| Passage | ...If you already have a loan, the bank will definitely consider the debt ratio, depending on the assets and repayment ability.... |
| Question | Will Credit Loan Affect Mortgage? |
| Response | It could affect your Mortgage, but it depends on your income level and assets. |
| Generated Response-One stage |
| T5 | Credit loan must affect Mortgage. |
| BART | It will affect your Mortgage. |
| Prompt-BART | Credit loan could affect mortgage, but depends on assets. |
| Generated Response-Two stage |
| T5 | Credit loan must affect Mortgage. |
| BART | Credit loan will affect mortgage |
| Prompt-BART | Credit loan may affect your Mortgage. |
",
+ "image_path": "a45db3a8f52187f070a482eaab47763578458fc935bcbe1d1f1e4a9c2401563f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 302,
+ 312,
+ 525,
+ 361
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 312,
+ 525,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 312,
+ 525,
+ 361
+ ],
+ "type": "text",
+ "content": "Table 6: A case study in our collected Penguin, and generated responses from the large version of baselines are provided. The text in bold refers to the expected response from human annotation."
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 368,
+ 525,
+ 502
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 368,
+ 525,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 368,
+ 525,
+ 502
+ ],
+ "type": "text",
+ "content": "each example in randomly selected 200 test samples from three views, as shown in Figure 6. We can observe that more than "
+ },
+ {
+ "bbox": [
+ 302,
+ 368,
+ 525,
+ 502
+ ],
+ "type": "inline_equation",
+ "content": "85\\%"
+ },
+ {
+ "bbox": [
+ 302,
+ 368,
+ 525,
+ 502
+ ],
+ "type": "text",
+ "content": " of the generated responses are superior to their counterpart answers across all evaluation perspectives, indicating they are closer to human-like responses than the generated answers. These Win/Lose results also prove the value of Penguin, that is, providing a training and testing bed for natural response generation to real-world scenarios."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 511,
+ 525,
+ 606
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 511,
+ 525,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 511,
+ 525,
+ 606
+ ],
+ "type": "text",
+ "content": "Case Study In addition, we showcase an example and its generated responses from Penguin in Table 6. We can see that almost all of the responses from the different models do a pretty good job of roughly capturing the intended meaning. Moreover, Prompt-BART could generate more complete and accurate responses, showing its efficiency."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 616,
+ 381,
+ 628
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 616,
+ 381,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 616,
+ 381,
+ 628
+ ],
+ "type": "text",
+ "content": "6 Conclusion"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "content": "In this paper, we summarize the limitation of current MRC efforts, and propose a new dataset named Penguin for natural response generation in Chinese MRC. Penguin requires the predicted responses not only be accurate but also natural-sounding, catering to real scenarios. Thereafter, we propose end-to-end and two-stage GRC pipelines to address the challenge in Penguin, achieving promising results. Moreover, we design prompt-tuning, which involves fine-tuning the pre-trained genera"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 312,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 312,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 312,
+ 791
+ ],
+ "type": "text",
+ "content": "11075"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 7
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 290,
+ 137
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 290,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 290,
+ 137
+ ],
+ "type": "text",
+ "content": "tive language models with a mix of different prefix prompts in Penguin, making the system perform better and more efficiently. As an extension of future work, we will explore other techs to build more human-like GRC models."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 148,
+ 131,
+ 161
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 148,
+ 131,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 148,
+ 131,
+ 161
+ ],
+ "type": "text",
+ "content": "Limitations"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 67,
+ 169,
+ 291,
+ 371
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 169,
+ 291,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 169,
+ 291,
+ 371
+ ],
+ "type": "text",
+ "content": "The main target of this paper is to generate natural responses for Chinese Reading Comprehension. We present a new benchmark named Penguin in the hope of creating an effective GRC model. More generally, we expect the proposed Penguin can facilitate the research of MRC, catering to more realistic QA scenarios. Our collected corpus will be soon made public for the whole research communities. Admittedly, one limitation of this work is that the proposed dataset is restricted to Chinese. Another limitation is that our labeled responses are not "
+ },
+ {
+ "bbox": [
+ 67,
+ 169,
+ 291,
+ 371
+ ],
+ "type": "inline_equation",
+ "content": "100\\%"
+ },
+ {
+ "bbox": [
+ 67,
+ 169,
+ 291,
+ 371
+ ],
+ "type": "text",
+ "content": " from human annotation due to the high annotation cost. These concerns warrant further research and consideration when utilizing this work to build effective GRC models."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 68,
+ 394,
+ 127,
+ 406
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 394,
+ 127,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 394,
+ 127,
+ 406
+ ],
+ "type": "text",
+ "content": "References"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 69,
+ 412,
+ 291,
+ 772
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 9,
+ "blocks": [
+ {
+ "bbox": [
+ 69,
+ 412,
+ 291,
+ 458
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 412,
+ 291,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 412,
+ 291,
+ 458
+ ],
+ "type": "text",
+ "content": "Prithviraj Ammanabrolu, Renee Jia, and Mark O. Riedl. 2022. Situated dialogue learning through procedural environment generation. In ACL (1), pages 8099-8116. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 69,
+ 463,
+ 290,
+ 509
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 463,
+ 290,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 463,
+ 290,
+ 509
+ ],
+ "type": "text",
+ "content": "Ashutosh Baheti, Alan Ritter, and Kevin Small. 2020. Fluent response generation for conversational question answering. In ACL, pages 191-207. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 69,
+ 515,
+ 291,
+ 581
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 515,
+ 291,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 515,
+ 291,
+ 581
+ ],
+ "type": "text",
+ "content": "Bin Bi, Chen Wu, Ming Yan, Wei Wang, Jiangnan Xia, and Chenliang Li. 2019. Incorporating external knowledge into machine reading for generative question answering. In EMNLP/IJCNLP (1), pages 2521-2530. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 69,
+ 588,
+ 291,
+ 634
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 588,
+ 291,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 588,
+ 291,
+ 634
+ ],
+ "type": "text",
+ "content": "Peter F. Brown, Stephen Della Pietra, Vincent J. Della Pietra, Jennifer C. Lai, and Robert L. Mercer. 1992. An estimate of an upper bound for the entropy of english. Comput. Linguistics, 18(1):31-40."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 69,
+ 640,
+ 291,
+ 772
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 640,
+ 291,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 640,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In NeurIPS."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 526,
+ 772
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 20,
+ "blocks": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 525,
+ 138
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 525,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 525,
+ 138
+ ],
+ "type": "text",
+ "content": "Nuo Chen, Hongguang Li, Yinan Bao, Junqing He, Xinshi Lin, Qi Yang, Jianfeng Liu, Ruyi Gan, Ji-axing Zhang, Baoyuan Wang, et al. 2023a. Orca: A few-shot benchmark for chinese conversational machine reading comprehension. arXiv preprint arXiv:2302.13619."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 304,
+ 147,
+ 526,
+ 224
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 147,
+ 526,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 147,
+ 526,
+ 224
+ ],
+ "type": "text",
+ "content": "Nuo Chen, Fenglin Liu, Chenyu You, Peilin Zhou, and Yuexian Zou. 2021. Adaptive bi-directional attention: Exploring multi-granularity representations for machine reading comprehension. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7833-7837. IEEE."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 304,
+ 233,
+ 526,
+ 289
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 233,
+ 526,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 233,
+ 526,
+ 289
+ ],
+ "type": "text",
+ "content": "Nuo Chen, Linjun Shou, Ming Gong, and Jian Pei. 2022a. From good to best: Two-stage training for cross-lingual machine reading comprehension. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10501-10508."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 304,
+ 297,
+ 526,
+ 387
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 297,
+ 526,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 297,
+ 526,
+ 387
+ ],
+ "type": "text",
+ "content": "Nuo Chen, Linjun Shou, Ming Gong, Jian Pei, and Daxin Jiang. 2022b. Bridging the gap between language models and cross-lingual sequence labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1909–1923, Seattle, United States. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 304,
+ 395,
+ 526,
+ 473
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 395,
+ 526,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 395,
+ 526,
+ 473
+ ],
+ "type": "text",
+ "content": "Nuo Chen, Linjun Shou, Tengtao Song, Ming Gong, Jian Pei, Jianhui Chang, Daxin Jiang, and Jia Li. 2023b. Structural contrastive pretraining for crosslingual comprehension. In Findings of the Association for Computational Linguistics: ACL 2023, pages 2042-2057, Toronto, Canada. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 304,
+ 481,
+ 526,
+ 526
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 481,
+ 526,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 481,
+ 526,
+ 526
+ ],
+ "type": "text",
+ "content": "Nuo Chen, Yan Wang, Haiyun Jiang, Deng Cai, Ziyang Chen, and Jia Li. 2022c. What would harry say? building dialogue agents for characters in a story. arXiv preprint arXiv:2211.06869."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 304,
+ 534,
+ 526,
+ 602
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 534,
+ 526,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 534,
+ 526,
+ 602
+ ],
+ "type": "text",
+ "content": "Yiming Cui, Ting Liu, Wanxiang Che, Li Xiao, Zhipeng Chen, Wentao Ma, Shijin Wang, and Guoping Hu. 2019. A span-extraction dataset for chinese machine reading comprehension. In EMNLP/IJCNLP (1), pages 5882-5888. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 304,
+ 610,
+ 526,
+ 666
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 610,
+ 526,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 610,
+ 526,
+ 666
+ ],
+ "type": "text",
+ "content": "Wei He, Kai Liu, Yajuan Lyu, Shiqi Zhao, Xinyan Xiao, Yuan Liu, Yizhong Wang, Hua Wu, Qiaoqiao She, Xuan Liu, et al. 2017. Dureader: a chinese machine reading comprehension dataset from real-world applications. arXiv preprint arXiv:1711.05073."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 304,
+ 674,
+ 526,
+ 708
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 674,
+ 526,
+ 708
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 674,
+ 526,
+ 708
+ ],
+ "type": "text",
+ "content": "Yanzhou Huang and Tao Zhong. 2018. Multitask learning for neural generative question answering. Mach. Vis. Appl., 29(6):1009-1017."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 304,
+ 717,
+ 526,
+ 772
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 717,
+ 526,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 717,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "content": "Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. 2022. Understanding and improving zero-shot multi-hop reasoning in generative question answering. In COLING, pages 1765-1775. International Committee on Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 312,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 312,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 312,
+ 791
+ ],
+ "type": "text",
+ "content": "11076"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 8
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 69,
+ 72,
+ 291,
+ 772
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 9,
+ "blocks": [
+ {
+ "bbox": [
+ 69,
+ 72,
+ 291,
+ 139
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 72,
+ 291,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 72,
+ 291,
+ 139
+ ],
+ "type": "text",
+ "content": "Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. Race: Large-scale reading comprehension dataset from examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 796-805. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 69,
+ 149,
+ 291,
+ 248
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 149,
+ 291,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 149,
+ 291,
+ 248
+ ],
+ "type": "text",
+ "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7871-7880. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 69,
+ 260,
+ 291,
+ 349
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 260,
+ 291,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 260,
+ 291,
+ 349
+ ],
+ "type": "text",
+ "content": "Chenliang Li, Bin Bi, Ming Yan, Wei Wang, and Songfang Huang. 2021. Addressing semantic drift in generative question answering with auxiliary extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 942-947, Online. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 69,
+ 360,
+ 291,
+ 449
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 360,
+ 291,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 360,
+ 291,
+ 449
+ ],
+ "type": "text",
+ "content": "Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110-119, San Diego, California. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 69,
+ 460,
+ 291,
+ 505
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 460,
+ 291,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 460,
+ 291,
+ 505
+ ],
+ "type": "text",
+ "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 69,
+ 516,
+ 291,
+ 571
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 516,
+ 291,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 516,
+ 291,
+ 571
+ ],
+ "type": "text",
+ "content": "Benjamin Muller, Luca Soldaini, Rik Koncel-Kedziorski, Eric Lind, and Alessandro Moschitti. 2021. Cross-lingual genqa: A language-agnostic generative question answering approach for opendomain question answering. CoRR, abs/2110.07150."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 69,
+ 582,
+ 291,
+ 638
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 582,
+ 291,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 582,
+ 291,
+ 638
+ ],
+ "type": "text",
+ "content": "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 69,
+ 649,
+ 291,
+ 705
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 649,
+ 291,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 649,
+ 291,
+ 705
+ ],
+ "type": "text",
+ "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 69,
+ 716,
+ 291,
+ 772
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 716,
+ 291,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 716,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 526,
+ 766
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 20,
+ "blocks": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 526,
+ 139
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 526,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 526,
+ 139
+ ],
+ "type": "text",
+ "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: "
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 526,
+ 139
+ ],
+ "type": "inline_equation",
+ "content": "100,000+"
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 526,
+ 139
+ ],
+ "type": "text",
+ "content": " questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 304,
+ 145,
+ 526,
+ 212
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 145,
+ 526,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 145,
+ 526,
+ 212
+ ],
+ "type": "text",
+ "content": "Tengtao Song, Nuo Chen, Ji Jiang, Zhihong Zhu, and Yuexian Zou. 2023. Improving retrieval-based dialogue system via syntax-informed attention. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 304,
+ 218,
+ 526,
+ 264
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 218,
+ 526,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 218,
+ 526,
+ 264
+ ],
+ "type": "text",
+ "content": "Adam Trischler, Tong Wang, Xingdi Yuan, Justin Harris, Alessandro Sordoni, Philip Bachman, and Kaheer Suleman. 2016. Newsqa: A machine comprehension dataset. arXiv preprint arXiv:1611.09830."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 304,
+ 269,
+ 526,
+ 380
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 269,
+ 526,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 269,
+ 526,
+ 380
+ ],
+ "type": "text",
+ "content": "Bingning Wang, Ting Yao, Qi Zhang, Jingfang Xu, and Xiaochuan Wang. 2020. Reco: A large scale chinese reading comprehension dataset on opinion. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 9146-9153. AAAI Press."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 304,
+ 387,
+ 526,
+ 443
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 387,
+ 526,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 387,
+ 526,
+ 443
+ ],
+ "type": "text",
+ "content": "Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. In ICLR. OpenReview.net."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 304,
+ 449,
+ 526,
+ 495
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 449,
+ 526,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 449,
+ 526,
+ 495
+ ],
+ "type": "text",
+ "content": "Jun Yin, Xin Jiang, Zhengdong Lu, Lifeng Shang, Hang Li, and Xiaoming Li. 2016. Neural generative question answering. In Proceedings of the Workshop on Human-Computer Question Answering, pages 36-42."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 304,
+ 501,
+ 526,
+ 556
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 501,
+ 526,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 501,
+ 526,
+ 556
+ ],
+ "type": "text",
+ "content": "Chenyu You, Nuo Chen, Fenglin Liu, Shen Ge, Xian Wu, and Yuexian Zou. 2022. End-to-end spoken conversational question answering: Task, dataset and model. In *NAACL-HLT* (Findings), pages 1219–1232. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 304,
+ 564,
+ 526,
+ 629
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 564,
+ 526,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 564,
+ 526,
+ 629
+ ],
+ "type": "text",
+ "content": "Chenyu You, Nuo Chen, and Yuexian Zou. 2021a. Knowledge distillation for improved accuracy in spoken question answering. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7793-7797. IEEE."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 304,
+ 637,
+ 526,
+ 703
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 637,
+ 526,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 637,
+ 526,
+ 703
+ ],
+ "type": "text",
+ "content": "Chenyu You, Nuo Chen, and Yuexian Zou. 2021b. Self-supervised contrastive cross-modality representation learning for spoken question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 28-39, Punta Cana, Dominican Republic. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 304,
+ 710,
+ 526,
+ 766
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 710,
+ 526,
+ 766
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 710,
+ 526,
+ 766
+ ],
+ "type": "text",
+ "content": "Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. 2018. SWAG: A large-scale adversarial dataset for grounded commonsense inference. In EMNLP, pages 93-104. Association for Computational Linguistics."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "type": "text",
+ "content": "11077"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 9
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 69,
+ 68,
+ 289,
+ 153
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 69,
+ 68,
+ 289,
+ 153
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 68,
+ 289,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 68,
+ 289,
+ 153
+ ],
+ "type": "table",
+ "html": "| Question Type | YES/NO (是否类) |
| Passage | ...Three-year-old baby eat two or three at a time on the line, the child's gastrointestinal function is relatively weak, eating more will cause gastrointestinal discomfort... ...三周岁宝宝每次吃两三个就行了,孩子的胃肠道功能比较弱,吃多了会引起肠胃不适... |
| Question | Is it good for three-year-old baby to litchi? 三周岁宝宝吃荔枝好吗? |
| Answer | Yes. 好 |
| Expected Response | Three-year-old babies can eat two or three litchis, but not too much. 三周岁的宝宝可以吃两三个荔枝,但是不能吃太多。 |
| Generated Response-One-Stage |
| T5 | They can not eat litchi. 他们不可以吃荔枝 |
| BART | They can eat litchi. 他们可以吃荔枝 |
| Prompt-BART | They can eat litchi. 他们可以吃荔枝 |
| Generated Response-Two-Stage |
| T5 | Three-year-old baby can not eat litchi. 三周岁的宝宝不可以吃荔枝 |
| BART | They can eat litchi. 他们可以吃荔枝 |
| Prompt-BART | Three-year-old baby can eat litchi. 三周岁的宝宝可以吃荔枝 |
| (a) A Generated Response example of Yes/No questions. |
| Question Type | Opinion (观点类) |
| Passage | ...It's not a matter of how fast you move, it's a matter of understanding. I have been learning the piano for 10 years and I have to practice these czernies and so on over and over again this year. It is recommended to learn Bayer first, which greatly improves the basic skills.... ...进度快不快不是问题,主要看你的理解能力。我学了10年琴,车尔尼这些等等今年都要反复的练习。推荐还是先学拜厄,对基本功有很大的提高。... |
| Question | I've already learned Faber, so do I still need to practice Baier? 我已经学习了菲伯尔,那还需要练习拜厄吗? |
| Answer | Of course. 需要 |
| Expected Response | You'd better practice Baier. It can improve your basic skills. 推荐还是练习拜厄,这对你的基本功有好处。 |
| Generated Response-One-Stage |
| T5 | You need to practice Baier. 还是要练习拜厄 |
| BART | You still need to practice Baier. 仍然要练习拜厄 |
| Prompt-BART | You still need to practice Baier. 仍然要练习拜厄 |
| Generated Response-Two-Stage |
| T5 | You still need to practice Baier. 仍然要练习拜厄 |
| BART | You still need to practice Baier. 仍然要练习拜厄 |
| Prompt-BART | He still needs to practice Baier. 他仍然要练习拜厄 |
| (b) A Generated Response example of Opinion questions. |
",
+ "image_path": "9e797ac4b64adbf7ec8db062485a2ebca39e9a3fae79d11c7d473f1e9fad3584.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 721,
+ 524,
+ 734
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 721,
+ 524,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 721,
+ 524,
+ 734
+ ],
+ "type": "text",
+ "content": "Table 11: Typical generated examples from each model in Penguin. English translation is also given for comparison."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "type": "text",
+ "content": "11079"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 11
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 488
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 488
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 488
+ ],
+ "type": "table",
+ "html": "| Question Type | Fact (事实类) |
| Passage | ...Zhang Lei is a Chinese football player and goalkeeper. Zhang Lei was selected to Guang-dong Hongyuan Youth Team at the age of 13, then played as the main goalkeeper for Dongguan Southern City before moving to Shenzhen in 2006. He was selected to the 2008 National Olympic Team and the 2005 World Youth Championships squad. In 2009, Beijing Guoan exchanged Cheng Yuelei, a backup goalkeeper, plus 400,000 yuan to Shenzhen in exchange for Zhang Lei, which was also the last domestic import transaction of Beijing Guoan in the first transfer market of the 2009 season. Zhang Lei did not get a chance to play because Beijing Guoan had Yang Zhi as their keeper. He was listed by Beijing Guoan and eventually moved to Changsha Jinde in 2010 after losing their main goalkeeper Song Zhenyu. In 2011, Zhang moved to Chongqing Lifan and became the team's main goalkeeper. Served for five years, the first three years to play the main force, the last two years gradually as a substitute. In 2016, he was transferred to Hangzhou Greentown. Zhang Lei joined the Lijiang Flying Tigers on loan from Hangzhou Greentown in 2017....张磊，中国足球运动员，司职守门员。张磊13岁时入选广东宏远少年队，随后担任东莞南城的主力门将，2006年转会到深圳队。他入选了08之星国奥队，并进入2005年世青赛的大名单。2009年北京国安用替补门将程月磊加上40万元，与深圳队交换张磊，这也是2009赛季中超第一次转会市场上北京国安队最后一笔内援引进交易。因为北京国安有杨智把关，所以张磊没有得到出场机会，结果被北京国安挂牌，最终在2010年转会到失去主力门将宋振瑜的长沙金德，成为球队新的主力门将。2011年，张磊转会到重庆力帆成为球队的主力门将。效力了五年期间，前三年打主力，后两年渐为替补。2016年转会到杭州绿城。2017年，张磊从杭州绿城租借到丽江飞虎。... |
| Question | In what year did Zhang Lei transfer to Shenzhen?张磊哪一年转会去的深圳? |
| Answer | 2006. 2006 |
| Expected Response | He transfers to Shenzhen in 2006. 他于2006年转会到深圳。 |
| Generated Response-One-Stage |
| T5BARTPrompt-BART | He transferred to Shenzhen in two years. 他于2年转会到深圳队He transferred to Shenzhen in two years. 他于2年转会到深圳队He transfers to Shenzhen in 2006. 他于2006年转会到深圳 |
| Generated Response-Two-Stage |
| T5BARTPrompt-BART | He transfers to Shenzhen in 2006. 他于2006年转会到深圳He transfers to Shenzhen in 2006. 他于2006年转会到深圳Zhang Lei transfers to Shenzhen in 2006. 张磊于2006年转会到深圳 |
",
+ "image_path": "5a88946cf8c50c27c7de13cec996a6a933bd7fff04e937bd61552d94019e35b2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 496,
+ 525,
+ 521
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 496,
+ 525,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 496,
+ 525,
+ 521
+ ],
+ "type": "text",
+ "content": "Table 12: A Generated Response example of Fact questions in Penguin. English translation is also given for comparison."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 67,
+ 541,
+ 155,
+ 555
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 541,
+ 155,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 541,
+ 155,
+ 555
+ ],
+ "type": "text",
+ "content": "G All Prompts"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 562,
+ 291,
+ 602
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 562,
+ 291,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 562,
+ 291,
+ 602
+ ],
+ "type": "text",
+ "content": "We design five prompts for each module in end-to-end and two-stage frameworks, which are presented in Table 13."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 313,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 313,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 313,
+ 791
+ ],
+ "type": "text",
+ "content": "11080"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 12
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 75,
+ 263,
+ 519,
+ 568
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 75,
+ 263,
+ 519,
+ 568
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 263,
+ 519,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 263,
+ 519,
+ 568
+ ],
+ "type": "table",
+ "html": "| Framework | Prompts |
| One-Stage | 1. Please generate a reasonable response based on the following question and passage. Question: Q.Passage: P.\n2. The question is Q, give the reasonable response based on the following paragraph: P.\n3. This is a reading comprehension problem, please generate the reasonable response based on the following question and passage. Question is Q. Passage is P.\n4. Please answer the question according to descriptions of the question and paragraph. Question is Q.Passage is P.\n5. The question is Q, Conduct question understanding and response generation, knowledge question and response task. Relevant paragraphs are as follows: P. |
| Answerer | 1. Please generate the reasonable answer based on the following question and passage. Question: Q.Passage: P.\n2. The question is Q, give the reasonable answer based on the following paragraph: P.\n3. This is a reading comprehension problem, please generate the reasonable answer based on the following question and passage. Question is Q. Passage is P.\n4. Please answer the question according to descriptions of the question and paragraph. Question is Q.Passage is P.\n5. The question is Q, Conduct question understanding and answer generation, knowledge question and answer task. Relevant paragraphs are as follows: P. |
| Responder | 1. Please generate the reasonable response based on the following question and answer. Question: Q.Answer: A.\n2. The question is Q, give the reasonable response based on the following answer: A.\n3. This is a answer rewriting problem, please generate the reasonable response based on the following question and answer. Question is Q. Answer is A.\n4. Please give the response according to descriptions of the question and answer. Question is Q. Answer is A.\n5. The question is Q, Conduct question understanding and response generation, answer rewriting task.Relevant answer is as follows: A. |
",
+ "image_path": "97e722c508ea6716f37dcd96b2b4f2d4f074c7d38667317d9b024539937bcc36.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 117,
+ 576,
+ 475,
+ 590
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 576,
+ 475,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 576,
+ 475,
+ 590
+ ],
+ "type": "text",
+ "content": "Table 13: All used prompts during training. English translation is only given for reading."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 311,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 311,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 780,
+ 311,
+ 791
+ ],
+ "type": "text",
+ "content": "11081"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 13
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/f2ea77f7-c2a2-49e4-b922-17fa39103dc4_content_list.json b/2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/f2ea77f7-c2a2-49e4-b922-17fa39103dc4_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..7deff9ea523bfd4b50a51fbdc973a8792bab009e
--- /dev/null
+++ b/2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/f2ea77f7-c2a2-49e4-b922-17fa39103dc4_content_list.json
@@ -0,0 +1,1580 @@
+[
+ {
+ "type": "text",
+ "text": "Neurosymbolic Sentiment Analysis with Dynamic Word Sense Disambiguation",
+ "text_level": 1,
+ "bbox": [
+ 275,
+ 90,
+ 722,
+ 130
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Xulang Zhang\\*, Rui Mao\\*, Kai He\\* and Erik Cambria\\*",
+ "bbox": [
+ 248,
+ 147,
+ 749,
+ 165
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "$\\spadesuit$ Nanyang Technological University, Singapore",
+ "bbox": [
+ 309,
+ 166,
+ 692,
+ 183
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "\\*National University of Singapore, Singapore",
+ "bbox": [
+ 314,
+ 183,
+ 687,
+ 200
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "xulang001@e.ntu.edu.sg; {rui.mao, cambria}@ntu.edu.sg; kai_he@nus.edu.sg",
+ "bbox": [
+ 141,
+ 200,
+ 858,
+ 216
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 260,
+ 252,
+ 339,
+ 268
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Sentiment analysis is a task that highly depends on the understanding of word senses. Traditional neural network models are black boxes that represent word senses as vectors that are uninterpretable for humans. On the other hand, the application of Word Sense Disambiguation (WSD) systems in downstream tasks poses challenges regarding i) which words need to be disambiguated, and ii) how to model explicit word senses into easily understandable terms for a downstream model. This work proposes a neurosymbolic framework that incorporates WSD by identifying and paraphrasing ambiguous words to improve the accuracy of sentiment predictions. The framework allows us to understand which words are paraphrased into which semantically unequivocal words, thus enabling a downstream task model to gain both accuracy and interpretability. To better fine-tune a lexical substitution model for WSD on a downstream task without ground-truth word sense labels, we leverage dynamic rewarding to jointly train sentiment analysis and lexical substitution models. Our framework proves to effectively improve the performance of sentiment analysis on corpora from different domains.",
+ "bbox": [
+ 144,
+ 282,
+ 460,
+ 652
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 Introduction",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 665,
+ 258,
+ 682
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "To improve the accuracy of sentiment analysis, it is crucial to disambiguate polysemous words according to the context (Cambria et al., 2023). A word may have senses that carry opposite sentiments, e.g., \"fine\" has the positive meaning of being delicate and refined, and the negative meaning of monetary penalty. Additionally, words may appear in positive, negative, or neutral senses in different domains. For instance, the word \"bearish\" is often used in a negative sense to describe the stock market in the financial domain, while more likely to be used neutrally to describe hulking figures in other domains.",
+ "bbox": [
+ 110,
+ 694,
+ 489,
+ 903
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "In traditional black-box sentiment analysis, the representations of word meanings are generated by neural networks (Mao and Li, 2021). As such, the disambiguation process is implicit and opaque. It is challenging to investigate what implicit senses are assigned to ambiguous words by a neural network model for it to make accurate predictions. Coupling symbolism and sub-symbolism, a neurosymbolic approach to sentiment analysis is transparent and interpretable, regarding the extraction and distinction of word meaning as a sub-problem (Cambria et al., 2017). However, research that explicitly deals with ambiguous words in sentiment analysis is limited (Xia et al., 2014, 2015; Cambria et al., 2015). Sense-annotated resources such as SentiWordNet (Baccianella et al., 2010) have been utilized, but only to detect and score sentimental words in the input (Hung and Lin, 2013; Nassirtoussi et al., 2015; Hung and Chen, 2016). The disambiguation of polysemous words in sentiment analysis inputs remains under-explored. Thus, we are motivated to develop a neurosymbolic framework for sentiment analysis with explicit Word Sense Disambiguation (WSD).",
