | 0.53 | 0.52 | 0.38 | 0.28 | 0.19 | 0.12 | 0.11 | 0.66 | 0.60 | 0.55 | 0.52 | 0.49 | 0.29 | 0.22 | 0.15 | 0.11 | 0.10 | |||
| 80 | 0.76 (+0.03) | 0.70 (+0.03) | 0.69 (+0.11) | 0.66 (+0.13) | 0.65 (+0.13) | 0.39 (+0.01) | 0.33 (+0.05) | 0.29 (+0.10) | 0.27 (+0.15) | 0.26 (+0.15) | 0.67 (+0.01) | 0.66 (+0.06) | 0.62 (+0.07) | 0.60 (+0.08) | 0.60 (+0.11) | 0.31 (+0.02) | 0.24 (+0.02) | 0.23 (+0.08) | 0.20 (+0.09) | 0.20 (+0.10) | ||||
| 40 | 0.75 (+0.02) | 0.72 (+0.05) | 0.70 (+0.12) | 0.67 (+0.14) | 0.67 (+0.15) | 0.41 (+0.02) | 0.36 (+0.08) | 0.33 (+0.14) | 0.31 (+0.19) | 0.29 (+0.18) | 0.69 (+0.02) | 0.66 (+0.06) | 0.65 (+0.10) | 0.63 (+0.11) | 0.61 (+0.12) | 0.31 (+0.02) | 0.26 (+0.04) | 0.24 (+0.09) | 0.23 (+0.12) | 0.23 (+0.13) | ||||
| Stablem Tuned (7b) | Static Attacker | all | 0.68 | 0.66 | 0.59 | 0.45 | 0.36 | 0.37 | 0.31 | 0.18 | 0.07 | 0.05 | 0.68 | 0.65 | 0.59 | 0.50 | 0.49 | 0.24 | 0.20 | 0.13 | 0.08 | 0.07 | ||
| 80 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.62 (+0.17) | 0.62 (+0.26) | 0.37 (+0.00) | 0.34 (+0.03) | 0.28 (+0.10) | 0.27 (+0.20) | 0.24 (+0.19) | 0.69 (+0.01) | 0.66 (+0.01) | 0.65 (+0.06) | 0.62 (+0.12) | 0.59 (+0.10) | 0.24 (+0.00) | 0.22 (+0.02) | 0.19 (+0.06) | 0.16 (+0.08) | |||||
| 40 | 0.69 (+0.01) | 0.67 (+0.01) | 0.66 (+0.07) | 0.64 (+0.19) | 0.64 (+0.28) | 0.38 (+0.01) | 0.34 (+0.03) | 0.30 (+0.12) | 0.27 (+0.20) | 0.27 (+0.22) | 0.69 (+0.01) | 0.69 (+0.04) | 0.61 (+0.02) | 0.65 (+0.15) | 0.65 (+0.16) | 0.25 (+0.01) | 0.22 (+0.02) | 0.20 (+0.07) | 0.19 (+0.11) | |||||
| Adaptive Attacker | all | 0.70 | 0.67 | 0.59 | 0.53 | 0.51 | 0.53 | 0.46 | 0.30 | 0.17 | 0.14 | 0.66 | 0.64 | 0.55 | 0.51 | 0.49 | 0.41 | 0.35 | 0.22 | 0.14 | 0.12 | |||
| 80 | 0.70 (+0.00) | 0.68 (+0.01) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.44 (+0.14) | 0.40 (+0.23) | 0.35 (+0.21) | 0.69 (+0.03) | 0.66 (+0.02) | 0.64 (+0.09) | 0.59 (+0.08) | 0.60 (+0.11) | 0.41 (+0.00) | 0.38 (+0.03) | 0.32 (+0.10) | 0.29 (+0.15) | |||||
| 40 | 0.70 (+0.00) | 0.69 (+0.02) | 0.67 (+0.08) | 0.64 (+0.11) | 0.63 (+0.12) | 0.53 (+0.00) | 0.49 (+0.03) | 0.45 (+0.15) | 0.42 (+0.25) | 0.39 (+0.25) | 0.69 (+0.03) | 0.67 (+0.03) | 0.64 (+0.09) | 0.60 (+0.09) | 0.62 (+0.13) | 0.42 (+0.01) | 0.37 (+0.02) | 0.34 (+0.12) | 0.32 (+0.18) |
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=2.0 | |
| Chat GPT | 0.882 | 0.879 | 0.863 | 0.805 | 0.765 | 0.89 | 0.887 | 0.876 | 0.83 | 0.777 |
| BERT Score | ||||||||||
| IMDB | Yelp | |||||||||
| t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | t=0.75 | t=1.0 | t=1.25 | t=1.5 | t=1.75 | |
| GPT-2 (xl) (fine tuned) | 0.838 | 0.822 | 0.794 | 0.762 | 0.748 | 0.852 | 0.837 | 0.804 | 0.765 | 0.750 |
| T5 (xl) | 0.831 | 0.812 | 0.790 | 0.775 | 0.763 | 0.834 | 0.817 | 0.797 | 0.782 | 0.769 |
| Stablelm-Base (3b) | 0.814 | 0.805 | 0.785 | 0.762 | 0.749 | 0.835 | 0.817 | 0.789 | 0.757 | 0.747 |
| Stablelm-Base (7b) | 0.813 | 0.803 | 0.779 | 0.759 | 0.748 | 0.835 | 0.819 | 0.788 | 0.757 | 0.746 |
| Flan T5 (xl) | 0.843 | 0.823 | 0.792 | 0.762 | 0.750 | 0.849 | 0.830 | 0.801 | 0.769 | 0.753 |
| Stablelm Tuned (3b) | 0.846 | 0.830 | 0.795 | 0.757 | 0.743 | 0.849 | 0.836 | 0.800 | 0.760 | 0.746 |
| Stablelm Tuned (7b) | 0.854 | 0.839 | 0.800 | 0.757 | 0.743 | 0.858 | 0.845 | 0.806 | 0.761 | 0.747 |
| Task | Origin | Size |
| Language to Logic | LOGICIN inference & FOLIO | 13,206 |
| One-Step Inference | LOGICIN inference & FOLIO | 23,943 |
| Inference Chain | LOGICIN inference & EntailmentBank | 26,228 |
| Multi-choice | LogiQA & ReClor | 5,606 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Size | 1572 | 1594 | 500 | 230 | 1354 | 805 | 35891 | 10071 | 52017 |
| LLaMA-7b-base | 18.04 | 19.06 | 15.83 | 13.91 | 20.25 | 32.40 | 25.20 | 37.35 | 20.22 |
| LLaMA-30b-supercot | 19.31 | 26.35 | 17.81 | 17.98 | 18.41 | 24.10 | 32.26 | 41.91 | 24.78 |
| Falcon-40b-instruct | 23.21 | 19.77 | 26.77 | 12.70 | 17.33 | 16.13 | 28.49 | 44.66 | 23.63 |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| Dataset | Target |
| LogiQA 2.0 test | 4-way multi-choice |
| LogiQA 2.0 zh test | 4-way multi-choice |
| ReClor dev | 4-way multi-choice |
| AR-LSAT test | 5-way multi-choice |
| LogiQA 2.0 OOD | 4-way multi-choice |
| ConTRoL test | E, C, N |
| HELP test | E, C, N |
| TaxiNLI test | E, C, N |
| Task | LLaMA-7b-base | LLaMA-7b-logicot |
| Math | 25.1 | 29.0 |
| Health | 34.0 | 42.9 |
| Physics | 29.4 | 34.2 |
| Business | 34.8 | 57.2 |
| Biology | 32.4 | 46.5 |
| Chemistry | 25.4 | 33.0 |
| Computer science | 28.2 | 40.0 |
| Economics | 26.5 | 38.5 |
| Engineering | 25.5 | 32.4 |
| Philosophy | 30.8 | 37.6 |
| Other | 40.4 | 50.8 |
| History | 38.2 | 55.2 |
| Geography | 28.8 | 52.5 |
| Politics | 32.1 | 53.4 |
| Psychology | 33.2 | 50.9 |
| Culture | 37.3 | 56.9 |
| Law | 29.3 | 39.6 |
| STEM | 27.6 | 34.8 |
| Humanities | 31.6 | 41.8 |
| Social Sciences | 31.5 | 49.2 |
| Other (misc.) | 36.4 | 47.7 |
| Average | 31.8 | 43.3 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| ChatGPT | 52.37 | 53.18 | 57.38 | 20.42 | 38.44 | 58.45 | 42.13 | 57.30 | 47.46 |
| GPT-4 | 72.25 | 70.56 | 87.20 | 33.48 | 58.49 | 56.40 | 46.01 | 60.08 | 60.56 |
| Removed | LogiEval | MMLU |
| None (Full data) | 40.7 | 43.3 |
| Language to Logic | 32.4 | 38.5 |
| One-step Inference | 38.1 | 37.7 |
| Inference Chain | 30.8 | 35.0 |
| Multi-choice | 35.6 | 30.9 |
| Instruction | Input | Output |
| Translate the following inference to logic notation | All squares have four sides. | ∀x (Square(x) → FourSides(x)) |
| What can be inferred from the following premises in a single inference step | David is working if and only if it is cloudy.\nDavid is not working. | If David works, then it is cloudy.\nIf it is cloudy, then David is working.\nIt is not cloudy. |
| What can be inferred from the following premises in a single inference step? Name the inference rule being used | For all x_9, x_9 is an artist if and only if x_9 will sleep.\nThere is at least one x_9 for which x_9 is an artist if and only if x_9 will sleep. | For all x_9, if x_9 were an artist, then x_9 is sleeping can be inferred via the universal biconditional elimination rule.\nThere is at least one x_9 for which x_9 were an artist, then x_9 is sleeping can be inferred via the existential biconditional elimination rule.\nFor all x_9, if x_9 sleeps, then x_9 is an artist can be inferred via the universal biconditional elimination rule.\nThere is at least one x_9 for which x_9 sleeps, then x_9 is an artist can be inferred via the existential biconditional elimination rule |
| Consider the following premises ... Can we infer ... from them? name the inference rules being used | Consider the following premises.\nIf Charles takes a plane, then it is raining.\nIf Patricia were curious, then Charles is taking a plane.\nIf it rains, then Patricia is curious.\nIf Charles takes a plane, then John is not reading a book.\nJohn is reading a book.\nCan we infer the following from them? If we can, name the inference rule being used.\nIt is not raining | From the fact that if it rains, then Patricia is curious, and that if Patricia were curious, then Charles is taking a plane we can infer that if it rains, then Charles is taking a plane via transitivity.\nFrom the fact that if it rains, then Charles is taking a plane, and that if Charles takes a plane, then it is raining we can infer that it is raining if and only if Charles is taking a plane via biconditional introduction.\nFrom the fact that if Charles takes a plane, then John is not reading a book, and that John is reading a book we can infer that Charles is not taking a plane via modus tollens.\nFinally, from the fact that it is raining if and only if Charles is taking a plane, and that Charles is not taking a plane we can infer that it is not raining via biconditional elimination\nTherefore, the answer is yes. |
| Instruction | Input | Output |
| identify the claim that must be true or is required in order for the argument to work. | Passage: Almost all of the books published in the past 150 years were printed on acidic paper. Unfortunately, every kind of acidic paper gradually destroys itself due to its very acidity. This process of deterioration can be slowed if the books are stored in a cool, dry environment. Techniques, which are now being developed, to deacidify books will probably be applied only to books with historical significance.\nQuestion: If all of the statements in the passage above are true, which one of the following must also be true?\nA. If a book was published in the past 150 years and is historically insignificant, it will probably deteriorate completely.\nB. Almost all of the books published in the past 150 years will gradually destroy themselves.\nC. Books published on acidic paper in 1900 should now all be at about the same state of deterioration.\nD. If a book is of historical significance and was printed before 150 years ago, it will be圾ified.\nAnswer and reasoning step by step: | Answer: B. Almost all of the books published in the past 150 years will gradually destroy themselves.\nReasoning: The passage states that almost all of the books published in the past 150 years were printed on acidic paper and that every kind of acidic paper gradually destroys itself due to its acidity. Therefore, it must be true that almost all of the books published in the past 150 years will gradually destroy themselves. |
| identify information that would strengthen an argument | Passage: The television show Henry was not widely watched until it was scheduled for Tuesday evenings immediately after That's Life, the most popular show on television. During the year after the move, Henry was consistently one of the ten most-watched shows on television. Since Henry's recent move to Wednesday evenings, however, he has been watched by far fewer people. We must conclude that Henry was widely watched before the move to Wednesday evenings because it followed That's Life and not because people especially liked it.\nQuestion: Which one of the following, if true, most strengthens the argument?\nA. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved.\nB. Henry has been on the air for three years, but That's Life has been on the air for only two years.\nC. After its recent move to Wednesday, Henry was aired at the same time as the second most popular show on television.\nD. That's Life was not widely watched during the first year it was aired.\nAnswer and reasoning step by step: | Answer: A. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved.\nReasoning: The argument suggests that Henry was widely watched not because people especially liked it, but because it followed That's Life, which was the most popular show on television. Option A strengthens this argument because it shows that another show that now follows That's Life on Tuesdays also experienced a significant increase in viewership after being moved to follow That's Life. This supports the idea that being scheduled after That's Life is the main factor contributing to a show's popularity, rather than the show's content or quality itself. |
| identify information that would weaken an argument | Passage: Several companies will soon offer personalized electronic news services, delivered via cable or telephone lines and displayed on a television. People using these services can view continually updated stories on those topics for which they subscribe. Since these services will provide people with the information they are looking for more quickly and efficiently than printed newspapers can, newspaper sales will decline drastically if these services become widely available.\nQuestion: Which one of the following, if true, most seriously weakens the argument?\nA. Approximately 30 percent of people have never relied on newspapers for information but instead have always relied on news programs broadcast on television and radio.\nB. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics.\nC. Companies offering personalized electronic news services will differ greatly in what they charge for access to their services, depending on how wide a range of topics they cover.\nD. The average monthly cost of subscribing to several channels on a personalized electronic news service will approximately equal the cost of a month's subscription to a newspaper.\nAnswer and reasoning step by step: | B. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics.\nReasoning: The argument is based on the assumption that these personalized electronic news services will replace newspapers because they provide information more quickly and efficiently. However, if most people also enjoy browsing through newspapers for amusing stories on unfamiliar or unusual topics (option B), then these electronic services may not provide the same experience for their users. This means that even if the electronic news services are more efficient in delivering specific news, they might not fully replace newspapers because people still enjoy the browsing experience that newspapers provide. This weakens the argument that newspaper sales will decline drastically if these services become widely available. |