+ "bbox": [
+ 507,
+ 253,
+ 884,
+ 638
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Traditional WSD aims to classify which sense of the target word is appropriate according to its context (Huang et al., 2019; Blevins and Zettlemoyer, 2020; Bevilacqua and Navigli, 2020; Song et al., 2021; Barba et al., 2021a,b). However, the limitations $(\\ell)$ of their task setups are that $(\\ell 1)$ the output of those WSD systems, whether sense label (Bevilacqua and Navigli, 2020) or gloss definition (Blevins and Zettlemoyer, 2020; Barba et al., 2021a), is difficult to integrate into downstream input; $(\\ell 2)$ the target words must be pre-defined for WSD systems to disambiguate, making it difficult to apply them directly to downstream tasks; and $(\\ell 3)$ the reliability of sense distinction of the benchmark has been questioned (Ramsey, 2017; Maru et al., 2022), as word senses are not discrete.",
+ "bbox": [
+ 507,
+ 642,
+ 884,
+ 898
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_number",
+ "text": "8772",
+ "bbox": [
+ 478,
+ 927,
+ 521,
+ 940
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "footer",
+ "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8772-8783 \nDecember 6-10, 2023 ©2023 Association for Computational Linguistics",
+ "bbox": [
+ 216,
+ 945,
+ 779,
+ 972
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Attempts have been made to incorporate the disambiguation of word senses into downstream applications. Nonetheless, they are either impractical in real-world settings (Farooq et al., 2015; Pu et al., 2018), or only applicable to a specific downstream task, i.e., machine translation (Hangya et al., 2021; Campolungo et al., 2022). To address these issues, we leverage dynamic WSD in our neurosymbolic framework. We regard WSD as a lexical substitution task (for $\\ell 1$ ), where ambiguous words in sentiment analysis input are paraphrased into ones that cause less sentiment confusion using a lexical substitution model with dynamic WSD.",
+ "bbox": [
+ 112,
+ 84,
+ 492,
+ 294
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Our dynamic WSD method comprises two distinct aspects: the dynamic selection of the target word for disambiguation, based on downstream contextual information (for $\\ell 2$ ); and the dynamic learning of the most suitable lexical substitutions, based on the prediction accuracy of the downstream model (for $\\ell 3$ ). To identify target words that need to be substituted, we first adopt an attention-based explainable encoder to rank input words by the level of influence they have on the output, motivated by the assumption that not all polysemous words affect sentiment predictions (Cao et al., 2015). Then, the top-ranking words with the most diverse senses are selected to be targets, based on our hypothesis that an ambiguous word may coincide with significant disparities in the semantic space between the word itself and its potential substitution candidates.",
+ "bbox": [
+ 115,
+ 300,
+ 490,
+ 575
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "As such, we can trace back which words are selected as targets, and which substitute words are used to disambiguate them, making our method interpretable and transparent. Furthermore, we fine-tune the lexical substitution model whilst training the sentiment analysis model by adopting a dynamic rewarding mechanism that does not require gold standard substitution labels. As such, the substitution model can be adjusted to provide more appropriate output for the downstream task as well as the target domain. Our model achieves more accurate sentiment analysis with explicit lexical-substitution-based WSD, despite not specifically trained on WSD and sentiment parallel datasets. Since the disambiguated word senses are represented as the lexical substitutions of the original words and an explainable encoder is used for retrieving the most contributing words, the neurosymbolic sentiment inference process is more interpretable for humans compared to black-box models.",
+ "bbox": [
+ 112,
+ 581,
+ 490,
+ 917
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Our model is examined with three publicly available datasets from movie review, finance, and Twitter domains, outperforming strong baselines by $1.23\\%$ on average. The contributions of our work are twofold: (1) We propose a neurosymbolic sentiment analysis framework that leverages WSD as an auxiliary lexical substitution task, achieving state-of-the-art performance on our examined sentiment analysis datasets from different domains; (2) We design a transparent mechanism to select WSD target words, based on a) the importance of words to the learning and inference of a downstream task, and b) word sense diversity.",
+ "bbox": [
+ 507,
+ 84,
+ 885,
+ 294
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 Related Work",
+ "text_level": 1,
+ "bbox": [
+ 509,
+ 309,
+ 665,
+ 325
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Word Sense Disambiguation is a task that assigns the appropriate meaning to an ambiguous word according to its context. Recent WSD systems attained F1 scores higher than $80\\%$ on the benchmark datasets (Song et al., 2021; Bevilacqua and Navigli, 2020; Barba et al., 2021a,b), which is regarded as the estimated human performance since it is the highest inter-annotator agreement (Edmonds and Kilgarriff, 2002; Palmer et al., 2007). However, research on the application of WSD systems on downstream tasks is limited. Pu et al. (2018) utilized clustering algorithms to obtain sense embeddings by inferring from the similarity of the context vectors of a target word. The heavy requirement for disambiguating every homographs prior to downstream applications makes their method less feasible in many real-world situations. Campolungo et al. (2022) achieved the disambiguation of word sense in neural machine translation by leveraging a multilingual WSD system to establish a sense-tagged parallel corpus. Nonetheless, their approach relied on cross-lingual word alignments for reliable annotations, meaning that it cannot be adopted by other types of downstream tasks.",
+ "bbox": [
+ 507,
+ 338,
+ 885,
+ 724
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Sentiment Analysis can be generally defined as the task that determines if a piece of text expresses positive, negative, or neutral opinion (Mao et al., 2023c). Word sense information has been applied in sentiment analysis in the form of sentiment-annotated lexicons (Devitt and Ahmad, 2007; Ohana and Tierney, 2009; Hung and Lin, 2013; Nassirtoussi et al., 2015), such as SentiWordNet. However, these methods incorporate the sentiment information of a word without disambiguating its word sense according to the context, but resorting to the emotional score of the most frequent",
+ "bbox": [
+ 507,
+ 726,
+ 885,
+ 919
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "8773",
+ "bbox": [
+ 480,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/8339731eaaa1ea3228dd3efe4658383c01c8f2bb2183ac4ccba471934e643f1c.jpg",
+ "image_caption": [
+ "Figure 1: Architecture of the proposed disambiguation framework for sentiment analysis. A dotted line means backpropagation. Green block means that the original word is replaced by a word provided by the lexical substitution model. The darker the green, the higher probability it is assigned to be the model. Orange block indicates the words with top ranking attention scores produced by HAN."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 114,
+ 80,
+ 885,
+ 325
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "sense, or the average score of all senses. Additionally, the main objective of this approach is to identify sentimental words in a sentence, instead of disambiguating polysemous words that might affect model predictions. Hung and Chen (2016); Farooq et al. (2015); Liu et al. (2018) attempted to disambiguate according to target domain and context, but either costed extensive human labour, or fail to incorporate word senses explicitly and transparently.",
+ "bbox": [
+ 112,
+ 420,
+ 487,
+ 580
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "To address these limitations, our neurosymbolic framework aims to identify ambiguous target words in the context and replace them with clarifying ones automatically, making them interpretable. We also leverage a dynamic rewarding mechanism so that the substitutions can be fine-tuned to better serve the downstream task and the text domain.",
+ "bbox": [
+ 112,
+ 582,
+ 487,
+ 694
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3 Task Definition",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 712,
+ 280,
+ 728
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In this work, we regard word sense ambiguity as the potential cause of errors in a downstream task, e.g., sentiment analysis. Inspired by the work of Mao et al. (2023b), we reframe WSD as a lexical substitution task, aiming to replace an ambiguous word with a synonym that not only represents the true sense of the original word in the context (Arthur et al., 2015), but also reduces misclassifications in the downstream task. Given an input sentence $w = (w_{1},w_{2},\\dots ,w_{L})$ , we first aim to identify the target word $t$ that is likely troublesome for",
+ "bbox": [
+ 112,
+ 741,
+ 489,
+ 917
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "the downstream prediction from $w$ . To resolve the potential ambiguity, we extract $M$ of its synonyms from WordNet (Fellbaum, 1998) as candidates $c = (c_{1},\\ldots ,c_{M})$ , forming the input $(s)$ as",
+ "bbox": [
+ 507,
+ 420,
+ 882,
+ 483
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\ns = \\langle s \\rangle , w _ {1}, w _ {2}, \\dots , t, \\dots , w _ {L}, \\langle / s \\rangle , \\tag {1}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 574,
+ 493,
+ 882,
+ 514
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "$< \\mathrm{s}>$ and $$ are special tokens that were defined by our employed pre-trained language model. The lexical substitution model should identify the best candidate from $c$ as the substitute $\\hat{c}$ which retains the original meaning of $t$ in the context by using contrast learning. Thus, $\\hat{c}$ is the symbolic representation of $t$ in context $w$ . Then, the paraphrased input $w^{f} = (w_{1},\\dots,\\hat{c},\\dots,w_{L})$ is fed into a neural network classifier to predict a sentiment label. The objective is that the paraphrased input $w^{f}$ increases the probability of correct sentiment prediction.",
+ "bbox": [
+ 507,
+ 529,
+ 882,
+ 706
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "4 Methodology",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 717,
+ 658,
+ 734
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Our proposed neurosymbolic framework can be viewed in Figure 1. There are two models in our framework, namely sentiment analysis and lexical substitution models. The sentiment analysis model consists of an encoder, an interpretable attention module, and multiple feedforward neural networks. First, through Figure 1 (1) and (2), the original input is fed into the encoder and the interpretable attention module to obtain the top $I$ tokens with the highest attention weights, which contribute the most to sentiment prediction.",
+ "bbox": [
+ 507,
+ 741,
+ 882,
+ 917
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "8774",
+ "bbox": [
+ 480,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In (3), the sense diversity of each token is computed as the average distance of the token's hidden state to those of its substitution candidates. Then, through (4), top $J$ tokens with the most diverse senses are selected as disambiguation targets, because these target tokens may have very different senses in different contexts. Subsequently, the pretrained lexical substitution model is applied to provide the best substitution to replace each target token, as shown in (5) and (6)a. The new input sentence is passed onto the sentiment analysis model for final prediction ((7)a and (8)a), which is used for the sentiment model backpropagation ((9)a).",
+ "bbox": [
+ 112,
+ 84,
+ 489,
+ 294
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "In order to fine-tune the lexical substitution model to provide better substitutions that improve the accuracy of sentiment analysis, we adapt a dynamic rewarding mechanism (Ge et al., 2022) that was proposed for multi-task learning. As shown in Figure 1 (5) and (6)b, for each target word, we choose the top $N$ candidates. Each of them is seen as a substitute candidate to calculate the probability of correct sentiment prediction after the replacement ((7)b and (8)b). Then, each probability is used to dynamically compute the loss weight when the corresponding candidate is learned by the lexical substitution model as a ground truth ((9)b and (10)b). If the replacement by a candidate leads to a higher chance of predicting the right sentiment, the candidate is more likely given a higher rank by the lexical substitution model. Otherwise, the model would be less likely to select the candidate as the best replacement. The important notations used in this paper are summarized in Table 1. The detailed training process can be seen in Algorithm 1.",
+ "bbox": [
+ 115,
+ 294,
+ 489,
+ 632
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "4.1 Sentiment Analysis with Interpretability",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 642,
+ 477,
+ 657
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Given input sentence $w = (w_{1},\\dots ,w_{L})$ , the goal is to predict the correct sentiment label $\\tilde{y}$ . The input is first fed into a pre-trained encoder:",
+ "bbox": [
+ 112,
+ 662,
+ 487,
+ 709
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nV = \\operatorname {E n c o d e r} (w), \\tag {2}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 226,
+ 721,
+ 487,
+ 738
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $V$ is hidden states. Next, we aim to find the tokens that contribute the most to sentiment inference. We adopt an interpretable attention module called Hierarchical Attention Network (HAN, Han et al., 2022), which effectively encodes hidden states with multiple non-linear projections and ranks the most influential tokens based on attention. Following the setup of Han et al. (2022), we stack two blocks of HAN to form our attention module.",
+ "bbox": [
+ 112,
+ 747,
+ 489,
+ 891
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nq, a = H A N (H A N (V)), \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 200,
+ 902,
+ 487,
+ 920
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "table",
+ "img_path": "images/3f931976f478537e282899b10e81e0924de510f5086401e4ef0f825eba60b565.jpg",
+ "table_caption": [
+ "Table 1: Notation table."
+ ],
+ "table_footnote": [],
+ "table_body": " | Description |
| L | Input sequence length |
| I | Number of words with top ranking attention scores to be potential target words |
| J | Number of words with top ranking diversity scores to be potential target words |
| N | Number of candidates used for finetuning lexical substitution via dynamic rewarding |
| K | Number of candidates for training lexical substitution model |
| w | The original input sentence |
| V | The hidden states of w |
| y | The ground-truth sentiment label of w |
| wf | The final input sentence where ambiguous words in w are substituted |
| ŷf | The predicted sentiment label with wf as input |
| wl | The l-th word in w |
| al | The attention weight of the l-th word in w |
| wiatt | The word in w with the i-th highest attention weight |
| viatt | The hidden state of wiatt |
| M | The number of synonyms of wiatt with the same part-of-speech that are provided by the WordNet |
| G | The hidden states of the M WordNet synonyms |
| gm | The hidden state of the m-th WordNet synonyms of wiatt |
| di | The average Euclidean distance between viatt and g1, . . . , gm, . . . , gM |
| tj | The j-th target word in w to be substituted |
| c | A set of candidate words from WordNet. |
| ˆc | The top N candidates ranked by the lexical substitution model |
| ˆcj,n | The top n-th candidate for replacing tj |
| wj,n | The resulting sentence when replacing tj in w withˆcj,n |
| ˆθj,n | The loss weight whenˆcj,n is treated as the gold standard substitution |
| s | The formatted input to lexical substitution model |
| U | The hidden state matrix of the input sentence in s |
| u1 | The hidden state of the l-th input word in s |
| R | The hidden state matrix of the candidates in s |
| rk | The hidden state of the k-th candidate in s |
| τ | The temperature hyper-parameter for training the lexical substitution model |
| β | The hyper-parameter for balancing the sentiment analysis loss and fine-tuning loss |
| L(sa) | The sentiment analysis loss |
| L(tune) | The loss for finetuning the pretrained language model into a lexical substitution model |
| L(ls) | The lexical substitution loss whenˆcj,n is treated as the gold standard substitution |
| L(ls) | The final weighted lexical substitution loss for backpropagation |
",
+ "bbox": [
+ 510,
+ 107,
+ 878,
+ 743
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where vector $q$ is the yielded hidden state. $a$ is the attention weights, indicating the contribution of each token to the final sentiment prediction. To obtain the sentiment prediction, $q$ is passed on to two layers of feed-forward neural networks (FNN) to obtain the probability distribution of sentiment prediction, with the first one being activated by ReLU (Agarap, 2018), and the second by softmax.",
+ "bbox": [
+ 507,
+ 790,
+ 884,
+ 919
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "8775",
+ "bbox": [
+ 480,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nh = R e L U \\left(F N N _ {1} (q)\\right) \\tag {4}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 196,
+ 112,
+ 487,
+ 129
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {y} = \\operatorname {s o f t m a x} \\left(F N N _ {2} (h)\\right) \\tag {5}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 200,
+ 131,
+ 487,
+ 149
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "We denote the prediction of the sentiment analysis model as $\\hat{y}^f$ when the input is $w^f$ , which denotes a paraphrased $w$ where all selected target tokens are replaced by sense-clarifying substitutions. Thus, the sentiment analysis loss is computed as:",
+ "bbox": [
+ 112,
+ 162,
+ 487,
+ 242
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} ^ {(s a)} = \\operatorname {C r o s s E n t r o p y} (\\hat {y} ^ {f}, \\tilde {y}) \\tag {6}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 181,
+ 253,
+ 487,
+ 272
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.2 Disambiguation by Lexical Substitution",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 294,
+ 470,
+ 311
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "To obtain a lexical substitution model, we fine-tune the pre-trained language model ALM (Mao et al., 2023a)| Setup | Content |
| Original | a delicious and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Paraphrased | a delighted and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Original | Le Lay succeeds Walter G++nter and will be based in Finland. |
| Paraphrased | Le Lay follows Walter G++nter and will be based in Finland. |
| Original | Strong brand visibility nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
| Paraphrased | Strong brand profile nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
",
+ "bbox": [
+ 124,
+ 80,
+ 480,
+ 298
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "6.4 Error Analysis",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 351,
+ 278,
+ 367
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "Finally, we perform a brief error analysis (Table 6). In the first sentence, the model wrongly paraphrases force into power, because it does not recognize the multi-word idiom tour de force. It shows the model's limitation in understanding the semantic meaning of infrequent multi-word idiom. In the second sentence, the word just is wrongly paraphrased as good, which negatively influenced the sentiment prediction. This is mainly due to the part-of-speech tagger wrongly classifying the word just into a adjective, instead of an adverb. Thus, the lexical substitution module might be subjected to error propagation of the part-of-speech tagger and introduce noise into the input. In the last sentence, the word skills is paraphrased as science. It is likely that the model mistakes the word major as a field of study instead of a student studying a field. Thus, the substitution makes sense in the local context the science of calculus major, but not in the whole context, meaning that the model fails to understand the semantic meaning of the sentence.",
+ "bbox": [
+ 112,
+ 375,
+ 489,
+ 712
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "7 Conclusion",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 728,
+ 247,
+ 744
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "In this work, we propose a neurosymbolic framework for sentiment analysis that incorporates the disambiguation of word senses by identifying and paraphrasing ambiguous words in the input, which is realized by leveraging a pre-trained lexical substitution model. The proposed framework not only improves the accuracy of sentiment analysis, but also achieves interpretability by enabling us to trace back which words are paraphrased and what words they are replaced with.",
+ "bbox": [
+ 112,
+ 758,
+ 489,
+ 917
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "table",
+ "img_path": "images/b912546ca4bfdea26d55ee3efc44ad13591b874ce31963783a2b0e27f8070c0e.jpg",
+ "table_caption": [
+ "Table 5: Case study of disambiguation as paraphrasing."
+ ],
+ "table_footnote": [],
+ "table_body": " | Description |
| L | Input sequence length |
| I | Number of words with top ranking attention scores to be potential target words |
| J | Number of words with top ranking diversity scores to be potential target words |
| N | Number of candidates used for finetuning lexical substitution via dynamic rewarding |
| K | Number of candidates for training lexical substitution model |
| w | The original input sentence |
| V | The hidden states of w |
| y | The ground-truth sentiment label of w |
| wf | The final input sentence where ambiguous words in w are substituted |
| ŷf | The predicted sentiment label with wf as input |
| wl | The l-th word in w |
| al | The attention weight of the l-th word in w |
| wiatt | The word in w with the i-th highest attention weight |
| viatt | The hidden state of wiatt |
| M | The number of synonyms of wiatt with the same part-of-speech that are provided by the WordNet |
| G | The hidden states of the M WordNet synonyms |
| gm | The hidden state of the m-th WordNet synonyms of wiatt |
| di | The average Euclidean distance between viatt and g1, . . . , gm, . . . , gM |
| tj | The j-th target word in w to be substituted |
| c | A set of candidate words from WordNet. |
| ˆc | The top N candidates ranked by the lexical substitution model |
| ˆcj,n | The top n-th candidate for replacing tj |
| wj,n | The resulting sentence when replacing tj in w withˆcj,n |
| ˆθj,n | The loss weight whenˆcj,n is treated as the gold standard substitution |
| s | The formatted input to lexical substitution model |
| U | The hidden state matrix of the input sentence in s |
| u1 | The hidden state of the l-th input word in s |
| R | The hidden state matrix of the candidates in s |
| rk | The hidden state of the k-th candidate in s |
| τ | The temperature hyper-parameter for training the lexical substitution model |
| β | The hyper-parameter for balancing the sentiment analysis loss and fine-tuning loss |
| L(sa) | The sentiment analysis loss |
| L(tune) | The loss for finetuning the pretrained language model into a lexical substitution model |
| L(ls) | The lexical substitution loss whenˆcj,n is treated as the gold standard substitution |
| L(ls) | The final weighted lexical substitution loss for backpropagation |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.791,
+ 0.885,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "where vector \\( q \\) is the yielded hidden state. \\( a \\) is the attention weights, indicating the contribution of each token to the final sentiment prediction. To obtain the sentiment prediction, \\( q \\) is passed on to two layers of feed-forward neural networks (FNN) to obtain the probability distribution of sentiment prediction, with the first one being activated by ReLU (Agarap, 2018), and the second by softmax."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "8775"
+ }
+ ],
+ [
+ {
+ "type": "equation",
+ "bbox": [
+ 0.198,
+ 0.113,
+ 0.488,
+ 0.13
+ ],
+ "angle": 0,
+ "content": "\\[\nh = R e L U \\left(F N N _ {1} (q)\\right) \\tag {4}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.2,
+ 0.133,
+ 0.488,
+ 0.151
+ ],
+ "angle": 0,
+ "content": "\\[\n\\hat {y} = \\operatorname {s o f t m a x} \\left(F N N _ {2} (h)\\right) \\tag {5}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.163,
+ 0.489,
+ 0.243
+ ],
+ "angle": 0,
+ "content": "We denote the prediction of the sentiment analysis model as \\(\\hat{y}^f\\) when the input is \\(w^f\\), which denotes a paraphrased \\(w\\) where all selected target tokens are replaced by sense-clarifying substitutions. Thus, the sentiment analysis loss is computed as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.182,
+ 0.254,
+ 0.488,
+ 0.273
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} ^ {(s a)} = \\operatorname {C r o s s E n t r o p y} (\\hat {y} ^ {f}, \\tilde {y}) \\tag {6}\n\\]"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.296,
+ 0.472,
+ 0.312
+ ],
+ "angle": 0,
+ "content": "4.2 Disambiguation by Lexical Substitution"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.316,
+ 0.489,
+ 0.445
+ ],
+ "angle": 0,
+ "content": "To obtain a lexical substitution model, we fine-tune the pre-trained language model ALM (Mao et al., 2023a)| Setup | Content |
| Original | a delicious and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Paraphrased | a delighted and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Original | Le Lay succeeds Walter G++nter and will be based in Finland. |
| Paraphrased | Le Lay follows Walter G++nter and will be based in Finland. |
| Original | Strong brand visibility nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
| Paraphrased | Strong brand profile nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.115,
+ 0.309,
+ 0.486,
+ 0.325
+ ],
+ "angle": 0,
+ "content": "Table 5: Case study of disambiguation as paraphrasing."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.352,
+ 0.279,
+ 0.368
+ ],
+ "angle": 0,
+ "content": "6.4 Error Analysis"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.376,
+ 0.49,
+ 0.713
+ ],
+ "angle": 0,
+ "content": "Finally, we perform a brief error analysis (Table 6). In the first sentence, the model wrongly paraphrases force into power, because it does not recognize the multi-word idiom tour de force. It shows the model's limitation in understanding the semantic meaning of infrequent multi-word idiom. In the second sentence, the word just is wrongly paraphrased as good, which negatively influenced the sentiment prediction. This is mainly due to the part-of-speech tagger wrongly classifying the word just into a adjective, instead of an adverb. Thus, the lexical substitution module might be subjected to error propagation of the part-of-speech tagger and introduce noise into the input. In the last sentence, the word skills is paraphrased as science. It is likely that the model mistakes the word major as a field of study instead of a student studying a field. Thus, the substitution makes sense in the local context the science of calculus major, but not in the whole context, meaning that the model fails to understand the semantic meaning of the sentence."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.73,
+ 0.248,
+ 0.745
+ ],
+ "angle": 0,
+ "content": "7 Conclusion"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.759,
+ 0.49,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "In this work, we propose a neurosymbolic framework for sentiment analysis that incorporates the disambiguation of word senses by identifying and paraphrasing ambiguous words in the input, which is realized by leveraging a pre-trained lexical substitution model. The proposed framework not only improves the accuracy of sentiment analysis, but also achieves interpretability by enabling us to trace back which words are paraphrased and what words they are replaced with."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.518,
+ 0.082,
+ 0.878,
+ 0.299
+ ],
+ "angle": 0,
+ "content": " | Description |
| L | Input sequence length |
| I | Number of words with top ranking attention scores to be potential target words |
| J | Number of words with top ranking diversity scores to be potential target words |
| N | Number of candidates used for finetuning lexical substitution via dynamic rewarding |
| K | Number of candidates for training lexical substitution model |
| w | The original input sentence |
| V | The hidden states of w |
| y | The ground-truth sentiment label of w |
| wf | The final input sentence where ambiguous words in w are substituted |
| ŷf | The predicted sentiment label with wf as input |
| wl | The l-th word in w |
| al | The attention weight of the l-th word in w |
| wiatt | The word in w with the i-th highest attention weight |
| viatt | The hidden state of wiatt |
| M | The number of synonyms of wiatt with the same part-of-speech that are provided by the WordNet |
| G | The hidden states of the M WordNet synonyms |
| gm | The hidden state of the m-th WordNet synonyms of wiatt |
| di | The average Euclidean distance between viatt and g1, . . . , gm, . . . , gM |
| tj | The j-th target word in w to be substituted |
| c | A set of candidate words from WordNet. |
| ˆc | The top N candidates ranked by the lexical substitution model |
| ˆcj,n | The top n-th candidate for replacing tj |
| wj,n | The resulting sentence when replacing tj in w withˆcj,n |
| ˆθj,n | The loss weight whenˆcj,n is treated as the gold standard substitution |
| s | The formatted input to lexical substitution model |
| U | The hidden state matrix of the input sentence in s |
| u1 | The hidden state of the l-th input word in s |
| R | The hidden state matrix of the candidates in s |
| rk | The hidden state of the k-th candidate in s |
| τ | The temperature hyper-parameter for training the lexical substitution model |
| β | The hyper-parameter for balancing the sentiment analysis loss and fine-tuning loss |
| L(sa) | The sentiment analysis loss |
| L(tune) | The loss for finetuning the pretrained language model into a lexical substitution model |
| L(ls) | The lexical substitution loss whenˆcj,n is treated as the gold standard substitution |
| L(ls) | The final weighted lexical substitution loss for backpropagation |
+
+where vector $q$ is the yielded hidden state. $a$ is the attention weights, indicating the contribution of each token to the final sentiment prediction. To obtain the sentiment prediction, $q$ is passed on to two layers of feed-forward neural networks (FNN) to obtain the probability distribution of sentiment prediction, with the first one being activated by ReLU (Agarap, 2018), and the second by softmax.
+
+$$
+h = R e L U \left(F N N _ {1} (q)\right) \tag {4}
+$$
+
+$$
+\hat {y} = \operatorname {s o f t m a x} \left(F N N _ {2} (h)\right) \tag {5}
+$$
+
+We denote the prediction of the sentiment analysis model as $\hat{y}^f$ when the input is $w^f$ , which denotes a paraphrased $w$ where all selected target tokens are replaced by sense-clarifying substitutions. Thus, the sentiment analysis loss is computed as:
+
+$$
+\mathcal {L} ^ {(s a)} = \operatorname {C r o s s E n t r o p y} (\hat {y} ^ {f}, \tilde {y}) \tag {6}
+$$
+
+# 4.2 Disambiguation by Lexical Substitution
+
+To obtain a lexical substitution model, we fine-tune the pre-trained language model ALM (Mao et al., 2023a)| Setup | Content |
| Original | a delicious and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Paraphrased | a delighted and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Original | Le Lay succeeds Walter G++nter and will be based in Finland. |
| Paraphrased | Le Lay follows Walter G++nter and will be based in Finland. |
| Original | Strong brand visibility nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
| Paraphrased | Strong brand profile nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
+
+# 6.4 Error Analysis
+
+Finally, we perform a brief error analysis (Table 6). In the first sentence, the model wrongly paraphrases force into power, because it does not recognize the multi-word idiom tour de force. It shows the model's limitation in understanding the semantic meaning of infrequent multi-word idiom. In the second sentence, the word just is wrongly paraphrased as good, which negatively influenced the sentiment prediction. This is mainly due to the part-of-speech tagger wrongly classifying the word just into a adjective, instead of an adverb. Thus, the lexical substitution module might be subjected to error propagation of the part-of-speech tagger and introduce noise into the input. In the last sentence, the word skills is paraphrased as science. It is likely that the model mistakes the word major as a field of study instead of a student studying a field. Thus, the substitution makes sense in the local context the science of calculus major, but not in the whole context, meaning that the model fails to understand the semantic meaning of the sentence.
+
+# 7 Conclusion
+
+In this work, we propose a neurosymbolic framework for sentiment analysis that incorporates the disambiguation of word senses by identifying and paraphrasing ambiguous words in the input, which is realized by leveraging a pre-trained lexical substitution model. The proposed framework not only improves the accuracy of sentiment analysis, but also achieves interpretability by enabling us to trace back which words are paraphrased and what words they are replaced with.
+
+Table 5: Case study of disambiguation as paraphrasing.
+
+ | Description |
| L | Input sequence length |
| I | Number of words with top ranking attention scores to be potential target words |
| J | Number of words with top ranking diversity scores to be potential target words |
| N | Number of candidates used for finetuning lexical substitution via dynamic rewarding |
| K | Number of candidates for training lexical substitution model |
| w | The original input sentence |
| V | The hidden states of w |
| y | The ground-truth sentiment label of w |
| wf | The final input sentence where ambiguous words in w are substituted |
| ŷf | The predicted sentiment label with wf as input |
| wl | The l-th word in w |
| al | The attention weight of the l-th word in w |
| wiatt | The word in w with the i-th highest attention weight |
| viatt | The hidden state of wiatt |
| M | The number of synonyms of wiatt with the same part-of-speech that are provided by the WordNet |
| G | The hidden states of the M WordNet synonyms |
| gm | The hidden state of the m-th WordNet synonyms of wiatt |
| di | The average Euclidean distance between viatt and g1, . . . , gm, . . . , gM |
| tj | The j-th target word in w to be substituted |
| c | A set of candidate words from WordNet. |
| ˆc | The top N candidates ranked by the lexical substitution model |
| ˆcj,n | The top n-th candidate for replacing tj |
| wj,n | The resulting sentence when replacing tj in w withˆcj,n |
| ˆθj,n | The loss weight whenˆcj,n is treated as the gold standard substitution |
| s | The formatted input to lexical substitution model |
| U | The hidden state matrix of the input sentence in s |
| u1 | The hidden state of the l-th input word in s |
| R | The hidden state matrix of the candidates in s |
| rk | The hidden state of the k-th candidate in s |
| τ | The temperature hyper-parameter for training the lexical substitution model |
| β | The hyper-parameter for balancing the sentiment analysis loss and fine-tuning loss |
| L(sa) | The sentiment analysis loss |
| L(tune) | The loss for finetuning the pretrained language model into a lexical substitution model |
| L(ls) | The lexical substitution loss whenˆcj,n is treated as the gold standard substitution |
| L(ls) | The final weighted lexical substitution loss for backpropagation |
",
+ "image_path": "3f931976f478537e282899b10e81e0924de510f5086401e4ef0f825eba60b565.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "text",
+ "content": "where vector "
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "text",
+ "content": " is the yielded hidden state. "
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "a"
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "text",
+ "content": " is the attention weights, indicating the contribution of each token to the final sentiment prediction. To obtain the sentiment prediction, "
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 526,
+ 773
+ ],
+ "type": "text",
+ "content": " is passed on to two layers of feed-forward neural networks (FNN) to obtain the probability distribution of sentiment prediction, with the first one being activated by ReLU (Agarap, 2018), and the second by softmax."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "8775"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 117,
+ 95,
+ 290,
+ 109
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 95,
+ 290,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 95,
+ 290,
+ 109
+ ],
+ "type": "interline_equation",
+ "content": "h = R e L U \\left(F N N _ {1} (q)\\right) \\tag {4}",
+ "image_path": "8307447682ac033ae5ef673b9ca47ab9f06ac8000219ffff2ddb8f6d4d674765.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 119,
+ 111,
+ 290,
+ 126
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 111,
+ 290,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 111,
+ 290,
+ 126
+ ],
+ "type": "interline_equation",
+ "content": "\\hat {y} = \\operatorname {s o f t m a x} \\left(F N N _ {2} (h)\\right) \\tag {5}",
+ "image_path": "4b5c44deab9a4dccc51f741fd654624046db80d8c22dd1c2a3f6b100656f6bba.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "text",
+ "content": "We denote the prediction of the sentiment analysis model as "
+ },
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{y}^f"
+ },
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "text",
+ "content": " when the input is "
+ },
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "inline_equation",
+ "content": "w^f"
+ },
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "text",
+ "content": ", which denotes a paraphrased "
+ },
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "inline_equation",
+ "content": "w"
+ },
+ {
+ "bbox": [
+ 67,
+ 137,
+ 290,
+ 204
+ ],
+ "type": "text",
+ "content": " where all selected target tokens are replaced by sense-clarifying substitutions. Thus, the sentiment analysis loss is computed as:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 108,
+ 213,
+ 290,
+ 229
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 213,
+ 290,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 213,
+ 290,
+ 229
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} ^ {(s a)} = \\operatorname {C r o s s E n t r o p y} (\\hat {y} ^ {f}, \\tilde {y}) \\tag {6}",
+ "image_path": "69cef144537dc4bfa71704dcfd85f94311e023a3d3ae1ccd1efd40923260e9c6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 248,
+ 280,
+ 262
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 248,
+ 280,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 248,
+ 280,
+ 262
+ ],
+ "type": "text",
+ "content": "4.2 Disambiguation by Lexical Substitution"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 265,
+ 290,
+ 374
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 265,
+ 290,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 265,
+ 290,
+ 374
+ ],
+ "type": "text",
+ "content": "To obtain a lexical substitution model, we fine-tune the pre-trained language model ALM (Mao et al., 2023a)| Setup | Content |
| Original | a delicious and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Paraphrased | a delighted and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them. |
| Original | Le Lay succeeds Walter G++nter and will be based in Finland. |
| Paraphrased | Le Lay follows Walter G++nter and will be based in Finland. |
| Original | Strong brand visibility nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
| Paraphrased | Strong brand profile nationally and regionally is of primary importance in home sales, vehicle and consumer advertising. |
",
+ "image_path": "ed434bca90666544c47056017d90d1f907c0500ad2e4c8fc856b66587fc576c6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 296,
+ 166,
+ 309
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 296,
+ 166,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 296,
+ 166,
+ 309
+ ],
+ "type": "text",
+ "content": "6.4 Error Analysis"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 316,
+ 291,
+ 599
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 316,
+ 291,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 316,
+ 291,
+ 599
+ ],
+ "type": "text",
+ "content": "Finally, we perform a brief error analysis (Table 6). In the first sentence, the model wrongly paraphrases force into power, because it does not recognize the multi-word idiom tour de force. It shows the model's limitation in understanding the semantic meaning of infrequent multi-word idiom. In the second sentence, the word just is wrongly paraphrased as good, which negatively influenced the sentiment prediction. This is mainly due to the part-of-speech tagger wrongly classifying the word just into a adjective, instead of an adverb. Thus, the lexical substitution module might be subjected to error propagation of the part-of-speech tagger and introduce noise into the input. In the last sentence, the word skills is paraphrased as science. It is likely that the model mistakes the word major as a field of study instead of a student studying a field. Thus, the substitution makes sense in the local context the science of calculus major, but not in the whole context, meaning that the model fails to understand the semantic meaning of the sentence."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 68,
+ 613,
+ 147,
+ 626
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 613,
+ 147,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 613,
+ 147,
+ 626
+ ],
+ "type": "text",
+ "content": "7 Conclusion"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 638,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 638,
+ 291,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 638,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": "In this work, we propose a neurosymbolic framework for sentiment analysis that incorporates the disambiguation of word senses by identifying and paraphrasing ambiguous words in the input, which is realized by leveraging a pre-trained lexical substitution model. The proposed framework not only improves the accuracy of sentiment analysis, but also achieves interpretability by enabling us to trace back which words are paraphrased and what words they are replaced with."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 308,
+ 68,
+ 522,
+ 251
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 68,
+ 259,
+ 289,
+ 273
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 259,
+ 289,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 259,
+ 289,
+ 273
+ ],
+ "type": "text",
+ "content": "Table 5: Case study of disambiguation as paraphrasing."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 308,
+ 68,
+ 522,
+ 251
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 68,
+ 522,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 68,
+ 522,
+ 251
+ ],
+ "type": "table",
+ "html": "| Method | Prompt Template |
| Zero-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag, you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
| Few-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag,you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible.Here are some in-context examples:(1), <input> SRC-1 </input>: <output>TGT-1 </output>(2), <input> SRC-1 </input>: <output>TGT-1 </output>; ... (n), <input> SRC-n </input>: <output>TGT-n </output>; Please feel free to refer to these examples. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
",
+ "bbox": [
+ 510,
+ 250,
+ 890,
+ 860
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Table 8: Zero-shot and few-shot prompt template for Chinese-English code-switching ASR error correction.",
+ "bbox": [
+ 507,
+ 869,
+ 882,
+ 897
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "page_number",
+ "text": "8086",
+ "bbox": [
+ 480,
+ 928,
+ 521,
+ 940
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "E Rule-based Method",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 83,
+ 321,
+ 98
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "The details of the five rules adopted in our proposed rule-based method are introduced in the following.",
+ "bbox": [
+ 112,
+ 109,
+ 487,
+ 140
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Delete. Randomly delete a token with a probability.",
+ "bbox": [
+ 112,
+ 141,
+ 487,
+ 172
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Add. Firstly, randomly select a word or Chinese character from the vocabulary of the ASR datasets, and then add the selected word or character to a random position.",
+ "bbox": [
+ 112,
+ 174,
+ 487,
+ 237
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Replace. Randomly replace a token with a word or Chinese character from the vocabulary of the ASR datasets. In particular, if the selected token is a Chinese character, we select the replacement from the homophones with a probability.",
+ "bbox": [
+ 112,
+ 239,
+ 487,
+ 317
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Shuffle. Shuffle the tokens by adding a normal distribution bias to the positions of the tokens.",
+ "bbox": [
+ 112,
+ 319,
+ 487,
+ 349
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Spell Error. Randomly apply spell error to an English word. We randomly perturb letters with operations, i.e. substitution, deletion, insertion or transposition of letters.",
+ "bbox": [
+ 112,
+ 351,
+ 487,
+ 414
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "F Prompt Template",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 426,
+ 302,
+ 443
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Zero-shot and few-shot prompt templates for Chinese-English code-switching ASR error correction are shown in Table 8.",
+ "bbox": [
+ 112,
+ 451,
+ 489,
+ 498
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "G Additional Experiment Results",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 511,
+ 421,
+ 527
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Results of the mT5 correction model on SEAME-C, when trained with outputs collected from different iterations of our method, are shown in Table 9.",
+ "bbox": [
+ 112,
+ 537,
+ 487,
+ 583
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "table",
+ "img_path": "images/5877df100337da3038d691ac1656f44416084fbea8bd03360c13266b5ab6e907.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Method | Prompt Template |
| Zero-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag, you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
| Few-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag,you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible.Here are some in-context examples:(1), <input> SRC-1 </input>: <output>TGT-1 </output>(2), <input> SRC-1 </input>: <output>TGT-1 </output>; ... (n), <input> SRC-n </input>: <output>TGT-n </output>; Please feel free to refer to these examples. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.509,
+ 0.87,
+ 0.884,
+ 0.898
+ ],
+ "angle": 0,
+ "content": "Table 8: Zero-shot and few-shot prompt template for Chinese-English code-switching ASR error correction."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.929,
+ 0.522,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "8086"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.084,
+ 0.322,
+ 0.099
+ ],
+ "angle": 0,
+ "content": "E Rule-based Method"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.11,
+ 0.488,
+ 0.141
+ ],
+ "angle": 0,
+ "content": "The details of the five rules adopted in our proposed rule-based method are introduced in the following."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.142,
+ 0.489,
+ 0.173
+ ],
+ "angle": 0,
+ "content": "Delete. Randomly delete a token with a probability."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.175,
+ 0.489,
+ 0.238
+ ],
+ "angle": 0,
+ "content": "Add. Firstly, randomly select a word or Chinese character from the vocabulary of the ASR datasets, and then add the selected word or character to a random position."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.24,
+ 0.489,
+ 0.318
+ ],
+ "angle": 0,
+ "content": "Replace. Randomly replace a token with a word or Chinese character from the vocabulary of the ASR datasets. In particular, if the selected token is a Chinese character, we select the replacement from the homophones with a probability."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.32,
+ 0.488,
+ 0.35
+ ],
+ "angle": 0,
+ "content": "Shuffle. Shuffle the tokens by adding a normal distribution bias to the positions of the tokens."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.352,
+ 0.489,
+ 0.415
+ ],
+ "angle": 0,
+ "content": "Spell Error. Randomly apply spell error to an English word. We randomly perturb letters with operations, i.e. substitution, deletion, insertion or transposition of letters."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.427,
+ 0.304,
+ 0.444
+ ],
+ "angle": 0,
+ "content": "F Prompt Template"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.453,
+ 0.49,
+ 0.499
+ ],
+ "angle": 0,
+ "content": "Zero-shot and few-shot prompt templates for Chinese-English code-switching ASR error correction are shown in Table 8."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.512,
+ 0.422,
+ 0.529
+ ],
+ "angle": 0,
+ "content": "G Additional Experiment Results"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.538,
+ 0.489,
+ 0.585
+ ],
+ "angle": 0,
+ "content": "Results of the mT5 correction model on SEAME-C, when trained with outputs collected from different iterations of our method, are shown in Table 9."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.125,
+ 0.595,
+ 0.481,
+ 0.684
+ ],
+ "angle": 0,
+ "content": "| Method | Prompt Template |
| Zero-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag, you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
| Few-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag,you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible.Here are some in-context examples:(1), <input> SRC-1 </input>: <output>TGT-1 </output>(2), <input> SRC-1 </input>: <output>TGT-1 </output>; ... (n), <input> SRC-n </input>: <output>TGT-n </output>; Please feel free to refer to these examples. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
+
+Table 8: Zero-shot and few-shot prompt template for Chinese-English code-switching ASR error correction.
+
+# E Rule-based Method
+
+The details of the five rules adopted in our proposed rule-based method are introduced in the following.
+
+Delete. Randomly delete a token with a probability.
+
+Add. Firstly, randomly select a word or Chinese character from the vocabulary of the ASR datasets, and then add the selected word or character to a random position.
+
+Replace. Randomly replace a token with a word or Chinese character from the vocabulary of the ASR datasets. In particular, if the selected token is a Chinese character, we select the replacement from the homophones with a probability.
+
+Shuffle. Shuffle the tokens by adding a normal distribution bias to the positions of the tokens.
+
+Spell Error. Randomly apply spell error to an English word. We randomly perturb letters with operations, i.e. substitution, deletion, insertion or transposition of letters.
+
+# F Prompt Template
+
+Zero-shot and few-shot prompt templates for Chinese-English code-switching ASR error correction are shown in Table 8.
+
+# G Additional Experiment Results
+
+Results of the mT5 correction model on SEAME-C, when trained with outputs collected from different iterations of our method, are shown in Table 9.
+
+| Method | Prompt Template |
| Zero-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag, you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
| Few-shot | "role": "system", "content": "You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence." "role": "user", "content": "Please identify and correct any errors in the following sentence indicated by <input>ERROR </input> tag,you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible.Here are some in-context examples:(1), <input> SRC-1 </input>: <output>TGT-1 </output>(2), <input> SRC-1 </input>: <output>TGT-1 </output>; ... (n), <input> SRC-n </input>: <output>TGT-n </output>; Please feel free to refer to these examples. Remember to format your corrected output results with the tag <output> Your Corrected Version </output>. Please start: <input>Input Sentence </input>:" |
",
+ "image_path": "d971f9b7833134aea161f655e045083925637fd550f0e430ab8f01dc00179331.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 731,
+ 525,
+ 755
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 731,
+ 525,
+ 755
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 731,
+ 525,
+ 755
+ ],
+ "type": "text",
+ "content": "Table 8: Zero-shot and few-shot prompt template for Chinese-English code-switching ASR error correction."
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 310,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 310,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 310,
+ 791
+ ],
+ "type": "text",
+ "content": "8086"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 11
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 68,
+ 70,
+ 191,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 70,
+ 191,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 70,
+ 191,
+ 83
+ ],
+ "type": "text",
+ "content": "E Rule-based Method"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 92,
+ 290,
+ 118
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 92,
+ 290,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 92,
+ 290,
+ 118
+ ],
+ "type": "text",
+ "content": "The details of the five rules adopted in our proposed rule-based method are introduced in the following."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 67,
+ 119,
+ 290,
+ 145
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 119,
+ 290,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 119,
+ 290,
+ 145
+ ],
+ "type": "text",
+ "content": "Delete. Randomly delete a token with a probability."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 147,
+ 290,
+ 200
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 147,
+ 290,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 147,
+ 290,
+ 200
+ ],
+ "type": "text",
+ "content": "Add. Firstly, randomly select a word or Chinese character from the vocabulary of the ASR datasets, and then add the selected word or character to a random position."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 201,
+ 290,
+ 267
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 201,
+ 290,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 201,
+ 290,
+ 267
+ ],
+ "type": "text",
+ "content": "Replace. Randomly replace a token with a word or Chinese character from the vocabulary of the ASR datasets. In particular, if the selected token is a Chinese character, we select the replacement from the homophones with a probability."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 269,
+ 290,
+ 294
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 269,
+ 290,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 269,
+ 290,
+ 294
+ ],
+ "type": "text",
+ "content": "Shuffle. Shuffle the tokens by adding a normal distribution bias to the positions of the tokens."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 67,
+ 296,
+ 290,
+ 349
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 296,
+ 290,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 296,
+ 290,
+ 349
+ ],
+ "type": "text",
+ "content": "Spell Error. Randomly apply spell error to an English word. We randomly perturb letters with operations, i.e. substitution, deletion, insertion or transposition of letters."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 68,
+ 359,
+ 180,
+ 373
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 359,
+ 180,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 359,
+ 180,
+ 373
+ ],
+ "type": "text",
+ "content": "F Prompt Template"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 67,
+ 380,
+ 291,
+ 419
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 380,
+ 291,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 380,
+ 291,
+ 419
+ ],
+ "type": "text",
+ "content": "Zero-shot and few-shot prompt templates for Chinese-English code-switching ASR error correction are shown in Table 8."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 68,
+ 430,
+ 251,
+ 444
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 430,
+ 251,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 430,
+ 251,
+ 444
+ ],
+ "type": "text",
+ "content": "G Additional Experiment Results"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 67,
+ 452,
+ 290,
+ 491
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 452,
+ 290,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 452,
+ 290,
+ 491
+ ],
+ "type": "text",
+ "content": "Results of the mT5 correction model on SEAME-C, when trained with outputs collected from different iterations of our method, are shown in Table 9."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 74,
+ 500,
+ 286,
+ 575
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 74,
+ 500,
+ 286,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 74,
+ 500,
+ 286,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 74,
+ 500,
+ 286,
+ 575
+ ],
+ "type": "table",
+ "html": "| Methods | StaQC | SO-DS |
| R@3 | R@10 | MRR | R@3 | R@10 | MRR |
| BM25 desc (Heyman and Van Cutsem, 2020) | 8.0 | 13.3 | 7.5 | 23.8 | 32.3 | 21.6 |
| NBOW (Heyman and Van Cutsem, 2020) | 10.9 | 16.6 | 9.5 | 27.7 | 38.0 | 24.7 |
| USE (Heyman and Van Cutsem, 2020) | 12.8 | 20.3 | 11.7 | 33.3 | 48.5 | 30.4 |
| CodeBERT (Feng et al., 2020) | - | - | 23.4 | - | - | 23.1 |
| GraphCodeBERT (Guo et al., 2021) | - | - | 24.1 | - | - | 25.2 |
| CodeRetriever (In-Batch Negative) (Li et al., 2022) | - | - | 25.5 | - | - | 27.1 |
| CodeRetriever (Hard Negative) (Li et al., 2022) | - | - | 24.6 | - | - | 31.8 |
| UniXcoder (In-Batch Negative) (Guo et al., 2022) | 29.98 | 47.47 | 28.04 | 31.90 | 51.21 | 28.29 |
| UniXcoder (Hard Negative) (Guo et al., 2022) | 31.18 | 48.38 | 28.63 | 33.42 | 53.37 | 29.97 |
| NPC (In-Batch Negative) | 33.07 | 50.35 | 30.39 | 35.58 | 54.54 | 30.96 |
| NPC (Hard Negative) | 34.38 | 52.20 | 31.36 | 38.00 | 56.51 | 32.49 |
",
+ "bbox": [
+ 137,
+ 80,
+ 858,
+ 296
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Table 1: Retrieval performance on StaQC and SO-DS, which are realistic-noisy datasets. The results of the first block are mainly borrowed from published papers (Heyman and Van Cutsem, 2020; Li et al., 2022). If the results are not provided, we mark them as “-”.",
+ "bbox": [
+ 112,
+ 306,
+ 882,
+ 350
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "been widely used on code summarization (Ped-damail et al., 2018) and code search (Heyman and Van Cutsem, 2020). SO-DS mines query-code pairs from the most upvoted Stack Overflow posts, mainly focuses on the data science domain. Following previous works (Heyman and Van Cutsem, 2020; Li et al., 2022), we resort to Recall of topk $(\\mathbb{R}@\\mathbb{k})$ and Mean Reciprocal Rank (MRR) as the evaluation metric. StaQC and SO-DS are constructed automatically without human annotation. Therefore, there are numerous mismatched pairs in training data.",
+ "bbox": [
+ 110,
+ 367,
+ 489,
+ 558
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Natural Questions (NQ) collects real queries from the Google search engine. Each question is paired with an answer span and golden passages from the Wikipedia pages. Trivia QA (TQ) is a reading comprehension dataset authored by trivia enthusiasts. During the retrieval stage of both datasets, the objective is to identify positive passages from a large collection. Positive pairs in these datasets are assessed based on strict rule, i.e., whether passages contain answers or not (Karpukhin et al., 2020a). Consequently, we consider these datasets to be of high quality. Thus, we leverage them for simulation experiments to quantitatively analyze the impact of varying proportions of noise. Drawing inspiration from the setting in the noisy classification task (Han et al., 2018), we simulate the mismatched-pair noise by randomly pairing queries with unrelated documents.",
+ "bbox": [
+ 110,
+ 561,
+ 489,
+ 851
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.2 Implementation Details",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 864,
+ 346,
+ 879
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "NPC is a general training paradigm that can be directly applied to almost all retrieval models. For",
+ "bbox": [
+ 112,
+ 887,
+ 489,
+ 917
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "StaQC and SO-DS, we adopt UniXcoder (Guo et al., 2022) as our backbone, which is the SoTA model for code representation. Following Guo et al. (2022), we adopt the cosine distance as a similarity function and set temperature $\\lambda$ to 20. We update model parameters using the Adam optimizer and perform early stopping on the development set. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 256, 5, and 10, respectively. In the \"Hard Negative\" setting, we adopt the same strategy as Li et al. (2022).",
+ "bbox": [
+ 507,
+ 367,
+ 884,
+ 543
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "For NQ and TQ, we adopt BERT (Devlin et al., 2019) as our initial model. Following Karpukhin et al. (2020a), we adopt inner-product as the similarity function and set temperature $\\lambda$ to 1. The max sequence length is 16 for query and 128 for passage. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 512, 10, and 40, respectively. We adopt \"BM25 Negative\" and \"Hard Negative\" strategies as described in the DPR toolkit. For a fair comparison, we implement DPR (Karpukhin et al., 2020a) with the same hyperparameters. All experiments are run on 8 NVIDIA Tesla A100 GPUs. The implementation of NPC is based on Huggingface (Wolf et al., 2020).",
+ "bbox": [
+ 507,
+ 545,
+ 884,
+ 770
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.3 Results",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 781,
+ 613,
+ 796
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Results on StaQC and SO-DS: Table 1 shows the results on the realistic-noisy datasets StaQC and SO-DS. Both datasets contain a large number of real noise pairs. The first block shows the results of previous SoTA methods. BM25 desc is a traditional",
+ "bbox": [
+ 507,
+ 802,
+ 882,
+ 883
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_footnote",
+ "text": "$\\dagger$ https://github.com/facebookresearch/DPR",
+ "bbox": [
+ 509,
+ 891,
+ 867,
+ 917
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "11443",
+ "bbox": [
+ 475,
+ 927,
+ 524,
+ 940
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "table",
+ "img_path": "images/508d1a78b29f8358f760df53799576b37b81fd7f214792d73f9106efa010ff5a.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": " | Question | Code |
| 1 | Split words in a nested list into letters | » [list(l[0]) for l in mylist] |
| 2 | Dictionary in python problem | » s = problem.getSuccessors( getStartState()) |
| 3 | Find the Common first name from Django Auth user Model | » import operator |
| 4 | Find all text files not containing some text string | » lst = [1,2,4,6,3,8,0,5] \n» for n in lst[:]: \n» if n % 2 == 0: \n»>>> lst.remove(n) \n» lst |
",
+ "bbox": [
+ 114,
+ 395,
+ 882,
+ 576
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Table 9: Some noisy pairs detected by NPC in StaQC training set.",
+ "bbox": [
+ 273,
+ 585,
+ 722,
+ 600
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "page_number",
+ "text": "11451",
+ "bbox": [
+ 477,
+ 928,
+ 522,
+ 940
+ ],
+ "page_idx": 12
+ }
+]
\ No newline at end of file
diff --git a/2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_model.json b/2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba12f38e9b6de3d3bde1d2f8aaf914f31ce871c4
--- /dev/null
+++ b/2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_model.json
@@ -0,0 +1,2789 @@
+[
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.286,
+ 0.091,
+ 0.713,
+ 0.111
+ ],
+ "angle": 0,
+ "content": "Noisy Pair Corrector for Dense Retrieval"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.242,
+ 0.125,
+ 0.755,
+ 0.159
+ ],
+ "angle": 0,
+ "content": "Hang Zhang\\(^{1}\\), Yeyun Gong, Xingwei He\\(^{†2}\\), Dayiheng Liu\\(^{1}\\), Daya Guo\\(^{3}\\), Jiancheng Lv\\(^{†1}\\), Jian Guo\\(^{†4}\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.295,
+ 0.159,
+ 0.71,
+ 0.175
+ ],
+ "angle": 0,
+ "content": "| Methods | StaQC | SO-DS |
| R@3 | R@10 | MRR | R@3 | R@10 | MRR |
| BM25 desc (Heyman and Van Cutsem, 2020) | 8.0 | 13.3 | 7.5 | 23.8 | 32.3 | 21.6 |
| NBOW (Heyman and Van Cutsem, 2020) | 10.9 | 16.6 | 9.5 | 27.7 | 38.0 | 24.7 |
| USE (Heyman and Van Cutsem, 2020) | 12.8 | 20.3 | 11.7 | 33.3 | 48.5 | 30.4 |
| CodeBERT (Feng et al., 2020) | - | - | 23.4 | - | - | 23.1 |
| GraphCodeBERT (Guo et al., 2021) | - | - | 24.1 | - | - | 25.2 |
| CodeRetriever (In-Batch Negative) (Li et al., 2022) | - | - | 25.5 | - | - | 27.1 |
| CodeRetriever (Hard Negative) (Li et al., 2022) | - | - | 24.6 | - | - | 31.8 |
| UniXcoder (In-Batch Negative) (Guo et al., 2022) | 29.98 | 47.47 | 28.04 | 31.90 | 51.21 | 28.29 |
| UniXcoder (Hard Negative) (Guo et al., 2022) | 31.18 | 48.38 | 28.63 | 33.42 | 53.37 | 29.97 |
| NPC (In-Batch Negative) | 33.07 | 50.35 | 30.39 | 35.58 | 54.54 | 30.96 |
| NPC (Hard Negative) | 34.38 | 52.20 | 31.36 | 38.00 | 56.51 | 32.49 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.113,
+ 0.307,
+ 0.884,
+ 0.351
+ ],
+ "angle": 0,
+ "content": "Table 1: Retrieval performance on StaQC and SO-DS, which are realistic-noisy datasets. The results of the first block are mainly borrowed from published papers (Heyman and Van Cutsem, 2020; Li et al., 2022). If the results are not provided, we mark them as “-”."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.112,
+ 0.368,
+ 0.49,
+ 0.56
+ ],
+ "angle": 0,
+ "content": "been widely used on code summarization (Ped-damail et al., 2018) and code search (Heyman and Van Cutsem, 2020). SO-DS mines query-code pairs from the most upvoted Stack Overflow posts, mainly focuses on the data science domain. Following previous works (Heyman and Van Cutsem, 2020; Li et al., 2022), we resort to Recall of topk \\((\\mathbb{R}@\\mathbb{k})\\) and Mean Reciprocal Rank (MRR) as the evaluation metric. StaQC and SO-DS are constructed automatically without human annotation. Therefore, there are numerous mismatched pairs in training data."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.112,
+ 0.562,
+ 0.49,
+ 0.852
+ ],
+ "angle": 0,
+ "content": "Natural Questions (NQ) collects real queries from the Google search engine. Each question is paired with an answer span and golden passages from the Wikipedia pages. Trivia QA (TQ) is a reading comprehension dataset authored by trivia enthusiasts. During the retrieval stage of both datasets, the objective is to identify positive passages from a large collection. Positive pairs in these datasets are assessed based on strict rule, i.e., whether passages contain answers or not (Karpukhin et al., 2020a). Consequently, we consider these datasets to be of high quality. Thus, we leverage them for simulation experiments to quantitatively analyze the impact of varying proportions of noise. Drawing inspiration from the setting in the noisy classification task (Han et al., 2018), we simulate the mismatched-pair noise by randomly pairing queries with unrelated documents."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.865,
+ 0.347,
+ 0.881
+ ],
+ "angle": 0,
+ "content": "4.2 Implementation Details"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.888,
+ 0.49,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "NPC is a general training paradigm that can be directly applied to almost all retrieval models. For"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.368,
+ 0.885,
+ 0.544
+ ],
+ "angle": 0,
+ "content": "StaQC and SO-DS, we adopt UniXcoder (Guo et al., 2022) as our backbone, which is the SoTA model for code representation. Following Guo et al. (2022), we adopt the cosine distance as a similarity function and set temperature \\(\\lambda\\) to 20. We update model parameters using the Adam optimizer and perform early stopping on the development set. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 256, 5, and 10, respectively. In the \"Hard Negative\" setting, we adopt the same strategy as Li et al. (2022)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.546,
+ 0.885,
+ 0.771
+ ],
+ "angle": 0,
+ "content": "For NQ and TQ, we adopt BERT (Devlin et al., 2019) as our initial model. Following Karpukhin et al. (2020a), we adopt inner-product as the similarity function and set temperature \\(\\lambda\\) to 1. The max sequence length is 16 for query and 128 for passage. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 512, 10, and 40, respectively. We adopt \"BM25 Negative\" and \"Hard Negative\" strategies as described in the DPR toolkit. For a fair comparison, we implement DPR (Karpukhin et al., 2020a) with the same hyperparameters. All experiments are run on 8 NVIDIA Tesla A100 GPUs. The implementation of NPC is based on Huggingface (Wolf et al., 2020)."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.782,
+ 0.615,
+ 0.797
+ ],
+ "angle": 0,
+ "content": "4.3 Results"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.803,
+ 0.884,
+ 0.884
+ ],
+ "angle": 0,
+ "content": "Results on StaQC and SO-DS: Table 1 shows the results on the realistic-noisy datasets StaQC and SO-DS. Both datasets contain a large number of real noise pairs. The first block shows the results of previous SoTA methods. BM25 desc is a traditional"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.51,
+ 0.892,
+ 0.868,
+ 0.918
+ ],
+ "angle": 0,
+ "content": "\\(\\dagger\\)https://github.com/facebookresearch/DPR"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.477,
+ 0.928,
+ 0.526,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "11443"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.139,
+ 0.082,
+ 0.86,
+ 0.462
+ ],
+ "angle": 0,
+ "content": " | Question | Code |
| 1 | Split words in a nested list into letters | » [list(l[0]) for l in mylist] |
| 2 | Dictionary in python problem | » s = problem.getSuccessors( getStartState()) |
| 3 | Find the Common first name from Django Auth user Model | » import operator |
| 4 | Find all text files not containing some text string | » lst = [1,2,4,6,3,8,0,5] \n» for n in lst[:]: \n» if n % 2 == 0: \n»>>> lst.remove(n) \n» lst |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.275,
+ 0.586,
+ 0.723,
+ 0.601
+ ],
+ "angle": 0,
+ "content": "Table 9: Some noisy pairs detected by NPC in StaQC training set."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.478,
+ 0.929,
+ 0.524,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "11451"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_origin.pdf b/2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..dee3a1dd9b11795473b6e5c891a0a3bc52c7120f
--- /dev/null
+++ b/2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:921869deba6db20b6db9ea4928326374e8e8b483585df1e9ef38de967e8314c9
+size 1539262
diff --git a/2023/Noisy Pair Corrector for Dense Retrieval/full.md b/2023/Noisy Pair Corrector for Dense Retrieval/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..584b53c05690d2997132477364d7ddb8b3bd41e5
--- /dev/null
+++ b/2023/Noisy Pair Corrector for Dense Retrieval/full.md
@@ -0,0 +1,388 @@
+# Noisy Pair Corrector for Dense Retrieval
+
+Hang Zhang $^{1}$ , Yeyun Gong, Xingwei He $^{†2}$ , Dayiheng Liu $^{1}$ , Daya Guo $^{3}$ , Jiancheng Lv $^{†1}$ , Jian Guo $^{†4}$
+
+| Methods | StaQC | SO-DS |
| R@3 | R@10 | MRR | R@3 | R@10 | MRR |
| BM25 desc (Heyman and Van Cutsem, 2020) | 8.0 | 13.3 | 7.5 | 23.8 | 32.3 | 21.6 |
| NBOW (Heyman and Van Cutsem, 2020) | 10.9 | 16.6 | 9.5 | 27.7 | 38.0 | 24.7 |
| USE (Heyman and Van Cutsem, 2020) | 12.8 | 20.3 | 11.7 | 33.3 | 48.5 | 30.4 |
| CodeBERT (Feng et al., 2020) | - | - | 23.4 | - | - | 23.1 |
| GraphCodeBERT (Guo et al., 2021) | - | - | 24.1 | - | - | 25.2 |
| CodeRetriever (In-Batch Negative) (Li et al., 2022) | - | - | 25.5 | - | - | 27.1 |
| CodeRetriever (Hard Negative) (Li et al., 2022) | - | - | 24.6 | - | - | 31.8 |
| UniXcoder (In-Batch Negative) (Guo et al., 2022) | 29.98 | 47.47 | 28.04 | 31.90 | 51.21 | 28.29 |
| UniXcoder (Hard Negative) (Guo et al., 2022) | 31.18 | 48.38 | 28.63 | 33.42 | 53.37 | 29.97 |
| NPC (In-Batch Negative) | 33.07 | 50.35 | 30.39 | 35.58 | 54.54 | 30.96 |
| NPC (Hard Negative) | 34.38 | 52.20 | 31.36 | 38.00 | 56.51 | 32.49 |
+
+Table 1: Retrieval performance on StaQC and SO-DS, which are realistic-noisy datasets. The results of the first block are mainly borrowed from published papers (Heyman and Van Cutsem, 2020; Li et al., 2022). If the results are not provided, we mark them as “-”.