| identify information that would explain or resolve a situation | Passage: On May first, in order to reduce the number of overdue books, a children's library instituted a policy of forgiving fines and giving bookmarks to children returning all of their overdue books. On July first there were twice as many overdue books as there had been on May first, although a record number of books had been returned during the interim.\nQuestion: Which of the following, if true, most helps to explain the apparent inconsistency in the results of the library's policy?\nA. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date.\nB. The children were allowed to borrow a maximum of five books for a two-week period, and hence each child could keep a maximum of fifteen books beyond their due date within a two-month period.\nC. Although the grace period entered some children to return all of their overdue books, it did not convince all of the children with overdue books to return all of their books.\nD. Although the library forgave overdue fines during the grace period, the amount previously charged the children was minimal; hence, the forgiveness of the fines did not provide enough incentive for them to return their overdue books.\nAnswer and reasoning step by step: | A. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date.\nReasoning: The situation presents an inconsistency: despite forgiving fines and giving bookmarks as incentives to return overdue books, the number of overdue books doubled. We need to identify which statement helps explain this inconsistency.\nA. If the bookmarks became popular and children borrowed more books than usual to get them, it is possible that they couldn't read and return all those books on time, causing an increase in overdue books.\nThis option explains the inconsistency, as the policy would have an unintended consequence - children borrowing too many books.\nB. This option only specifies the borrowing limit and duration, but it does not explain why the number of overdue books would increase despite the incentives.\nC. This option implies that not all children returned their overdue books but it doesn't explain the doubling of the overdue books in the given period. It only suggests that the policy didn't work to its full potential.\nD. This option suggests that the fines were minimal, but it doesn't explain why the number of overdue books increased. If fines were minimal, we could assume that the situation would remain stable, not worsen.\nTherefore, option A most logically helps to explain the inconsistency in the results of the library's policy. |
| identify a flaw in an argument's reasoning | Passage: If you study history, then you will appreciate the vast differences among past civilizations, and you will appreciate these differences provided that you reflect on your own civilization. Hence, if you study history you will reflect on your own civilization.\nQuestion: Which one of the following is most closely parallel in its flawed reasoning to the flawed reasoning in the argument above?\nA. One can become informed about the world provided that one reads the newspaper daily. If one is informed about the world, then one has an appreciation of other cultures. So if one reads the newspaper daily, then one can come to appreciate other cultures.\nB. If you learn Latin, you can improve your vocabulary, and you can improve your vocabulary if you study great works of literature. So you will study great works of literature if you learn Latin.\nC. Traveling to other countries deepens one's appreciation for their cultures, and this appreciation often encourages one to study the history of those lands. So the study of history increases one's desire to travel.\nD. By studying ancient art you begin to appreciate how much was accomplished with limited materials.\nAppreciation of ancient art can engender a profound new appreciation for modern art. Therefore, studying ancient art can engender a profound new appreciation for modern art.\nAnswer and reasoning step by step: | B. The flaw reasoning in the original argument is that it assumes that if one thing leads to another, then the reverse is also true. In the original argument, studying history is assumed to lead to reflecting on one's own civilization, and this flaw is parallel in option B, where learning Latin is assumed to lead to studying great works of literature. Both assume a reverse causation that isn't necessarily true. |
| Verb | Noun object | Count |
| consider | premise | 24388 |
| translate | inference | 13206 |
| answer | question | 1840 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Accuracy | 43.83 | 28.94 | 45.40 | 19.82 | 29.57 | 33.09 | 31.47 | 53.00 | 35.64 |
| Task | Origin | Size |
| Language to Logic | LOGICIN inference & FOLIO | 13,206 |
| One-Step Inference | LOGICIN inference & FOLIO | 23,943 |
| Inference Chain | LOGICIN inference & EntailmentBank | 26,228 |
| Multi-choice | LogiQA & ReClor | 5,606 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Size | 1572 | 1594 | 500 | 230 | 1354 | 805 | 35891 | 10071 | 52017 |
| LLaMA-7b-base | 18.04 | 19.06 | 15.83 | 13.91 | 20.25 | 32.40 | 25.20 | 37.35 | 20.22 |
| LLaMA-30b-supercot | 19.31 | 26.35 | 17.81 | 17.98 | 18.41 | 24.10 | 32.26 | 41.91 | 24.78 |
| Falcon-40b-instruct | 23.21 | 19.77 | 26.77 | 12.70 | 17.33 | 16.13 | 28.49 | 44.66 | 23.63 |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| Dataset | Target |
| LogiQA 2.0 test | 4-way multi-choice |
| LogiQA 2.0 zh test | 4-way multi-choice |
| ReClor dev | 4-way multi-choice |
| AR-LSAT test | 5-way multi-choice |
| LogiQA 2.0 OOD | 4-way multi-choice |
| ConTRoL test | E, C, N |
| HELP test | E, C, N |
| TaxiNLI test | E, C, N |
| Task | LLaMA-7b-base | LLaMA-7b-logicot |
| Math | 25.1 | 29.0 |
| Health | 34.0 | 42.9 |
| Physics | 29.4 | 34.2 |
| Business | 34.8 | 57.2 |
| Biology | 32.4 | 46.5 |
| Chemistry | 25.4 | 33.0 |
| Computer science | 28.2 | 40.0 |
| Economics | 26.5 | 38.5 |
| Engineering | 25.5 | 32.4 |
| Philosophy | 30.8 | 37.6 |
| Other | 40.4 | 50.8 |
| History | 38.2 | 55.2 |
| Geography | 28.8 | 52.5 |
| Politics | 32.1 | 53.4 |
| Psychology | 33.2 | 50.9 |
| Culture | 37.3 | 56.9 |
| Law | 29.3 | 39.6 |
| STEM | 27.6 | 34.8 |
| Humanities | 31.6 | 41.8 |
| Social Sciences | 31.5 | 49.2 |
| Other (misc.) | 36.4 | 47.7 |
| Average | 31.8 | 43.3 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| ChatGPT | 52.37 | 53.18 | 57.38 | 20.42 | 38.44 | 58.45 | 42.13 | 57.30 | 47.46 |
| GPT-4 | 72.25 | 70.56 | 87.20 | 33.48 | 58.49 | 56.40 | 46.01 | 60.08 | 60.56 |
| Removed | LogiEval | MMLU |
| None (Full data) | 40.7 | 43.3 |
| Language to Logic | 32.4 | 38.5 |
| One-step Inference | 38.1 | 37.7 |
| Inference Chain | 30.8 | 35.0 |
| Multi-choice | 35.6 | 30.9 |
| Instruction | Input | Output |
| Translate the following inference to logic notation | All squares have four sides. | ∀x (Square(x) → FourSides(x)) |
| What can be inferred from the following premises in a single inference step | David is working if and only if it is cloudy.\nDavid is not working. | If David works, then it is cloudy.\nIf it is cloudy, then David is working.\nIt is not cloudy. |
| What can be inferred from the following premises in a single inference step? Name the inference rule being used | For all x_9, x_9 is an artist if and only if x_9 will sleep.\nThere is at least one x_9 for which x_9 is an artist if and only if x_9 will sleep. | For all x_9, if x_9 were an artist, then x_9 is sleeping can be inferred via the universal biconditional elimination rule.\nThere is at least one x_9 for which x_9 were an artist, then x_9 is sleeping can be inferred via the existential biconditional elimination rule.\nFor all x_9, if x_9 sleeps, then x_9 is an artist can be inferred via the universal biconditional elimination rule.\nThere is at least one x_9 for which x_9 sleeps, then x_9 is an artist can be inferred via the existential biconditional elimination rule |
| Consider the following premises ... Can we infer ... from them? name the inference rules being used | Consider the following premises.\nIf Charles takes a plane, then it is raining.\nIf Patricia were curious, then Charles is taking a plane.\nIf it rains, then Patricia is curious.\nIf Charles takes a plane, then John is not reading a book.\nJohn is reading a book.\nCan we infer the following from them? If we can, name the inference rule being used.\nIt is not raining | From the fact that if it rains, then Patricia is curious, and that if Patricia were curious, then Charles is taking a plane we can infer that if it rains, then Charles is taking a plane via transitivity.\nFrom the fact that if it rains, then Charles is taking a plane, and that if Charles takes a plane, then it is raining we can infer that it is raining if and only if Charles is taking a plane via biconditional introduction.\nFrom the fact that if Charles takes a plane, then John is not reading a book, and that John is reading a book we can infer that Charles is not taking a plane via modus tollens.\nFinally, from the fact that it is raining if and only if Charles is taking a plane, and that Charles is not taking a plane we can infer that it is not raining via biconditional elimination\nTherefore, the answer is yes. |
| Instruction | Input | Output |
| identify the claim that must be true or is required in order for the argument to work. | Passage: Almost all of the books published in the past 150 years were printed on acidic paper. Unfortunately, every kind of acidic paper gradually destroys itself due to its very acidity. This process of deterioration can be slowed if the books are stored in a cool, dry environment. Techniques, which are now being developed, to deacidify books will probably be applied only to books with historical significance.\nQuestion: If all of the statements in the passage above are true, which one of the following must also be true?\nA. If a book was published in the past 150 years and is historically insignificant, it will probably deteriorate completely.\nB. Almost all of the books published in the past 150 years will gradually destroy themselves.\nC. Books published on acidic paper in 1900 should now all be at about the same state of deterioration.\nD. If a book is of historical significance and was printed before 150 years ago, it will be圾ified.\nAnswer and reasoning step by step: | Answer: B. Almost all of the books published in the past 150 years will gradually destroy themselves.\nReasoning: The passage states that almost all of the books published in the past 150 years were printed on acidic paper and that every kind of acidic paper gradually destroys itself due to its acidity. Therefore, it must be true that almost all of the books published in the past 150 years will gradually destroy themselves. |
| identify information that would strengthen an argument | Passage: The television show Henry was not widely watched until it was scheduled for Tuesday evenings immediately after That's Life, the most popular show on television. During the year after the move, Henry was consistently one of the ten most-watched shows on television. Since Henry's recent move to Wednesday evenings, however, he has been watched by far fewer people. We must conclude that Henry was widely watched before the move to Wednesday evenings because it followed That's Life and not because people especially liked it.\nQuestion: Which one of the following, if true, most strengthens the argument?\nA. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved.\nB. Henry has been on the air for three years, but That's Life has been on the air for only two years.\nC. After its recent move to Wednesday, Henry was aired at the same time as the second most popular show on television.\nD. That's Life was not widely watched during the first year it was aired.\nAnswer and reasoning step by step: | Answer: A. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved.\nReasoning: The argument suggests that Henry was widely watched not because people especially liked it, but because it followed That's Life, which was the most popular show on television. Option A strengthens this argument because it shows that another show that now follows That's Life on Tuesdays also experienced a significant increase in viewership after being moved to follow That's Life. This supports the idea that being scheduled after That's Life is the main factor contributing to a show's popularity, rather than the show's content or quality itself. |
| identify information that would weaken an argument | Passage: Several companies will soon offer personalized electronic news services, delivered via cable or telephone lines and displayed on a television. People using these services can view continually updated stories on those topics for which they subscribe. Since these services will provide people with the information they are looking for more quickly and efficiently than printed newspapers can, newspaper sales will decline drastically if these services become widely available.\nQuestion: Which one of the following, if true, most seriously weakens the argument?\nA. Approximately 30 percent of people have never relied on newspapers for information but instead have always relied on news programs broadcast on television and radio.\nB. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics.\nC. Companies offering personalized electronic news services will differ greatly in what they charge for access to their services, depending on how wide a range of topics they cover.\nD. The average monthly cost of subscribing to several channels on a personalized electronic news service will approximately equal the cost of a month's subscription to a newspaper.\nAnswer and reasoning step by step: | B. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics.\nReasoning: The argument is based on the assumption that these personalized electronic news services will replace newspapers because they provide information more quickly and efficiently. However, if most people also enjoy browsing through newspapers for amusing stories on unfamiliar or unusual topics (option B), then these electronic services may not provide the same experience for their users. This means that even if the electronic news services are more efficient in delivering specific news, they might not fully replace newspapers because people still enjoy the browsing experience that newspapers provide. This weakens the argument that newspaper sales will decline drastically if these services become widely available. |