+
+been widely used on code summarization (Ped-damail et al., 2018) and code search (Heyman and Van Cutsem, 2020). SO-DS mines query-code pairs from the most upvoted Stack Overflow posts, mainly focuses on the data science domain. Following previous works (Heyman and Van Cutsem, 2020; Li et al., 2022), we resort to Recall of topk $(\mathbb{R}@\mathbb{k})$ and Mean Reciprocal Rank (MRR) as the evaluation metric. StaQC and SO-DS are constructed automatically without human annotation. Therefore, there are numerous mismatched pairs in training data.
+
+Natural Questions (NQ) collects real queries from the Google search engine. Each question is paired with an answer span and golden passages from the Wikipedia pages. Trivia QA (TQ) is a reading comprehension dataset authored by trivia enthusiasts. During the retrieval stage of both datasets, the objective is to identify positive passages from a large collection. Positive pairs in these datasets are assessed based on strict rule, i.e., whether passages contain answers or not (Karpukhin et al., 2020a). Consequently, we consider these datasets to be of high quality. Thus, we leverage them for simulation experiments to quantitatively analyze the impact of varying proportions of noise. Drawing inspiration from the setting in the noisy classification task (Han et al., 2018), we simulate the mismatched-pair noise by randomly pairing queries with unrelated documents.
+
+# 4.2 Implementation Details
+
+NPC is a general training paradigm that can be directly applied to almost all retrieval models. For
+
+StaQC and SO-DS, we adopt UniXcoder (Guo et al., 2022) as our backbone, which is the SoTA model for code representation. Following Guo et al. (2022), we adopt the cosine distance as a similarity function and set temperature $\lambda$ to 20. We update model parameters using the Adam optimizer and perform early stopping on the development set. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 256, 5, and 10, respectively. In the "Hard Negative" setting, we adopt the same strategy as Li et al. (2022).
+
+For NQ and TQ, we adopt BERT (Devlin et al., 2019) as our initial model. Following Karpukhin et al. (2020a), we adopt inner-product as the similarity function and set temperature $\lambda$ to 1. The max sequence length is 16 for query and 128 for passage. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 512, 10, and 40, respectively. We adopt "BM25 Negative" and "Hard Negative" strategies as described in the DPR toolkit. For a fair comparison, we implement DPR (Karpukhin et al., 2020a) with the same hyperparameters. All experiments are run on 8 NVIDIA Tesla A100 GPUs. The implementation of NPC is based on Huggingface (Wolf et al., 2020).
+
+# 4.3 Results
+
+Results on StaQC and SO-DS: Table 1 shows the results on the realistic-noisy datasets StaQC and SO-DS. Both datasets contain a large number of real noise pairs. The first block shows the results of previous SoTA methods. BM25 desc is a traditional
+
+ | Question | Code |
| 1 | Split words in a nested list into letters | » [list(l[0]) for l in mylist] |
| 2 | Dictionary in python problem | » s = problem.getSuccessors( getStartState()) |
| 3 | Find the Common first name from Django Auth user Model | » import operator |
| 4 | Find all text files not containing some text string | » lst = [1,2,4,6,3,8,0,5]
+» for n in lst[:]:
+» if n % 2 == 0:
+»>>> lst.remove(n)
+» lst |
+
+Table 9: Some noisy pairs detected by NPC in StaQC training set.
\ No newline at end of file
diff --git a/2023/Noisy Pair Corrector for Dense Retrieval/images.zip b/2023/Noisy Pair Corrector for Dense Retrieval/images.zip
new file mode 100644
index 0000000000000000000000000000000000000000..a449dd6117f512e8fe16a0ffa7235864c2e402d7
--- /dev/null
+++ b/2023/Noisy Pair Corrector for Dense Retrieval/images.zip
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94cb83dba96e64b17ba5ed2b2bb4a7573acdc1d14841a0d265f201c487e5443d
+size 668527
diff --git a/2023/Noisy Pair Corrector for Dense Retrieval/layout.json b/2023/Noisy Pair Corrector for Dense Retrieval/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..0100fc24689b62fa899458fc806eefd3db3c2409
--- /dev/null
+++ b/2023/Noisy Pair Corrector for Dense Retrieval/layout.json
@@ -0,0 +1,10550 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 170,
+ 76,
+ 424,
+ 93
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 170,
+ 76,
+ 424,
+ 93
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 170,
+ 76,
+ 424,
+ 93
+ ],
+ "type": "text",
+ "content": "Noisy Pair Corrector for Dense Retrieval"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "text",
+ "content": "Hang Zhang"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "text",
+ "content": ", Yeyun Gong, Xingwei He"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{†2}"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "text",
+ "content": ", Dayiheng Liu"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "text",
+ "content": ", Daya Guo"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{3}"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "text",
+ "content": ", Jiancheng Lv"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{†1}"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "text",
+ "content": ", Jian Guo"
+ },
+ {
+ "bbox": [
+ 143,
+ 105,
+ 449,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{†4}"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 175,
+ 133,
+ 422,
+ 147
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 175,
+ 133,
+ 422,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 175,
+ 133,
+ 422,
+ 147
+ ],
+ "type": "text",
+ "content": "| Methods | StaQC | SO-DS |
| R@3 | R@10 | MRR | R@3 | R@10 | MRR |
| BM25 desc (Heyman and Van Cutsem, 2020) | 8.0 | 13.3 | 7.5 | 23.8 | 32.3 | 21.6 |
| NBOW (Heyman and Van Cutsem, 2020) | 10.9 | 16.6 | 9.5 | 27.7 | 38.0 | 24.7 |
| USE (Heyman and Van Cutsem, 2020) | 12.8 | 20.3 | 11.7 | 33.3 | 48.5 | 30.4 |
| CodeBERT (Feng et al., 2020) | - | - | 23.4 | - | - | 23.1 |
| GraphCodeBERT (Guo et al., 2021) | - | - | 24.1 | - | - | 25.2 |
| CodeRetriever (In-Batch Negative) (Li et al., 2022) | - | - | 25.5 | - | - | 27.1 |
| CodeRetriever (Hard Negative) (Li et al., 2022) | - | - | 24.6 | - | - | 31.8 |
| UniXcoder (In-Batch Negative) (Guo et al., 2022) | 29.98 | 47.47 | 28.04 | 31.90 | 51.21 | 28.29 |
| UniXcoder (Hard Negative) (Guo et al., 2022) | 31.18 | 48.38 | 28.63 | 33.42 | 53.37 | 29.97 |
| NPC (In-Batch Negative) | 33.07 | 50.35 | 30.39 | 35.58 | 54.54 | 30.96 |
| NPC (Hard Negative) | 34.38 | 52.20 | 31.36 | 38.00 | 56.51 | 32.49 |
",
+ "image_path": "fa72da3b4f438baa439f55cc23d4578fc68d5c88c479adada14949b5f8d94c77.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 258,
+ 525,
+ 295
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 258,
+ 525,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 258,
+ 525,
+ 295
+ ],
+ "type": "text",
+ "content": "Table 1: Retrieval performance on StaQC and SO-DS, which are realistic-noisy datasets. The results of the first block are mainly borrowed from published papers (Heyman and Van Cutsem, 2020; Li et al., 2022). If the results are not provided, we mark them as “-”."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 66,
+ 309,
+ 291,
+ 470
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 66,
+ 309,
+ 291,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 66,
+ 309,
+ 291,
+ 470
+ ],
+ "type": "text",
+ "content": "been widely used on code summarization (Ped-damail et al., 2018) and code search (Heyman and Van Cutsem, 2020). SO-DS mines query-code pairs from the most upvoted Stack Overflow posts, mainly focuses on the data science domain. Following previous works (Heyman and Van Cutsem, 2020; Li et al., 2022), we resort to Recall of topk "
+ },
+ {
+ "bbox": [
+ 66,
+ 309,
+ 291,
+ 470
+ ],
+ "type": "inline_equation",
+ "content": "(\\mathbb{R}@\\mathbb{k})"
+ },
+ {
+ "bbox": [
+ 66,
+ 309,
+ 291,
+ 470
+ ],
+ "type": "text",
+ "content": " and Mean Reciprocal Rank (MRR) as the evaluation metric. StaQC and SO-DS are constructed automatically without human annotation. Therefore, there are numerous mismatched pairs in training data."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 66,
+ 472,
+ 291,
+ 716
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 66,
+ 472,
+ 291,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 66,
+ 472,
+ 291,
+ 716
+ ],
+ "type": "text",
+ "content": "Natural Questions (NQ) collects real queries from the Google search engine. Each question is paired with an answer span and golden passages from the Wikipedia pages. Trivia QA (TQ) is a reading comprehension dataset authored by trivia enthusiasts. During the retrieval stage of both datasets, the objective is to identify positive passages from a large collection. Positive pairs in these datasets are assessed based on strict rule, i.e., whether passages contain answers or not (Karpukhin et al., 2020a). Consequently, we consider these datasets to be of high quality. Thus, we leverage them for simulation experiments to quantitatively analyze the impact of varying proportions of noise. Drawing inspiration from the setting in the noisy classification task (Han et al., 2018), we simulate the mismatched-pair noise by randomly pairing queries with unrelated documents."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 727,
+ 206,
+ 740
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 727,
+ 206,
+ 740
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 727,
+ 206,
+ 740
+ ],
+ "type": "text",
+ "content": "4.2 Implementation Details"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 746,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 746,
+ 291,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 746,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": "NPC is a general training paradigm that can be directly applied to almost all retrieval models. For"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 302,
+ 309,
+ 526,
+ 457
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 309,
+ 526,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 309,
+ 526,
+ 457
+ ],
+ "type": "text",
+ "content": "StaQC and SO-DS, we adopt UniXcoder (Guo et al., 2022) as our backbone, which is the SoTA model for code representation. Following Guo et al. (2022), we adopt the cosine distance as a similarity function and set temperature "
+ },
+ {
+ "bbox": [
+ 302,
+ 309,
+ 526,
+ 457
+ ],
+ "type": "inline_equation",
+ "content": "\\lambda"
+ },
+ {
+ "bbox": [
+ 302,
+ 309,
+ 526,
+ 457
+ ],
+ "type": "text",
+ "content": " to 20. We update model parameters using the Adam optimizer and perform early stopping on the development set. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 256, 5, and 10, respectively. In the \"Hard Negative\" setting, we adopt the same strategy as Li et al. (2022)."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 302,
+ 459,
+ 526,
+ 648
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 459,
+ 526,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 459,
+ 526,
+ 648
+ ],
+ "type": "text",
+ "content": "For NQ and TQ, we adopt BERT (Devlin et al., 2019) as our initial model. Following Karpukhin et al. (2020a), we adopt inner-product as the similarity function and set temperature "
+ },
+ {
+ "bbox": [
+ 302,
+ 459,
+ 526,
+ 648
+ ],
+ "type": "inline_equation",
+ "content": "\\lambda"
+ },
+ {
+ "bbox": [
+ 302,
+ 459,
+ 526,
+ 648
+ ],
+ "type": "text",
+ "content": " to 1. The max sequence length is 16 for query and 128 for passage. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 512, 10, and 40, respectively. We adopt \"BM25 Negative\" and \"Hard Negative\" strategies as described in the DPR toolkit. For a fair comparison, we implement DPR (Karpukhin et al., 2020a) with the same hyperparameters. All experiments are run on 8 NVIDIA Tesla A100 GPUs. The implementation of NPC is based on Huggingface (Wolf et al., 2020)."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 302,
+ 657,
+ 365,
+ 670
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 657,
+ 365,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 657,
+ 365,
+ 670
+ ],
+ "type": "text",
+ "content": "4.3 Results"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 675,
+ 525,
+ 743
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 675,
+ 525,
+ 743
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 675,
+ 525,
+ 743
+ ],
+ "type": "text",
+ "content": "Results on StaQC and SO-DS: Table 1 shows the results on the realistic-noisy datasets StaQC and SO-DS. Both datasets contain a large number of real noise pairs. The first block shows the results of previous SoTA methods. BM25 desc is a traditional"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 303,
+ 750,
+ 516,
+ 772
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 303,
+ 750,
+ 516,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 750,
+ 516,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\dagger"
+ },
+ {
+ "bbox": [
+ 303,
+ 750,
+ 516,
+ 772
+ ],
+ "type": "text",
+ "content": "https://github.com/facebookresearch/DPR"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 283,
+ 780,
+ 312,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 780,
+ 312,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 780,
+ 312,
+ 791
+ ],
+ "type": "text",
+ "content": "11443"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 388
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 82,
+ 68,
+ 511,
+ 388
+ ],
+ "type": "table",
+ "html": " | Question | Code |
| 1 | Split words in a nested list into letters | » [list(l[0]) for l in mylist] |
| 2 | Dictionary in python problem | » s = problem.getSuccessors( getStartState()) |
| 3 | Find the Common first name from Django Auth user Model | » import operator |
| 4 | Find all text files not containing some text string | » lst = [1,2,4,6,3,8,0,5] \n» for n in lst[:]: \n» if n % 2 == 0: \n»>>> lst.remove(n) \n» lst |
",
+ "image_path": "013248572b5114b398641d8ed44af23db091529f116f4c49f6b702063a6f4128.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 163,
+ 492,
+ 430,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 163,
+ 492,
+ 430,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 163,
+ 492,
+ 430,
+ 505
+ ],
+ "type": "text",
+ "content": "Table 9: Some noisy pairs detected by NPC in StaQC training set."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 311,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 311,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 311,
+ 791
+ ],
+ "type": "text",
+ "content": "11451"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 12
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2023/Noisy Self-Training with Synthetic Queries for Dense Retrieval/260bb510-850d-4d78-a651-3f79c3ed0b9d_content_list.json b/2023/Noisy Self-Training with Synthetic Queries for Dense Retrieval/260bb510-850d-4d78-a651-3f79c3ed0b9d_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..ef22fc13e5bd5557764d6584a435c95ea423eecf
--- /dev/null
+++ b/2023/Noisy Self-Training with Synthetic Queries for Dense Retrieval/260bb510-850d-4d78-a651-3f79c3ed0b9d_content_list.json
@@ -0,0 +1,2207 @@
+[
+ {
+ "type": "text",
+ "text": "Noisy Self-Training with Synthetic Queries for Dense Retrieval",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 89,
+ 823,
+ 111
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Fan Jiang and Tom Drummond and Trevor Cohn *",
+ "bbox": [
+ 277,
+ 130,
+ 729,
+ 147
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "School of Computing and Information Systems",
+ "bbox": [
+ 307,
+ 148,
+ 695,
+ 164
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "The University of Melbourne, Victoria, Australia",
+ "bbox": [
+ 302,
+ 165,
+ 699,
+ 181
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "fan.jiang1@student.unimelb.edu.au",
+ "bbox": [
+ 332,
+ 181,
+ 670,
+ 197
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "{tom.drummond, trevor.cohn}@unimelb.edu.au",
+ "bbox": [
+ 289,
+ 198,
+ 712,
+ 214
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 260,
+ 252,
+ 339,
+ 268
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Although existing neural retrieval models reveal promising results when training data is abundant and the performance keeps improving as training data increases, collecting high-quality annotated data is prohibitively costly. To this end, we introduce a novel noisy self-training framework combined with synthetic queries, showing that neural retrievers can be improved in a self-evolution manner with no reliance on any external models. Experimental results show that our method improves consistently over existing methods on both general-domain (e.g., MS-MARCO) and out-of-domain (i.e., BEIR) retrieval benchmarks. Extra analysis on low-resource settings reveals that our method is data efficient and outperforms competitive baselines, with as little as $30\\%$ of labelled training data. Further extending the framework for reranker training demonstrates that the proposed method is general and yields additional gains on tasks of diverse domains. $^{1}$",
+ "bbox": [
+ 141,
+ 279,
+ 460,
+ 579
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 Introduction",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 590,
+ 260,
+ 606
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "As an important task, information retrieval (IR) refers to the task of finding relevant texts from a large collection of passages or documents to satisfy the specific information needs of users. The information need is usually expressed as a short textual query, and the task is formulated as retrieving texts that are most relevant to the given query.",
+ "bbox": [
+ 112,
+ 615,
+ 489,
+ 728
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Recently, impressive achievements have been made in neural retrieval models through adopting large-scale pre-trained models (Devlin et al., 2019). Dual-encoders typically serve as the backbone architecture, which enables retrieving relevant knowledge from collections with millions or billions of passages in a fraction of time (Karpukhin et al., 2020). Unlike traditional term-matching-based lexical retrievers, such as TF-IDF (Manning et al.,",
+ "bbox": [
+ 112,
+ 728,
+ 489,
+ 873
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "2008) or BM25 (Robertson and Zaragoza, 2009), which can be applied without any training, neural retrieval methods normally require training on a sufficient number of human-labelled query-passage pairs to work well. Nevertheless, due to the high cost of human annotations, the number of available query-passage pairs is way smaller compared to the size of passage collections (500k at most (Bajaj et al., 2016) v.s. 21M passages (Kwiatkowski et al., 2019)). Moreover, the situation becomes increasingly worse in out-of-domain applications, where only a few or no training examples are available. Directly applying neural retrieval models trained on high-resource datasets typically achieves low out-of-domain performance and lags behind their lexical counterparts (Thakur et al., 2021).",
+ "bbox": [
+ 507,
+ 252,
+ 884,
+ 510
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "In this work, we aim to improve the performance of state-of-the-art neural retrieval models on both general domain and out-of-domain datasets, by using automatically synthesised queries. For this purpose, we use a query generator to automatically synthesise queries for each passage in the target dataset, which we then use for pre-training, based on a self-training objective (Scudder, 1965). More specifically, instead of directly training the model on synthetic query-passage pairs (Ma et al., 2021), we use a neural retrieval model that was trained on the labelled dataset as the teacher to generate soft labels for each synthetic query, providing a supervision signal which is more robust to noise in the generated data. This ameliorates issues with the synthetic queries which are often generic or ambiguous, meaning that the query is only weakly related to its originating passage, and accordingly, other passages may match the synthetic query equally well or better. Table 1 shows a motivating example, where the synthetic query is spuriously-related to its originating passage with only the keyword achievement matched. However, the top-2 sampled \"negatives\" are highly related to the query in topics and semantics, according to both human judgment",
+ "bbox": [
+ 507,
+ 517,
+ 884,
+ 919
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_footnote",
+ "text": "*Now at Google DeepMind",
+ "bbox": [
+ 139,
+ 879,
+ 315,
+ 892
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_footnote",
+ "text": "| Method | MS-MARCO | Natural Questions |
| MRR@10 | R@50 | R@1k | R@5 | R@20 | R@100 |
| BM25 (Robertson and Zaragoza, 2009)* | 18.7 | 59.2 | 85.7 | - | 59.1 | |
| DeepCT (Dai and Callan, 2020)* | 24.3 | 69.0 | 91.0 | - | - | - |
| docT5query (Nogueira and Lin, 2019)* | 27.7 | 75.6 | 94.7 | - | - | - |
| GAR (Mao et al., 2021)* | - | - | - | 60.9 | 74.4 | 85.3 |
| DPR (Karpukhin et al., 2020)* | - | - | - | - | 74.4 | 85.3 |
| ANCE (Xiong et al., 2021)* | 33.0 | - | 95.9 | - | 81.9 | 87.5 |
| ColBERT (Khattab and Zaharia, 2020)* | 36.0 | 82.9 | 96.8 | - | - | - |
| DPR-PAQ (Oguz et al., 2022)* | | | | | | |
| - BERTbase | 31.4 | - | - | 74.5 | 83.7 | 88.6 |
| - RoBERTa(base) | 32.3 | - | - | 74.2 | 84.0 | 89.2 |
| Condenser (Gao and Callan, 2021)* | 36.6 | - | 97.4 | - | 83.2 | 88.4 |
| coCondenser (Gao and Callan, 2022)* | 38.2 | - | 98.4 | 75.8 | 84.3 | 89.0 |
| Our coCondenser (Teacher) (Gao and Callan, 2022)† | 37.8 | 85.5 | 98.2 | 75.9 | 85.0 | 89.4 |
| Student coCondenser | 39.4 | 86.7 | 98.5 | 77.0 | 85.5 | 89.5 |
",
+ "bbox": [
+ 132,
+ 80,
+ 867,
+ 310
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/435276f9c36238ea120a7a1e5ec9fe53c983ddd312d4a106263f1ff155b9ca25.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 131,
+ 376,
+ 470,
+ 455
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/18c5824ab0eaea25b401bd159cadbb6d6e175cfb06dcf2f623d6d020613d9524.jpg",
+ "image_caption": [
+ "Figure 2: Results on MS-MARCO dev and SCIDOCS test sets when using different pre-trained models."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 129,
+ 458,
+ 470,
+ 536
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "table",
+ "img_path": "images/899f5e99168a4b55322379e3d16aefdae65d2e9fefedbc956d7b5a7f1e9117ae.jpg",
+ "table_caption": [
+ "Table 2: Results on MS-MARCO dev set and Natural Questions test set. * indicates results directly copied from Gao and Callan (2022) and Khattab and Zaharia (2020). † indicates our implementation. The best results are marked bold and unavailable results are left blank."
+ ],
+ "table_footnote": [],
+ "table_body": " | Hyperparameters | MS-MARCO | Natural Questions | BEIR |
| Shared | Max length | 156 | 156 | 351 |
| Optimizer | AdamW | AdamW | AdamW |
| Scheduler | Linear warmup | Linear warmup | Linear Warmup |
| Warmup proportion | 0.1 | 0.1 | 0.1 |
| Weight decay | 0.1 | 0.1 | 0.1 |
| #Params | 110M | 110M | 110M |
| Teacher Reranker | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | - |
| Batch size | 12 | 12 | - |
| #hard negatives | 40 | 15 | - |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | - |
| #Epochs | 2 | 10 | - |
| Self Training |
| Pre-train | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) |
| Batch size | 48 | 128 | 32 |
| #hard negatives | 40 | 15 | 7 |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | stage 2 teacher retriever |
| #Epochs | 1 | 1 | 5 |
| Finetune | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | - |
| Batch size | 12 | 12 | - |
| #hard negatives | 40 | 15 | - |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | - |
| #Epochs | 2 | 10 | - - |
",
+ "bbox": [
+ 115,
+ 560,
+ 882,
+ 881
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Table 9: Hyperparameter settings for reranker training.",
+ "bbox": [
+ 310,
+ 890,
+ 682,
+ 904
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "page_number",
+ "text": "12006",
+ "bbox": [
+ 477,
+ 928,
+ 524,
+ 940
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "table",
+ "img_path": "images/0dd50a85fc8fdf9a25db2d7875a9d171dc539aa94418b1343ab5f9218ff6a1b9.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Method | MS-MARCO | Natural Questions |
| MRR@10 | R@50 | R@1k | R@5 | R@20 | R@100 |
| BM25 (Robertson and Zaragoza, 2009)* | 18.7 | 59.2 | 85.7 | - | 59.1 | |
| DeepCT (Dai and Callan, 2020)* | 24.3 | 69.0 | 91.0 | - | - | - |
| docT5query (Nogueira and Lin, 2019)* | 27.7 | 75.6 | 94.7 | - | - | - |
| GAR (Mao et al., 2021)* | - | - | - | 60.9 | 74.4 | 85.3 |
| DPR (Karpukhin et al., 2020)* | - | - | - | - | 74.4 | 85.3 |
| ANCE (Xiong et al., 2021)* | 33.0 | - | 95.9 | - | 81.9 | 87.5 |
| ColBERT (Khattab and Zaharia, 2020)* | 36.0 | 82.9 | 96.8 | - | - | - |
| DPR-PAQ (Oguz et al., 2022)* | | | | | | |
| - BERTbase | 31.4 | - | - | 74.5 | 83.7 | 88.6 |
| - RoBERTa(base) | 32.3 | - | - | 74.2 | 84.0 | 89.2 |
| Condenser (Gao and Callan, 2021)* | 36.6 | - | 97.4 | - | 83.2 | 88.4 |
| coCondenser (Gao and Callan, 2022)* | 38.2 | - | 98.4 | 75.8 | 84.3 | 89.0 |
| Our coCondenser (Teacher) (Gao and Callan, 2022)† | 37.8 | 85.5 | 98.2 | 75.9 | 85.0 | 89.4 |
| Student coCondenser | 39.4 | 86.7 | 98.5 | 77.0 | 85.5 | 89.5 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.113,
+ 0.32,
+ 0.884,
+ 0.364
+ ],
+ "angle": 0,
+ "content": "Table 2: Results on MS-MARCO dev set and Natural Questions test set. * indicates results directly copied from Gao and Callan (2022) and Khattab and Zaharia (2020). † indicates our implementation. The best results are marked bold and unavailable results are left blank."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.132,
+ 0.378,
+ 0.472,
+ 0.456
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.131,
+ 0.459,
+ 0.472,
+ 0.537
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.114,
+ 0.549,
+ 0.487,
+ 0.579
+ ],
+ "angle": 0,
+ "content": "Figure 2: Results on MS-MARCO dev and SCIDOCS test sets when using different pre-trained models."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.122,
+ 0.585,
+ 0.482,
+ 0.656
+ ],
+ "angle": 0,
+ "content": " | Hyperparameters | MS-MARCO | Natural Questions | BEIR |
| Shared | Max length | 156 | 156 | 351 |
| Optimizer | AdamW | AdamW | AdamW |
| Scheduler | Linear warmup | Linear warmup | Linear Warmup |
| Warmup proportion | 0.1 | 0.1 | 0.1 |
| Weight decay | 0.1 | 0.1 | 0.1 |
| #Params | 110M | 110M | 110M |
| Teacher Reranker | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | - |
| Batch size | 12 | 12 | - |
| #hard negatives | 40 | 15 | - |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | - |
| #Epochs | 2 | 10 | - |
| Self Training |
| Pre-train | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) |
| Batch size | 48 | 128 | 32 |
| #hard negatives | 40 | 15 | 7 |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | stage 2 teacher retriever |
| #Epochs | 1 | 1 | 5 |
| Finetune | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | - |
| Batch size | 12 | 12 | - |
| #hard negatives | 40 | 15 | - |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | - |
| #Epochs | 2 | 10 | - - |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.312,
+ 0.891,
+ 0.684,
+ 0.906
+ ],
+ "angle": 0,
+ "content": "Table 9: Hyperparameter settings for reranker training."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.478,
+ 0.929,
+ 0.526,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "12006"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.121,
+ 0.13,
+ 0.879,
+ 0.476
+ ],
+ "angle": 0,
+ "content": "| Method | MS-MARCO | Natural Questions |
| MRR@10 | R@50 | R@1k | R@5 | R@20 | R@100 |
| BM25 (Robertson and Zaragoza, 2009)* | 18.7 | 59.2 | 85.7 | - | 59.1 | |
| DeepCT (Dai and Callan, 2020)* | 24.3 | 69.0 | 91.0 | - | - | - |
| docT5query (Nogueira and Lin, 2019)* | 27.7 | 75.6 | 94.7 | - | - | - |
| GAR (Mao et al., 2021)* | - | - | - | 60.9 | 74.4 | 85.3 |
| DPR (Karpukhin et al., 2020)* | - | - | - | - | 74.4 | 85.3 |
| ANCE (Xiong et al., 2021)* | 33.0 | - | 95.9 | - | 81.9 | 87.5 |
| ColBERT (Khattab and Zaharia, 2020)* | 36.0 | 82.9 | 96.8 | - | - | - |
| DPR-PAQ (Oguz et al., 2022)* | | | | | | |
| - BERTbase | 31.4 | - | - | 74.5 | 83.7 | 88.6 |
| - RoBERTa(base) | 32.3 | - | - | 74.2 | 84.0 | 89.2 |
| Condenser (Gao and Callan, 2021)* | 36.6 | - | 97.4 | - | 83.2 | 88.4 |
| coCondenser (Gao and Callan, 2022)* | 38.2 | - | 98.4 | 75.8 | 84.3 | 89.0 |
| Our coCondenser (Teacher) (Gao and Callan, 2022)† | 37.8 | 85.5 | 98.2 | 75.9 | 85.0 | 89.4 |
| Student coCondenser | 39.4 | 86.7 | 98.5 | 77.0 | 85.5 | 89.5 |
+
+
+
+
+Figure 2: Results on MS-MARCO dev and SCIDOCS test sets when using different pre-trained models.
+
+Table 2: Results on MS-MARCO dev set and Natural Questions test set. * indicates results directly copied from Gao and Callan (2022) and Khattab and Zaharia (2020). † indicates our implementation. The best results are marked bold and unavailable results are left blank.
+
+| Method | MS-MARCO | Natural Questions |
| MRR@10 | R@50 | R@1k | R@5 | R@20 | R@100 |
| BM25 (Robertson and Zaragoza, 2009)* | 18.7 | 59.2 | 85.7 | - | 59.1 | |
| DeepCT (Dai and Callan, 2020)* | 24.3 | 69.0 | 91.0 | - | - | - |
| docT5query (Nogueira and Lin, 2019)* | 27.7 | 75.6 | 94.7 | - | - | - |
| GAR (Mao et al., 2021)* | - | - | - | 60.9 | 74.4 | 85.3 |
| DPR (Karpukhin et al., 2020)* | - | - | - | - | 74.4 | 85.3 |
| ANCE (Xiong et al., 2021)* | 33.0 | - | 95.9 | - | 81.9 | 87.5 |
| ColBERT (Khattab and Zaharia, 2020)* | 36.0 | 82.9 | 96.8 | - | - | - |
| DPR-PAQ (Oguz et al., 2022)* | | | | | | |
| - BERTbase | 31.4 | - | - | 74.5 | 83.7 | 88.6 |
| - RoBERTa(base) | 32.3 | - | - | 74.2 | 84.0 | 89.2 |
| Condenser (Gao and Callan, 2021)* | 36.6 | - | 97.4 | - | 83.2 | 88.4 |
| coCondenser (Gao and Callan, 2022)* | 38.2 | - | 98.4 | 75.8 | 84.3 | 89.0 |
| Our coCondenser (Teacher) (Gao and Callan, 2022)† | 37.8 | 85.5 | 98.2 | 75.9 | 85.0 | 89.4 |
| Student coCondenser | 39.4 | 86.7 | 98.5 | 77.0 | 85.5 | 89.5 |
",
+ "image_path": "c0f3950481205cbd5d5ad99e9e5ae0c818fe0efa889af8ad939823a1d8bd3a66.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 78,
+ 317,
+ 280,
+ 383
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 78,
+ 317,
+ 280,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 317,
+ 280,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 317,
+ 280,
+ 383
+ ],
+ "type": "image",
+ "image_path": "435276f9c36238ea120a7a1e5ec9fe53c983ddd312d4a106263f1ff155b9ca25.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 77,
+ 386,
+ 280,
+ 451
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 77,
+ 386,
+ 280,
+ 451
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 386,
+ 280,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 386,
+ 280,
+ 451
+ ],
+ "type": "image",
+ "image_path": "18c5824ab0eaea25b401bd159cadbb6d6e175cfb06dcf2f623d6d020613d9524.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 67,
+ 461,
+ 289,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 461,
+ 289,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 461,
+ 289,
+ 486
+ ],
+ "type": "text",
+ "content": "Figure 2: Results on MS-MARCO dev and SCIDOCS test sets when using different pre-trained models."
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 72,
+ 491,
+ 286,
+ 551
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 67,
+ 269,
+ 525,
+ 306
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 269,
+ 525,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 269,
+ 525,
+ 306
+ ],
+ "type": "text",
+ "content": "Table 2: Results on MS-MARCO dev set and Natural Questions test set. * indicates results directly copied from Gao and Callan (2022) and Khattab and Zaharia (2020). † indicates our implementation. The best results are marked bold and unavailable results are left blank."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 72,
+ 491,
+ 286,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 491,
+ 286,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 491,
+ 286,
+ 551
+ ],
+ "type": "table",
+ "html": " | Hyperparameters | MS-MARCO | Natural Questions | BEIR |
| Shared | Max length | 156 | 156 | 351 |
| Optimizer | AdamW | AdamW | AdamW |
| Scheduler | Linear warmup | Linear warmup | Linear Warmup |
| Warmup proportion | 0.1 | 0.1 | 0.1 |
| Weight decay | 0.1 | 0.1 | 0.1 |
| #Params | 110M | 110M | 110M |
| Teacher Reranker | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | - |
| Batch size | 12 | 12 | - |
| #hard negatives | 40 | 15 | - |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | - |
| #Epochs | 2 | 10 | - |
| Self Training |
| Pre-train | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) |
| Batch size | 48 | 128 | 32 |
| #hard negatives | 40 | 15 | 7 |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | stage 2 teacher retriever |
| #Epochs | 1 | 1 | 5 |
| Finetune | Learning rate | \\( 1 \\times 10^{-5} \\) | \\( 1 \\times 10^{-5} \\) | - |
| Batch size | 12 | 12 | - |
| #hard negatives | 40 | 15 | - |
| Hard negative source | stage 2 teacher retriever | stage 2 teacher retriever | - |
| #Epochs | 2 | 10 | - - |
",
+ "image_path": "d7c911f72ed607a28a345b9a18ecf386d0ff02c59c45ec45e49ca0ee9081a4c2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 185,
+ 749,
+ 406,
+ 761
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 185,
+ 749,
+ 406,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 185,
+ 749,
+ 406,
+ 761
+ ],
+ "type": "text",
+ "content": "Table 9: Hyperparameter settings for reranker training."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 781,
+ 312,
+ 791
+ ],
+ "type": "text",
+ "content": "12006"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 15
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 71,
+ 109,
+ 523,
+ 400
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 71,
+ 109,
+ 523,
+ 400
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 109,
+ 523,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 109,
+ 523,
+ 400
+ ],
+ "type": "table",
+ "html": "| Method | TED (d-BLEU) | News (d-BLEU) | Europarl (d-BLEU) | Average (d-BLEU) | Speed -up |
| Raw | KD | Raw | KD | Raw | KD | Raw | KD |
| AT Baselines |
| G-Transformer (Bao et al., 2021) | 27.23 | 26.42 | 27.22 | 26.38 | 34.09 | 32.87 | 29.51 | 28.56 | - |
| Transformer (Vaswani et al., 2017) | 0.69 | 26.04 | 0.23 | 0.43 | 33.41 | 26.43 | 11.44 | 17.63 | 1.0x |
| Existing NAT Models |
| Vanilla NAT (Gu et al., 2018) | 0.45 | 0.34 | 0.12 | 0.61 | 1.30 | 2.43 | 0.62 | 1.13 | 41.2x |
| GLAT (Qian et al., 2021) | 1.77 | 0.03 | 0.01 | 2.56 | 2.13 | 8.17 | 1.30 | 3.59 | 40.0x |
| Latent-GLAT (Bao et al., 2022) | 2.10 | 1.87 | 0.30 | 2.30 | 5.26 | 4.25 | 2.55 | 2.81 | 29.2x |
| NAT+CTC (Libovický and Helcl, 2018) | 21.54 | 24.98 | 15.95 | 24.03 | 0.00 | 31.58 | 12.50 | 26.86 | 27.7x |
| GLAT+CTC (Qian et al., 2021) | 18.49 | 25.31 | 10.23 | 20.01 | 0.00 | 29.47 | 9.57 | 24.93 | 26.3x |
| DA-Transformer (Huang et al., 2022) | 20.47 | 25.02 | 13.99 | 23.37 | 30.34 | 32.14 | 21.60 | 26.84 | 30.8x |
| NAT Models with Sentence Alignment |
| G-Trans+GLAT (ours) | 19.96* | 24.23* | 15.14* | 23.04* | 25.67* | 31.45* | 20.26 | 26.24 | 30.0x |
| G-Trans+GLAT+CTC (ours) | 24.09* | 26.31* | 21.68* | 25.61* | 30.35* | 32.24* | 25.37 | 28.05 | 20.0x |
| G-Trans+DA-Trans (ours) | 23.45* | 25.73* | 21.43* | 24.70* | 30.36 | 32.29 | 25.08 | 27.57 | 25.1x |
",
+ "bbox": [
+ 114,
+ 80,
+ 884,
+ 291
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Table 1: Main results on raw data (Raw) and knowledge distilled data (KD), where the failures are marked and investigated in section 5.1. We use the official code from the respective papers of NAT models, except NAT+CTC, which is implemented by ourselves based on the GLAT code. “*” denotes a statistical significance at the level of $p < 0.01$ using a t-test, compared to the corresponding baseline NAT model without sentence alignment.",
+ "bbox": [
+ 110,
+ 299,
+ 882,
+ 357
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "G-Trans+GLAT+CTC integrates the sentence alignment with the CTC alignment. The default CTC algorithm aggregates all possible latent alignments across the entire sequence, which may align a source sentence to a wrong target sentence (e.g., the first source sentence to the second target sentence). Such global alignment not only slows down the training process but also causes unstable model performance. Different from the global alignment in Eq. 3, we apply CTC alignment on each sentence",
+ "bbox": [
+ 112,
+ 382,
+ 487,
+ 542
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\np _ {\\theta} (y | x) = \\prod_ {j = 1} ^ {K} \\left[ \\sum_ {a _ {j} \\in \\beta \\left(y _ {j}\\right)} \\prod_ {i = 1} ^ {M _ {j}} p _ {\\theta} \\left(a _ {j, i} \\mid x _ {j}, x\\right) \\right], \\tag {8}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 154,
+ 556,
+ 487,
+ 599
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $M_{j}$ denotes the reserved token positions for the $j$ -th target sentence. Since CTC alignment is restricted inside each target sentence, the alignment space $\\beta(y_{j})$ is enormously reduced compared to the $\\beta(y)$ in Eq. 3.",
+ "bbox": [
+ 112,
+ 611,
+ 487,
+ 690
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "G-Trans+DA-Trans introduces the sentence alignment into the output DAG of the DA-Transformer. The default DAG models the whole sequence, which enables the transition from a vertex in one sentence to a vertex in another sentence, making the transition space linearly increase with the enlargement of the sequence length.",
+ "bbox": [
+ 112,
+ 693,
+ 487,
+ 804
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "To address the issue, we enforce a constraint to isolate the vertex transitions of each sentence, forcing the path $a$ on each sentence starting with a special vertex $\" < s > \"$ and ending with a special vertex $\" < / s > \"$ . The transition between sentences only happens from the vertex $\" < / s > \"$ of a previous sentence to the vertex $\" < s > \"$ of the current sentence.",
+ "bbox": [
+ 112,
+ 806,
+ 487,
+ 917
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Formally, we use the same factorization as Eq. 4 but with a different collection $\\beta(y)$ of paths. At the implementation level, we simply mask the transition matrix to disable transitions from one sentence to another sentence, so that the dynamic programming algorithm keeps unchanged.",
+ "bbox": [
+ 507,
+ 382,
+ 882,
+ 479
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4 Experiments",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 489,
+ 653,
+ 506
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.1 Experimental Settings",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 514,
+ 729,
+ 530
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "We evaluate the models using document-level MT benchmark (Maruf et al., 2019), which includes three datasets TED, News, and Europarl, representing three domains and various data scales for English-German translation. More details about each dataset and the preprocessing are in Appendix A.1. We follow Liu et al. (2020), evaluating the model performance in sentence-level BLEU score $(s\\text{-BLEU})$ and document-level BLEU score $(d\\text{-BLEU})$ , which are explained in detail as Appendix A.2. We experiment on Base model (where Big model does not provide a stronger baseline) and evaluate the speedup using 1 GPU and 4 CPUs of a Tesla V100 environment as Appendix A.3.",
+ "bbox": [
+ 507,
+ 533,
+ 882,
+ 760
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.2 Overall Results",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 769,
+ 678,
+ 784
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "As shown in Table 1, NAT models achieve high speedup but still suffer from a significant performance gap with their AT counterparts.",
+ "bbox": [
+ 507,
+ 790,
+ 882,
+ 838
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Speedup. The inference accelerations of NAT models on the document level are between 25x and 41x, which surpasses the accelerations on the sentence level (between 2x to 15x) by a big margin. Specifically, NAT and GLAT provide the biggest",
+ "bbox": [
+ 507,
+ 839,
+ 882,
+ 917
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "acceleration by around $40\\mathrm{x}$ . More complex models, such as Latent-GLAT, DA-Transformer, and G-Trans+GLAT, accelerate inference by approximately $30\\mathrm{x}$ . CTC-based models have lower accelerations between $20\\mathrm{x}$ and $27\\mathrm{x}$ . On average, the acceleration on the document level is about $30\\mathrm{x}$ , which means that document-level MT systems could potentially save $96\\%$ computational resources and energy consumption by using NAT models.",
+ "bbox": [
+ 112,
+ 84,
+ 489,
+ 228
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Performance. Though NAT models underperform G-Transformer, some outpace the Transformer. The lengthy text sequence challenges both the AT models and the NAT models, resulting in exceptionally low d-BLEU scores (<10) in some settings. We treat these low scores as model failures and investigate them in section 5.1.",
+ "bbox": [
+ 112,
+ 230,
+ 489,
+ 343
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "With the help of sentence alignment, the performance gap between NAT models and AT baselines is largely reduced, especially when trained on KD data. For example, G-Trans+GLAT+CTC achieves an average d-BLEU of 28.05 on KD data, which is only 0.51 points lower than the d-BLEU of 28.56 of G-Transformer. However, its performance on Raw data is 4.14 points lower than G-Transformer, suggesting that NAT models experience severe challenges with raw data. These results demonstrate that even though knowledge distillation and sentence alignment enhance NAT models largely, there is still a gap between the NAT models and the strongest AT baseline.",
+ "bbox": [
+ 115,
+ 344,
+ 489,
+ 570
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.3 Breakdown Results",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 587,
+ 314,
+ 602
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Sentence-level alignment. Sentence alignment substantially enhances the performance of NAT models. Specifically, for GLAT, sentence alignment enhances the performance on KD data from an average of 3.59 to 26.24 d-BLEU of G-Trans+GLAT. For GLAT+CTC, sentence alignment elevates the performance on KD data from 24.93 to 28.05 d-BLEU on average. This score is comparable to G-Transformer on KD data, leaving a minor gap of 0.51 d-BLEU on average. The better results of G-Trans+GLAT+CTC than G-Trans+GLAT suggest that sentence alignment complements output alignment techniques such as CTC.",
+ "bbox": [
+ 112,
+ 611,
+ 489,
+ 820
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "The sentence alignment brings more benefits to models trained on raw data. As the average scores in Table 1 show, G-Trans+GLAT+CTC outperforms GLAT+CTC by 15.80 points on raw data and 3.12 points on KD data. G-Trans+DA-Trans outperforms DA-Transformer by 3.48 points on",
+ "bbox": [
+ 112,
+ 822,
+ 489,
+ 920
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "raw data and 0.73 points on KD data. These results suggest that sentence alignment mitigates the challenge brought by more modalities in raw data.",
+ "bbox": [
+ 507,
+ 84,
+ 884,
+ 131
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Token-level alignment. NAT models with implicit alignment such as vanilla NAT, GLAT, and Latent-GLAT fail on document-level MT, resulting in messy translations (repetitions, wrong words, and meaningless phrases). Conversely, NAT models with explicit alignments, such as NAT+CTC, GLAT+CTC, and DA-Transformer, produce superior document-level performance. DA-Transformer delivers the best overall performance among existing NAT models, even outperforming Transformer on Europarl (KD). These results suggest that token-level alignment plays an important role in NAT models to achieve good performance.",
+ "bbox": [
+ 507,
+ 135,
+ 885,
+ 344
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "However, we also observe that training with CTC loss on documents is not always stable, leading to occasional failure and a high variance in their performance (e.g., +/-2.66 d-BLEU on raw data and +/-0.84 on KD data for GLAT+CTC). The failure happens more frequently on raw data than on KD data. We speculate that this instability is caused by the increased alignment space on long sequences, which can be mitigated by sentence alignment. Experiments on G-Trans+GLAT+CTC show stable training and lower variance in performance (+/- 0.19 d-BLEU). These results suggest that token-level alignment alone is not enough to achieve stable training and good performance.",
+ "bbox": [
+ 507,
+ 347,
+ 885,
+ 571
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Knowledge distillation. Intuitively, documents have more modalities because they are much longer than sentences. If it is the case, knowledge distillation will become more critical for training NAT models on document-level MT than on sentence-level MT. We compare NAT models trained with and without knowledge distillation.",
+ "bbox": [
+ 507,
+ 575,
+ 884,
+ 688
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "As Table 1 shows, NAT models trained on KD data generally outperform the same model trained on raw data. The improvement is especially significant for NAT models with explicit alignment. For example, DA-Transformer obtains an average of 26.84 d-BLEU on KD data, which is 5.24 points higher than 21.60 d-BLEU on raw data. In contrast, DA-Transformer on sentence-level MT achieves similar results on raw data and on KD data (Huang et al., 2022). The results suggest that compared to sentences, documents have more severe multimodality issues.",
+ "bbox": [
+ 507,
+ 690,
+ 885,
+ 883
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Although knowledge distillation enhances the performance of NAT models, it also sets a ceiling",
+ "bbox": [
+ 507,
+ 887,
+ 884,
+ 920
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/5ee81d0b79c72258487b0ab4486f91c348b87b205008822352882329a3e312bb.jpg",
+ "image_caption": [
+ "Figure 3: Speedup rises as long as the sequence length grows, evaluated on TED and News."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 171,
+ 82,
+ 421,
+ 195
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "to the performance on document-level MT. Comparing the performance of G-Transformer on KD data to raw data, we see that KD data downgrades the performance by about 1 d-BLEU on average (from 29.51 to 28.56). These results differ from previous results on sentence-level MT (Huang et al., 2022), where knowledge distillation maintains or even enhances the performance of AT models. The discrepancy calls for further research on techniques to reduce the modalities on the document level.",
+ "bbox": [
+ 112,
+ 266,
+ 489,
+ 426
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Speedup on different sequence lengths. We evaluate GLAT and GLAT+CTC on various sequence lengths, including the single sentence and segments with a maximum length of 64, 128, 256, and 512. As Figure 3 shows, the speedup displays consistent trends among different datasets and NAT models. The GLAT generally has a higher speedup than GLAT-CTC, especially when the segment length goes above 256. The speedups on TED and News are almost identical, which is expected because the time costs are supposed to be irrelevant to the data domains. The trends suggest that we benefit more from the inference acceleration on longer documents, where document-level MT tasks provide NAT models with the best scenario.",
+ "bbox": [
+ 115,
+ 430,
+ 489,
+ 671
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Speedup on different batch sizes. The previous study (Helcl et al., 2022) reports that NAT models have limited acceleration on the sentence-level MT in parallel inference settings. We evaluate the models in the settings with different batch sizes, including 1, 2, 4, and 8 instances per batch. Given that document-level MT sequences are much longer than sentence-level MT sequences, we do not evaluate using a batch size larger than 8.",
+ "bbox": [
+ 112,
+ 675,
+ 489,
+ 819
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "As Figure 4 shows, when we increase the batch size from 1 to 8, the overall speedup of GLAT decreases from $40\\mathrm{x}$ to $9\\mathrm{x}$ . However, if we consider a more strict calculation of the speedup, excluding the time for initializing the models (which takes almost constant time) from the total evaluation time,",
+ "bbox": [
+ 112,
+ 822,
+ 489,
+ 917
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/bf8650d2403274b90126466ea4a7406a816f9a9c72e32a7c5552cfde358763da.jpg",
+ "image_caption": [
+ "Figure 4: Speedup with various batch sizes evaluated on News, where the “/ex” conducts a strict calculation of speedup by excluding the time cost for model initialization from both the AT baseline and the NAT models."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 566,
+ 82,
+ 818,
+ 195
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "the inference speedup of GLAT is $125\\mathrm{x}$ and $44\\mathrm{x}$ for the batch size of 1 and 8, respectively. These results suggest that although the speedup ratio decreases for bigger batch sizes, NAT models on document-level MT still show significant acceleration in the parallel inference settings.",
+ "bbox": [
+ 507,
+ 287,
+ 884,
+ 385
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5 The Challenge of Long Sequence",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 395,
+ 826,
+ 413
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "The long input and output in document-level MT bring unexplored challenges to NAT models. In section 4, we learn that NAT models have a significant performance gap with their AT counterparts, and various NAT models even fail in some settings. We investigate these challenges in this section, leaving other discussions in Appendix D.",
+ "bbox": [
+ 507,
+ 420,
+ 884,
+ 533
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5.1 The Failure of NAT Models",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 544,
+ 769,
+ 558
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Previous study (Bao et al., 2021) suggests that Transformer fails on long sequences because of local minima of loss values during the training process. We first investigate this possibility for the cause of the failure on NAT models. We evaluate GLAT, GLAT+CTC, Transformer, and GTransformer on different lengths of sequences, as shown in Figure 5a. Overall, the four models produce different patterns of trends. The d-BLEU score of the Transformer rises as long as the sequence length increases until the failure happens at the length of 512 tokens. In contrast, the d-BLEU scores of the GLAT and GLAT+CTC descend progressively as long as the sequence length increases, which suggests a performance decline instead of a training failure on NAT models.",
+ "bbox": [
+ 507,
+ 565,
+ 884,
+ 821
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Further evidence on the bigger dataset Europarl confirms that the low performance of NAT models is different from the failure of Transformer. Bao et al. (2021) suggests that the local minima encountered by Transformer can be escaped by using a bigger dataset, resulting in normal scores on Europarl.",
+ "bbox": [
+ 507,
+ 822,
+ 885,
+ 919
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/d58843f75fcbc9506a34733eaa2356f44d310e8a4feeb40cdc6c84c76a17a0b4.jpg",
+ "image_caption": [
+ "(a) NAT and AT performance on different sequence length."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 115,
+ 80,
+ 349,
+ 187
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/9f141243d8aa4c436811c51bbc230d887438983f828d307b2059a4b74eed3174.jpg",
+ "image_caption": [
+ "(b) GLAT detailed BLEU scores on different sequence lengths.",
+ "Figure 5: NAT vs. AT models on trends evaluated on News, where Transformer fails abruptly at the length of 512 while NAT models decrease progressively."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 381,
+ 80,
+ 613,
+ 186
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/9e84efc128d6b4550bae17b17f5d0bea87828d989f05db40bbf2fc78271511b6.jpg",
+ "image_caption": [
+ "(c) $GLAT + CTC$ detailed BLEU scores on different sequence lengths."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 648,
+ 80,
+ 880,
+ 186
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We evaluate Transformer, GLAT, and GLAT+CTC on Europarl, obtaining d-BLEU scores of 33.41, 2.13, and 0.00, respectively. We could see that GLAT and GLAT+CTC still give low scores on the bigger dataset, suggesting a different cause from the local minima.",
+ "bbox": [
+ 112,
+ 282,
+ 487,
+ 378
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We look into the d-BLEU score details, where BLEU- $n$ ( $n \\in [1,2,3,4]$ ) denotes the precision on $n$ -gram. As the d-BLEU scores on GLAT in Figure 5b show, the BLEU-3/4 decreases rapidly, reaching almost 0.0 at the lengths of 256 and 512. The BLEU-1 remains relatively high at 27.3 on the length of 256, resulting in messy translations, where the generated tokens are related to the content but are repeated, disordered, or in the wrong collocation as shown in Table 2. In comparison, the d-BLEU scores on GLAT+CTC in Figure 5c decrease slowly, where the BLEU-1 even increases on the length of 512. We speculate the rapid decrease in BLEU-3/4 on GLAT is caused by the multi-modality and misalignment issues, which can be mitigated by explicit alignments, such as CTC.",
+ "bbox": [
+ 112,
+ 382,
+ 489,
+ 640
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "The exceptional zero scores. Table 1 reveals some unexpected results. Both NAT+CTC and GLAT+CTC score a 0.00 d-BLEU on the raw Europarl dataset, a performance notably inferior to that of NAT and GLAT. This unprecedented anomaly may stem from the challenges of applying CTC loss on long sequences. The table further indicates that both combinations, NAT/GLAT+CTC and NAT/GLAT, experience training failures on the raw Europarl dataset. These failures are likely due to multi-modality and misalignment issues. These results empirically demonstrate that, while CTC loss can be effective, its application to document-level training is not consistently stable. We hypothesize that this instability arises from the expanded alignment space between lengthy input and output sequences, as detailed in the token-level alignment",
+ "bbox": [
+ 112,
+ 645,
+ 489,
+ 919
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Source: companies with a high percentage of floating rate debt stand to lose the most, Goldman said. outside pure stock plays, consumers stand to benefit as well through the rising dollar.",
+ "bbox": [
+ 517,
+ 282,
+ 873,
+ 332
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Target: Unternehmen mit einem hohen Anteil an flexiblen Zinsen werden am meisten verlieren, sagte Goldman. außerhalb der reinen Aktienspiele werden Verbraucher ebenfalls durch den steigenden Dollar profitieren.",
+ "bbox": [
+ 517,
+ 337,
+ 875,
+ 386
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "GLAT: Unternehmen mit einem hohen, freiem freiem freien erverlieren die die meisten meisten, . . . te te dazu kräften profitieren profitieren profitieren profitieren durch durch steigenden steigenden profitieren profitieren.",
+ "bbox": [
+ 517,
+ 391,
+ 873,
+ 451
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "GLAT+CTC: Unternehmen mit hohen der Schulden, die verlieren. außerhalb spielten Verbrauchener den profitieren.",
+ "bbox": [
+ 517,
+ 456,
+ 873,
+ 482
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Table 2: GLAT and GLAT+CTC produce good translation at the beginning but downgrade later with repetitions and missing translations.",
+ "bbox": [
+ 507,
+ 494,
+ 884,
+ 538
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "discussion in Section 4.3.",
+ "bbox": [
+ 507,
+ 564,
+ 700,
+ 579
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "5.2 Document Context",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 592,
+ 702,
+ 606
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We compare AT and NAT models with document context to the Transformer baseline without document context (trained on sentence) in Appendix B. The results suggest that document context enhances AT model (G-Transformer) by 0.70 s-BLEU. However, NAT models with document context still underperform the Transformer baseline, indicating that current NAT models have a limited ability to utilize document context.",
+ "bbox": [
+ 505,
+ 612,
+ 884,
+ 756
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We further apply an ablation study on the document context to quantify its contribution. As Table 3 illustrates, the performance of G-Trans+GLAT+CTC without document context drops by approximately 0.28 s-BLEU on average over the three benchmarks. Specifically, the target-side context contributes merely 0.01, while the source-side context contributes 0.27. The context contributions in G-Trans+GLAT+CTC are less than that in G-Transformer (0.23 and 0.70 s-BLEU on",
+ "bbox": [
+ 507,
+ 758,
+ 884,
+ 917
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "table",
+ "img_path": "images/6ec6a5ab82f0c1eea9bef9325f6994ddb6ccdf731bf829c9011d670ce2535cb5.jpg",
+ "table_caption": [],
+ "table_footnote": [
+ "Table 3: Impact of document context, evaluated in sBLEU. $\\diamond$ - the scores are from the paper report, where the model is trained on the raw datasets. The NAT models are trained on the KD datasets."