| identify information that would explain or resolve a situation | Passage: On May first, in order to reduce the number of overdue books, a children's library instituted a policy of forgiving fines and giving bookmarks to children returning all of their overdue books. On July first there were twice as many overdue books as there had been on May first, although a record number of books had been returned during the interim.\nQuestion: Which of the following, if true, most helps to explain the apparent inconsistency in the results of the library's policy?\nA. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date.\nB. The children were allowed to borrow a maximum of five books for a two-week period, and hence each child could keep a maximum of fifteen books beyond their due date within a two-month period.\nC. Although the grace period entered some children to return all of their overdue books, it did not convince all of the children with overdue books to return all of their books.\nD. Although the library forgave overdue fines during the grace period, the amount previously charged the children was minimal; hence, the forgiveness of the fines did not provide enough incentive for them to return their overdue books.\nAnswer and reasoning step by step: | A. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date.\nReasoning: The situation presents an inconsistency: despite forgiving fines and giving bookmarks as incentives to return overdue books, the number of overdue books doubled. We need to identify which statement helps explain this inconsistency.\nA. If the bookmarks became popular and children borrowed more books than usual to get them, it is possible that they couldn't read and return all those books on time, causing an increase in overdue books.\nThis option explains the inconsistency, as the policy would have an unintended consequence - children borrowing too many books.\nB. This option only specifies the borrowing limit and duration, but it does not explain why the number of overdue books would increase despite the incentives.\nC. This option implies that not all children returned their overdue books but it doesn't explain the doubling of the overdue books in the given period. It only suggests that the policy didn't work to its full potential.\nD. This option suggests that the fines were minimal, but it doesn't explain why the number of overdue books increased. If fines were minimal, we could assume that the situation would remain stable, not worsen.\nTherefore, option A most logically helps to explain the inconsistency in the results of the library's policy. |
| identify a flaw in an argument's reasoning | Passage: If you study history, then you will appreciate the vast differences among past civilizations, and you will appreciate these differences provided that you reflect on your own civilization. Hence, if you study history you will reflect on your own civilization.\nQuestion: Which one of the following is most closely parallel in its flawed reasoning to the flawed reasoning in the argument above?\nA. One can become informed about the world provided that one reads the newspaper daily. If one is informed about the world, then one has an appreciation of other cultures. So if one reads the newspaper daily, then one can come to appreciate other cultures.\nB. If you learn Latin, you can improve your vocabulary, and you can improve your vocabulary if you study great works of literature. So you will study great works of literature if you learn Latin.\nC. Traveling to other countries deepens one's appreciation for their cultures, and this appreciation often encourages one to study the history of those lands. So the study of history increases one's desire to travel.\nD. By studying ancient art you begin to appreciate how much was accomplished with limited materials.\nAppreciation of ancient art can engender a profound new appreciation for modern art. Therefore, studying ancient art can engender a profound new appreciation for modern art.\nAnswer and reasoning step by step: | B. The flaw reasoning in the original argument is that it assumes that if one thing leads to another, then the reverse is also true. In the original argument, studying history is assumed to lead to reflecting on one's own civilization, and this flaw is parallel in option B, where learning Latin is assumed to lead to studying great works of literature. Both assume a reverse causation that isn't necessarily true. |
| Verb | Noun object | Count |
| consider | premise | 24388 |
| translate | inference | 13206 |
| answer | question | 1840 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Accuracy | 43.83 | 28.94 | 45.40 | 19.82 | 29.57 | 33.09 | 31.47 | 53.00 | 35.64 |
| Task | Origin | Size |
| Language to Logic | LOGICIN inference & FOLIO | 13,206 |
| One-Step Inference | LOGICIN inference & FOLIO | 23,943 |
| Inference Chain | LOGICIN inference & EntailmentBank | 26,228 |
| Multi-choice | LogiQA & ReClor | 5,606 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Size | 1572 | 1594 | 500 | 230 | 1354 | 805 | 35891 | 10071 | 52017 |
| LLaMA-7b-base | 18.04 | 19.06 | 15.83 | 13.91 | 20.25 | 32.40 | 25.20 | 37.35 | 20.22 |
| LLaMA-30b-supercot | 19.31 | 26.35 | 17.81 | 17.98 | 18.41 | 24.10 | 32.26 | 41.91 | 24.78 |
| Falcon-40b-instruct | 23.21 | 19.77 | 26.77 | 12.70 | 17.33 | 16.13 | 28.49 | 44.66 | 23.63 |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| Dataset | Target |
| LogiQA 2.0 test | 4-way multi-choice |
| LogiQA 2.0 zh test | 4-way multi-choice |
| ReClor dev | 4-way multi-choice |
| AR-LSAT test | 5-way multi-choice |
| LogiQA 2.0 OOD | 4-way multi-choice |
| ConTRoL test | E, C, N |
| HELP test | E, C, N |
| TaxiNLI test | E, C, N |
| Task | LLaMA-7b-base | LLaMA-7b-logicot |
| Math | 25.1 | 29.0 |
| Health | 34.0 | 42.9 |
| Physics | 29.4 | 34.2 |
| Business | 34.8 | 57.2 |
| Biology | 32.4 | 46.5 |
| Chemistry | 25.4 | 33.0 |
| Computer science | 28.2 | 40.0 |
| Economics | 26.5 | 38.5 |
| Engineering | 25.5 | 32.4 |
| Philosophy | 30.8 | 37.6 |
| Other | 40.4 | 50.8 |
| History | 38.2 | 55.2 |
| Geography | 28.8 | 52.5 |
| Politics | 32.1 | 53.4 |
| Psychology | 33.2 | 50.9 |
| Culture | 37.3 | 56.9 |
| Law | 29.3 | 39.6 |
| STEM | 27.6 | 34.8 |
| Humanities | 31.6 | 41.8 |
| Social Sciences | 31.5 | 49.2 |
| Other (misc.) | 36.4 | 47.7 |
| Average | 31.8 | 43.3 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| ChatGPT | 52.37 | 53.18 | 57.38 | 20.42 | 38.44 | 58.45 | 42.13 | 57.30 | 47.46 |
| GPT-4 | 72.25 | 70.56 | 87.20 | 33.48 | 58.49 | 56.40 | 46.01 | 60.08 | 60.56 |
| Removed | LogiEval | MMLU |
| None (Full data) | 40.7 | 43.3 |
| Language to Logic | 32.4 | 38.5 |
| One-step Inference | 38.1 | 37.7 |
| Inference Chain | 30.8 | 35.0 |
| Multi-choice | 35.6 | 30.9 |
| Instruction | Input | Output |
| Translate the following inference to logic notation | All squares have four sides. | ∀x (Square(x) → FourSides(x)) |
| What can be inferred from the following premises in a single inference step | David is working if and only if it is cloudy. +David is not working. | If David works, then it is cloudy. +If it is cloudy, then David is working. +It is not cloudy. |
| What can be inferred from the following premises in a single inference step? Name the inference rule being used | For all x_9, x_9 is an artist if and only if x_9 will sleep. +There is at least one x_9 for which x_9 is an artist if and only if x_9 will sleep. | For all x_9, if x_9 were an artist, then x_9 is sleeping can be inferred via the universal biconditional elimination rule. +There is at least one x_9 for which x_9 were an artist, then x_9 is sleeping can be inferred via the existential biconditional elimination rule. +For all x_9, if x_9 sleeps, then x_9 is an artist can be inferred via the universal biconditional elimination rule. +There is at least one x_9 for which x_9 sleeps, then x_9 is an artist can be inferred via the existential biconditional elimination rule |
| Consider the following premises ... Can we infer ... from them? name the inference rules being used | Consider the following premises. +If Charles takes a plane, then it is raining. +If Patricia were curious, then Charles is taking a plane. +If it rains, then Patricia is curious. +If Charles takes a plane, then John is not reading a book. +John is reading a book. +Can we infer the following from them? If we can, name the inference rule being used. +It is not raining | From the fact that if it rains, then Patricia is curious, and that if Patricia were curious, then Charles is taking a plane we can infer that if it rains, then Charles is taking a plane via transitivity. +From the fact that if it rains, then Charles is taking a plane, and that if Charles takes a plane, then it is raining we can infer that it is raining if and only if Charles is taking a plane via biconditional introduction. +From the fact that if Charles takes a plane, then John is not reading a book, and that John is reading a book we can infer that Charles is not taking a plane via modus tollens. +Finally, from the fact that it is raining if and only if Charles is taking a plane, and that Charles is not taking a plane we can infer that it is not raining via biconditional elimination +Therefore, the answer is yes. |
| Instruction | Input | Output |
| identify the claim that must be true or is required in order for the argument to work. | Passage: Almost all of the books published in the past 150 years were printed on acidic paper. Unfortunately, every kind of acidic paper gradually destroys itself due to its very acidity. This process of deterioration can be slowed if the books are stored in a cool, dry environment. Techniques, which are now being developed, to deacidify books will probably be applied only to books with historical significance. +Question: If all of the statements in the passage above are true, which one of the following must also be true? +A. If a book was published in the past 150 years and is historically insignificant, it will probably deteriorate completely. +B. Almost all of the books published in the past 150 years will gradually destroy themselves. +C. Books published on acidic paper in 1900 should now all be at about the same state of deterioration. +D. If a book is of historical significance and was printed before 150 years ago, it will be圾ified. +Answer and reasoning step by step: | Answer: B. Almost all of the books published in the past 150 years will gradually destroy themselves. +Reasoning: The passage states that almost all of the books published in the past 150 years were printed on acidic paper and that every kind of acidic paper gradually destroys itself due to its acidity. Therefore, it must be true that almost all of the books published in the past 150 years will gradually destroy themselves. |
| identify information that would strengthen an argument | Passage: The television show Henry was not widely watched until it was scheduled for Tuesday evenings immediately after That's Life, the most popular show on television. During the year after the move, Henry was consistently one of the ten most-watched shows on television. Since Henry's recent move to Wednesday evenings, however, he has been watched by far fewer people. We must conclude that Henry was widely watched before the move to Wednesday evenings because it followed That's Life and not because people especially liked it. +Question: Which one of the following, if true, most strengthens the argument? +A. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved. +B. Henry has been on the air for three years, but That's Life has been on the air for only two years. +C. After its recent move to Wednesday, Henry was aired at the same time as the second most popular show on television. +D. That's Life was not widely watched during the first year it was aired. +Answer and reasoning step by step: | Answer: A. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved. +Reasoning: The argument suggests that Henry was widely watched not because people especially liked it, but because it followed That's Life, which was the most popular show on television. Option A strengthens this argument because it shows that another show that now follows That's Life on Tuesdays also experienced a significant increase in viewership after being moved to follow That's Life. This supports the idea that being scheduled after That's Life is the main factor contributing to a show's popularity, rather than the show's content or quality itself. |
| identify information that would weaken an argument | Passage: Several companies will soon offer personalized electronic news services, delivered via cable or telephone lines and displayed on a television. People using these services can view continually updated stories on those topics for which they subscribe. Since these services will provide people with the information they are looking for more quickly and efficiently than printed newspapers can, newspaper sales will decline drastically if these services become widely available. +Question: Which one of the following, if true, most seriously weakens the argument? +A. Approximately 30 percent of people have never relied on newspapers for information but instead have always relied on news programs broadcast on television and radio. +B. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics. +C. Companies offering personalized electronic news services will differ greatly in what they charge for access to their services, depending on how wide a range of topics they cover. +D. The average monthly cost of subscribing to several channels on a personalized electronic news service will approximately equal the cost of a month's subscription to a newspaper. +Answer and reasoning step by step: | B. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics. +Reasoning: The argument is based on the assumption that these personalized electronic news services will replace newspapers because they provide information more quickly and efficiently. However, if most people also enjoy browsing through newspapers for amusing stories on unfamiliar or unusual topics (option B), then these electronic services may not provide the same experience for their users. This means that even if the electronic news services are more efficient in delivering specific news, they might not fully replace newspapers because people still enjoy the browsing experience that newspapers provide. This weakens the argument that newspaper sales will decline drastically if these services become widely available. |
| identify information that would explain or resolve a situation | Passage: On May first, in order to reduce the number of overdue books, a children's library instituted a policy of forgiving fines and giving bookmarks to children returning all of their overdue books. On July first there were twice as many overdue books as there had been on May first, although a record number of books had been returned during the interim. +Question: Which of the following, if true, most helps to explain the apparent inconsistency in the results of the library's policy? +A. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date. +B. The children were allowed to borrow a maximum of five books for a two-week period, and hence each child could keep a maximum of fifteen books beyond their due date within a two-month period. +C. Although the grace period entered some children to return all of their overdue books, it did not convince all of the children with overdue books to return all of their books. +D. Although the library forgave overdue fines during the grace period, the amount previously charged the children was minimal; hence, the forgiveness of the fines did not provide enough incentive for them to return their overdue books. +Answer and reasoning step by step: | A. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date. +Reasoning: The situation presents an inconsistency: despite forgiving fines and giving bookmarks as incentives to return overdue books, the number of overdue books doubled. We need to identify which statement helps explain this inconsistency. +A. If the bookmarks became popular and children borrowed more books than usual to get them, it is possible that they couldn't read and return all those books on time, causing an increase in overdue books. +This option explains the inconsistency, as the policy would have an unintended consequence - children borrowing too many books. +B. This option only specifies the borrowing limit and duration, but it does not explain why the number of overdue books would increase despite the incentives. +C. This option implies that not all children returned their overdue books but it doesn't explain the doubling of the overdue books in the given period. It only suggests that the policy didn't work to its full potential. +D. This option suggests that the fines were minimal, but it doesn't explain why the number of overdue books increased. If fines were minimal, we could assume that the situation would remain stable, not worsen. +Therefore, option A most logically helps to explain the inconsistency in the results of the library's policy. |
| identify a flaw in an argument's reasoning | Passage: If you study history, then you will appreciate the vast differences among past civilizations, and you will appreciate these differences provided that you reflect on your own civilization. Hence, if you study history you will reflect on your own civilization. +Question: Which one of the following is most closely parallel in its flawed reasoning to the flawed reasoning in the argument above? +A. One can become informed about the world provided that one reads the newspaper daily. If one is informed about the world, then one has an appreciation of other cultures. So if one reads the newspaper daily, then one can come to appreciate other cultures. +B. If you learn Latin, you can improve your vocabulary, and you can improve your vocabulary if you study great works of literature. So you will study great works of literature if you learn Latin. +C. Traveling to other countries deepens one's appreciation for their cultures, and this appreciation often encourages one to study the history of those lands. So the study of history increases one's desire to travel. +D. By studying ancient art you begin to appreciate how much was accomplished with limited materials. +Appreciation of ancient art can engender a profound new appreciation for modern art. Therefore, studying ancient art can engender a profound new appreciation for modern art. +Answer and reasoning step by step: | B. The flaw reasoning in the original argument is that it assumes that if one thing leads to another, then the reverse is also true. In the original argument, studying history is assumed to lead to reflecting on one's own civilization, and this flaw is parallel in option B, where learning Latin is assumed to lead to studying great works of literature. Both assume a reverse causation that isn't necessarily true. |
| Verb | Noun object | Count |
| consider | premise | 24388 |
| translate | inference | 13206 |
| answer | question | 1840 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Accuracy | 43.83 | 28.94 | 45.40 | 19.82 | 29.57 | 33.09 | 31.47 | 53.00 | 35.64 |
| Task | Origin | Size |
| Language to Logic | LOGICIN inference & FOLIO | 13,206 |
| One-Step Inference | LOGICIN inference & FOLIO | 23,943 |
| Inference Chain | LOGICIN inference & EntailmentBank | 26,228 |
| Multi-choice | LogiQA & ReClor | 5,606 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Size | 1572 | 1594 | 500 | 230 | 1354 | 805 | 35891 | 10071 | 52017 |
| LLaMA-7b-base | 18.04 | 19.06 | 15.83 | 13.91 | 20.25 | 32.40 | 25.20 | 37.35 | 20.22 |
| LLaMA-30b-supercot | 19.31 | 26.35 | 17.81 | 17.98 | 18.41 | 24.10 | 32.26 | 41.91 | 24.78 |
| Falcon-40b-instruct | 23.21 | 19.77 | 26.77 | 12.70 | 17.33 | 16.13 | 28.49 | 44.66 | 23.63 |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| Dataset | Target |
| LogiQA 2.0 test | 4-way multi-choice |
| LogiQA 2.0 zh test | 4-way multi-choice |
| ReClor dev | 4-way multi-choice |
| AR-LSAT test | 5-way multi-choice |
| LogiQA 2.0 OOD | 4-way multi-choice |
| ConTRoL test | E, C, N |
| HELP test | E, C, N |
| TaxiNLI test | E, C, N |
| Task | LLaMA-7b-base | LLaMA-7b-logicot |
| Math | 25.1 | 29.0 |
| Health | 34.0 | 42.9 |
| Physics | 29.4 | 34.2 |
| Business | 34.8 | 57.2 |
| Biology | 32.4 | 46.5 |
| Chemistry | 25.4 | 33.0 |
| Computer science | 28.2 | 40.0 |
| Economics | 26.5 | 38.5 |
| Engineering | 25.5 | 32.4 |
| Philosophy | 30.8 | 37.6 |
| Other | 40.4 | 50.8 |
| History | 38.2 | 55.2 |
| Geography | 28.8 | 52.5 |
| Politics | 32.1 | 53.4 |
| Psychology | 33.2 | 50.9 |
| Culture | 37.3 | 56.9 |
| Law | 29.3 | 39.6 |
| STEM | 27.6 | 34.8 |
| Humanities | 31.6 | 41.8 |
| Social Sciences | 31.5 | 49.2 |
| Other (misc.) | 36.4 | 47.7 |
| Average | 31.8 | 43.3 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| LLaMA-7b-logicot | 50.25 | 32.77 | 57.60 | 16.96 | 38.79 | 35.68 | 35.44 | 58.05 | 40.69 |
| ChatGPT | 52.37 | 53.18 | 57.38 | 20.42 | 38.44 | 58.45 | 42.13 | 57.30 | 47.46 |
| GPT-4 | 72.25 | 70.56 | 87.20 | 33.48 | 58.49 | 56.40 | 46.01 | 60.08 | 60.56 |
| Removed | LogiEval | MMLU |
| None (Full data) | 40.7 | 43.3 |
| Language to Logic | 32.4 | 38.5 |
| One-step Inference | 38.1 | 37.7 |
| Inference Chain | 30.8 | 35.0 |
| Multi-choice | 35.6 | 30.9 |
| Instruction | Input | Output |
| Translate the following inference to logic notation | All squares have four sides. | ∀x (Square(x) → FourSides(x)) |
| What can be inferred from the following premises in a single inference step | David is working if and only if it is cloudy.\nDavid is not working. | If David works, then it is cloudy.\nIf it is cloudy, then David is working.\nIt is not cloudy. |
| What can be inferred from the following premises in a single inference step? Name the inference rule being used | For all x_9, x_9 is an artist if and only if x_9 will sleep.\nThere is at least one x_9 for which x_9 is an artist if and only if x_9 will sleep. | For all x_9, if x_9 were an artist, then x_9 is sleeping can be inferred via the universal biconditional elimination rule.\nThere is at least one x_9 for which x_9 were an artist, then x_9 is sleeping can be inferred via the existential biconditional elimination rule.\nFor all x_9, if x_9 sleeps, then x_9 is an artist can be inferred via the universal biconditional elimination rule.\nThere is at least one x_9 for which x_9 sleeps, then x_9 is an artist can be inferred via the existential biconditional elimination rule |
| Consider the following premises ... Can we infer ... from them? name the inference rules being used | Consider the following premises.\nIf Charles takes a plane, then it is raining.\nIf Patricia were curious, then Charles is taking a plane.\nIf it rains, then Patricia is curious.\nIf Charles takes a plane, then John is not reading a book.\nJohn is reading a book.\nCan we infer the following from them? If we can, name the inference rule being used.\nIt is not raining | From the fact that if it rains, then Patricia is curious, and that if Patricia were curious, then Charles is taking a plane we can infer that if it rains, then Charles is taking a plane via transitivity.\nFrom the fact that if it rains, then Charles is taking a plane, and that if Charles takes a plane, then it is raining we can infer that it is raining if and only if Charles is taking a plane via biconditional introduction.\nFrom the fact that if Charles takes a plane, then John is not reading a book, and that John is reading a book we can infer that Charles is not taking a plane via modus tollens.\nFinally, from the fact that it is raining if and only if Charles is taking a plane, and that Charles is not taking a plane we can infer that it is not raining via biconditional elimination\nTherefore, the answer is yes. |
| Instruction | Input | Output |
| identify the claim that must be true or is required in order for the argument to work. | Passage: Almost all of the books published in the past 150 years were printed on acidic paper. Unfortunately, every kind of acidic paper gradually destroys itself due to its very acidity. This process of deterioration can be slowed if the books are stored in a cool, dry environment. Techniques, which are now being developed, to deacidify books will probably be applied only to books with historical significance.\nQuestion: If all of the statements in the passage above are true, which one of the following must also be true?\nA. If a book was published in the past 150 years and is historically insignificant, it will probably deteriorate completely.\nB. Almost all of the books published in the past 150 years will gradually destroy themselves.\nC. Books published on acidic paper in 1900 should now all be at about the same state of deterioration.\nD. If a book is of historical significance and was printed before 150 years ago, it will be圾ified.\nAnswer and reasoning step by step: | Answer: B. Almost all of the books published in the past 150 years will gradually destroy themselves.\nReasoning: The passage states that almost all of the books published in the past 150 years were printed on acidic paper and that every kind of acidic paper gradually destroys itself due to its acidity. Therefore, it must be true that almost all of the books published in the past 150 years will gradually destroy themselves. |
| identify information that would strengthen an argument | Passage: The television show Henry was not widely watched until it was scheduled for Tuesday evenings immediately after That's Life, the most popular show on television. During the year after the move, Henry was consistently one of the ten most-watched shows on television. Since Henry's recent move to Wednesday evenings, however, he has been watched by far fewer people. We must conclude that Henry was widely watched before the move to Wednesday evenings because it followed That's Life and not because people especially liked it.\nQuestion: Which one of the following, if true, most strengthens the argument?\nA. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved.\nB. Henry has been on the air for three years, but That's Life has been on the air for only two years.\nC. After its recent move to Wednesday, Henry was aired at the same time as the second most popular show on television.\nD. That's Life was not widely watched during the first year it was aired.\nAnswer and reasoning step by step: | Answer: A. The show that now follows That's Life on Tuesdays has double the number of viewers it had before being moved.\nReasoning: The argument suggests that Henry was widely watched not because people especially liked it, but because it followed That's Life, which was the most popular show on television. Option A strengthens this argument because it shows that another show that now follows That's Life on Tuesdays also experienced a significant increase in viewership after being moved to follow That's Life. This supports the idea that being scheduled after That's Life is the main factor contributing to a show's popularity, rather than the show's content or quality itself. |
| identify information that would weaken an argument | Passage: Several companies will soon offer personalized electronic news services, delivered via cable or telephone lines and displayed on a television. People using these services can view continually updated stories on those topics for which they subscribe. Since these services will provide people with the information they are looking for more quickly and efficiently than printed newspapers can, newspaper sales will decline drastically if these services become widely available.\nQuestion: Which one of the following, if true, most seriously weakens the argument?\nA. Approximately 30 percent of people have never relied on newspapers for information but instead have always relied on news programs broadcast on television and radio.\nB. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics.\nC. Companies offering personalized electronic news services will differ greatly in what they charge for access to their services, depending on how wide a range of topics they cover.\nD. The average monthly cost of subscribing to several channels on a personalized electronic news service will approximately equal the cost of a month's subscription to a newspaper.\nAnswer and reasoning step by step: | B. In reading newspapers, most people not only look for stories on specific topics but also like to idly browse through headlines or pictures for amusing stories on unfamiliar or unusual topics.\nReasoning: The argument is based on the assumption that these personalized electronic news services will replace newspapers because they provide information more quickly and efficiently. However, if most people also enjoy browsing through newspapers for amusing stories on unfamiliar or unusual topics (option B), then these electronic services may not provide the same experience for their users. This means that even if the electronic news services are more efficient in delivering specific news, they might not fully replace newspapers because people still enjoy the browsing experience that newspapers provide. This weakens the argument that newspaper sales will decline drastically if these services become widely available. |