+ ],
+ "table_body": "| Method | TED (d-BLEU) | News (d-BLEU) | Europarl (d-BLEU) | Average (d-BLEU) | Speed -up |
| Raw | KD | Raw | KD | Raw | KD | Raw | KD |
| AT Baselines |
| G-Transformer (Bao et al., 2021) | 27.23 | 26.42 | 27.22 | 26.38 | 34.09 | 32.87 | 29.51 | 28.56 | - |
| Transformer (Vaswani et al., 2017) | 0.69 | 26.04 | 0.23 | 0.43 | 33.41 | 26.43 | 11.44 | 17.63 | 1.0x |
| Existing NAT Models |
| Vanilla NAT (Gu et al., 2018) | 0.45 | 0.34 | 0.12 | 0.61 | 1.30 | 2.43 | 0.62 | 1.13 | 41.2x |
| GLAT (Qian et al., 2021) | 1.77 | 0.03 | 0.01 | 2.56 | 2.13 | 8.17 | 1.30 | 3.59 | 40.0x |
| Latent-GLAT (Bao et al., 2022) | 2.10 | 1.87 | 0.30 | 2.30 | 5.26 | 4.25 | 2.55 | 2.81 | 29.2x |
| NAT+CTC (Libovický and Helcl, 2018) | 21.54 | 24.98 | 15.95 | 24.03 | 0.00 | 31.58 | 12.50 | 26.86 | 27.7x |
| GLAT+CTC (Qian et al., 2021) | 18.49 | 25.31 | 10.23 | 20.01 | 0.00 | 29.47 | 9.57 | 24.93 | 26.3x |
| DA-Transformer (Huang et al., 2022) | 20.47 | 25.02 | 13.99 | 23.37 | 30.34 | 32.14 | 21.60 | 26.84 | 30.8x |
| NAT Models with Sentence Alignment |
| G-Trans+GLAT (ours) | 19.96* | 24.23* | 15.14* | 23.04* | 25.67* | 31.45* | 20.26 | 26.24 | 30.0x |
| G-Trans+GLAT+CTC (ours) | 24.09* | 26.31* | 21.68* | 25.61* | 30.35* | 32.24* | 25.37 | 28.05 | 20.0x |
| G-Trans+DA-Trans (ours) | 23.45* | 25.73* | 21.43* | 24.70* | 30.36 | 32.29 | 25.08 | 27.57 | 25.1x |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.112,
+ 0.3,
+ 0.884,
+ 0.358
+ ],
+ "angle": 0,
+ "content": "Table 1: Main results on raw data (Raw) and knowledge distilled data (KD), where the failures are marked and investigated in section 5.1. We use the official code from the respective papers of NAT models, except NAT+CTC, which is implemented by ourselves based on the GLAT code. “*” denotes a statistical significance at the level of \\( p < 0.01 \\) using a t-test, compared to the corresponding baseline NAT model without sentence alignment."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.383,
+ 0.489,
+ 0.543
+ ],
+ "angle": 0,
+ "content": "G-Trans+GLAT+CTC integrates the sentence alignment with the CTC alignment. The default CTC algorithm aggregates all possible latent alignments across the entire sequence, which may align a source sentence to a wrong target sentence (e.g., the first source sentence to the second target sentence). Such global alignment not only slows down the training process but also causes unstable model performance. Different from the global alignment in Eq. 3, we apply CTC alignment on each sentence"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.156,
+ 0.557,
+ 0.488,
+ 0.6
+ ],
+ "angle": 0,
+ "content": "\\[\np _ {\\theta} (y | x) = \\prod_ {j = 1} ^ {K} \\left[ \\sum_ {a _ {j} \\in \\beta \\left(y _ {j}\\right)} \\prod_ {i = 1} ^ {M _ {j}} p _ {\\theta} \\left(a _ {j, i} \\mid x _ {j}, x\\right) \\right], \\tag {8}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.612,
+ 0.489,
+ 0.692
+ ],
+ "angle": 0,
+ "content": "where \\(M_{j}\\) denotes the reserved token positions for the \\(j\\)-th target sentence. Since CTC alignment is restricted inside each target sentence, the alignment space \\(\\beta(y_{j})\\) is enormously reduced compared to the \\(\\beta(y)\\) in Eq. 3."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.694,
+ 0.489,
+ 0.806
+ ],
+ "angle": 0,
+ "content": "G-Trans+DA-Trans introduces the sentence alignment into the output DAG of the DA-Transformer. The default DAG models the whole sequence, which enables the transition from a vertex in one sentence to a vertex in another sentence, making the transition space linearly increase with the enlargement of the sequence length."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.807,
+ 0.489,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "To address the issue, we enforce a constraint to isolate the vertex transitions of each sentence, forcing the path \\(a\\) on each sentence starting with a special vertex \\(\" < s > \"\\) and ending with a special vertex \\(\" < / s > \"\\). The transition between sentences only happens from the vertex \\(\" < / s > \"\\) of a previous sentence to the vertex \\(\" < s > \"\\) of the current sentence."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.383,
+ 0.884,
+ 0.48
+ ],
+ "angle": 0,
+ "content": "Formally, we use the same factorization as Eq. 4 but with a different collection \\(\\beta(y)\\) of paths. At the implementation level, we simply mask the transition matrix to disable transitions from one sentence to another sentence, so that the dynamic programming algorithm keeps unchanged."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.49,
+ 0.655,
+ 0.507
+ ],
+ "angle": 0,
+ "content": "4 Experiments"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.515,
+ 0.731,
+ 0.531
+ ],
+ "angle": 0,
+ "content": "4.1 Experimental Settings"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.535,
+ 0.884,
+ 0.761
+ ],
+ "angle": 0,
+ "content": "We evaluate the models using document-level MT benchmark (Maruf et al., 2019), which includes three datasets TED, News, and Europarl, representing three domains and various data scales for English-German translation. More details about each dataset and the preprocessing are in Appendix A.1. We follow Liu et al. (2020), evaluating the model performance in sentence-level BLEU score \\((s\\text{-BLEU})\\) and document-level BLEU score \\((d\\text{-BLEU})\\), which are explained in detail as Appendix A.2. We experiment on Base model (where Big model does not provide a stronger baseline) and evaluate the speedup using 1 GPU and 4 CPUs of a Tesla V100 environment as Appendix A.3."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.77,
+ 0.679,
+ 0.785
+ ],
+ "angle": 0,
+ "content": "4.2 Overall Results"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.791,
+ 0.883,
+ 0.839
+ ],
+ "angle": 0,
+ "content": "As shown in Table 1, NAT models achieve high speedup but still suffer from a significant performance gap with their AT counterparts."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.84,
+ 0.884,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "Speedup. The inference accelerations of NAT models on the document level are between 25x and 41x, which surpasses the accelerations on the sentence level (between 2x to 15x) by a big margin. Specifically, NAT and GLAT provide the biggest"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.085,
+ 0.49,
+ 0.229
+ ],
+ "angle": 0,
+ "content": "acceleration by around \\(40\\mathrm{x}\\). More complex models, such as Latent-GLAT, DA-Transformer, and G-Trans+GLAT, accelerate inference by approximately \\(30\\mathrm{x}\\). CTC-based models have lower accelerations between \\(20\\mathrm{x}\\) and \\(27\\mathrm{x}\\). On average, the acceleration on the document level is about \\(30\\mathrm{x}\\), which means that document-level MT systems could potentially save \\(96\\%\\) computational resources and energy consumption by using NAT models."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.231,
+ 0.49,
+ 0.344
+ ],
+ "angle": 0,
+ "content": "Performance. Though NAT models underperform G-Transformer, some outpace the Transformer. The lengthy text sequence challenges both the AT models and the NAT models, resulting in exceptionally low d-BLEU scores (<10) in some settings. We treat these low scores as model failures and investigate them in section 5.1."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.346,
+ 0.49,
+ 0.571
+ ],
+ "angle": 0,
+ "content": "With the help of sentence alignment, the performance gap between NAT models and AT baselines is largely reduced, especially when trained on KD data. For example, G-Trans+GLAT+CTC achieves an average d-BLEU of 28.05 on KD data, which is only 0.51 points lower than the d-BLEU of 28.56 of G-Transformer. However, its performance on Raw data is 4.14 points lower than G-Transformer, suggesting that NAT models experience severe challenges with raw data. These results demonstrate that even though knowledge distillation and sentence alignment enhance NAT models largely, there is still a gap between the NAT models and the strongest AT baseline."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.588,
+ 0.315,
+ 0.603
+ ],
+ "angle": 0,
+ "content": "4.3 Breakdown Results"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.612,
+ 0.49,
+ 0.821
+ ],
+ "angle": 0,
+ "content": "Sentence-level alignment. Sentence alignment substantially enhances the performance of NAT models. Specifically, for GLAT, sentence alignment enhances the performance on KD data from an average of 3.59 to 26.24 d-BLEU of G-Trans+GLAT. For GLAT+CTC, sentence alignment elevates the performance on KD data from 24.93 to 28.05 d-BLEU on average. This score is comparable to G-Transformer on KD data, leaving a minor gap of 0.51 d-BLEU on average. The better results of G-Trans+GLAT+CTC than G-Trans+GLAT suggest that sentence alignment complements output alignment techniques such as CTC."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.823,
+ 0.49,
+ 0.921
+ ],
+ "angle": 0,
+ "content": "The sentence alignment brings more benefits to models trained on raw data. As the average scores in Table 1 show, G-Trans+GLAT+CTC outperforms GLAT+CTC by 15.80 points on raw data and 3.12 points on KD data. G-Trans+DA-Trans outperforms DA-Transformer by 3.48 points on"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.085,
+ 0.885,
+ 0.133
+ ],
+ "angle": 0,
+ "content": "raw data and 0.73 points on KD data. These results suggest that sentence alignment mitigates the challenge brought by more modalities in raw data."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.136,
+ 0.886,
+ 0.345
+ ],
+ "angle": 0,
+ "content": "Token-level alignment. NAT models with implicit alignment such as vanilla NAT, GLAT, and Latent-GLAT fail on document-level MT, resulting in messy translations (repetitions, wrong words, and meaningless phrases). Conversely, NAT models with explicit alignments, such as NAT+CTC, GLAT+CTC, and DA-Transformer, produce superior document-level performance. DA-Transformer delivers the best overall performance among existing NAT models, even outperforming Transformer on Europarl (KD). These results suggest that token-level alignment plays an important role in NAT models to achieve good performance."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.348,
+ 0.886,
+ 0.573
+ ],
+ "angle": 0,
+ "content": "However, we also observe that training with CTC loss on documents is not always stable, leading to occasional failure and a high variance in their performance (e.g., +/-2.66 d-BLEU on raw data and +/-0.84 on KD data for GLAT+CTC). The failure happens more frequently on raw data than on KD data. We speculate that this instability is caused by the increased alignment space on long sequences, which can be mitigated by sentence alignment. Experiments on G-Trans+GLAT+CTC show stable training and lower variance in performance (+/- 0.19 d-BLEU). These results suggest that token-level alignment alone is not enough to achieve stable training and good performance."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.576,
+ 0.885,
+ 0.689
+ ],
+ "angle": 0,
+ "content": "Knowledge distillation. Intuitively, documents have more modalities because they are much longer than sentences. If it is the case, knowledge distillation will become more critical for training NAT models on document-level MT than on sentence-level MT. We compare NAT models trained with and without knowledge distillation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.692,
+ 0.886,
+ 0.884
+ ],
+ "angle": 0,
+ "content": "As Table 1 shows, NAT models trained on KD data generally outperform the same model trained on raw data. The improvement is especially significant for NAT models with explicit alignment. For example, DA-Transformer obtains an average of 26.84 d-BLEU on KD data, which is 5.24 points higher than 21.60 d-BLEU on raw data. In contrast, DA-Transformer on sentence-level MT achieves similar results on raw data and on KD data (Huang et al., 2022). The results suggest that compared to sentences, documents have more severe multimodality issues."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.509,
+ 0.888,
+ 0.885,
+ 0.921
+ ],
+ "angle": 0,
+ "content": "Although knowledge distillation enhances the performance of NAT models, it also sets a ceiling"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.173,
+ 0.083,
+ 0.422,
+ 0.196
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.114,
+ 0.206,
+ 0.488,
+ 0.234
+ ],
+ "angle": 0,
+ "content": "Figure 3: Speedup rises as long as the sequence length grows, evaluated on TED and News."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.267,
+ 0.49,
+ 0.427
+ ],
+ "angle": 0,
+ "content": "to the performance on document-level MT. Comparing the performance of G-Transformer on KD data to raw data, we see that KD data downgrades the performance by about 1 d-BLEU on average (from 29.51 to 28.56). These results differ from previous results on sentence-level MT (Huang et al., 2022), where knowledge distillation maintains or even enhances the performance of AT models. The discrepancy calls for further research on techniques to reduce the modalities on the document level."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.431,
+ 0.49,
+ 0.672
+ ],
+ "angle": 0,
+ "content": "Speedup on different sequence lengths. We evaluate GLAT and GLAT+CTC on various sequence lengths, including the single sentence and segments with a maximum length of 64, 128, 256, and 512. As Figure 3 shows, the speedup displays consistent trends among different datasets and NAT models. The GLAT generally has a higher speedup than GLAT-CTC, especially when the segment length goes above 256. The speedups on TED and News are almost identical, which is expected because the time costs are supposed to be irrelevant to the data domains. The trends suggest that we benefit more from the inference acceleration on longer documents, where document-level MT tasks provide NAT models with the best scenario."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.676,
+ 0.49,
+ 0.82
+ ],
+ "angle": 0,
+ "content": "Speedup on different batch sizes. The previous study (Helcl et al., 2022) reports that NAT models have limited acceleration on the sentence-level MT in parallel inference settings. We evaluate the models in the settings with different batch sizes, including 1, 2, 4, and 8 instances per batch. Given that document-level MT sequences are much longer than sentence-level MT sequences, we do not evaluate using a batch size larger than 8."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.823,
+ 0.49,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "As Figure 4 shows, when we increase the batch size from 1 to 8, the overall speedup of GLAT decreases from \\(40\\mathrm{x}\\) to \\(9\\mathrm{x}\\). However, if we consider a more strict calculation of the speedup, excluding the time for initializing the models (which takes almost constant time) from the total evaluation time,"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.568,
+ 0.083,
+ 0.819,
+ 0.196
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.508,
+ 0.206,
+ 0.885,
+ 0.264
+ ],
+ "angle": 0,
+ "content": "Figure 4: Speedup with various batch sizes evaluated on News, where the “/ex” conducts a strict calculation of speedup by excluding the time cost for model initialization from both the AT baseline and the NAT models."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.288,
+ 0.885,
+ 0.386
+ ],
+ "angle": 0,
+ "content": "the inference speedup of GLAT is \\(125\\mathrm{x}\\) and \\(44\\mathrm{x}\\) for the batch size of 1 and 8, respectively. These results suggest that although the speedup ratio decreases for bigger batch sizes, NAT models on document-level MT still show significant acceleration in the parallel inference settings."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.397,
+ 0.828,
+ 0.414
+ ],
+ "angle": 0,
+ "content": "5 The Challenge of Long Sequence"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.422,
+ 0.885,
+ 0.534
+ ],
+ "angle": 0,
+ "content": "The long input and output in document-level MT bring unexplored challenges to NAT models. In section 4, we learn that NAT models have a significant performance gap with their AT counterparts, and various NAT models even fail in some settings. We investigate these challenges in this section, leaving other discussions in Appendix D."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.545,
+ 0.771,
+ 0.56
+ ],
+ "angle": 0,
+ "content": "5.1 The Failure of NAT Models"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.566,
+ 0.885,
+ 0.822
+ ],
+ "angle": 0,
+ "content": "Previous study (Bao et al., 2021) suggests that Transformer fails on long sequences because of local minima of loss values during the training process. We first investigate this possibility for the cause of the failure on NAT models. We evaluate GLAT, GLAT+CTC, Transformer, and GTransformer on different lengths of sequences, as shown in Figure 5a. Overall, the four models produce different patterns of trends. The d-BLEU score of the Transformer rises as long as the sequence length increases until the failure happens at the length of 512 tokens. In contrast, the d-BLEU scores of the GLAT and GLAT+CTC descend progressively as long as the sequence length increases, which suggests a performance decline instead of a training failure on NAT models."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.824,
+ 0.887,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "Further evidence on the bigger dataset Europarl confirms that the low performance of NAT models is different from the failure of Transformer. Bao et al. (2021) suggests that the local minima encountered by Transformer can be escaped by using a bigger dataset, resulting in normal scores on Europarl."
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.117,
+ 0.082,
+ 0.35,
+ 0.188
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.127,
+ 0.193,
+ 0.337,
+ 0.219
+ ],
+ "angle": 0,
+ "content": "(a) NAT and AT performance on different sequence length."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.382,
+ 0.082,
+ 0.614,
+ 0.187
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.393,
+ 0.193,
+ 0.604,
+ 0.219
+ ],
+ "angle": 0,
+ "content": "(b) GLAT detailed BLEU scores on different sequence lengths."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.649,
+ 0.082,
+ 0.882,
+ 0.187
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.649,
+ 0.193,
+ 0.882,
+ 0.219
+ ],
+ "angle": 0,
+ "content": "(c) \\(GLAT + CTC\\) detailed BLEU scores on different sequence lengths."
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.113,
+ 0.229,
+ 0.882,
+ 0.259
+ ],
+ "angle": 0,
+ "content": "Figure 5: NAT vs. AT models on trends evaluated on News, where Transformer fails abruptly at the length of 512 while NAT models decrease progressively."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.283,
+ 0.489,
+ 0.379
+ ],
+ "angle": 0,
+ "content": "We evaluate Transformer, GLAT, and GLAT+CTC on Europarl, obtaining d-BLEU scores of 33.41, 2.13, and 0.00, respectively. We could see that GLAT and GLAT+CTC still give low scores on the bigger dataset, suggesting a different cause from the local minima."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.384,
+ 0.49,
+ 0.642
+ ],
+ "angle": 0,
+ "content": "We look into the d-BLEU score details, where BLEU-\\(n\\) (\\(n \\in [1,2,3,4]\\)) denotes the precision on \\(n\\)-gram. As the d-BLEU scores on GLAT in Figure 5b show, the BLEU-3/4 decreases rapidly, reaching almost 0.0 at the lengths of 256 and 512. The BLEU-1 remains relatively high at 27.3 on the length of 256, resulting in messy translations, where the generated tokens are related to the content but are repeated, disordered, or in the wrong collocation as shown in Table 2. In comparison, the d-BLEU scores on GLAT+CTC in Figure 5c decrease slowly, where the BLEU-1 even increases on the length of 512. We speculate the rapid decrease in BLEU-3/4 on GLAT is caused by the multi-modality and misalignment issues, which can be mitigated by explicit alignments, such as CTC."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.646,
+ 0.49,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "The exceptional zero scores. Table 1 reveals some unexpected results. Both NAT+CTC and GLAT+CTC score a 0.00 d-BLEU on the raw Europarl dataset, a performance notably inferior to that of NAT and GLAT. This unprecedented anomaly may stem from the challenges of applying CTC loss on long sequences. The table further indicates that both combinations, NAT/GLAT+CTC and NAT/GLAT, experience training failures on the raw Europarl dataset. These failures are likely due to multi-modality and misalignment issues. These results empirically demonstrate that, while CTC loss can be effective, its application to document-level training is not consistently stable. We hypothesize that this instability arises from the expanded alignment space between lengthy input and output sequences, as detailed in the token-level alignment"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.284,
+ 0.875,
+ 0.334
+ ],
+ "angle": 0,
+ "content": "Source: companies with a high percentage of floating rate debt stand to lose the most, Goldman said. outside pure stock plays, consumers stand to benefit as well through the rising dollar."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.338,
+ 0.876,
+ 0.387
+ ],
+ "angle": 0,
+ "content": "Target: Unternehmen mit einem hohen Anteil an flexiblen Zinsen werden am meisten verlieren, sagte Goldman. außerhalb der reinen Aktienspiele werden Verbraucher ebenfalls durch den steigenden Dollar profitieren."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.392,
+ 0.875,
+ 0.452
+ ],
+ "angle": 0,
+ "content": "GLAT: Unternehmen mit einem hohen, freiem freiem freien erverlieren die die meisten meisten, . . . te te dazu kräften profitieren profitieren profitieren profitieren durch durch steigenden steigenden profitieren profitieren."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.518,
+ 0.457,
+ 0.874,
+ 0.483
+ ],
+ "angle": 0,
+ "content": "GLAT+CTC: Unternehmen mit hohen der Schulden, die verlieren. außerhalb spielten Verbrauchener den profitieren."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.495,
+ 0.885,
+ 0.539
+ ],
+ "angle": 0,
+ "content": "Table 2: GLAT and GLAT+CTC produce good translation at the beginning but downgrade later with repetitions and missing translations."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.509,
+ 0.565,
+ 0.702,
+ 0.58
+ ],
+ "angle": 0,
+ "content": "discussion in Section 4.3."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.593,
+ 0.704,
+ 0.607
+ ],
+ "angle": 0,
+ "content": "5.2 Document Context"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.507,
+ 0.613,
+ 0.885,
+ 0.757
+ ],
+ "angle": 0,
+ "content": "We compare AT and NAT models with document context to the Transformer baseline without document context (trained on sentence) in Appendix B. The results suggest that document context enhances AT model (G-Transformer) by 0.70 s-BLEU. However, NAT models with document context still underperform the Transformer baseline, indicating that current NAT models have a limited ability to utilize document context."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.759,
+ 0.885,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "We further apply an ablation study on the document context to quantify its contribution. As Table 3 illustrates, the performance of G-Trans+GLAT+CTC without document context drops by approximately 0.28 s-BLEU on average over the three benchmarks. Specifically, the target-side context contributes merely 0.01, while the source-side context contributes 0.27. The context contributions in G-Trans+GLAT+CTC are less than that in G-Transformer (0.23 and 0.70 s-BLEU on"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.115,
+ 0.082,
+ 0.488,
+ 0.169
+ ],
+ "angle": 0,
+ "content": "| Method | TED (d-BLEU) | News (d-BLEU) | Europarl (d-BLEU) | Average (d-BLEU) | Speed -up |
| Raw | KD | Raw | KD | Raw | KD | Raw | KD |
| AT Baselines |
| G-Transformer (Bao et al., 2021) | 27.23 | 26.42 | 27.22 | 26.38 | 34.09 | 32.87 | 29.51 | 28.56 | - |
| Transformer (Vaswani et al., 2017) | 0.69 | 26.04 | 0.23 | 0.43 | 33.41 | 26.43 | 11.44 | 17.63 | 1.0x |
| Existing NAT Models |
| Vanilla NAT (Gu et al., 2018) | 0.45 | 0.34 | 0.12 | 0.61 | 1.30 | 2.43 | 0.62 | 1.13 | 41.2x |
| GLAT (Qian et al., 2021) | 1.77 | 0.03 | 0.01 | 2.56 | 2.13 | 8.17 | 1.30 | 3.59 | 40.0x |
| Latent-GLAT (Bao et al., 2022) | 2.10 | 1.87 | 0.30 | 2.30 | 5.26 | 4.25 | 2.55 | 2.81 | 29.2x |
| NAT+CTC (Libovický and Helcl, 2018) | 21.54 | 24.98 | 15.95 | 24.03 | 0.00 | 31.58 | 12.50 | 26.86 | 27.7x |
| GLAT+CTC (Qian et al., 2021) | 18.49 | 25.31 | 10.23 | 20.01 | 0.00 | 29.47 | 9.57 | 24.93 | 26.3x |
| DA-Transformer (Huang et al., 2022) | 20.47 | 25.02 | 13.99 | 23.37 | 30.34 | 32.14 | 21.60 | 26.84 | 30.8x |
| NAT Models with Sentence Alignment |
| G-Trans+GLAT (ours) | 19.96* | 24.23* | 15.14* | 23.04* | 25.67* | 31.45* | 20.26 | 26.24 | 30.0x |
| G-Trans+GLAT+CTC (ours) | 24.09* | 26.31* | 21.68* | 25.61* | 30.35* | 32.24* | 25.37 | 28.05 | 20.0x |
| G-Trans+DA-Trans (ours) | 23.45* | 25.73* | 21.43* | 24.70* | 30.36 | 32.29 | 25.08 | 27.57 | 25.1x |
+
+Table 1: Main results on raw data (Raw) and knowledge distilled data (KD), where the failures are marked and investigated in section 5.1. We use the official code from the respective papers of NAT models, except NAT+CTC, which is implemented by ourselves based on the GLAT code. “*” denotes a statistical significance at the level of $p < 0.01$ using a t-test, compared to the corresponding baseline NAT model without sentence alignment.
+
+G-Trans+GLAT+CTC integrates the sentence alignment with the CTC alignment. The default CTC algorithm aggregates all possible latent alignments across the entire sequence, which may align a source sentence to a wrong target sentence (e.g., the first source sentence to the second target sentence). Such global alignment not only slows down the training process but also causes unstable model performance. Different from the global alignment in Eq. 3, we apply CTC alignment on each sentence
+
+$$
+p _ {\theta} (y | x) = \prod_ {j = 1} ^ {K} \left[ \sum_ {a _ {j} \in \beta \left(y _ {j}\right)} \prod_ {i = 1} ^ {M _ {j}} p _ {\theta} \left(a _ {j, i} \mid x _ {j}, x\right) \right], \tag {8}
+$$
+
+where $M_{j}$ denotes the reserved token positions for the $j$ -th target sentence. Since CTC alignment is restricted inside each target sentence, the alignment space $\beta(y_{j})$ is enormously reduced compared to the $\beta(y)$ in Eq. 3.
+
+G-Trans+DA-Trans introduces the sentence alignment into the output DAG of the DA-Transformer. The default DAG models the whole sequence, which enables the transition from a vertex in one sentence to a vertex in another sentence, making the transition space linearly increase with the enlargement of the sequence length.
+
+To address the issue, we enforce a constraint to isolate the vertex transitions of each sentence, forcing the path $a$ on each sentence starting with a special vertex $" < s > "$ and ending with a special vertex $" < / s > "$ . The transition between sentences only happens from the vertex $" < / s > "$ of a previous sentence to the vertex $" < s > "$ of the current sentence.
+
+Formally, we use the same factorization as Eq. 4 but with a different collection $\beta(y)$ of paths. At the implementation level, we simply mask the transition matrix to disable transitions from one sentence to another sentence, so that the dynamic programming algorithm keeps unchanged.
+
+# 4 Experiments
+
+# 4.1 Experimental Settings
+
+We evaluate the models using document-level MT benchmark (Maruf et al., 2019), which includes three datasets TED, News, and Europarl, representing three domains and various data scales for English-German translation. More details about each dataset and the preprocessing are in Appendix A.1. We follow Liu et al. (2020), evaluating the model performance in sentence-level BLEU score $(s\text{-BLEU})$ and document-level BLEU score $(d\text{-BLEU})$ , which are explained in detail as Appendix A.2. We experiment on Base model (where Big model does not provide a stronger baseline) and evaluate the speedup using 1 GPU and 4 CPUs of a Tesla V100 environment as Appendix A.3.
+
+# 4.2 Overall Results
+
+As shown in Table 1, NAT models achieve high speedup but still suffer from a significant performance gap with their AT counterparts.
+
+Speedup. The inference accelerations of NAT models on the document level are between 25x and 41x, which surpasses the accelerations on the sentence level (between 2x to 15x) by a big margin. Specifically, NAT and GLAT provide the biggest
+
+acceleration by around $40\mathrm{x}$ . More complex models, such as Latent-GLAT, DA-Transformer, and G-Trans+GLAT, accelerate inference by approximately $30\mathrm{x}$ . CTC-based models have lower accelerations between $20\mathrm{x}$ and $27\mathrm{x}$ . On average, the acceleration on the document level is about $30\mathrm{x}$ , which means that document-level MT systems could potentially save $96\%$ computational resources and energy consumption by using NAT models.
+
+Performance. Though NAT models underperform G-Transformer, some outpace the Transformer. The lengthy text sequence challenges both the AT models and the NAT models, resulting in exceptionally low d-BLEU scores (<10) in some settings. We treat these low scores as model failures and investigate them in section 5.1.
+
+With the help of sentence alignment, the performance gap between NAT models and AT baselines is largely reduced, especially when trained on KD data. For example, G-Trans+GLAT+CTC achieves an average d-BLEU of 28.05 on KD data, which is only 0.51 points lower than the d-BLEU of 28.56 of G-Transformer. However, its performance on Raw data is 4.14 points lower than G-Transformer, suggesting that NAT models experience severe challenges with raw data. These results demonstrate that even though knowledge distillation and sentence alignment enhance NAT models largely, there is still a gap between the NAT models and the strongest AT baseline.
+
+# 4.3 Breakdown Results
+
+Sentence-level alignment. Sentence alignment substantially enhances the performance of NAT models. Specifically, for GLAT, sentence alignment enhances the performance on KD data from an average of 3.59 to 26.24 d-BLEU of G-Trans+GLAT. For GLAT+CTC, sentence alignment elevates the performance on KD data from 24.93 to 28.05 d-BLEU on average. This score is comparable to G-Transformer on KD data, leaving a minor gap of 0.51 d-BLEU on average. The better results of G-Trans+GLAT+CTC than G-Trans+GLAT suggest that sentence alignment complements output alignment techniques such as CTC.
+
+The sentence alignment brings more benefits to models trained on raw data. As the average scores in Table 1 show, G-Trans+GLAT+CTC outperforms GLAT+CTC by 15.80 points on raw data and 3.12 points on KD data. G-Trans+DA-Trans outperforms DA-Transformer by 3.48 points on
+
+raw data and 0.73 points on KD data. These results suggest that sentence alignment mitigates the challenge brought by more modalities in raw data.
+
+Token-level alignment. NAT models with implicit alignment such as vanilla NAT, GLAT, and Latent-GLAT fail on document-level MT, resulting in messy translations (repetitions, wrong words, and meaningless phrases). Conversely, NAT models with explicit alignments, such as NAT+CTC, GLAT+CTC, and DA-Transformer, produce superior document-level performance. DA-Transformer delivers the best overall performance among existing NAT models, even outperforming Transformer on Europarl (KD). These results suggest that token-level alignment plays an important role in NAT models to achieve good performance.
+
+However, we also observe that training with CTC loss on documents is not always stable, leading to occasional failure and a high variance in their performance (e.g., +/-2.66 d-BLEU on raw data and +/-0.84 on KD data for GLAT+CTC). The failure happens more frequently on raw data than on KD data. We speculate that this instability is caused by the increased alignment space on long sequences, which can be mitigated by sentence alignment. Experiments on G-Trans+GLAT+CTC show stable training and lower variance in performance (+/- 0.19 d-BLEU). These results suggest that token-level alignment alone is not enough to achieve stable training and good performance.
+
+Knowledge distillation. Intuitively, documents have more modalities because they are much longer than sentences. If it is the case, knowledge distillation will become more critical for training NAT models on document-level MT than on sentence-level MT. We compare NAT models trained with and without knowledge distillation.
+
+As Table 1 shows, NAT models trained on KD data generally outperform the same model trained on raw data. The improvement is especially significant for NAT models with explicit alignment. For example, DA-Transformer obtains an average of 26.84 d-BLEU on KD data, which is 5.24 points higher than 21.60 d-BLEU on raw data. In contrast, DA-Transformer on sentence-level MT achieves similar results on raw data and on KD data (Huang et al., 2022). The results suggest that compared to sentences, documents have more severe multimodality issues.
+
+Although knowledge distillation enhances the performance of NAT models, it also sets a ceiling
+
+
+Figure 3: Speedup rises as long as the sequence length grows, evaluated on TED and News.
+
+to the performance on document-level MT. Comparing the performance of G-Transformer on KD data to raw data, we see that KD data downgrades the performance by about 1 d-BLEU on average (from 29.51 to 28.56). These results differ from previous results on sentence-level MT (Huang et al., 2022), where knowledge distillation maintains or even enhances the performance of AT models. The discrepancy calls for further research on techniques to reduce the modalities on the document level.
+
+Speedup on different sequence lengths. We evaluate GLAT and GLAT+CTC on various sequence lengths, including the single sentence and segments with a maximum length of 64, 128, 256, and 512. As Figure 3 shows, the speedup displays consistent trends among different datasets and NAT models. The GLAT generally has a higher speedup than GLAT-CTC, especially when the segment length goes above 256. The speedups on TED and News are almost identical, which is expected because the time costs are supposed to be irrelevant to the data domains. The trends suggest that we benefit more from the inference acceleration on longer documents, where document-level MT tasks provide NAT models with the best scenario.
+
+Speedup on different batch sizes. The previous study (Helcl et al., 2022) reports that NAT models have limited acceleration on the sentence-level MT in parallel inference settings. We evaluate the models in the settings with different batch sizes, including 1, 2, 4, and 8 instances per batch. Given that document-level MT sequences are much longer than sentence-level MT sequences, we do not evaluate using a batch size larger than 8.
+
+As Figure 4 shows, when we increase the batch size from 1 to 8, the overall speedup of GLAT decreases from $40\mathrm{x}$ to $9\mathrm{x}$ . However, if we consider a more strict calculation of the speedup, excluding the time for initializing the models (which takes almost constant time) from the total evaluation time,
+
+
+Figure 4: Speedup with various batch sizes evaluated on News, where the “/ex” conducts a strict calculation of speedup by excluding the time cost for model initialization from both the AT baseline and the NAT models.