| identify information that would explain or resolve a situation | Passage: On May first, in order to reduce the number of overdue books, a children's library instituted a policy of forgiving fines and giving bookmarks to children returning all of their overdue books. On July first there were twice as many overdue books as there had been on May first, although a record number of books had been returned during the interim.\nQuestion: Which of the following, if true, most helps to explain the apparent inconsistency in the results of the library's policy?\nA. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date.\nB. The children were allowed to borrow a maximum of five books for a two-week period, and hence each child could keep a maximum of fifteen books beyond their due date within a two-month period.\nC. Although the grace period entered some children to return all of their overdue books, it did not convince all of the children with overdue books to return all of their books.\nD. Although the library forgave overdue fines during the grace period, the amount previously charged the children was minimal; hence, the forgiveness of the fines did not provide enough incentive for them to return their overdue books.\nAnswer and reasoning step by step: | A. The bookmarks became popular among the children, so in order to collect the bookmarks, many children borrowed many more books than they usually did and kept them past their due date.\nReasoning: The situation presents an inconsistency: despite forgiving fines and giving bookmarks as incentives to return overdue books, the number of overdue books doubled. We need to identify which statement helps explain this inconsistency.\nA. If the bookmarks became popular and children borrowed more books than usual to get them, it is possible that they couldn't read and return all those books on time, causing an increase in overdue books.\nThis option explains the inconsistency, as the policy would have an unintended consequence - children borrowing too many books.\nB. This option only specifies the borrowing limit and duration, but it does not explain why the number of overdue books would increase despite the incentives.\nC. This option implies that not all children returned their overdue books but it doesn't explain the doubling of the overdue books in the given period. It only suggests that the policy didn't work to its full potential.\nD. This option suggests that the fines were minimal, but it doesn't explain why the number of overdue books increased. If fines were minimal, we could assume that the situation would remain stable, not worsen.\nTherefore, option A most logically helps to explain the inconsistency in the results of the library's policy. |
| identify a flaw in an argument's reasoning | Passage: If you study history, then you will appreciate the vast differences among past civilizations, and you will appreciate these differences provided that you reflect on your own civilization. Hence, if you study history you will reflect on your own civilization.\nQuestion: Which one of the following is most closely parallel in its flawed reasoning to the flawed reasoning in the argument above?\nA. One can become informed about the world provided that one reads the newspaper daily. If one is informed about the world, then one has an appreciation of other cultures. So if one reads the newspaper daily, then one can come to appreciate other cultures.\nB. If you learn Latin, you can improve your vocabulary, and you can improve your vocabulary if you study great works of literature. So you will study great works of literature if you learn Latin.\nC. Traveling to other countries deepens one's appreciation for their cultures, and this appreciation often encourages one to study the history of those lands. So the study of history increases one's desire to travel.\nD. By studying ancient art you begin to appreciate how much was accomplished with limited materials.\nAppreciation of ancient art can engender a profound new appreciation for modern art. Therefore, studying ancient art can engender a profound new appreciation for modern art.\nAnswer and reasoning step by step: | B. The flaw reasoning in the original argument is that it assumes that if one thing leads to another, then the reverse is also true. In the original argument, studying history is assumed to lead to reflecting on one's own civilization, and this flaw is parallel in option B, where learning Latin is assumed to lead to studying great works of literature. Both assume a reverse causation that isn't necessarily true. |
| Verb | Noun object | Count |
| consider | premise | 24388 |
| translate | inference | 13206 |
| answer | question | 1840 |
| Dataset | LQ | LQ zh | RC | AL | LQ ood | CT | HL | TN | Overall |
| Accuracy | 43.83 | 28.94 | 45.40 | 19.82 | 29.57 | 33.09 | 31.47 | 53.00 | 35.64 |
| Model | False Negative Ratio |
| Random | 3.29 |
| Semi (Li et al., 2019) | 7.16 |
| CIR (Penha and Hauff, 2020) | 7.03 |
| Grey (Lin et al., 2020) | 7.42 |
| HCL (Su et al., 2020) | 6.95 |
| Mask-and-fill (Gupta et al., 2021) | 5.63 |
| Model | False Negative Preventing | Negative Updating |
| Semi (Li et al., 2019) | X | X |
| CIR(Penha and Hauff, 2020) | X | ✓ |
| HCL(Su et al., 2020) | X | ✓ |
| Grey(Lin et al., 2020) | X | X |
| Mask-and-fill(Gupta et al., 2021) | ✓ | X |
| Ours | ✓ | ✓ |
| Model | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT (Devlin et al., 2018) | 80.8 | 89.7 | 97.5 | 59.1 | 63.3 | 45.4 | 28.0 | 47.0 |
| +Semi (Li et al., 2019) | 81.0 | 89.8 | 97.5 | 60.3 | 63.8 | 46.0 | 28.2 | 47.9 |
| +CIR (Penha and Hauff, 2020) | 81.2 | 89.9 | 97.6 | 60.7 | 64.2 | 46.4 | 28.5 | 49.1 |
| +Gray (Lin et al., 2020) | 81.5 | 90.1 | 97.5 | 61.5 | 64.8 | 47.1 | 29.1 | 50.7 |
| +HCL (Su et al., 2020) | 81.5 | 90.2 | 97.6 | 61.7 | 65.9 | 48.0 | 30.4 | 51.0 |
| +MF (Gupta et al., 2021) | 81.6 | 90.0 | 97.5 | 61.9 | 66.0 | 48.1 | 30.5 | 51.3 |
| +TRIGGER (Ours) | 81.9† | 90.4 | 97.5 | 62.9† | 67.8† | 50.2† | 32.5† | 52.1† |
| SA-BERT (Gu et al., 2020) | 85.5 | 92.8 | 98.3 | 61.9 | 65.9 | 49.6 | 31.3 | 48.1 |
| +Semi (Li et al., 2019) | 85.8 | 93.1 | 98.9 | 62.3 | 66.4 | 50.0 | 31.7 | 49.0 |
| +CIR (Penha and Hauff, 2020) | 86.0 | 93.5 | 99.0 | 62.4 | 66.6 | 50.3 | 31.8 | 49.7 |
| +Gray (Lin et al., 2020) | 86.1 | 93.4 | 99.1 | 62.8 | 67.0 | 50.3 | 32.0 | 50.3 |
| +HCL (Su et al., 2020) | 86.7 | 94.0 | 99.2 | 63.9 | 68.1 | 51.4 | 33.0 | 53.1 |
| +MF (Gupta et al., 2021) | 86.9 | 93.7 | 98.9 | 63.6 | 68.4 | 52.0 | 33.5 | 52.9 |
| +TRIGGER (Ours) | 87.1† | 93.8 | 98.7 | 64.1 | 69.5† | 52.9† | 34.3† | 53.7† |
| BERT-FP (Han et al., 2021) | 91.1 | 96.2 | 99.4 | 64.4 | 68.0 | 51.2 | 32.4 | 54.2 |
| +Semi (Li et al., 2019) | 91.2 | 96.2 | 99.3 | 65.8 | 68.7 | 53.6 | 34.3 | 54.7 |
| +CIR (Penha and Hauff, 2020) | 91.2 | 96.4 | 99.3 | 66.5 | 70.6 | 54.8 | 35.5 | 55.0 |
| +Gray (Lin et al., 2020) | 91.3 | 96.3 | 99.4 | 66.7 | 70.5 | 55.0 | 36.3 | 54.8 |
| +HCL (Su et al., 2020) | 91.2 | 96.3 | 99.4 | 67.1 | 70.9 | 55.3 | 36.4 | 55.7 |
| +MF (Gupta et al., 2021) | 91.3 | 96.4 | 99.3 | 67.2 | 71.4 | 55.4 | 36.2 | 56.7 |
| +TRIGGER (Ours) | 91.7† | 96.6 | 99.4 | 67.9† | 72.4† | 56.5† | 36.7 | 59.5† |
| Rank | Transition | Probability |
| 1 | 233→115 | 14.25% |
| 2 | 265→211 | 12.90% |
| 3 | 211→265 | 10.45% |
| 4 | 274→210 | 2.20% |
| 5 | 110→274 | 2.22% |
| Rank | Latent Label | Implication |
| 1 | 233 | Short instructions about paste operation. |
| 2 | 115 | Reference to a URL. |
| 3 | 211 | Audio and sound configuration in Ubuntu. |
| 4 | 265 | Sound system breakdown. |
| 5 | 110 | Brief yes/no answer. |
| Models | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT(Devlin et al., 2018) | 2.0 | 4.0 | 17.0 | 22.3 | 25.0 | 8.7 | 3.2 | 6.5 |
| BERT+TRIGGER | 13.4† | 20.2† | 48.1† | 37.6† | 42.0† | 23.5† | 13.7† | 21.1† |
| BERT-FP (Han et al., 2021) | 52.5 | 57.4 | 69.2 | 32.5 | 36.8 | 21.9 | 12.9 | 17.3 |
| BERT-FP+TRIGGER | 76.2† | 81.9† | 90.1† | 38.8† | 43.0† | 25.4† | 16.2† | 23.5† |
| Model | False Negative Ratio |
| Random | 3.29 |
| Semi (Li et al., 2019) | 7.16 |
| CIR (Penha and Hauff, 2020) | 7.03 |
| Grey (Lin et al., 2020) | 7.42 |
| HCL (Su et al., 2020) | 6.95 |
| Mask-and-fill (Gupta et al., 2021) | 5.63 |
| Model | False Negative Preventing | Negative Updating |
| Semi (Li et al., 2019) | X | X |
| CIR(Penha and Hauff, 2020) | X | ✓ |
| HCL(Su et al., 2020) | X | ✓ |
| Grey(Lin et al., 2020) | X | X |
| Mask-and-fill(Gupta et al., 2021) | ✓ | X |
| Ours | ✓ | ✓ |
| Model | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT (Devlin et al., 2018) | 80.8 | 89.7 | 97.5 | 59.1 | 63.3 | 45.4 | 28.0 | 47.0 |
| +Semi (Li et al., 2019) | 81.0 | 89.8 | 97.5 | 60.3 | 63.8 | 46.0 | 28.2 | 47.9 |
| +CIR (Penha and Hauff, 2020) | 81.2 | 89.9 | 97.6 | 60.7 | 64.2 | 46.4 | 28.5 | 49.1 |
| +Gray (Lin et al., 2020) | 81.5 | 90.1 | 97.5 | 61.5 | 64.8 | 47.1 | 29.1 | 50.7 |
| +HCL (Su et al., 2020) | 81.5 | 90.2 | 97.6 | 61.7 | 65.9 | 48.0 | 30.4 | 51.0 |
| +MF (Gupta et al., 2021) | 81.6 | 90.0 | 97.5 | 61.9 | 66.0 | 48.1 | 30.5 | 51.3 |
| +TRIGGER (Ours) | 81.9† | 90.4 | 97.5 | 62.9† | 67.8† | 50.2† | 32.5† | 52.1† |
| SA-BERT (Gu et al., 2020) | 85.5 | 92.8 | 98.3 | 61.9 | 65.9 | 49.6 | 31.3 | 48.1 |
| +Semi (Li et al., 2019) | 85.8 | 93.1 | 98.9 | 62.3 | 66.4 | 50.0 | 31.7 | 49.0 |
| +CIR (Penha and Hauff, 2020) | 86.0 | 93.5 | 99.0 | 62.4 | 66.6 | 50.3 | 31.8 | 49.7 |
| +Gray (Lin et al., 2020) | 86.1 | 93.4 | 99.1 | 62.8 | 67.0 | 50.3 | 32.0 | 50.3 |
| +HCL (Su et al., 2020) | 86.7 | 94.0 | 99.2 | 63.9 | 68.1 | 51.4 | 33.0 | 53.1 |
| +MF (Gupta et al., 2021) | 86.9 | 93.7 | 98.9 | 63.6 | 68.4 | 52.0 | 33.5 | 52.9 |
| +TRIGGER (Ours) | 87.1† | 93.8 | 98.7 | 64.1 | 69.5† | 52.9† | 34.3† | 53.7† |
| BERT-FP (Han et al., 2021) | 91.1 | 96.2 | 99.4 | 64.4 | 68.0 | 51.2 | 32.4 | 54.2 |
| +Semi (Li et al., 2019) | 91.2 | 96.2 | 99.3 | 65.8 | 68.7 | 53.6 | 34.3 | 54.7 |
| +CIR (Penha and Hauff, 2020) | 91.2 | 96.4 | 99.3 | 66.5 | 70.6 | 54.8 | 35.5 | 55.0 |
| +Gray (Lin et al., 2020) | 91.3 | 96.3 | 99.4 | 66.7 | 70.5 | 55.0 | 36.3 | 54.8 |
| +HCL (Su et al., 2020) | 91.2 | 96.3 | 99.4 | 67.1 | 70.9 | 55.3 | 36.4 | 55.7 |
| +MF (Gupta et al., 2021) | 91.3 | 96.4 | 99.3 | 67.2 | 71.4 | 55.4 | 36.2 | 56.7 |
| +TRIGGER (Ours) | 91.7† | 96.6 | 99.4 | 67.9† | 72.4† | 56.5† | 36.7 | 59.5† |
| Rank | Transition | Probability |
| 1 | 233→115 | 14.25% |
| 2 | 265→211 | 12.90% |
| 3 | 211→265 | 10.45% |
| 4 | 274→210 | 2.20% |
| 5 | 110→274 | 2.22% |
| Rank | Latent Label | Implication |
| 1 | 233 | Short instructions about paste operation. |
| 2 | 115 | Reference to a URL. |
| 3 | 211 | Audio and sound configuration in Ubuntu. |
| 4 | 265 | Sound system breakdown. |
| 5 | 110 | Brief yes/no answer. |
| Models | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT(Devlin et al., 2018) | 2.0 | 4.0 | 17.0 | 22.3 | 25.0 | 8.7 | 3.2 | 6.5 |
| BERT+TRIGGER | 13.4† | 20.2† | 48.1† | 37.6† | 42.0† | 23.5† | 13.7† | 21.1† |
| BERT-FP (Han et al., 2021) | 52.5 | 57.4 | 69.2 | 32.5 | 36.8 | 21.9 | 12.9 | 17.3 |
| BERT-FP+TRIGGER | 76.2† | 81.9† | 90.1† | 38.8† | 43.0† | 25.4† | 16.2† | 23.5† |
| Model | False Negative Ratio |
| Random | 3.29 |
| Semi (Li et al., 2019) | 7.16 |
| CIR (Penha and Hauff, 2020) | 7.03 |
| Grey (Lin et al., 2020) | 7.42 |
| HCL (Su et al., 2020) | 6.95 |
| Mask-and-fill (Gupta et al., 2021) | 5.63 |
| Model | False Negative Preventing | Negative Updating |
| Semi (Li et al., 2019) | X | X |
| CIR(Penha and Hauff, 2020) | X | ✓ |
| HCL(Su et al., 2020) | X | ✓ |
| Grey(Lin et al., 2020) | X | X |
| Mask-and-fill(Gupta et al., 2021) | ✓ | X |
| Ours | ✓ | ✓ |
| Model | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT (Devlin et al., 2018) | 80.8 | 89.7 | 97.5 | 59.1 | 63.3 | 45.4 | 28.0 | 47.0 |
| +Semi (Li et al., 2019) | 81.0 | 89.8 | 97.5 | 60.3 | 63.8 | 46.0 | 28.2 | 47.9 |
| +CIR (Penha and Hauff, 2020) | 81.2 | 89.9 | 97.6 | 60.7 | 64.2 | 46.4 | 28.5 | 49.1 |
| +Gray (Lin et al., 2020) | 81.5 | 90.1 | 97.5 | 61.5 | 64.8 | 47.1 | 29.1 | 50.7 |
| +HCL (Su et al., 2020) | 81.5 | 90.2 | 97.6 | 61.7 | 65.9 | 48.0 | 30.4 | 51.0 |
| +MF (Gupta et al., 2021) | 81.6 | 90.0 | 97.5 | 61.9 | 66.0 | 48.1 | 30.5 | 51.3 |
| +TRIGGER (Ours) | 81.9† | 90.4 | 97.5 | 62.9† | 67.8† | 50.2† | 32.5† | 52.1† |
| SA-BERT (Gu et al., 2020) | 85.5 | 92.8 | 98.3 | 61.9 | 65.9 | 49.6 | 31.3 | 48.1 |
| +Semi (Li et al., 2019) | 85.8 | 93.1 | 98.9 | 62.3 | 66.4 | 50.0 | 31.7 | 49.0 |
| +CIR (Penha and Hauff, 2020) | 86.0 | 93.5 | 99.0 | 62.4 | 66.6 | 50.3 | 31.8 | 49.7 |
| +Gray (Lin et al., 2020) | 86.1 | 93.4 | 99.1 | 62.8 | 67.0 | 50.3 | 32.0 | 50.3 |
| +HCL (Su et al., 2020) | 86.7 | 94.0 | 99.2 | 63.9 | 68.1 | 51.4 | 33.0 | 53.1 |
| +MF (Gupta et al., 2021) | 86.9 | 93.7 | 98.9 | 63.6 | 68.4 | 52.0 | 33.5 | 52.9 |
| +TRIGGER (Ours) | 87.1† | 93.8 | 98.7 | 64.1 | 69.5† | 52.9† | 34.3† | 53.7† |
| BERT-FP (Han et al., 2021) | 91.1 | 96.2 | 99.4 | 64.4 | 68.0 | 51.2 | 32.4 | 54.2 |
| +Semi (Li et al., 2019) | 91.2 | 96.2 | 99.3 | 65.8 | 68.7 | 53.6 | 34.3 | 54.7 |
| +CIR (Penha and Hauff, 2020) | 91.2 | 96.4 | 99.3 | 66.5 | 70.6 | 54.8 | 35.5 | 55.0 |
| +Gray (Lin et al., 2020) | 91.3 | 96.3 | 99.4 | 66.7 | 70.5 | 55.0 | 36.3 | 54.8 |
| +HCL (Su et al., 2020) | 91.2 | 96.3 | 99.4 | 67.1 | 70.9 | 55.3 | 36.4 | 55.7 |
| +MF (Gupta et al., 2021) | 91.3 | 96.4 | 99.3 | 67.2 | 71.4 | 55.4 | 36.2 | 56.7 |
| +TRIGGER (Ours) | 91.7† | 96.6 | 99.4 | 67.9† | 72.4† | 56.5† | 36.7 | 59.5† |