+
+the inference speedup of GLAT is $125\mathrm{x}$ and $44\mathrm{x}$ for the batch size of 1 and 8, respectively. These results suggest that although the speedup ratio decreases for bigger batch sizes, NAT models on document-level MT still show significant acceleration in the parallel inference settings.
+
+# 5 The Challenge of Long Sequence
+
+The long input and output in document-level MT bring unexplored challenges to NAT models. In section 4, we learn that NAT models have a significant performance gap with their AT counterparts, and various NAT models even fail in some settings. We investigate these challenges in this section, leaving other discussions in Appendix D.
+
+# 5.1 The Failure of NAT Models
+
+Previous study (Bao et al., 2021) suggests that Transformer fails on long sequences because of local minima of loss values during the training process. We first investigate this possibility for the cause of the failure on NAT models. We evaluate GLAT, GLAT+CTC, Transformer, and GTransformer on different lengths of sequences, as shown in Figure 5a. Overall, the four models produce different patterns of trends. The d-BLEU score of the Transformer rises as long as the sequence length increases until the failure happens at the length of 512 tokens. In contrast, the d-BLEU scores of the GLAT and GLAT+CTC descend progressively as long as the sequence length increases, which suggests a performance decline instead of a training failure on NAT models.
+
+Further evidence on the bigger dataset Europarl confirms that the low performance of NAT models is different from the failure of Transformer. Bao et al. (2021) suggests that the local minima encountered by Transformer can be escaped by using a bigger dataset, resulting in normal scores on Europarl.
+
+
+(a) NAT and AT performance on different sequence length.
+
+
+(b) GLAT detailed BLEU scores on different sequence lengths.
+Figure 5: NAT vs. AT models on trends evaluated on News, where Transformer fails abruptly at the length of 512 while NAT models decrease progressively.
+
+
+(c) $GLAT + CTC$ detailed BLEU scores on different sequence lengths.
+
+We evaluate Transformer, GLAT, and GLAT+CTC on Europarl, obtaining d-BLEU scores of 33.41, 2.13, and 0.00, respectively. We could see that GLAT and GLAT+CTC still give low scores on the bigger dataset, suggesting a different cause from the local minima.
+
+We look into the d-BLEU score details, where BLEU- $n$ ( $n \in [1,2,3,4]$ ) denotes the precision on $n$ -gram. As the d-BLEU scores on GLAT in Figure 5b show, the BLEU-3/4 decreases rapidly, reaching almost 0.0 at the lengths of 256 and 512. The BLEU-1 remains relatively high at 27.3 on the length of 256, resulting in messy translations, where the generated tokens are related to the content but are repeated, disordered, or in the wrong collocation as shown in Table 2. In comparison, the d-BLEU scores on GLAT+CTC in Figure 5c decrease slowly, where the BLEU-1 even increases on the length of 512. We speculate the rapid decrease in BLEU-3/4 on GLAT is caused by the multi-modality and misalignment issues, which can be mitigated by explicit alignments, such as CTC.
+
+The exceptional zero scores. Table 1 reveals some unexpected results. Both NAT+CTC and GLAT+CTC score a 0.00 d-BLEU on the raw Europarl dataset, a performance notably inferior to that of NAT and GLAT. This unprecedented anomaly may stem from the challenges of applying CTC loss on long sequences. The table further indicates that both combinations, NAT/GLAT+CTC and NAT/GLAT, experience training failures on the raw Europarl dataset. These failures are likely due to multi-modality and misalignment issues. These results empirically demonstrate that, while CTC loss can be effective, its application to document-level training is not consistently stable. We hypothesize that this instability arises from the expanded alignment space between lengthy input and output sequences, as detailed in the token-level alignment
+
+Source: companies with a high percentage of floating rate debt stand to lose the most, Goldman said. outside pure stock plays, consumers stand to benefit as well through the rising dollar.
+
+Target: Unternehmen mit einem hohen Anteil an flexiblen Zinsen werden am meisten verlieren, sagte Goldman. außerhalb der reinen Aktienspiele werden Verbraucher ebenfalls durch den steigenden Dollar profitieren.
+
+GLAT: Unternehmen mit einem hohen, freiem freiem freien erverlieren die die meisten meisten, . . . te te dazu kräften profitieren profitieren profitieren profitieren durch durch steigenden steigenden profitieren profitieren.
+
+GLAT+CTC: Unternehmen mit hohen der Schulden, die verlieren. außerhalb spielten Verbrauchener den profitieren.
+
+Table 2: GLAT and GLAT+CTC produce good translation at the beginning but downgrade later with repetitions and missing translations.
+
+discussion in Section 4.3.
+
+# 5.2 Document Context
+
+We compare AT and NAT models with document context to the Transformer baseline without document context (trained on sentence) in Appendix B. The results suggest that document context enhances AT model (G-Transformer) by 0.70 s-BLEU. However, NAT models with document context still underperform the Transformer baseline, indicating that current NAT models have a limited ability to utilize document context.
+
+We further apply an ablation study on the document context to quantify its contribution. As Table 3 illustrates, the performance of G-Trans+GLAT+CTC without document context drops by approximately 0.28 s-BLEU on average over the three benchmarks. Specifically, the target-side context contributes merely 0.01, while the source-side context contributes 0.27. The context contributions in G-Trans+GLAT+CTC are less than that in G-Transformer (0.23 and 0.70 s-BLEU on
+
+| Method | TED (d-BLEU) | News (d-BLEU) | Europarl (d-BLEU) | Average (d-BLEU) | Speed -up |
| Raw | KD | Raw | KD | Raw | KD | Raw | KD |
| AT Baselines |
| G-Transformer (Bao et al., 2021) | 27.23 | 26.42 | 27.22 | 26.38 | 34.09 | 32.87 | 29.51 | 28.56 | - |
| Transformer (Vaswani et al., 2017) | 0.69 | 26.04 | 0.23 | 0.43 | 33.41 | 26.43 | 11.44 | 17.63 | 1.0x |
| Existing NAT Models |
| Vanilla NAT (Gu et al., 2018) | 0.45 | 0.34 | 0.12 | 0.61 | 1.30 | 2.43 | 0.62 | 1.13 | 41.2x |
| GLAT (Qian et al., 2021) | 1.77 | 0.03 | 0.01 | 2.56 | 2.13 | 8.17 | 1.30 | 3.59 | 40.0x |
| Latent-GLAT (Bao et al., 2022) | 2.10 | 1.87 | 0.30 | 2.30 | 5.26 | 4.25 | 2.55 | 2.81 | 29.2x |
| NAT+CTC (Libovický and Helcl, 2018) | 21.54 | 24.98 | 15.95 | 24.03 | 0.00 | 31.58 | 12.50 | 26.86 | 27.7x |
| GLAT+CTC (Qian et al., 2021) | 18.49 | 25.31 | 10.23 | 20.01 | 0.00 | 29.47 | 9.57 | 24.93 | 26.3x |
| DA-Transformer (Huang et al., 2022) | 20.47 | 25.02 | 13.99 | 23.37 | 30.34 | 32.14 | 21.60 | 26.84 | 30.8x |
| NAT Models with Sentence Alignment |
| G-Trans+GLAT (ours) | 19.96* | 24.23* | 15.14* | 23.04* | 25.67* | 31.45* | 20.26 | 26.24 | 30.0x |
| G-Trans+GLAT+CTC (ours) | 24.09* | 26.31* | 21.68* | 25.61* | 30.35* | 32.24* | 25.37 | 28.05 | 20.0x |
| G-Trans+DA-Trans (ours) | 23.45* | 25.73* | 21.43* | 24.70* | 30.36 | 32.29 | 25.08 | 27.57 | 25.1x |
",
+ "image_path": "4c63ecfc31fb0ca25562e43d855fb184178afac08c75a28218c748d7a66140c1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 66,
+ 252,
+ 525,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 66,
+ 252,
+ 525,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 66,
+ 252,
+ 525,
+ 301
+ ],
+ "type": "text",
+ "content": "Table 1: Main results on raw data (Raw) and knowledge distilled data (KD), where the failures are marked and investigated in section 5.1. We use the official code from the respective papers of NAT models, except NAT+CTC, which is implemented by ourselves based on the GLAT code. “*” denotes a statistical significance at the level of "
+ },
+ {
+ "bbox": [
+ 66,
+ 252,
+ 525,
+ 301
+ ],
+ "type": "inline_equation",
+ "content": "p < 0.01"
+ },
+ {
+ "bbox": [
+ 66,
+ 252,
+ 525,
+ 301
+ ],
+ "type": "text",
+ "content": " using a t-test, compared to the corresponding baseline NAT model without sentence alignment."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 290,
+ 456
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 322,
+ 290,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 322,
+ 290,
+ 456
+ ],
+ "type": "text",
+ "content": "G-Trans+GLAT+CTC integrates the sentence alignment with the CTC alignment. The default CTC algorithm aggregates all possible latent alignments across the entire sequence, which may align a source sentence to a wrong target sentence (e.g., the first source sentence to the second target sentence). Such global alignment not only slows down the training process but also causes unstable model performance. Different from the global alignment in Eq. 3, we apply CTC alignment on each sentence"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 92,
+ 468,
+ 290,
+ 504
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 92,
+ 468,
+ 290,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 92,
+ 468,
+ 290,
+ 504
+ ],
+ "type": "interline_equation",
+ "content": "p _ {\\theta} (y | x) = \\prod_ {j = 1} ^ {K} \\left[ \\sum_ {a _ {j} \\in \\beta \\left(y _ {j}\\right)} \\prod_ {i = 1} ^ {M _ {j}} p _ {\\theta} \\left(a _ {j, i} \\mid x _ {j}, x\\right) \\right], \\tag {8}",
+ "image_path": "ce61ffd3e830fb6934b663e13d3cb36cf52c283015540d82d4d60203b4b1c3e0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "inline_equation",
+ "content": "M_{j}"
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "text",
+ "content": " denotes the reserved token positions for the "
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "inline_equation",
+ "content": "j"
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "text",
+ "content": "-th target sentence. Since CTC alignment is restricted inside each target sentence, the alignment space "
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "inline_equation",
+ "content": "\\beta(y_{j})"
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "text",
+ "content": " is enormously reduced compared to the "
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "inline_equation",
+ "content": "\\beta(y)"
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 290,
+ 581
+ ],
+ "type": "text",
+ "content": " in Eq. 3."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 583,
+ 290,
+ 677
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 583,
+ 290,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 583,
+ 290,
+ 677
+ ],
+ "type": "text",
+ "content": "G-Trans+DA-Trans introduces the sentence alignment into the output DAG of the DA-Transformer. The default DAG models the whole sequence, which enables the transition from a vertex in one sentence to a vertex in another sentence, making the transition space linearly increase with the enlargement of the sequence length."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "content": "To address the issue, we enforce a constraint to isolate the vertex transitions of each sentence, forcing the path "
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "a"
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "content": " on each sentence starting with a special vertex "
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\" < s > \""
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "content": " and ending with a special vertex "
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\" < / s > \""
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "content": ". The transition between sentences only happens from the vertex "
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\" < / s > \""
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "content": " of a previous sentence to the vertex "
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\" < s > \""
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "content": " of the current sentence."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 302,
+ 322,
+ 525,
+ 403
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 322,
+ 525,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 322,
+ 525,
+ 403
+ ],
+ "type": "text",
+ "content": "Formally, we use the same factorization as Eq. 4 but with a different collection "
+ },
+ {
+ "bbox": [
+ 302,
+ 322,
+ 525,
+ 403
+ ],
+ "type": "inline_equation",
+ "content": "\\beta(y)"
+ },
+ {
+ "bbox": [
+ 302,
+ 322,
+ 525,
+ 403
+ ],
+ "type": "text",
+ "content": " of paths. At the implementation level, we simply mask the transition matrix to disable transitions from one sentence to another sentence, so that the dynamic programming algorithm keeps unchanged."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 302,
+ 412,
+ 389,
+ 426
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 412,
+ 389,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 412,
+ 389,
+ 426
+ ],
+ "type": "text",
+ "content": "4 Experiments"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 433,
+ 434,
+ 446
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 433,
+ 434,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 433,
+ 434,
+ 446
+ ],
+ "type": "text",
+ "content": "4.1 Experimental Settings"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 449,
+ 525,
+ 640
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 449,
+ 525,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 449,
+ 525,
+ 640
+ ],
+ "type": "text",
+ "content": "We evaluate the models using document-level MT benchmark (Maruf et al., 2019), which includes three datasets TED, News, and Europarl, representing three domains and various data scales for English-German translation. More details about each dataset and the preprocessing are in Appendix A.1. We follow Liu et al. (2020), evaluating the model performance in sentence-level BLEU score "
+ },
+ {
+ "bbox": [
+ 302,
+ 449,
+ 525,
+ 640
+ ],
+ "type": "inline_equation",
+ "content": "(s\\text{-BLEU})"
+ },
+ {
+ "bbox": [
+ 302,
+ 449,
+ 525,
+ 640
+ ],
+ "type": "text",
+ "content": " and document-level BLEU score "
+ },
+ {
+ "bbox": [
+ 302,
+ 449,
+ 525,
+ 640
+ ],
+ "type": "inline_equation",
+ "content": "(d\\text{-BLEU})"
+ },
+ {
+ "bbox": [
+ 302,
+ 449,
+ 525,
+ 640
+ ],
+ "type": "text",
+ "content": ", which are explained in detail as Appendix A.2. We experiment on Base model (where Big model does not provide a stronger baseline) and evaluate the speedup using 1 GPU and 4 CPUs of a Tesla V100 environment as Appendix A.3."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 647,
+ 404,
+ 660
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 647,
+ 404,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 647,
+ 404,
+ 660
+ ],
+ "type": "text",
+ "content": "4.2 Overall Results"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 302,
+ 665,
+ 525,
+ 705
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 665,
+ 525,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 665,
+ 525,
+ 705
+ ],
+ "type": "text",
+ "content": "As shown in Table 1, NAT models achieve high speedup but still suffer from a significant performance gap with their AT counterparts."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 302,
+ 706,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 706,
+ 525,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 706,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": "Speedup. The inference accelerations of NAT models on the document level are between 25x and 41x, which surpasses the accelerations on the sentence level (between 2x to 15x) by a big margin. Specifically, NAT and GLAT provide the biggest"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "discarded_blocks": [],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": "acceleration by around "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "40\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": ". More complex models, such as Latent-GLAT, DA-Transformer, and G-Trans+GLAT, accelerate inference by approximately "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "30\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": ". CTC-based models have lower accelerations between "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "20\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "27\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": ". On average, the acceleration on the document level is about "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "30\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": ", which means that document-level MT systems could potentially save "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "96\\%"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " computational resources and energy consumption by using NAT models."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 194,
+ 291,
+ 289
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 194,
+ 291,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 194,
+ 291,
+ 289
+ ],
+ "type": "text",
+ "content": "Performance. Though NAT models underperform G-Transformer, some outpace the Transformer. The lengthy text sequence challenges both the AT models and the NAT models, resulting in exceptionally low d-BLEU scores (<10) in some settings. We treat these low scores as model failures and investigate them in section 5.1."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 69,
+ 290,
+ 291,
+ 480
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 290,
+ 291,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 290,
+ 291,
+ 480
+ ],
+ "type": "text",
+ "content": "With the help of sentence alignment, the performance gap between NAT models and AT baselines is largely reduced, especially when trained on KD data. For example, G-Trans+GLAT+CTC achieves an average d-BLEU of 28.05 on KD data, which is only 0.51 points lower than the d-BLEU of 28.56 of G-Transformer. However, its performance on Raw data is 4.14 points lower than G-Transformer, suggesting that NAT models experience severe challenges with raw data. These results demonstrate that even though knowledge distillation and sentence alignment enhance NAT models largely, there is still a gap between the NAT models and the strongest AT baseline."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 494,
+ 187,
+ 507
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 494,
+ 187,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 494,
+ 187,
+ 507
+ ],
+ "type": "text",
+ "content": "4.3 Breakdown Results"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 291,
+ 690
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 514,
+ 291,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 514,
+ 291,
+ 690
+ ],
+ "type": "text",
+ "content": "Sentence-level alignment. Sentence alignment substantially enhances the performance of NAT models. Specifically, for GLAT, sentence alignment enhances the performance on KD data from an average of 3.59 to 26.24 d-BLEU of G-Trans+GLAT. For GLAT+CTC, sentence alignment elevates the performance on KD data from 24.93 to 28.05 d-BLEU on average. This score is comparable to G-Transformer on KD data, leaving a minor gap of 0.51 d-BLEU on average. The better results of G-Trans+GLAT+CTC than G-Trans+GLAT suggest that sentence alignment complements output alignment techniques such as CTC."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 774
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 774
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 774
+ ],
+ "type": "text",
+ "content": "The sentence alignment brings more benefits to models trained on raw data. As the average scores in Table 1 show, G-Trans+GLAT+CTC outperforms GLAT+CTC by 15.80 points on raw data and 3.12 points on KD data. G-Trans+DA-Trans outperforms DA-Transformer by 3.48 points on"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 111
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 111
+ ],
+ "type": "text",
+ "content": "raw data and 0.73 points on KD data. These results suggest that sentence alignment mitigates the challenge brought by more modalities in raw data."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 302,
+ 114,
+ 527,
+ 290
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 114,
+ 527,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 114,
+ 527,
+ 290
+ ],
+ "type": "text",
+ "content": "Token-level alignment. NAT models with implicit alignment such as vanilla NAT, GLAT, and Latent-GLAT fail on document-level MT, resulting in messy translations (repetitions, wrong words, and meaningless phrases). Conversely, NAT models with explicit alignments, such as NAT+CTC, GLAT+CTC, and DA-Transformer, produce superior document-level performance. DA-Transformer delivers the best overall performance among existing NAT models, even outperforming Transformer on Europarl (KD). These results suggest that token-level alignment plays an important role in NAT models to achieve good performance."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 302,
+ 292,
+ 527,
+ 481
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 292,
+ 527,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 292,
+ 527,
+ 481
+ ],
+ "type": "text",
+ "content": "However, we also observe that training with CTC loss on documents is not always stable, leading to occasional failure and a high variance in their performance (e.g., +/-2.66 d-BLEU on raw data and +/-0.84 on KD data for GLAT+CTC). The failure happens more frequently on raw data than on KD data. We speculate that this instability is caused by the increased alignment space on long sequences, which can be mitigated by sentence alignment. Experiments on G-Trans+GLAT+CTC show stable training and lower variance in performance (+/- 0.19 d-BLEU). These results suggest that token-level alignment alone is not enough to achieve stable training and good performance."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 484,
+ 526,
+ 579
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 484,
+ 526,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 484,
+ 526,
+ 579
+ ],
+ "type": "text",
+ "content": "Knowledge distillation. Intuitively, documents have more modalities because they are much longer than sentences. If it is the case, knowledge distillation will become more critical for training NAT models on document-level MT than on sentence-level MT. We compare NAT models trained with and without knowledge distillation."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 581,
+ 527,
+ 743
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 581,
+ 527,
+ 743
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 581,
+ 527,
+ 743
+ ],
+ "type": "text",
+ "content": "As Table 1 shows, NAT models trained on KD data generally outperform the same model trained on raw data. The improvement is especially significant for NAT models with explicit alignment. For example, DA-Transformer obtains an average of 26.84 d-BLEU on KD data, which is 5.24 points higher than 21.60 d-BLEU on raw data. In contrast, DA-Transformer on sentence-level MT achieves similar results on raw data and on KD data (Huang et al., 2022). The results suggest that compared to sentences, documents have more severe multimodality issues."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 526,
+ 774
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 746,
+ 526,
+ 774
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 746,
+ 526,
+ 774
+ ],
+ "type": "text",
+ "content": "Although knowledge distillation enhances the performance of NAT models, it also sets a ceiling"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "discarded_blocks": [],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 102,
+ 69,
+ 251,
+ 164
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 102,
+ 69,
+ 251,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 69,
+ 251,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 69,
+ 251,
+ 164
+ ],
+ "type": "image",
+ "image_path": "5ee81d0b79c72258487b0ab4486f91c348b87b205008822352882329a3e312bb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 67,
+ 173,
+ 290,
+ 196
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 173,
+ 290,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 173,
+ 290,
+ 196
+ ],
+ "type": "text",
+ "content": "Figure 3: Speedup rises as long as the sequence length grows, evaluated on TED and News."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 224,
+ 291,
+ 359
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 224,
+ 291,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 224,
+ 291,
+ 359
+ ],
+ "type": "text",
+ "content": "to the performance on document-level MT. Comparing the performance of G-Transformer on KD data to raw data, we see that KD data downgrades the performance by about 1 d-BLEU on average (from 29.51 to 28.56). These results differ from previous results on sentence-level MT (Huang et al., 2022), where knowledge distillation maintains or even enhances the performance of AT models. The discrepancy calls for further research on techniques to reduce the modalities on the document level."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 69,
+ 362,
+ 291,
+ 565
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 362,
+ 291,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 362,
+ 291,
+ 565
+ ],
+ "type": "text",
+ "content": "Speedup on different sequence lengths. We evaluate GLAT and GLAT+CTC on various sequence lengths, including the single sentence and segments with a maximum length of 64, 128, 256, and 512. As Figure 3 shows, the speedup displays consistent trends among different datasets and NAT models. The GLAT generally has a higher speedup than GLAT-CTC, especially when the segment length goes above 256. The speedups on TED and News are almost identical, which is expected because the time costs are supposed to be irrelevant to the data domains. The trends suggest that we benefit more from the inference acceleration on longer documents, where document-level MT tasks provide NAT models with the best scenario."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 568,
+ 291,
+ 689
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 568,
+ 291,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 568,
+ 291,
+ 689
+ ],
+ "type": "text",
+ "content": "Speedup on different batch sizes. The previous study (Helcl et al., 2022) reports that NAT models have limited acceleration on the sentence-level MT in parallel inference settings. We evaluate the models in the settings with different batch sizes, including 1, 2, 4, and 8 instances per batch. Given that document-level MT sequences are much longer than sentence-level MT sequences, we do not evaluate using a batch size larger than 8."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": "As Figure 4 shows, when we increase the batch size from 1 to 8, the overall speedup of GLAT decreases from "
+ },
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "40\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": " to "
+ },
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "9\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 67,
+ 692,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": ". However, if we consider a more strict calculation of the speedup, excluding the time for initializing the models (which takes almost constant time) from the total evaluation time,"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 337,
+ 69,
+ 487,
+ 164
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 337,
+ 69,
+ 487,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 337,
+ 69,
+ 487,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 337,
+ 69,
+ 487,
+ 164
+ ],
+ "type": "image",
+ "image_path": "bf8650d2403274b90126466ea4a7406a816f9a9c72e32a7c5552cfde358763da.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 302,
+ 173,
+ 526,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 173,
+ 526,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 173,
+ 526,
+ 222
+ ],
+ "type": "text",
+ "content": "Figure 4: Speedup with various batch sizes evaluated on News, where the “/ex” conducts a strict calculation of speedup by excluding the time cost for model initialization from both the AT baseline and the NAT models."
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 302,
+ 242,
+ 526,
+ 324
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 242,
+ 526,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 242,
+ 526,
+ 324
+ ],
+ "type": "text",
+ "content": "the inference speedup of GLAT is "
+ },
+ {
+ "bbox": [
+ 302,
+ 242,
+ 526,
+ 324
+ ],
+ "type": "inline_equation",
+ "content": "125\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 302,
+ 242,
+ 526,
+ 324
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 242,
+ 526,
+ 324
+ ],
+ "type": "inline_equation",
+ "content": "44\\mathrm{x}"
+ },
+ {
+ "bbox": [
+ 302,
+ 242,
+ 526,
+ 324
+ ],
+ "type": "text",
+ "content": " for the batch size of 1 and 8, respectively. These results suggest that although the speedup ratio decreases for bigger batch sizes, NAT models on document-level MT still show significant acceleration in the parallel inference settings."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 333,
+ 492,
+ 348
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 333,
+ 492,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 333,
+ 492,
+ 348
+ ],
+ "type": "text",
+ "content": "5 The Challenge of Long Sequence"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 354,
+ 526,
+ 449
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 354,
+ 526,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 354,
+ 526,
+ 449
+ ],
+ "type": "text",
+ "content": "The long input and output in document-level MT bring unexplored challenges to NAT models. In section 4, we learn that NAT models have a significant performance gap with their AT counterparts, and various NAT models even fail in some settings. We investigate these challenges in this section, leaving other discussions in Appendix D."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 458,
+ 458,
+ 470
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 458,
+ 458,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 458,
+ 458,
+ 470
+ ],
+ "type": "text",
+ "content": "5.1 The Failure of NAT Models"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 302,
+ 476,
+ 526,
+ 691
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 476,
+ 526,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 476,
+ 526,
+ 691
+ ],
+ "type": "text",
+ "content": "Previous study (Bao et al., 2021) suggests that Transformer fails on long sequences because of local minima of loss values during the training process. We first investigate this possibility for the cause of the failure on NAT models. We evaluate GLAT, GLAT+CTC, Transformer, and GTransformer on different lengths of sequences, as shown in Figure 5a. Overall, the four models produce different patterns of trends. The d-BLEU score of the Transformer rises as long as the sequence length increases until the failure happens at the length of 512 tokens. In contrast, the d-BLEU scores of the GLAT and GLAT+CTC descend progressively as long as the sequence length increases, which suggests a performance decline instead of a training failure on NAT models."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 302,
+ 692,
+ 527,
+ 773
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 692,
+ 527,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 692,
+ 527,
+ 773
+ ],
+ "type": "text",
+ "content": "Further evidence on the bigger dataset Europarl confirms that the low performance of NAT models is different from the failure of Transformer. Bao et al. (2021) suggests that the local minima encountered by Transformer can be escaped by using a bigger dataset, resulting in normal scores on Europarl."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "discarded_blocks": [],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 69,
+ 68,
+ 208,
+ 158
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 69,
+ 68,
+ 208,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 68,
+ 208,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 68,
+ 208,
+ 158
+ ],
+ "type": "image",
+ "image_path": "d58843f75fcbc9506a34733eaa2356f44d310e8a4feeb40cdc6c84c76a17a0b4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 75,
+ 162,
+ 200,
+ 184
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 162,
+ 200,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 162,
+ 200,
+ 184
+ ],
+ "type": "text",
+ "content": "(a) NAT and AT performance on different sequence length."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 227,
+ 68,
+ 365,
+ 157
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 227,
+ 68,
+ 365,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 227,
+ 68,
+ 365,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 227,
+ 68,
+ 365,
+ 157
+ ],
+ "type": "image",
+ "image_path": "9f141243d8aa4c436811c51bbc230d887438983f828d307b2059a4b74eed3174.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 233,
+ 162,
+ 359,
+ 184
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 233,
+ 162,
+ 359,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 233,
+ 162,
+ 359,
+ 184
+ ],
+ "type": "text",
+ "content": "(b) GLAT detailed BLEU scores on different sequence lengths."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 67,
+ 192,
+ 524,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 192,
+ 524,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 192,
+ 524,
+ 217
+ ],
+ "type": "text",
+ "content": "Figure 5: NAT vs. AT models on trends evaluated on News, where Transformer fails abruptly at the length of 512 while NAT models decrease progressively."
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 386,
+ 68,
+ 524,
+ 157
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 386,
+ 68,
+ 524,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 386,
+ 68,
+ 524,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 68,
+ 524,
+ 157
+ ],
+ "type": "image",
+ "image_path": "9e84efc128d6b4550bae17b17f5d0bea87828d989f05db40bbf2fc78271511b6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 386,
+ 162,
+ 524,
+ 184
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 386,
+ 162,
+ 524,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 162,
+ 524,
+ 184
+ ],
+ "type": "text",
+ "content": "(c) "
+ },
+ {
+ "bbox": [
+ 386,
+ 162,
+ 524,
+ 184
+ ],
+ "type": "inline_equation",
+ "content": "GLAT + CTC"
+ },
+ {
+ "bbox": [
+ 386,
+ 162,
+ 524,
+ 184
+ ],
+ "type": "text",
+ "content": " detailed BLEU scores on different sequence lengths."