| Rank | Transition | Probability |
| 1 | 233→115 | 14.25% |
| 2 | 265→211 | 12.90% |
| 3 | 211→265 | 10.45% |
| 4 | 274→210 | 2.20% |
| 5 | 110→274 | 2.22% |
| Rank | Latent Label | Implication |
| 1 | 233 | Short instructions about paste operation. |
| 2 | 115 | Reference to a URL. |
| 3 | 211 | Audio and sound configuration in Ubuntu. |
| 4 | 265 | Sound system breakdown. |
| 5 | 110 | Brief yes/no answer. |
| Models | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT(Devlin et al., 2018) | 2.0 | 4.0 | 17.0 | 22.3 | 25.0 | 8.7 | 3.2 | 6.5 |
| BERT+TRIGGER | 13.4† | 20.2† | 48.1† | 37.6† | 42.0† | 23.5† | 13.7† | 21.1† |
| BERT-FP (Han et al., 2021) | 52.5 | 57.4 | 69.2 | 32.5 | 36.8 | 21.9 | 12.9 | 17.3 |
| BERT-FP+TRIGGER | 76.2† | 81.9† | 90.1† | 38.8† | 43.0† | 25.4† | 16.2† | 23.5† |
| Model | False Negative Ratio |
| Random | 3.29 |
| Semi (Li et al., 2019) | 7.16 |
| CIR (Penha and Hauff, 2020) | 7.03 |
| Grey (Lin et al., 2020) | 7.42 |
| HCL (Su et al., 2020) | 6.95 |
| Mask-and-fill (Gupta et al., 2021) | 5.63 |
| Model | False Negative Preventing | Negative Updating |
| Semi (Li et al., 2019) | X | X |
| CIR(Penha and Hauff, 2020) | X | ✓ |
| HCL(Su et al., 2020) | X | ✓ |
| Grey(Lin et al., 2020) | X | X |
| Mask-and-fill(Gupta et al., 2021) | ✓ | X |
| Ours | ✓ | ✓ |
| Model | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT (Devlin et al., 2018) | 80.8 | 89.7 | 97.5 | 59.1 | 63.3 | 45.4 | 28.0 | 47.0 |
| +Semi (Li et al., 2019) | 81.0 | 89.8 | 97.5 | 60.3 | 63.8 | 46.0 | 28.2 | 47.9 |
| +CIR (Penha and Hauff, 2020) | 81.2 | 89.9 | 97.6 | 60.7 | 64.2 | 46.4 | 28.5 | 49.1 |
| +Gray (Lin et al., 2020) | 81.5 | 90.1 | 97.5 | 61.5 | 64.8 | 47.1 | 29.1 | 50.7 |
| +HCL (Su et al., 2020) | 81.5 | 90.2 | 97.6 | 61.7 | 65.9 | 48.0 | 30.4 | 51.0 |
| +MF (Gupta et al., 2021) | 81.6 | 90.0 | 97.5 | 61.9 | 66.0 | 48.1 | 30.5 | 51.3 |
| +TRIGGER (Ours) | 81.9† | 90.4 | 97.5 | 62.9† | 67.8† | 50.2† | 32.5† | 52.1† |
| SA-BERT (Gu et al., 2020) | 85.5 | 92.8 | 98.3 | 61.9 | 65.9 | 49.6 | 31.3 | 48.1 |
| +Semi (Li et al., 2019) | 85.8 | 93.1 | 98.9 | 62.3 | 66.4 | 50.0 | 31.7 | 49.0 |
| +CIR (Penha and Hauff, 2020) | 86.0 | 93.5 | 99.0 | 62.4 | 66.6 | 50.3 | 31.8 | 49.7 |
| +Gray (Lin et al., 2020) | 86.1 | 93.4 | 99.1 | 62.8 | 67.0 | 50.3 | 32.0 | 50.3 |
| +HCL (Su et al., 2020) | 86.7 | 94.0 | 99.2 | 63.9 | 68.1 | 51.4 | 33.0 | 53.1 |
| +MF (Gupta et al., 2021) | 86.9 | 93.7 | 98.9 | 63.6 | 68.4 | 52.0 | 33.5 | 52.9 |
| +TRIGGER (Ours) | 87.1† | 93.8 | 98.7 | 64.1 | 69.5† | 52.9† | 34.3† | 53.7† |
| BERT-FP (Han et al., 2021) | 91.1 | 96.2 | 99.4 | 64.4 | 68.0 | 51.2 | 32.4 | 54.2 |
| +Semi (Li et al., 2019) | 91.2 | 96.2 | 99.3 | 65.8 | 68.7 | 53.6 | 34.3 | 54.7 |
| +CIR (Penha and Hauff, 2020) | 91.2 | 96.4 | 99.3 | 66.5 | 70.6 | 54.8 | 35.5 | 55.0 |
| +Gray (Lin et al., 2020) | 91.3 | 96.3 | 99.4 | 66.7 | 70.5 | 55.0 | 36.3 | 54.8 |
| +HCL (Su et al., 2020) | 91.2 | 96.3 | 99.4 | 67.1 | 70.9 | 55.3 | 36.4 | 55.7 |
| +MF (Gupta et al., 2021) | 91.3 | 96.4 | 99.3 | 67.2 | 71.4 | 55.4 | 36.2 | 56.7 |
| +TRIGGER (Ours) | 91.7† | 96.6 | 99.4 | 67.9† | 72.4† | 56.5† | 36.7 | 59.5† |
| Rank | Transition | Probability |
| 1 | 233→115 | 14.25% |
| 2 | 265→211 | 12.90% |
| 3 | 211→265 | 10.45% |
| 4 | 274→210 | 2.20% |
| 5 | 110→274 | 2.22% |
| Rank | Latent Label | Implication |
| 1 | 233 | Short instructions about paste operation. |
| 2 | 115 | Reference to a URL. |
| 3 | 211 | Audio and sound configuration in Ubuntu. |
| 4 | 265 | Sound system breakdown. |
| 5 | 110 | Brief yes/no answer. |
| Models | Ubuntu | Douban | ||||||
| R10@1 | R10@2 | R10@5 | MAP | MRR | P@1 | R10@1 | R10@2 | |
| BERT(Devlin et al., 2018) | 2.0 | 4.0 | 17.0 | 22.3 | 25.0 | 8.7 | 3.2 | 6.5 |
| BERT+TRIGGER | 13.4† | 20.2† | 48.1† | 37.6† | 42.0† | 23.5† | 13.7† | 21.1† |
| BERT-FP (Han et al., 2021) | 52.5 | 57.4 | 69.2 | 32.5 | 36.8 | 21.9 | 12.9 | 17.3 |
| BERT-FP+TRIGGER | 76.2† | 81.9† | 90.1† | 38.8† | 43.0† | 25.4† | 16.2† | 23.5† |
| Problem | Formulation | Example | Solver | Dataset | |
| NL Sentence | Symbolic Formulation | ||||
| Deductive Reasoning | LP | If the circuit is complete and the circuit has the light bulb then the light bulb is glowing. | Complete(Circuit, True) ∧ Has(Circuit, LightBulb) → Glowing(LightBulb, True) | Pyke | ProntoQA, ProofWriter |
| First-Order Logic | FOL | A Czech person wrote a book in 1946. | ∃x2∃x1(Czech(x1) ∧ Author(x2,x1) ∧ Book(x2) ∧ Publish(x2,1946)) | Prover9 | FOLIO |
| Constraint Satisfaction | CSP | On a shelf, there are five books. The blue book is to the right of the yellow book. | blue_book ∈ {1,2,3,4,5} \nyellow_book ∈ {1,2,3,4,5} \nblue_book > yellow_book | python-constraint | LogicalDeduction |
| Analytical Reasoning | SAT | Xena and exactly three other technicians repair radios | repairs(Xena, radios) ∧ Count([t:technicians], t ≠ Xena ∧ repairs(t, radios)) == 3) | Z3 | AR-LSAT |
| Dataset | ChatGPT (gpt-3.5-turbo) | GPT-3.5 (text-davinci-003) | GPT-4 (gpt-4) | ||||||
| Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | |
| PrOntoQA | 47.40 | 67.80 | 61.00 | 51.80 | 83.00 | 85.00 | 77.40 | 98.79 | 83.20 |
| ProofWriter | 35.50 | 49.17 | 58.33 | 36.16 | 48.33 | 71.45 | 52.67 | 68.11 | 79.66 |
| FOLIO | 45.09 | 57.35 | 62.74 | 54.60 | 57.84 | 61.27 | 69.11 | 70.58 | 78.92 |
| LogicalDeduction | 40.00 | 42.33 | 65.67 | 41.33 | 48.33 | 62.00 | 71.33 | 75.25 | 87.63 |
| AR-LSAT | 20.34 | 17.31 | 26.41 | 22.51 | 22.51 | 25.54 | 33.33 | 35.06 | 43.04 |
| Dataset | SR | GPT-3.5 | GPT-4 | ||
| Exe_Rate | Exe_Acc | Exe_Rate | Exe_Acc | ||
| ProntoQA | - | 99.4% | 84.9 | 100.0% | 83.2 |
| + | 100.0% ↑0.6 | 85.0 ↑0.1 | 100.0% | 83.2 | |
| ProofWriter | - | 87.3% | 73.6 | 99.0% | 79.6 |
| + | 95.6% ↑8.3 | 74.1 ↑0.5 | 99.0% | 79.6 | |
| FOLIO | - | 66.7% | 61.8 | 79.9% | 80.4 |
| + | 84.3% ↑17.6 | 64.3 ↑2.5 | 85.8% ↑5.9 | 79.9 ↓0.5 | |
| Logical | - | 100.0% | 62.0 | 100.0% | 87.6 |
| Deduction | + | 100.0% | 62.0 | 100.0% | 87.6 |
| AR-LSAT | - | 11.3% | 57.7 | 32.6% | 60.0 |
| + | 21.8% ↑10.5 | 60.3 ↑2.6 | 39.8% ↑7.2 | 58.8 ↓1.2 |
| Name | FOL Notation |
| Constant | lowercase letters |
| Variable | x, y, z, ··· |
| Atom | P(a1, ···, an) |
| Negation | ¬P |
| Conjunction | P1 ∧ P2\nP1 ∧ ···, ∧ Pn |
| Disjunction | P1 ∨ P2\nP1 ∨ ···, ∨ Pn |
| Implication | P1 → P2 |
| Equivalence | P1 ↔ P2 |
| Existential Quantifier | ∃xP(x, ···) |
| Universal Quantifier | ∀xP(x, ···) |
| Dataset | Reasoning | Test Size | #OptS |
| PrOntoQA | Deductive | 500 | 2 |
| ProofWriter | Deductive | 600 | 3 |
| FOLIO | FOL | 204 | 3 |
| LogicalDeduction | CSP | 300 | 3,5,7 |
| AR-LSAT | AR | 230 | 5 |
| Context: Jompuses are not shy. Jompuses are yumpuses. \n(··· more context here···) \nZumpuses are rompuses. Max is a yumpus. |
| Question: Is the following statement true or false? \nMax is sour. |
| Options: \nA) True \nB) False |
| The correct option is: B |
| Problem | Formulation | Example | Solver | Dataset | |
| NL Sentence | Symbolic Formulation | ||||
| Deductive Reasoning | LP | If the circuit is complete and the circuit has the light bulb then the light bulb is glowing. | Complete(Circuit, True) ∧ Has(Circuit, LightBulb) → Glowing(LightBulb, True) | Pyke | ProntoQA, ProofWriter |
| First-Order Logic | FOL | A Czech person wrote a book in 1946. | ∃x2∃x1(Czech(x1) ∧ Author(x2,x1) ∧ Book(x2) ∧ Publish(x2,1946)) | Prover9 | FOLIO |
| Constraint Satisfaction | CSP | On a shelf, there are five books. The blue book is to the right of the yellow book. | blue_book ∈ {1,2,3,4,5} \nyellow_book ∈ {1,2,3,4,5} \nblue_book > yellow_book | python-constraint | LogicalDeduction |
| Analytical Reasoning | SAT | Xena and exactly three other technicians repair radios | repairs(Xena, radios) ∧ Count([t:technicians], t ≠ Xena ∧ repairs(t, radios)) == 3) | Z3 | AR-LSAT |
| Dataset | ChatGPT (gpt-3.5-turbo) | GPT-3.5 (text-davinci-003) | GPT-4 (gpt-4) | ||||||
| Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | |
| PrOntoQA | 47.40 | 67.80 | 61.00 | 51.80 | 83.00 | 85.00 | 77.40 | 98.79 | 83.20 |
| ProofWriter | 35.50 | 49.17 | 58.33 | 36.16 | 48.33 | 71.45 | 52.67 | 68.11 | 79.66 |
| FOLIO | 45.09 | 57.35 | 62.74 | 54.60 | 57.84 | 61.27 | 69.11 | 70.58 | 78.92 |
| LogicalDeduction | 40.00 | 42.33 | 65.67 | 41.33 | 48.33 | 62.00 | 71.33 | 75.25 | 87.63 |
| AR-LSAT | 20.34 | 17.31 | 26.41 | 22.51 | 22.51 | 25.54 | 33.33 | 35.06 | 43.04 |
| Dataset | SR | GPT-3.5 | GPT-4 | ||
| Exe_Rate | Exe_Acc | Exe_Rate | Exe_Acc | ||
| ProntoQA | - | 99.4% | 84.9 | 100.0% | 83.2 |
| + | 100.0% ↑0.6 | 85.0 ↑0.1 | 100.0% | 83.2 | |
| ProofWriter | - | 87.3% | 73.6 | 99.0% | 79.6 |
| + | 95.6% ↑8.3 | 74.1 ↑0.5 | 99.0% | 79.6 | |
| FOLIO | - | 66.7% | 61.8 | 79.9% | 80.4 |
| + | 84.3% ↑17.6 | 64.3 ↑2.5 | 85.8% ↑5.9 | 79.9 ↓0.5 | |
| Logical | - | 100.0% | 62.0 | 100.0% | 87.6 |
| Deduction | + | 100.0% | 62.0 | 100.0% | 87.6 |
| AR-LSAT | - | 11.3% | 57.7 | 32.6% | 60.0 |
| + | 21.8% ↑10.5 | 60.3 ↑2.6 | 39.8% ↑7.2 | 58.8 ↓1.2 |
| Name | FOL Notation |
| Constant | lowercase letters |
| Variable | x, y, z, ··· |
| Atom | P(a1, ···, an) |
| Negation | ¬P |
| Conjunction | P1 ∧ P2\nP1 ∧ ···, ∧ Pn |
| Disjunction | P1 ∨ P2\nP1 ∨ ···, ∨ Pn |
| Implication | P1 → P2 |
| Equivalence | P1 ↔ P2 |
| Existential Quantifier | ∃xP(x, ···) |
| Universal Quantifier | ∀xP(x, ···) |
| Dataset | Reasoning | Test Size | #OptS |
| PrOntoQA | Deductive | 500 | 2 |
| ProofWriter | Deductive | 600 | 3 |
| FOLIO | FOL | 204 | 3 |
| LogicalDeduction | CSP | 300 | 3,5,7 |
| AR-LSAT | AR | 230 | 5 |
| Context: Jompuses are not shy. Jompuses are yumpuses. \n(··· more context here···) \nZumpuses are rompuses. Max is a yumpus. |
| Question: Is the following statement true or false? \nMax is sour. |
| Options: \nA) True \nB) False |
| The correct option is: B |
| Problem | Formulation | Example | Solver | Dataset | |
| NL Sentence | Symbolic Formulation | ||||
| Deductive Reasoning | LP | If the circuit is complete and the circuit has the light bulb then the light bulb is glowing. | Complete(Circuit, True) ∧ Has(Circuit, LightBulb) → Glowing(LightBulb, True) | Pyke | ProntoQA, ProofWriter |
| First-Order Logic | FOL | A Czech person wrote a book in 1946. | ∃x2∃x1(Czech(x1) ∧ Author(x2,x1) ∧ Book(x2) ∧ Publish(x2,1946)) | Prover9 | FOLIO |
| Constraint Satisfaction | CSP | On a shelf, there are five books. The blue book is to the right of the yellow book. | blue_book ∈ {1,2,3,4,5} +yellow_book ∈ {1,2,3,4,5} +blue_book > yellow_book | python-constraint | LogicalDeduction |
| Analytical Reasoning | SAT | Xena and exactly three other technicians repair radios | repairs(Xena, radios) ∧ Count([t:technicians], t ≠ Xena ∧ repairs(t, radios)) == 3) | Z3 | AR-LSAT |
| Dataset | ChatGPT (gpt-3.5-turbo) | GPT-3.5 (text-davinci-003) | GPT-4 (gpt-4) | ||||||
| Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | |
| PrOntoQA | 47.40 | 67.80 | 61.00 | 51.80 | 83.00 | 85.00 | 77.40 | 98.79 | 83.20 |
| ProofWriter | 35.50 | 49.17 | 58.33 | 36.16 | 48.33 | 71.45 | 52.67 | 68.11 | 79.66 |
| FOLIO | 45.09 | 57.35 | 62.74 | 54.60 | 57.84 | 61.27 | 69.11 | 70.58 | 78.92 |
| LogicalDeduction | 40.00 | 42.33 | 65.67 | 41.33 | 48.33 | 62.00 | 71.33 | 75.25 | 87.63 |
| AR-LSAT | 20.34 | 17.31 | 26.41 | 22.51 | 22.51 | 25.54 | 33.33 | 35.06 | 43.04 |
| Dataset | SR | GPT-3.5 | GPT-4 | ||
| Exe_Rate | Exe_Acc | Exe_Rate | Exe_Acc | ||
| ProntoQA | - | 99.4% | 84.9 | 100.0% | 83.2 |
| + | 100.0% ↑0.6 | 85.0 ↑0.1 | 100.0% | 83.2 | |
| ProofWriter | - | 87.3% | 73.6 | 99.0% | 79.6 |
| + | 95.6% ↑8.3 | 74.1 ↑0.5 | 99.0% | 79.6 | |
| FOLIO | - | 66.7% | 61.8 | 79.9% | 80.4 |
| + | 84.3% ↑17.6 | 64.3 ↑2.5 | 85.8% ↑5.9 | 79.9 ↓0.5 | |
| Logical | - | 100.0% | 62.0 | 100.0% | 87.6 |
| Deduction | + | 100.0% | 62.0 | 100.0% | 87.6 |
| AR-LSAT | - | 11.3% | 57.7 | 32.6% | 60.0 |
| + | 21.8% ↑10.5 | 60.3 ↑2.6 | 39.8% ↑7.2 | 58.8 ↓1.2 | |
| Name | FOL Notation |
| Constant | lowercase letters |
| Variable | x, y, z, ··· |
| Atom | P(a1, ···, an) |
| Negation | ¬P |
| Conjunction | P1 ∧ P2 +P1 ∧ ···, ∧ Pn |
| Disjunction | P1 ∨ P2 +P1 ∨ ···, ∨ Pn |
| Implication | P1 → P2 |
| Equivalence | P1 ↔ P2 |
| Existential Quantifier | ∃xP(x, ···) |
| Universal Quantifier | ∀xP(x, ···) |
| Dataset | Reasoning | Test Size | #OptS |
| PrOntoQA | Deductive | 500 | 2 |
| ProofWriter | Deductive | 600 | 3 |
| FOLIO | FOL | 204 | 3 |
| LogicalDeduction | CSP | 300 | 3,5,7 |
| AR-LSAT | AR | 230 | 5 |
| Context: Jompuses are not shy. Jompuses are yumpuses. +(··· more context here···) +Zumpuses are rompuses. Max is a yumpus. |
| Question: Is the following statement true or false? +Max is sour. |
| Options: +A) True +B) False |
| The correct option is: B |
| Problem | Formulation | Example | Solver | Dataset | |
| NL Sentence | Symbolic Formulation | ||||
| Deductive Reasoning | LP | If the circuit is complete and the circuit has the light bulb then the light bulb is glowing. | Complete(Circuit, True) ∧ Has(Circuit, LightBulb) → Glowing(LightBulb, True) | Pyke | ProntoQA, ProofWriter |
| First-Order Logic | FOL | A Czech person wrote a book in 1946. | ∃x2∃x1(Czech(x1) ∧ Author(x2,x1) ∧ Book(x2) ∧ Publish(x2,1946)) | Prover9 | FOLIO |
| Constraint Satisfaction | CSP | On a shelf, there are five books. The blue book is to the right of the yellow book. | blue_book ∈ {1,2,3,4,5} \nyellow_book ∈ {1,2,3,4,5} \nblue_book > yellow_book | python-constraint | LogicalDeduction |
| Analytical Reasoning | SAT | Xena and exactly three other technicians repair radios | repairs(Xena, radios) ∧ Count([t:technicians], t ≠ Xena ∧ repairs(t, radios)) == 3) | Z3 | AR-LSAT |
| Dataset | ChatGPT (gpt-3.5-turbo) | GPT-3.5 (text-davinci-003) | GPT-4 (gpt-4) | ||||||
| Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | Standard | CoT | Logic-LM | |
| PrOntoQA | 47.40 | 67.80 | 61.00 | 51.80 | 83.00 | 85.00 | 77.40 | 98.79 | 83.20 |
| ProofWriter | 35.50 | 49.17 | 58.33 | 36.16 | 48.33 | 71.45 | 52.67 | 68.11 | 79.66 |
| FOLIO | 45.09 | 57.35 | 62.74 | 54.60 | 57.84 | 61.27 | 69.11 | 70.58 | 78.92 |
| LogicalDeduction | 40.00 | 42.33 | 65.67 | 41.33 | 48.33 | 62.00 | 71.33 | 75.25 | 87.63 |
| AR-LSAT | 20.34 | 17.31 | 26.41 | 22.51 | 22.51 | 25.54 | 33.33 | 35.06 | 43.04 |
| Dataset | SR | GPT-3.5 | GPT-4 | ||
| Exe_Rate | Exe_Acc | Exe_Rate | Exe_Acc | ||
| ProntoQA | - | 99.4% | 84.9 | 100.0% | 83.2 |
| + | 100.0% ↑0.6 | 85.0 ↑0.1 | 100.0% | 83.2 | |
| ProofWriter | - | 87.3% | 73.6 | 99.0% | 79.6 |
| + | 95.6% ↑8.3 | 74.1 ↑0.5 | 99.0% | 79.6 | |
| FOLIO | - | 66.7% | 61.8 | 79.9% | 80.4 |
| + | 84.3% ↑17.6 | 64.3 ↑2.5 | 85.8% ↑5.9 | 79.9 ↓0.5 | |
| Logical | - | 100.0% | 62.0 | 100.0% | 87.6 |
| Deduction | + | 100.0% | 62.0 | 100.0% | 87.6 |
| AR-LSAT | - | 11.3% | 57.7 | 32.6% | 60.0 |
| + | 21.8% ↑10.5 | 60.3 ↑2.6 | 39.8% ↑7.2 | 58.8 ↓1.2 |
| Name | FOL Notation |
| Constant | lowercase letters |
| Variable | x, y, z, ··· |
| Atom | P(a1, ···, an) |
| Negation | ¬P |
| Conjunction | P1 ∧ P2\nP1 ∧ ···, ∧ Pn |
| Disjunction | P1 ∨ P2\nP1 ∨ ···, ∨ Pn |
| Implication | P1 → P2 |
| Equivalence | P1 ↔ P2 |
| Existential Quantifier | ∃xP(x, ···) |
| Universal Quantifier | ∀xP(x, ···) |
| Dataset | Reasoning | Test Size | #OptS |
| PrOntoQA | Deductive | 500 | 2 |
| ProofWriter | Deductive | 600 | 3 |
| FOLIO | FOL | 204 | 3 |
| LogicalDeduction | CSP | 300 | 3,5,7 |
| AR-LSAT | AR | 230 | 5 |
| Context: Jompuses are not shy. Jompuses are yumpuses. \n(··· more context here···) \nZumpuses are rompuses. Max is a yumpus. |
| Question: Is the following statement true or false? \nMax is sour. |
| Options: \nA) True \nB) False |
| The correct option is: B |
| Name | Formal Expressions | Attack Examples |
| Original 1 | (p1→h1) | Premise: An older man wearing a salon drape getting a haircut.\nHypothesis: A man gets a haircut. |
| Original 2 | (p2→h2) | Premise: The dog runs towards the ball.\nHypothesis: A dog runs. |
| Modus Tollens | ((p1→h1) ∧ ¬h1) ⊨ ¬p1 | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nNo man gets a haircut.\nHypothesis: No older man wearing a salon drape is getting a haircut. |
| Constructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ p2)) ⊢ (h1 ∨ h2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs. But either an older man wearing a salon drape getting a haircut or the dog runs towards the ball.\nHypothesis: Either a man gets a haircut or a dog runs. |
| Destructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (¬h1 ∨ ¬h2)) ⊢ (¬p1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs. But either no man gets a haircut or no dog runs.\nHypothesis: Either no older man wearing a salon drape is getting a haircut or the dog does not run towards the ball. |
| Bidirectional Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ ¬h2)) ⊢ (h1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs.\nBut either an older man wearing a salon drape getting a haircut or no dog runs.\nHypothesis: Either a man gets a haircut or the dog does not run towards the ball. |
| Transportation 1 | ((p1→h1) ⊢ (¬h1 → ¬p1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nHypothesis: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut. |
| Transportation 2 | (¬h1 → ¬p1) ⊢ ((p1 → h1) | Premise: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut.\nHypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Material Implication 1 | (p1→h1) ⊢ (¬p1 ∨ h1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nHypothesis: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut. |
| Material Implication 2 | (¬p1 ∨ h1) ⊢ (p1 → h1) | Premise: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut.\nHypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Negate Hypothesis | (p1 → ¬h1) | Premise: An older man wearing a salon drape getting a haircut.\nHypothesis: No man gets a haircut. |
| Dataset | Attack Type | Single-Task | Multi-Task | Prompting | F1 value | ||||
| RoBERTa | RoBERTa | BART | FLAN-T5 | GPT-3 | ChatGPT | GPT-4 | |||
| SNLI | Modus Tollens | 99.7 | 94.1 | 45.5 | 99.9 | 97.8 | 96.3 | 50.6 | 61.0 |
| Constructive Dilemma | 34.7 | 3.9 | 0.1 | 0.6 | 34.7 | 11.3 | 9.0 | 60.4 | |
| Destructive Dilemma | 93.6 | 30.2 | 24.9 | 87.3 | 90.4 | 46.7 | 18.8 | 57.9 | |
| Bidirectional Dilemma | 73.4 | 67.0 | 22.5 | 15.6 | 63.5 | 16.3 | 11.7 | 59.1 | |
| Transposition 1 | 99.7 | 82.5 | 99.1 | 82.0 | 79.6 | 25.0 | 22.7 | 58.6 | |
| Transposition 2 | 95.4 | 40.0 | 94.4 | 94.2 | 46.2 | 44.0 | 16.2 | 58.6 | |
| Material Implication 1 | 98.6 | 85.8 | 94.8 | 94.5 | 54.6 | 25.5 | 11.1 | 59.8 | |
| Material Implication 2 | 85.6 | 93.5 | 46.1 | 33.0 | 35.4 | 12.9 | 29.8 | 59.8 | |
| Negate Hypothesis | 1.1 | 3.2 | 1.9 | 0.6 | 35.0 | 15.4 | 9.7 | 95.2 | |
| Avg. | 76.1 | 56.2 | 48.7 | 56.4 | 59.7 | 32.7 | 20.0 | 63.4 | |
| MNLI | Modus Tollens | 96.6 | 81.1 | 68.1 | 94.8 | 94.6 | 96.1 | 72.3 | 62.3 |
| Constructive Dilemma | 23.8 | 2.8 | 0.7 | 7.0 | 26.9 | 4.6 | 0.0 | 62.2 | |
| Destructive Dilemma | 90.5 | 58.4 | 64.0 | 86.9 | 96.5 | 94.8 | 63.2 | 60.4 | |
| Bidirectional Dilemma | 85.6 | 68.9 | 52.8 | 49.4 | 70.1 | 58.4 | 30.6 | 61.3 | |
| Transposition 1 | 96.5 | 87.1 | 91.5 | 78.1 | 67.1 | 57.3 | 47.1 | 60.5 | |
| Transposition 2 | 90.4 | 74.1 | 92.3 | 95.4 | 33.6 | 32.7 | 42.9 | 60.5 | |
| Material Implication 1 | 88.9 | 70.4 | 72.9 | 78.6 | 45.6 | 53.1 | 40.7 | 61.4 | |
| Material Implication 2 | 75.8 | 79.3 | 64.9 | 43.2 | 18.9 | 7.6 | 69.9 | 61.4 | |
| Negate Hypothesis | 8.8 | 11.0 | 10.0 | 5.9 | 11.6 | 9.1 | 5.6 | 94.2 | |
| Avg. | 73.5 | 59.9 | 58.3 | 59.9 | 51.7 | 46.0 | 41.4 | 64.9 |
| Dataset | Attack Strategies | SimCSE | BERTScore |
| SNLI | Modus Tollens | 0.839 | 0.919 |
| Constructive Dilemma | 0.935 | 0.908 | |
| Destructive Dilemma | 0.869 | 0.906 | |
| Bidirectional Dilemma | 0.909 | 0.906 | |
| Transposition 1 | 0.866 | 0.912 | |
| Transposition 2 | 0.848 | 0.904 | |
| Material Implication 1 | 0.901 | 0.917 | |
| Material Implication 2 | 0.880 | 0.910 | |
| Negate Hypothesis | 0.868 | 0.967 | |
| Avg. | 0.879 | 0.917 | |
| MNLI | Modus Tollens | 0.846 | 0.926 |
| Constructive Dilemma | 0.937 | 0.914 | |
| Destructive Dilemma | 0.892 | 0.914 | |
| Bidirectional Dilemma | 0.914 | 0.914 | |
| Transposition 1 | 0.864 | 0.920 | |
| Transposition 2 | 0.864 | 0.911 | |
| Material Implication 1 | 0.905 | 0.925 | |
| Material Implication 2 | 0.891 | 0.918 | |
| Negate Hypothesis | 0.879 | 0.963 | |
| Avg. | 0.888 | 0.923 |
| Inference Rules | GPT-3 | ChatGPT | GPT-4 | |||
| 3-shots | 6-shots | 3-shots | 6-shots | 3-shots | 6-shots | |
| Modus Tollens | 69.4 | 67.2 | 90.9 | 92.8 | 12.3 | 11.4 |
| Constructive Dilemma | 2.4 | 4.2 | 3.7 | 3.6 | 5.3 | 5.8 |
| Destructive Dilemma | 15.9 | 15.6 | 14.7 | 17.7 | 13.8 | 13.6 |
| Bidirectional Dilemma | 4.3 | 6.5 | 5.4 | 6.1 | 8.3 | 8.6 |
| Transposition 1 | 71.5 | 68.7 | 4.7 | 5.3 | 13.2 | 10.0 |
| Transposition 2 | 11.5 | 16.1 | 10.3 | 11.7 | 12.8 | 11.1 |
| Material Implication 1 | 5.1 | 6.5 | 3.7 | 4.1 | 5.6 | 5.9 |
| Material Implication 2 | 4.1 | 7 | 3.7 | 3.7 | 23.5 | 15.9 |
| Negate Hypothesis | 4.4 | 5.6 | 7.9 | 8.7 | 5.7 | 6.4 |
| Avg. | 19.5 | 20.8 | 16.1 | 17.1 | 11.2 | 9.9 |
| Inference Rules | # of negations | # of logical AND | # of logical OR | # of implications | # of contained original sentences |
| Modus Tollens | 2 | 1 | 0 | 1 | 4 |
| Constructive Dilemma | 0 | 2 | 2 | 2 | 8 |
| Destructive Dilemma | 4 | 2 | 2 | 2 | 8 |
| Bidirectional Dilemma | 2 | 2 | 2 | 2 | 8 |
| Transposition 1 | 2 | 0 | 0 | 2 | 4 |
| Transposition 2 | 2 | 0 | 0 | 2 | 4 |
| Material Implication 1 | 1 | 0 | 1 | 1 | 4 |
| Material Implication 2 | 1 | 0 | 1 | 1 | 4 |
| Negate Hypothesis | 1 | 0 | 0 | 0 | 2 |
| Name | Formal Expressions | Attack Examples |
| Original 1 | (p1→h1) | Premise: An older man wearing a salon drape getting a haircut.\nHypothesis: A man gets a haircut. |
| Original 2 | (p2→h2) | Premise: The dog runs towards the ball.\nHypothesis: A dog runs. |
| Modus Tollens | ((p1→h1) ∧ ¬h1) ⊨ ¬p1 | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nNo man gets a haircut.\nHypothesis: No older man wearing a salon drape is getting a haircut. |
| Constructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ p2)) ⊢ (h1 ∨ h2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs. But either an older man wearing a salon drape getting a haircut or the dog runs towards the ball.\nHypothesis: Either a man gets a haircut or a dog runs. |
| Destructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (¬h1 ∨ ¬h2)) ⊢ (¬p1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs. But either no man gets a haircut or no dog runs.\nHypothesis: Either no older man wearing a salon drape is getting a haircut or the dog does not run towards the ball. |
| Bidirectional Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ ¬h2)) ⊢ (h1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs.\nBut either an older man wearing a salon drape getting a haircut or no dog runs.\nHypothesis: Either a man gets a haircut or the dog does not run towards the ball. |
| Transportation 1 | ((p1→h1) ⊢ (¬h1 → ¬p1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nHypothesis: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut. |
| Transportation 2 | (¬h1 → ¬p1) ⊢ ((p1 → h1) | Premise: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut.\nHypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Material Implication 1 | (p1→h1) ⊢ (¬p1 ∨ h1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nHypothesis: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut. |
| Material Implication 2 | (¬p1 ∨ h1) ⊢ (p1 → h1) | Premise: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut.\nHypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Negate Hypothesis | (p1 → ¬h1) | Premise: An older man wearing a salon drape getting a haircut.\nHypothesis: No man gets a haircut. |
| Dataset | Attack Type | Single-Task | Multi-Task | Prompting | F1 value | ||||
| RoBERTa | RoBERTa | BART | FLAN-T5 | GPT-3 | ChatGPT | GPT-4 | |||
| SNLI | Modus Tollens | 99.7 | 94.1 | 45.5 | 99.9 | 97.8 | 96.3 | 50.6 | 61.0 |
| Constructive Dilemma | 34.7 | 3.9 | 0.1 | 0.6 | 34.7 | 11.3 | 9.0 | 60.4 | |
| Destructive Dilemma | 93.6 | 30.2 | 24.9 | 87.3 | 90.4 | 46.7 | 18.8 | 57.9 | |
| Bidirectional Dilemma | 73.4 | 67.0 | 22.5 | 15.6 | 63.5 | 16.3 | 11.7 | 59.1 | |
| Transposition 1 | 99.7 | 82.5 | 99.1 | 82.0 | 79.6 | 25.0 | 22.7 | 58.6 | |
| Transposition 2 | 95.4 | 40.0 | 94.4 | 94.2 | 46.2 | 44.0 | 16.2 | 58.6 | |
| Material Implication 1 | 98.6 | 85.8 | 94.8 | 94.5 | 54.6 | 25.5 | 11.1 | 59.8 | |
| Material Implication 2 | 85.6 | 93.5 | 46.1 | 33.0 | 35.4 | 12.9 | 29.8 | 59.8 | |
| Negate Hypothesis | 1.1 | 3.2 | 1.9 | 0.6 | 35.0 | 15.4 | 9.7 | 95.2 | |
| Avg. | 76.1 | 56.2 | 48.7 | 56.4 | 59.7 | 32.7 | 20.0 | 63.4 | |
| MNLI | Modus Tollens | 96.6 | 81.1 | 68.1 | 94.8 | 94.6 | 96.1 | 72.3 | 62.3 |
| Constructive Dilemma | 23.8 | 2.8 | 0.7 | 7.0 | 26.9 | 4.6 | 0.0 | 62.2 | |
| Destructive Dilemma | 90.5 | 58.4 | 64.0 | 86.9 | 96.5 | 94.8 | 63.2 | 60.4 | |
| Bidirectional Dilemma | 85.6 | 68.9 | 52.8 | 49.4 | 70.1 | 58.4 | 30.6 | 61.3 | |
| Transposition 1 | 96.5 | 87.1 | 91.5 | 78.1 | 67.1 | 57.3 | 47.1 | 60.5 | |
| Transposition 2 | 90.4 | 74.1 | 92.3 | 95.4 | 33.6 | 32.7 | 42.9 | 60.5 | |
| Material Implication 1 | 88.9 | 70.4 | 72.9 | 78.6 | 45.6 | 53.1 | 40.7 | 61.4 | |
| Material Implication 2 | 75.8 | 79.3 | 64.9 | 43.2 | 18.9 | 7.6 | 69.9 | 61.4 | |
| Negate Hypothesis | 8.8 | 11.0 | 10.0 | 5.9 | 11.6 | 9.1 | 5.6 | 94.2 | |
| Avg. | 73.5 | 59.9 | 58.3 | 59.9 | 51.7 | 46.0 | 41.4 | 64.9 |
| Dataset | Attack Strategies | SimCSE | BERTScore |
| SNLI | Modus Tollens | 0.839 | 0.919 |
| Constructive Dilemma | 0.935 | 0.908 | |
| Destructive Dilemma | 0.869 | 0.906 | |
| Bidirectional Dilemma | 0.909 | 0.906 | |
| Transposition 1 | 0.866 | 0.912 | |
| Transposition 2 | 0.848 | 0.904 | |
| Material Implication 1 | 0.901 | 0.917 | |
| Material Implication 2 | 0.880 | 0.910 | |
| Negate Hypothesis | 0.868 | 0.967 | |
| Avg. | 0.879 | 0.917 | |
| MNLI | Modus Tollens | 0.846 | 0.926 |
| Constructive Dilemma | 0.937 | 0.914 | |
| Destructive Dilemma | 0.892 | 0.914 | |
| Bidirectional Dilemma | 0.914 | 0.914 | |
| Transposition 1 | 0.864 | 0.920 | |
| Transposition 2 | 0.864 | 0.911 | |
| Material Implication 1 | 0.905 | 0.925 | |
| Material Implication 2 | 0.891 | 0.918 | |
| Negate Hypothesis | 0.879 | 0.963 | |
| Avg. | 0.888 | 0.923 |
| Inference Rules | GPT-3 | ChatGPT | GPT-4 | |||
| 3-shots | 6-shots | 3-shots | 6-shots | 3-shots | 6-shots | |
| Modus Tollens | 69.4 | 67.2 | 90.9 | 92.8 | 12.3 | 11.4 |
| Constructive Dilemma | 2.4 | 4.2 | 3.7 | 3.6 | 5.3 | 5.8 |
| Destructive Dilemma | 15.9 | 15.6 | 14.7 | 17.7 | 13.8 | 13.6 |
| Bidirectional Dilemma | 4.3 | 6.5 | 5.4 | 6.1 | 8.3 | 8.6 |
| Transposition 1 | 71.5 | 68.7 | 4.7 | 5.3 | 13.2 | 10.0 |
| Transposition 2 | 11.5 | 16.1 | 10.3 | 11.7 | 12.8 | 11.1 |
| Material Implication 1 | 5.1 | 6.5 | 3.7 | 4.1 | 5.6 | 5.9 |
| Material Implication 2 | 4.1 | 7 | 3.7 | 3.7 | 23.5 | 15.9 |
| Negate Hypothesis | 4.4 | 5.6 | 7.9 | 8.7 | 5.7 | 6.4 |
| Avg. | 19.5 | 20.8 | 16.1 | 17.1 | 11.2 | 9.9 |
| Inference Rules | # of negations | # of logical AND | # of logical OR | # of implications | # of contained original sentences |
| Modus Tollens | 2 | 1 | 0 | 1 | 4 |
| Constructive Dilemma | 0 | 2 | 2 | 2 | 8 |
| Destructive Dilemma | 4 | 2 | 2 | 2 | 8 |
| Bidirectional Dilemma | 2 | 2 | 2 | 2 | 8 |
| Transposition 1 | 2 | 0 | 0 | 2 | 4 |
| Transposition 2 | 2 | 0 | 0 | 2 | 4 |
| Material Implication 1 | 1 | 0 | 1 | 1 | 4 |
| Material Implication 2 | 1 | 0 | 1 | 1 | 4 |
| Negate Hypothesis | 1 | 0 | 0 | 0 | 2 |
| Name | Formal Expressions | Attack Examples |
| Original 1 | (p1→h1) | Premise: An older man wearing a salon drape getting a haircut. +Hypothesis: A man gets a haircut. |
| Original 2 | (p2→h2) | Premise: The dog runs towards the ball. +Hypothesis: A dog runs. |
| Modus Tollens | ((p1→h1) ∧ ¬h1) ⊨ ¬p1 | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. +No man gets a haircut. +Hypothesis: No older man wearing a salon drape is getting a haircut. |
| Constructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ p2)) ⊢ (h1 ∨ h2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. +And if the dog runs towards the ball, then a dog runs. But either an older man wearing a salon drape getting a haircut or the dog runs towards the ball. +Hypothesis: Either a man gets a haircut or a dog runs. |
| Destructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (¬h1 ∨ ¬h2)) ⊢ (¬p1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. +And if the dog runs towards the ball, then a dog runs. But either no man gets a haircut or no dog runs. +Hypothesis: Either no older man wearing a salon drape is getting a haircut or the dog does not run towards the ball. |
| Bidirectional Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ ¬h2)) ⊢ (h1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. +And if the dog runs towards the ball, then a dog runs. +But either an older man wearing a salon drape getting a haircut or no dog runs. +Hypothesis: Either a man gets a haircut or the dog does not run towards the ball. |
| Transportation 1 | ((p1→h1) ⊢ (¬h1 → ¬p1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. +Hypothesis: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut. |
| Transportation 2 | (¬h1 → ¬p1) ⊢ ((p1 → h1) | Premise: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut. +Hypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Material Implication 1 | (p1→h1) ⊢ (¬p1 ∨ h1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. +Hypothesis: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut. |
| Material Implication 2 | (¬p1 ∨ h1) ⊢ (p1 → h1) | Premise: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut. +Hypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Negate Hypothesis | (p1 → ¬h1) | Premise: An older man wearing a salon drape getting a haircut. +Hypothesis: No man gets a haircut. |
| Dataset | Attack Type | Single-Task | Multi-Task | Prompting | F1 value | ||||
| RoBERTa | RoBERTa | BART | FLAN-T5 | GPT-3 | ChatGPT | GPT-4 | |||
| SNLI | Modus Tollens | 99.7 | 94.1 | 45.5 | 99.9 | 97.8 | 96.3 | 50.6 | 61.0 |
| Constructive Dilemma | 34.7 | 3.9 | 0.1 | 0.6 | 34.7 | 11.3 | 9.0 | 60.4 | |
| Destructive Dilemma | 93.6 | 30.2 | 24.9 | 87.3 | 90.4 | 46.7 | 18.8 | 57.9 | |
| Bidirectional Dilemma | 73.4 | 67.0 | 22.5 | 15.6 | 63.5 | 16.3 | 11.7 | 59.1 | |
| Transposition 1 | 99.7 | 82.5 | 99.1 | 82.0 | 79.6 | 25.0 | 22.7 | 58.6 | |
| Transposition 2 | 95.4 | 40.0 | 94.4 | 94.2 | 46.2 | 44.0 | 16.2 | 58.6 | |
| Material Implication 1 | 98.6 | 85.8 | 94.8 | 94.5 | 54.6 | 25.5 | 11.1 | 59.8 | |
| Material Implication 2 | 85.6 | 93.5 | 46.1 | 33.0 | 35.4 | 12.9 | 29.8 | 59.8 | |
| Negate Hypothesis | 1.1 | 3.2 | 1.9 | 0.6 | 35.0 | 15.4 | 9.7 | 95.2 | |
| Avg. | 76.1 | 56.2 | 48.7 | 56.4 | 59.7 | 32.7 | 20.0 | 63.4 | |
| MNLI | Modus Tollens | 96.6 | 81.1 | 68.1 | 94.8 | 94.6 | 96.1 | 72.3 | 62.3 |
| Constructive Dilemma | 23.8 | 2.8 | 0.7 | 7.0 | 26.9 | 4.6 | 0.0 | 62.2 | |
| Destructive Dilemma | 90.5 | 58.4 | 64.0 | 86.9 | 96.5 | 94.8 | 63.2 | 60.4 | |
| Bidirectional Dilemma | 85.6 | 68.9 | 52.8 | 49.4 | 70.1 | 58.4 | 30.6 | 61.3 | |
| Transposition 1 | 96.5 | 87.1 | 91.5 | 78.1 | 67.1 | 57.3 | 47.1 | 60.5 | |
| Transposition 2 | 90.4 | 74.1 | 92.3 | 95.4 | 33.6 | 32.7 | 42.9 | 60.5 | |
| Material Implication 1 | 88.9 | 70.4 | 72.9 | 78.6 | 45.6 | 53.1 | 40.7 | 61.4 | |
| Material Implication 2 | 75.8 | 79.3 | 64.9 | 43.2 | 18.9 | 7.6 | 69.9 | 61.4 | |
| Negate Hypothesis | 8.8 | 11.0 | 10.0 | 5.9 | 11.6 | 9.1 | 5.6 | 94.2 | |
| Avg. | 73.5 | 59.9 | 58.3 | 59.9 | 51.7 | 46.0 | 41.4 | 64.9 | |
| Dataset | Attack Strategies | SimCSE | BERTScore |
| SNLI | Modus Tollens | 0.839 | 0.919 |
| Constructive Dilemma | 0.935 | 0.908 | |
| Destructive Dilemma | 0.869 | 0.906 | |
| Bidirectional Dilemma | 0.909 | 0.906 | |
| Transposition 1 | 0.866 | 0.912 | |
| Transposition 2 | 0.848 | 0.904 | |
| Material Implication 1 | 0.901 | 0.917 | |
| Material Implication 2 | 0.880 | 0.910 | |
| Negate Hypothesis | 0.868 | 0.967 | |
| Avg. | 0.879 | 0.917 | |
| MNLI | Modus Tollens | 0.846 | 0.926 |
| Constructive Dilemma | 0.937 | 0.914 | |
| Destructive Dilemma | 0.892 | 0.914 | |
| Bidirectional Dilemma | 0.914 | 0.914 | |
| Transposition 1 | 0.864 | 0.920 | |
| Transposition 2 | 0.864 | 0.911 | |
| Material Implication 1 | 0.905 | 0.925 | |
| Material Implication 2 | 0.891 | 0.918 | |
| Negate Hypothesis | 0.879 | 0.963 | |
| Avg. | 0.888 | 0.923 |
| Inference Rules | GPT-3 | ChatGPT | GPT-4 | |||
| 3-shots | 6-shots | 3-shots | 6-shots | 3-shots | 6-shots | |
| Modus Tollens | 69.4 | 67.2 | 90.9 | 92.8 | 12.3 | 11.4 |
| Constructive Dilemma | 2.4 | 4.2 | 3.7 | 3.6 | 5.3 | 5.8 |
| Destructive Dilemma | 15.9 | 15.6 | 14.7 | 17.7 | 13.8 | 13.6 |
| Bidirectional Dilemma | 4.3 | 6.5 | 5.4 | 6.1 | 8.3 | 8.6 |
| Transposition 1 | 71.5 | 68.7 | 4.7 | 5.3 | 13.2 | 10.0 |
| Transposition 2 | 11.5 | 16.1 | 10.3 | 11.7 | 12.8 | 11.1 |
| Material Implication 1 | 5.1 | 6.5 | 3.7 | 4.1 | 5.6 | 5.9 |
| Material Implication 2 | 4.1 | 7 | 3.7 | 3.7 | 23.5 | 15.9 |
| Negate Hypothesis | 4.4 | 5.6 | 7.9 | 8.7 | 5.7 | 6.4 |
| Avg. | 19.5 | 20.8 | 16.1 | 17.1 | 11.2 | 9.9 |
| Inference Rules | # of negations | # of logical AND | # of logical OR | # of implications | # of contained original sentences |
| Modus Tollens | 2 | 1 | 0 | 1 | 4 |
| Constructive Dilemma | 0 | 2 | 2 | 2 | 8 |
| Destructive Dilemma | 4 | 2 | 2 | 2 | 8 |
| Bidirectional Dilemma | 2 | 2 | 2 | 2 | 8 |
| Transposition 1 | 2 | 0 | 0 | 2 | 4 |
| Transposition 2 | 2 | 0 | 0 | 2 | 4 |
| Material Implication 1 | 1 | 0 | 1 | 1 | 4 |
| Material Implication 2 | 1 | 0 | 1 | 1 | 4 |
| Negate Hypothesis | 1 | 0 | 0 | 0 | 2 |
| Name | Formal Expressions | Attack Examples |
| Original 1 | (p1→h1) | Premise: An older man wearing a salon drape getting a haircut.\nHypothesis: A man gets a haircut. |
| Original 2 | (p2→h2) | Premise: The dog runs towards the ball.\nHypothesis: A dog runs. |
| Modus Tollens | ((p1→h1) ∧ ¬h1) ⊨ ¬p1 | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nNo man gets a haircut.\nHypothesis: No older man wearing a salon drape is getting a haircut. |
| Constructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ p2)) ⊢ (h1 ∨ h2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs. But either an older man wearing a salon drape getting a haircut or the dog runs towards the ball.\nHypothesis: Either a man gets a haircut or a dog runs. |
| Destructive Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (¬h1 ∨ ¬h2)) ⊢ (¬p1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs. But either no man gets a haircut or no dog runs.\nHypothesis: Either no older man wearing a salon drape is getting a haircut or the dog does not run towards the ball. |
| Bidirectional Dilemma | ((p1→h1) ∧ (p2→h2) ∧ (p1 ∨ ¬h2)) ⊢ (h1 ∨ ¬p2) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nAnd if the dog runs towards the ball, then a dog runs.\nBut either an older man wearing a salon drape getting a haircut or no dog runs.\nHypothesis: Either a man gets a haircut or the dog does not run towards the ball. |
| Transportation 1 | ((p1→h1) ⊢ (¬h1 → ¬p1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nHypothesis: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut. |
| Transportation 2 | (¬h1 → ¬p1) ⊢ ((p1 → h1) | Premise: If no man gets a haircut, then no older man wearing a salon drape is getting a haircut.\nHypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Material Implication 1 | (p1→h1) ⊢ (¬p1 ∨ h1) | Premise: If an older man wearing a salon drape getting a haircut, then a man gets a haircut.\nHypothesis: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut. |
| Material Implication 2 | (¬p1 ∨ h1) ⊢ (p1 → h1) | Premise: Either no older man wearing a salon drape is getting a haircut or a man gets a haircut.\nHypothesis: If an older man wearing a salon drape getting a haircut, then a man gets a haircut. |
| Negate Hypothesis | (p1 → ¬h1) | Premise: An older man wearing a salon drape getting a haircut.\nHypothesis: No man gets a haircut. |
| Dataset | Attack Type | Single-Task | Multi-Task | Prompting | F1 value | ||||
| RoBERTa | RoBERTa | BART | FLAN-T5 | GPT-3 | ChatGPT | GPT-4 | |||
| SNLI | Modus Tollens | 99.7 | 94.1 | 45.5 | 99.9 | 97.8 | 96.3 | 50.6 | 61.0 |
| Constructive Dilemma | 34.7 | 3.9 | 0.1 | 0.6 | 34.7 | 11.3 | 9.0 | 60.4 | |
| Destructive Dilemma | 93.6 | 30.2 | 24.9 | 87.3 | 90.4 | 46.7 | 18.8 | 57.9 | |
| Bidirectional Dilemma | 73.4 | 67.0 | 22.5 | 15.6 | 63.5 | 16.3 | 11.7 | 59.1 | |
| Transposition 1 | 99.7 | 82.5 | 99.1 | 82.0 | 79.6 | 25.0 | 22.7 | 58.6 | |
| Transposition 2 | 95.4 | 40.0 | 94.4 | 94.2 | 46.2 | 44.0 | 16.2 | 58.6 | |
| Material Implication 1 | 98.6 | 85.8 | 94.8 | 94.5 | 54.6 | 25.5 | 11.1 | 59.8 | |
| Material Implication 2 | 85.6 | 93.5 | 46.1 | 33.0 | 35.4 | 12.9 | 29.8 | 59.8 | |
| Negate Hypothesis | 1.1 | 3.2 | 1.9 | 0.6 | 35.0 | 15.4 | 9.7 | 95.2 | |
| Avg. | 76.1 | 56.2 | 48.7 | 56.4 | 59.7 | 32.7 | 20.0 | 63.4 | |
| MNLI | Modus Tollens | 96.6 | 81.1 | 68.1 | 94.8 | 94.6 | 96.1 | 72.3 | 62.3 |
| Constructive Dilemma | 23.8 | 2.8 | 0.7 | 7.0 | 26.9 | 4.6 | 0.0 | 62.2 | |
| Destructive Dilemma | 90.5 | 58.4 | 64.0 | 86.9 | 96.5 | 94.8 | 63.2 | 60.4 | |
| Bidirectional Dilemma | 85.6 | 68.9 | 52.8 | 49.4 | 70.1 | 58.4 | 30.6 | 61.3 | |
| Transposition 1 | 96.5 | 87.1 | 91.5 | 78.1 | 67.1 | 57.3 | 47.1 | 60.5 | |
| Transposition 2 | 90.4 | 74.1 | 92.3 | 95.4 | 33.6 | 32.7 | 42.9 | 60.5 | |
| Material Implication 1 | 88.9 | 70.4 | 72.9 | 78.6 | 45.6 | 53.1 | 40.7 | 61.4 | |
| Material Implication 2 | 75.8 | 79.3 | 64.9 | 43.2 | 18.9 | 7.6 | 69.9 | 61.4 | |
| Negate Hypothesis | 8.8 | 11.0 | 10.0 | 5.9 | 11.6 | 9.1 | 5.6 | 94.2 | |
| Avg. | 73.5 | 59.9 | 58.3 | 59.9 | 51.7 | 46.0 | 41.4 | 64.9 |
| Dataset | Attack Strategies | SimCSE | BERTScore |
| SNLI | Modus Tollens | 0.839 | 0.919 |
| Constructive Dilemma | 0.935 | 0.908 | |
| Destructive Dilemma | 0.869 | 0.906 | |
| Bidirectional Dilemma | 0.909 | 0.906 | |
| Transposition 1 | 0.866 | 0.912 | |
| Transposition 2 | 0.848 | 0.904 | |
| Material Implication 1 | 0.901 | 0.917 | |
| Material Implication 2 | 0.880 | 0.910 | |
| Negate Hypothesis | 0.868 | 0.967 | |
| Avg. | 0.879 | 0.917 | |
| MNLI | Modus Tollens | 0.846 | 0.926 |
| Constructive Dilemma | 0.937 | 0.914 | |
| Destructive Dilemma | 0.892 | 0.914 | |
| Bidirectional Dilemma | 0.914 | 0.914 | |
| Transposition 1 | 0.864 | 0.920 | |
| Transposition 2 | 0.864 | 0.911 | |
| Material Implication 1 | 0.905 | 0.925 | |
| Material Implication 2 | 0.891 | 0.918 | |
| Negate Hypothesis | 0.879 | 0.963 | |
| Avg. | 0.888 | 0.923 |
| Inference Rules | GPT-3 | ChatGPT | GPT-4 | |||
| 3-shots | 6-shots | 3-shots | 6-shots | 3-shots | 6-shots | |
| Modus Tollens | 69.4 | 67.2 | 90.9 | 92.8 | 12.3 | 11.4 |
| Constructive Dilemma | 2.4 | 4.2 | 3.7 | 3.6 | 5.3 | 5.8 |
| Destructive Dilemma | 15.9 | 15.6 | 14.7 | 17.7 | 13.8 | 13.6 |
| Bidirectional Dilemma | 4.3 | 6.5 | 5.4 | 6.1 | 8.3 | 8.6 |
| Transposition 1 | 71.5 | 68.7 | 4.7 | 5.3 | 13.2 | 10.0 |
| Transposition 2 | 11.5 | 16.1 | 10.3 | 11.7 | 12.8 | 11.1 |
| Material Implication 1 | 5.1 | 6.5 | 3.7 | 4.1 | 5.6 | 5.9 |
| Material Implication 2 | 4.1 | 7 | 3.7 | 3.7 | 23.5 | 15.9 |
| Negate Hypothesis | 4.4 | 5.6 | 7.9 | 8.7 | 5.7 | 6.4 |
| Avg. | 19.5 | 20.8 | 16.1 | 17.1 | 11.2 | 9.9 |
| Inference Rules | # of negations | # of logical AND | # of logical OR | # of implications | # of contained original sentences |
| Modus Tollens | 2 | 1 | 0 | 1 | 4 |
| Constructive Dilemma | 0 | 2 | 2 | 2 | 8 |
| Destructive Dilemma | 4 | 2 | 2 | 2 | 8 |
| Bidirectional Dilemma | 2 | 2 | 2 | 2 | 8 |
| Transposition 1 | 2 | 0 | 0 | 2 | 4 |
| Transposition 2 | 2 | 0 | 0 | 2 | 4 |
| Material Implication 1 | 1 | 0 | 1 | 1 | 4 |
| Material Implication 2 | 1 | 0 | 1 | 1 | 4 |
| Negate Hypothesis | 1 | 0 | 0 | 0 | 2 |
| MODEL | CONSTRAINT | SEARCH | WELLFORMED | F1 |
| T5 BASE | UNCONSTRAINED | GREEDY | 99.4% | - |
| BEAM=4 | 99.4% | - | ||
| LEVENShteIN | GREEDY | 100.0% | 0.786 | |
| FINE TUNED | BEAM=4 | 100.0% | 0.788 | |
| FST | GREEDY | 100.0% | 0.786 | |
| BEAM=4 | 100.0% | 0.788 | ||
| PALM 8B | UNCONSTRAINED | GREEDY | 14.5% | - |
| BEAM=4 | 52.7% | - | ||
| LEVENShteIN | GREEDY | 100.0% | 0.715 | |
| PROMPT TUNED | BEAM=4 | 100.0% | 0.689 | |
| FST | GREEDY | 100.0% | 0.717 | |
| BEAM=4 | 100.0% | 0.727 | ||
| PALM 62B | UNCONSTRAINED | GREEDY | 85.9% | - |
| BEAM=4 | 89.0% | - | ||
| LEVENShteIN | GREEDY | 100.0% | 0.735 | |
| PROMPT TUNED | BEAM=4 | 100.0% | 0.737 | |
| FST | GREEDY | 100.0% | 0.761 | |
| BEAM=4 | 100.0% | 0.764 |
| Policy | F1 | EN-DE | EN-ES | EN-AR | Avg | ||||||
| TED 2014 | 2014 | 2015 | 2018 | 2014 | 2015 | 2018 | 2014 | 2015 | 2018 | ||
| Baselines and Oracle: | |||||||||||
| ORACLE | 1.000 | 26.66 | 30.24 | 25.21 | 40.38 | 41.72 | 41.84 | 15.66 | 18.18 | 17.59 | 29.62 |
| FIXEDLENGTH | 0.041 | 20.82 | 23.45 | 19.66 | 32.76 | 34.03 | 34.01 | 12.64 | 14.79 | 13.92 | 23.66 |
| LI ET AL. (2021) | - | - | 27.00 | 22.00 | - | - | - | - | - | - | - |
| Small Structured Prediction Models: | |||||||||||
| PUNCTUATE | - | 22.80 | 26.30 | 21.60 | 35.70 | 36.90 | 36.70 | 13.70 | 15.80 | 15.40 | 25.81 |
| BIGRU F.T. | 0.697 | 24.55 | 28.10 | 23.14 | 37.31 | 39.08 | 38.64 | 14.41 | 16.77 | 16.19 | 27.39 |
| LLMs: | |||||||||||
| T5-BASE | 0.788 | 25.28 | 29.14 | 24.05 | 38.75 | 40.23 | 39.96 | 14.94 | 17.32 | 16.57 | 28.33 |
| T5-11B | 0.821 | 25.63 | 29.63 | 24.27 | 39.16 | 40.64 | 40.05 | 15.31 | 17.60 | 16.48 | 28.66 |
| T5-11B-ASR | 0.836 | 25.71 | 29.28 | 24.22 | 39.11 | 40.47 | 40.02 | 15.24 | 17.58 | 16.66 | 28.59 |
| PALM-PROMPTUNED-62B | 0.764 | 25.10 | 28.69 | 23.92 | 38.52 | 40.01 | 39.22 | 15.03 | 17.13 | 16.58 | 28.08 |
| PALM-PROMPTUNED-62B-ASR | 0.781 | 25.15 | 29.09 | 23.71 | 38.69 | 40.07 | 39.31 | 15.13 | 17.21 | 16.76 | 28.17 |
| PALM-FINETUNED-62B | 0.820 | 25.71 | 29.19 | 23.97 | 38.96 | 40.56 | 39.74 | 15.07 | 17.66 | 16.90 | 28.51 |
| PALM-FINETUNED-62B-ASR | 0.832 | 25.84 | 29.37 | 24.13 | 39.02 | 40.46 | 39.89 | 15.17 | 17.80 | 16.65 | 28.61 |
| PALM-PROMPTUNED-540B | 0.816 | 25.44 | 29.29 | 24.23 | 38.95 | 40.70 | 39.74 | 15.03 | 17.61 | 16.86 | 28.49 |
| PALM-PROMPTUNED-540B-ASR | 0.835 | 25.52 | 29.37 | 24.15 | 39.08 | 40.67 | 39.98 | 15.11 | 17.64 | 16.61 | 28.56 |