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 238,
+ 290,
+ 318
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 238,
+ 290,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 238,
+ 290,
+ 318
+ ],
+ "type": "text",
+ "content": "We evaluate Transformer, GLAT, and GLAT+CTC on Europarl, obtaining d-BLEU scores of 33.41, 2.13, and 0.00, respectively. We could see that GLAT and GLAT+CTC still give low scores on the bigger dataset, suggesting a different cause from the local minima."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "text",
+ "content": "We look into the d-BLEU score details, where BLEU-"
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "inline_equation",
+ "content": "n"
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "text",
+ "content": " ("
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "inline_equation",
+ "content": "n \\in [1,2,3,4]"
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "text",
+ "content": ") denotes the precision on "
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "inline_equation",
+ "content": "n"
+ },
+ {
+ "bbox": [
+ 67,
+ 322,
+ 291,
+ 539
+ ],
+ "type": "text",
+ "content": "-gram. As the d-BLEU scores on GLAT in Figure 5b show, the BLEU-3/4 decreases rapidly, reaching almost 0.0 at the lengths of 256 and 512. The BLEU-1 remains relatively high at 27.3 on the length of 256, resulting in messy translations, where the generated tokens are related to the content but are repeated, disordered, or in the wrong collocation as shown in Table 2. In comparison, the d-BLEU scores on GLAT+CTC in Figure 5c decrease slowly, where the BLEU-1 even increases on the length of 512. We speculate the rapid decrease in BLEU-3/4 on GLAT is caused by the multi-modality and misalignment issues, which can be mitigated by explicit alignments, such as CTC."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 67,
+ 543,
+ 291,
+ 773
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 543,
+ 291,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 543,
+ 291,
+ 773
+ ],
+ "type": "text",
+ "content": "The exceptional zero scores. Table 1 reveals some unexpected results. Both NAT+CTC and GLAT+CTC score a 0.00 d-BLEU on the raw Europarl dataset, a performance notably inferior to that of NAT and GLAT. This unprecedented anomaly may stem from the challenges of applying CTC loss on long sequences. The table further indicates that both combinations, NAT/GLAT+CTC and NAT/GLAT, experience training failures on the raw Europarl dataset. These failures are likely due to multi-modality and misalignment issues. These results empirically demonstrate that, while CTC loss can be effective, its application to document-level training is not consistently stable. We hypothesize that this instability arises from the expanded alignment space between lengthy input and output sequences, as detailed in the token-level alignment"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 308,
+ 238,
+ 520,
+ 280
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 238,
+ 520,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 238,
+ 520,
+ 280
+ ],
+ "type": "text",
+ "content": "Source: companies with a high percentage of floating rate debt stand to lose the most, Goldman said. outside pure stock plays, consumers stand to benefit as well through the rising dollar."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 308,
+ 284,
+ 521,
+ 325
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 284,
+ 521,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 284,
+ 521,
+ 325
+ ],
+ "type": "text",
+ "content": "Target: Unternehmen mit einem hohen Anteil an flexiblen Zinsen werden am meisten verlieren, sagte Goldman. außerhalb der reinen Aktienspiele werden Verbraucher ebenfalls durch den steigenden Dollar profitieren."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 308,
+ 329,
+ 520,
+ 380
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 329,
+ 520,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 329,
+ 520,
+ 380
+ ],
+ "type": "text",
+ "content": "GLAT: Unternehmen mit einem hohen, freiem freiem freien erverlieren die die meisten meisten, . . . te te dazu kräften profitieren profitieren profitieren profitieren durch durch steigenden steigenden profitieren profitieren."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 308,
+ 384,
+ 520,
+ 406
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 384,
+ 520,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 384,
+ 520,
+ 406
+ ],
+ "type": "text",
+ "content": "GLAT+CTC: Unternehmen mit hohen der Schulden, die verlieren. außerhalb spielten Verbrauchener den profitieren."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 302,
+ 416,
+ 526,
+ 453
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 416,
+ 526,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 416,
+ 526,
+ 453
+ ],
+ "type": "text",
+ "content": "Table 2: GLAT and GLAT+CTC produce good translation at the beginning but downgrade later with repetitions and missing translations."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 302,
+ 475,
+ 417,
+ 487
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 475,
+ 417,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 475,
+ 417,
+ 487
+ ],
+ "type": "text",
+ "content": "discussion in Section 4.3."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 302,
+ 498,
+ 418,
+ 510
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 498,
+ 418,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 498,
+ 418,
+ 510
+ ],
+ "type": "text",
+ "content": "5.2 Document Context"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 301,
+ 515,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 515,
+ 526,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 515,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "content": "We compare AT and NAT models with document context to the Transformer baseline without document context (trained on sentence) in Appendix B. The results suggest that document context enhances AT model (G-Transformer) by 0.70 s-BLEU. However, NAT models with document context still underperform the Transformer baseline, indicating that current NAT models have a limited ability to utilize document context."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "content": "We further apply an ablation study on the document context to quantify its contribution. As Table 3 illustrates, the performance of G-Trans+GLAT+CTC without document context drops by approximately 0.28 s-BLEU on average over the three benchmarks. Specifically, the target-side context contributes merely 0.01, while the source-side context contributes 0.27. The context contributions in G-Trans+GLAT+CTC are less than that in G-Transformer (0.23 and 0.70 s-BLEU on"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "discarded_blocks": [],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 7
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 68,
+ 68,
+ 290,
+ 142
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 68,
+ 68,
+ 290,
+ 142
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 68,
+ 290,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 68,
+ 290,
+ 142
+ ],
+ "type": "table",
+ "html": "| Unordered Sentences | Coherent Paragraph |
| S1. She poured herself a glass. | S2. Kelsi opened up a bottle of grape juice. |
| S2. Kelsi opened up a bottle of grape juice. | S1. She poured herself a glass. |
| S3. The carpet was now covered in juice. | S4. Accidentally she spilled some. |
| S4. Accidentally she spilled some. | S3. The carpet was now covered in juice. |
| S5. Kelsi just had to call the cleaners. | S5. Kelsi just had to call the cleaners. |
",
+ "bbox": [
+ 515,
+ 250,
+ 877,
+ 326
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Figure 1: An example of sentence ordering sourced from SIND dataset. The goal of sentence ordering is to understand the semantics and logic of a set of unordered sentences (the left box), and reorganize them to a coherent paragraph (the right box).",
+ "bbox": [
+ 507,
+ 338,
+ 884,
+ 411
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "sophisticated designs are costly in labor and time, expertise-required, and even cannot be well generalized to other scenarios. In the past couple of years, inspired by the great success of deep learning, dozens of deep neural methods for sentence ordering have been proposed and achieved great success (Cui et al., 2018; Kumar et al., 2020; Wang and Wan, 2019; Prabhumoye et al., 2020; Basu Roy Chowdhury et al., 2021).",
+ "bbox": [
+ 507,
+ 435,
+ 882,
+ 580
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "There exist two paradigms of neural sentence ordering: 1) ranking model via the scores of relative positions between paired sentences, and 2) generation model based on an encoder-decoder framework to predict the sentence order with sequence generation. The former one, ranking model (Prabhumoye et al., 2020; Chen et al., 2016), starts from calculating a score for each sentence pair to indicate their relative position in a coherent paragraph, then employs ranking or searching strategies to derive out the gold order. Since the score between paired sentences only measures the one-to-one relative position, which may fail to capture the one-to-many interactions on the global context level. Besides, the naive ranking process is somehow brute-force and time-consuming, and highly depends on the quality of sentence representation.",
+ "bbox": [
+ 507,
+ 581,
+ 882,
+ 853
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "The latter one, generation model regards sentence ordering as a sequence prediction problem to explore the relations between sentences and positions, which takes a set of sentences and generates",
+ "bbox": [
+ 507,
+ 854,
+ 884,
+ 917
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_number",
+ "text": "4198",
+ "bbox": [
+ 478,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "footer",
+ "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4198-4214",
+ "bbox": [
+ 216,
+ 945,
+ 779,
+ 957
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "footer",
+ "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+ "bbox": [
+ 277,
+ 957,
+ 719,
+ 971
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "the coherent sentence sequence. At the early stage, several encoder-decoder based approaches (Gong et al., 2016; Logeswaran et al., 2018) treat the unordered input sentences as a permutation sequence and encode them sequentially. Such sequential modeling encodes incorrect sentence order and semantic logic between sentences, which may mislead the decoder to predict an incoherent paragraph. Specifically, with sequential modeling, different permutations of the same paragraph may derive different paragraph representations and result in different output sentence order, which is not reasonable and inconsistent with the intuition of humans. To address this issue, Cui et al. (2018) first propose a deep attentive sentence ordering network equipped with self-attention to learn a reliable and consistent paragraph representation. Their model implements order-independent encoding with a non-positional variant of Transformer (Vaswani et al., 2017), but still predicts sentences via a pointer network as most previous works do. Nevertheless, the RNN-based pointer networks predict the sentence order step-by-step, which only explores the unilateral dependencies with past predictions and fails to leverage the comprehensive bilateral interactions between sentences for ordering.",
+ "bbox": [
+ 112,
+ 84,
+ 492,
+ 502
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "To overcome the above limitations, in this paper, we propose a novel Non-Autoregressive Ordering Network, dubbed NAON, which explores the bilateral dependencies between sentences and predicts the sentence for each position in parallel. Specifically, our NAON consists of a basic sentence encoder, an order-independent contextual sentence encoder and a non-autoregressive Transformer (NAT) (Gu et al., 2018) decoder. We first employ a sequence encoder, e.g., BERT in this work, to map discrete words in each sentence to a compact representation. The obtained sentence representations are injected to a Transformer (Vaswani et al., 2017) encoder to exploit the interaction and relation between sentences, and export contextual sentence representations. Note that we remove the positional embedding in the Transformer to achieve the order-invariant encoding for the sentences. In contrast to obtaining the paragraph representation via pooling to initialize the memory of auto-regressive decoder (Cui et al., 2018), we make full use of contextual sentence representations through multi-head attention in non-autoregressive decoder. We design a non-autoregressive decoder (NAD) that takes positions as input and predicts",
+ "bbox": [
+ 112,
+ 517,
+ 490,
+ 921
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "the sentence for each position utilizing a pointer network (Vinyals et al., 2015). In particular, we first employ multi-head attention to explore the relations between positions. Then the output distribution over the original sentence representation set for each position will be modelled by a pointer network. As mentioned in (Gu et al., 2018), naive non-autoregressive decoding suffers from severe problem of output repetition due to the complete conditional independence. To address this issue, we elaborate an exclusive loss constraining the exclusiveness between each sentence and position pair during training, and a greedy selective and removing strategy for inference. In summary, our main contributions are as follows:",
+ "bbox": [
+ 507,
+ 84,
+ 885,
+ 325
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- We propose a novel approach, non-autoregressive ordering network (NAON), for sentence ordering, which implements self-attention to explore the bilateral interactions between sentences, going beyond the unilateral dependencies of RNN-based step-by-step decoding. To the best of our knowledge, our approach is one of the first attempts using non-autoregressive decoders for the sentence ordering problem.",
+ "- To alleviate the repetition problem of vanilla non-autoregressive models, we design an exclusive loss under the implication of exclusive constraint between positions and sentences during training. For inference, we simply design a greedy selective and removing strategy to match each sentence-position pair.",
+ "- Extensive experiments are conducted on several common-used datasets, and the results demonstrate the effectiveness of the proposed approach."
+ ],
+ "bbox": [
+ 509,
+ 332,
+ 885,
+ 659
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 The Proposed Approach",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 677,
+ 754,
+ 694
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2.1 Preliminary",
+ "text_level": 1,
+ "bbox": [
+ 509,
+ 702,
+ 650,
+ 718
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Sentence ordering aims to capture the coherence and recover the gold order of a set of unordered sentences. Specifically, given a sentence set $S = \\{s_1, s_2, \\ldots, s_N\\}$ with $N$ sentences, and each sentence $s_i = [w_{i1}, w_{i2}, \\ldots, w_{iL_i}]$ contains a sequence of $L_i$ words. The goal of sentence ordering is to find an order $O^* = [o_1^*, o_2^*, \\ldots, o_N^*]$ conveying the coherent semantics, which can be modelled by:",
+ "bbox": [
+ 507,
+ 722,
+ 885,
+ 852
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nP (O ^ {*} | S) \\geq P (O | S), \\forall O \\in \\Omega , \\tag {1}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 579,
+ 860,
+ 882,
+ 878
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "where $P(O|S)$ is the probability of order $O$ given $S$ , and $\\Omega$ denotes the exhaustive set of all potential",
+ "bbox": [
+ 507,
+ 887,
+ 882,
+ 919
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "4199",
+ "bbox": [
+ 480,
+ 927,
+ 521,
+ 940
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/3c0cce24d4272f86916a212ca75dd8e54d094ff552edd61d96ac65d379f59edc.jpg",
+ "image_caption": [
+ "Figure 2: The overall flowchart of the proposed Non-Autoregressive Ordering Network (NAON), which consists of a basic sentence encoder, a contextual sentence encoder and a non-autoregressive decoder."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 174,
+ 80,
+ 831,
+ 300
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "order with the size of $A_N^N$ . Existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019) generate the coherent paragraph sentence-by-sentence to recover the gold order, and optimize the model by maximizing:",
+ "bbox": [
+ 112,
+ 369,
+ 489,
+ 450
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\sum_ {j = 1} ^ {N} \\log P \\left(s _ {j} \\mid s _ {1}, s _ {2}, \\dots , s _ {j - 1}, S\\right). \\tag {2}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 176,
+ 456,
+ 487,
+ 501
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "While our non-autoregressive ordering network implements the optimization in parallel by maximizing the logarithmic probabilities for all sentences:",
+ "bbox": [
+ 112,
+ 506,
+ 489,
+ 555
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\sum_ {j = 1} ^ {N} \\log P (s _ {j} | S). \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 233,
+ 560,
+ 487,
+ 604
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2.2 Basic Sentence Representation",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 611,
+ 401,
+ 626
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "To map discrete words of a sentence into compact representations, as shown on the left of Figure 2, a sequential encoder is first employed, e.g., an LSTM or BERT. Inspired by the success in language modeling of BERT (Devlin et al., 2019), we adopt a pre-trained BERT-base model for basic sentence representation. In particular, following previous work (Kumar et al., 2020), we concatenate the [CLS] token and word sequence as BERT encoder input and average the outputs of all the tokens to denote the entire sentence representation. To make the representation more compatible with sentence ordering task, we also fine-tune it in an end-to-end fashion on each dataset.",
+ "bbox": [
+ 112,
+ 631,
+ 489,
+ 854
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2.3 Contextual Sentence Representation",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 866,
+ 445,
+ 881
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Most existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019; Logeswaran et al.,",
+ "bbox": [
+ 112,
+ 887,
+ 489,
+ 917
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2018) provide a \"paragraph representation\" to initialize the hidden state and memory cell of the decoder. A straightforward way is assuming that the given order is a \"pseudo order\" of the unordered sentences, and implementing an RNN-based encoder (Wang and Wan, 2019), e.g., LSTM or GRU, to map sentence representations to a dense feature vector for the paragraph. While such pseudo order assumption introduces incorrect sentence order, which may induce semantic incoherence and mislead the decoder in recovering the ground-truth order. To address this issue, Cui et al. (2018) proposed a deep attentive ordering network integrating a self-attention mechanism to derive the order invariant sentence representations, and adopted mean pooling across sentences to obtain the paragraph representation. Different from existing autoregressive approaches which need to initialize the decoder memory, our non-autoregressive decoder does not require such \"paragraph representation\". We design a contextual sentence encoder to interact each sentence with all the sentences in the set and injects all the contextual sentence representations into the decoder to find the correct order.",
+ "bbox": [
+ 507,
+ 369,
+ 884,
+ 755
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Obviously, it is important to exploit semantic relations between sentences for recovering coherence. Towards this end, we design a contextual sentence encoder, shown in the middle of Figure 2, to interact each sentence with other sentences. Specifically, after obtaining the basic sentence representations $E^{b} = \\{e_{1}^{b}, e_{2}^{b}, \\dots, e_{N}^{b}\\}$ , we employ a Transformer-based architecture similar to (Cui et al., 2018), equipped with multi-head self-attention mechanism without the positional embedding, to exploit the",
+ "bbox": [
+ 507,
+ 758,
+ 884,
+ 917
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "4200",
+ "bbox": [
+ 480,
+ 927,
+ 521,
+ 940
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "contextual information as follows:",
+ "bbox": [
+ 112,
+ 84,
+ 371,
+ 99
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {4}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 129,
+ 105,
+ 489,
+ 142
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathrm {M H} (Q, K, V) = \\left[ H _ {1}, H _ {2}, \\dots , H _ {h} \\right] W, \\tag {5}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 161,
+ 149,
+ 487,
+ 167
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nH _ {i} = \\operatorname {A t t e n t i o n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right). \\tag {6}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 142,
+ 170,
+ 487,
+ 191
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "To explore deep contextual interaction, we duplicate the multi-head self-attention block multiple times, e.g., 4 for NSF and arXiv datasets in this work. Finally, the contextual sentence representations $E^{c} = \\{e_{1}^{c}, e_{2}^{c}, \\ldots, e_{N}^{c}\\}$ can be derived from the output of the last layer.",
+ "bbox": [
+ 112,
+ 196,
+ 489,
+ 294
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.4 Non-Autoregressive Prediction",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 303,
+ 401,
+ 318
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Almost all the previous generation approaches (Logeswaran et al., 2018; Cui et al., 2018; Basu Roy Chowdhury et al., 2021) adopt a recurrent-based architecture composing with a pointer network (Vinyals et al., 2015) to predict the ordered sentence step-by-step, which only leverages the unilateral dependencies with the past prediction, and cannot capture the coherent dependencies between all the sentences. Besides, as argued in (Bengio et al., 2015), the discrepancy between training and inference of RNN-based generation strategy may decrease the inference performance. Inspired by the potentials of the non-autoregressive transformer in neural machine translation, we devise a non-autoregressive decoder (NAD) for recovering the correct order of sentences, with the exploration of bilateral dependencies. Our NAD is designed based on a Non-Autoregressive Transformer (NAT) decoder in machine translation, which removes the autoregressive connections between steps and generates all the target words in parallel, rather than step-by-step (i.e., word-by-word). Beyond the aforementioned superiority of NAT in translation, NAD is especially suitable for sentence ordering from two aspects: (1) the length of the decoder is determined by the input set of sentences, and (2) the sentences and positions match exclusively.",
+ "bbox": [
+ 112,
+ 323,
+ 489,
+ 758
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "As suggested in (Gu et al., 2018), different inputs of non-autoregressive decoder may lead to quite different outputs resulting in a remarkable performance gap. Thus it is crucial to choose a proper information source for the input of NAD. Towards this end, we meticulously adopt positional information as the input of our non-autoregressive decoder. Because the sentence ordering problem could be interpreted as matching the unordered sentences to the right positions, and it is straightforward to",
+ "bbox": [
+ 112,
+ 758,
+ 489,
+ 919
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "take the position information as input and predict the correct sentence for each position in parallel. Specifically, we follow previous work (Vaswani et al., 2017) to project each discrete position into a compact representation $P$ via:",
+ "bbox": [
+ 507,
+ 84,
+ 884,
+ 165
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\np _ {i, 2 j} = \\sin \\left(i / 1 0 0 0 0 ^ {2 j / d _ {k}}\\right), \\tag {7}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 589,
+ 173,
+ 882,
+ 192
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\np _ {i, 2 j + 1} = \\cos (i / 1 0 0 0 0 ^ {2 j / d _ {k}}), \\tag {8}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 583,
+ 200,
+ 882,
+ 221
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $i$ denotes the position and $j$ is the $j$ -th dimension in $p_i$ . We then exploit the interaction and relative information between positions implementing a multi-head self-attention as:",
+ "bbox": [
+ 507,
+ 225,
+ 884,
+ 290
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\bar {P} = \\mathrm {M H A t t} \\left(P W ^ {Q}, P W ^ {K}, P W ^ {V}\\right), \\tag {9}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 557,
+ 298,
+ 882,
+ 319
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where MHAtt denotes multi-head self-attention block comprised of Equation 4-6. To connect the positions and the elements in the unordered context set, we further implement multi-head interattention between positions and sentences to exploit the correlations between them. Specifically, we query the contextual sentence representation set using $\\overline{P}$ , to obtain the attentive representations of sentences as:",
+ "bbox": [
+ 507,
+ 328,
+ 882,
+ 470
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nE ^ {p} = \\operatorname {s o f t m a x} \\left(\\frac {\\bar {P} (E ^ {c}) ^ {T}}{\\sqrt {d _ {k}}}\\right) E ^ {c}, \\tag {10}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 579,
+ 478,
+ 882,
+ 514
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $E^c$ is the contextual sentence representation derived from the contextual encoder. With such an operation, the obtained $E^p$ now is position sensitive representation, and could be used to select the most relevant sentence for each position. Specifically, the probability of the sentence $j$ to be in the position $i$ can be obtained by:",
+ "bbox": [
+ 507,
+ 521,
+ 882,
+ 634
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\omega_ {i j} = u ^ {T} \\tanh \\left(W _ {p} e _ {i} ^ {p} + W _ {b} e _ {j} ^ {b}\\right), \\tag {11}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 574,
+ 643,
+ 882,
+ 663
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nP t r _ {i} = \\text {s o f t m a x} (\\omega_ {i}), \\tag {12}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 610,
+ 671,
+ 880,
+ 688
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $W_{*}$ are learned parameters, and $u$ is a column weight vector. $e_j^b$ and $e_j^p$ are the basic and positional sensitive representations of sentence $j$ . $P_{tr_i}$ is the probabilistic distribution across all the sentences for $i$ -th position.",
+ "bbox": [
+ 507,
+ 694,
+ 884,
+ 775
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.5 Training and Inference",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 785,
+ 737,
+ 800
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Training: Following previous ordering works (Cui et al., 2018; Chen et al., 2016), the entire model could be trained to maximize the coherence probability by minimizing the cross entropy loss as:",
+ "bbox": [
+ 507,
+ 806,
+ 882,
+ 870
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nL _ {c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log P \\left(o _ {i} \\mid p _ {i}; \\Theta\\right), \\tag {13}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 576,
+ 879,
+ 882,
+ 921
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "4201",
+ "bbox": [
+ 480,
+ 927,
+ 517,
+ 940
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "image",
+ "img_path": "images/627068c9fcaac382cb196d0322e9dcc66362e23883e10f71487297c90d47ab52.jpg",
+ "image_caption": [
+ "Figure 3: Illustration of our proposed exclusive loss, which simultaneously constrain the sentence selection and position matching."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 181,
+ 87,
+ 428,
+ 287
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $o_i$ is the ground-truth sentence for given $i$ -th position $p_i$ , and $\\Theta$ denotes all the learnable parameters in the model. As each sentence in the ordering problem is allowed to be selected only once, to avoid repeat selection, the RNN-based autoregressive model adopts a mask to remove selected sentences (Yin et al., 2020). However, the non-autoregressive model suffers from the repetition problem for its parallel decoding fashion (Gu et al., 2018). We have observed that the traditional pointer loss only penalizes the repetition for sentence choosing. Considering the characteristic of the ordering task, as depicted in Figure 3, every position should also choose one and only one sentence. Towards this end, we introduce an elegant exclusive loss to constrain the mutual exclusiveness of optimal matching between positions and elements. Specifically, we simultaneously calculate two pointers for sentence choosing and position choosing when we obtain the pointer map by Equation 11. Given the pointer map $\\omega$ , we implement softmax across column and row similar to Equation 12, where the column one interprets choosing sentence for each position and the row one assigns a position for each sentence. Then we optimize the entire model by minimizing the combined exclusive loss as follows:",
+ "bbox": [
+ 112,
+ 369,
+ 489,
+ 803
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nL _ {e x} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\log P \\left(o _ {i} \\mid p _ {i}\\right) + \\log P \\left(p _ {i} \\mid o _ {i}\\right)\\right). \\tag {14}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 127,
+ 814,
+ 487,
+ 869
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Through such an exclusive constraint, the repetition problem could be distinctly alleviated.",
+ "bbox": [
+ 112,
+ 870,
+ 485,
+ 901
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Inference: Obviously, the proposed exclusive loss",
+ "bbox": [
+ 112,
+ 903,
+ 485,
+ 917
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "cannot completely avoid repetition problem during decoding. Similar with the masking strategy used in recurrent pointer networks, we introduce a simple yet effective greedy selective and removing strategy to ensure the exclusive matching for inference. Given the probability distribution array, as described in Algorithm 1, the proposed strategy greedily selects the largest probability and returns its index indicating the matching position and sentence. It then removes all the probability values of the corresponding column and row by setting them to zero. The process is repeated until all the values of the array are set to zero, which means that the matching process between positions and sentences is completely finished.",
+ "bbox": [
+ 507,
+ 84,
+ 884,
+ 326
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "code",
+ "sub_type": "algorithm",
+ "code_caption": [
+ "Algorithm 1 Pseudo-code of greedy selective and removing in a PyTorch-like style."
+ ],
+ "code_body": "$\\# \\text{Ptr}$ .. $N\\times N$ output probability distribution array of NAD \nINPUTPtr \noutput $= []$ \nwhile $\\mathrm{sum}(\\mathrm{Ptr}) > 0$ . row_idx,col_idx $=$ max_index(Ptr) #max_index(): get the index of the maximum value output.append((row_idx, col_idx)) Ptr[row_idx, :] $= 0$ Ptr[:,col_idx] $= 0$ \nRETURN output",
+ "bbox": [
+ 524,
+ 373,
+ 882,
+ 495
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3 Experiments",
+ "text_level": 1,
+ "bbox": [
+ 509,
+ 524,
+ 655,
+ 539
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "To verify the effectiveness of our NAON, we conduct extensive experiments on six datasets, and evaluate the performance with Acc, PMR, and $\\tau$ . More detailed experimental settings are described in the Appendix B.",
+ "bbox": [
+ 507,
+ 550,
+ 882,
+ 630
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.1 Comparing with Baselines",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 640,
+ 761,
+ 656
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "As aforementioned, existing works lie in two categories: ranking-based and generation-based methods. To assess the proposed NAON, we compare it with the recent state-of-the-art methods, including Pairwise model via pairwise ranking (Chen et al., 2016), RankTxNet via deep attentive ranking (Kumar et al., 2020), and B-TSort searching correct order with topological sort (Prabhumoye et al., 2020) of ranking model, and the first architecture using pointer network LSTM-Ptr (Gong et al., 2016), SetLSTM designing an LSTM based attention to process the set encoding (Logeswaran et al., 2018), AON with a deep attentive order invariant paragraph encoder (Cui et al., 2018), HAN equipped with hierarchical attention (Wang and Wan, 2019), TGCM modeling topic-guide coherence (Oh et al.,",
+ "bbox": [
+ 507,
+ 661,
+ 884,
+ 917
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "4202",
+ "bbox": [
+ 480,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "table",
+ "img_path": "images/2a5f7870cca137223e5b52b06f66cb32571c333967005662b25c6f0a9d451c73.jpg",
+ "table_caption": [
+ "Table 1: Comparison with baselines. The best and 2nd-best results are in bold and underlined, respectively. NAON-NE indicates no exclusive loss. Besides, we also take an attempt at relative order exploration and BART enhancement, namely NAON-RO and BART-NAON respectively, and compare them with BERSON and RE-BART, listed in the last two blocks in blue."
+ ],
+ "table_footnote": [],
+ "table_body": "| Unordered Sentences | Coherent Paragraph |
| S1. She poured herself a glass. | S2. Kelsi opened up a bottle of grape juice. |
| S2. Kelsi opened up a bottle of grape juice. | S1. She poured herself a glass. |
| S3. The carpet was now covered in juice. | S4. Accidentally she spilled some. |
| S4. Accidentally she spilled some. | S3. The carpet was now covered in juice. |
| S5. Kelsi just had to call the cleaners. | S5. Kelsi just had to call the cleaners. |
"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.509,
+ 0.339,
+ 0.885,
+ 0.412
+ ],
+ "angle": 0,
+ "content": "Figure 1: An example of sentence ordering sourced from SIND dataset. The goal of sentence ordering is to understand the semantics and logic of a set of unordered sentences (the left box), and reorganize them to a coherent paragraph (the right box)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.436,
+ 0.884,
+ 0.581
+ ],
+ "angle": 0,
+ "content": "sophisticated designs are costly in labor and time, expertise-required, and even cannot be well generalized to other scenarios. In the past couple of years, inspired by the great success of deep learning, dozens of deep neural methods for sentence ordering have been proposed and achieved great success (Cui et al., 2018; Kumar et al., 2020; Wang and Wan, 2019; Prabhumoye et al., 2020; Basu Roy Chowdhury et al., 2021)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.582,
+ 0.884,
+ 0.854
+ ],
+ "angle": 0,
+ "content": "There exist two paradigms of neural sentence ordering: 1) ranking model via the scores of relative positions between paired sentences, and 2) generation model based on an encoder-decoder framework to predict the sentence order with sequence generation. The former one, ranking model (Prabhumoye et al., 2020; Chen et al., 2016), starts from calculating a score for each sentence pair to indicate their relative position in a coherent paragraph, then employs ranking or searching strategies to derive out the gold order. Since the score between paired sentences only measures the one-to-one relative position, which may fail to capture the one-to-many interactions on the global context level. Besides, the naive ranking process is somehow brute-force and time-consuming, and highly depends on the quality of sentence representation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.855,
+ 0.885,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "The latter one, generation model regards sentence ordering as a sequence prediction problem to explore the relations between sentences and positions, which takes a set of sentences and generates"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.48,
+ 0.928,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "4198"
+ },
+ {
+ "type": "footer",
+ "bbox": [
+ 0.218,
+ 0.946,
+ 0.78,
+ 0.958
+ ],
+ "angle": 0,
+ "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4198-4214"
+ },
+ {
+ "type": "footer",
+ "bbox": [
+ 0.278,
+ 0.958,
+ 0.72,
+ 0.972
+ ],
+ "angle": 0,
+ "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.085,
+ 0.493,
+ 0.504
+ ],
+ "angle": 0,
+ "content": "the coherent sentence sequence. At the early stage, several encoder-decoder based approaches (Gong et al., 2016; Logeswaran et al., 2018) treat the unordered input sentences as a permutation sequence and encode them sequentially. Such sequential modeling encodes incorrect sentence order and semantic logic between sentences, which may mislead the decoder to predict an incoherent paragraph. Specifically, with sequential modeling, different permutations of the same paragraph may derive different paragraph representations and result in different output sentence order, which is not reasonable and inconsistent with the intuition of humans. To address this issue, Cui et al. (2018) first propose a deep attentive sentence ordering network equipped with self-attention to learn a reliable and consistent paragraph representation. Their model implements order-independent encoding with a non-positional variant of Transformer (Vaswani et al., 2017), but still predicts sentences via a pointer network as most previous works do. Nevertheless, the RNN-based pointer networks predict the sentence order step-by-step, which only explores the unilateral dependencies with past predictions and fails to leverage the comprehensive bilateral interactions between sentences for ordering."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.518,
+ 0.492,
+ 0.922
+ ],
+ "angle": 0,
+ "content": "To overcome the above limitations, in this paper, we propose a novel Non-Autoregressive Ordering Network, dubbed NAON, which explores the bilateral dependencies between sentences and predicts the sentence for each position in parallel. Specifically, our NAON consists of a basic sentence encoder, an order-independent contextual sentence encoder and a non-autoregressive Transformer (NAT) (Gu et al., 2018) decoder. We first employ a sequence encoder, e.g., BERT in this work, to map discrete words in each sentence to a compact representation. The obtained sentence representations are injected to a Transformer (Vaswani et al., 2017) encoder to exploit the interaction and relation between sentences, and export contextual sentence representations. Note that we remove the positional embedding in the Transformer to achieve the order-invariant encoding for the sentences. In contrast to obtaining the paragraph representation via pooling to initialize the memory of auto-regressive decoder (Cui et al., 2018), we make full use of contextual sentence representations through multi-head attention in non-autoregressive decoder. We design a non-autoregressive decoder (NAD) that takes positions as input and predicts"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.085,
+ 0.887,
+ 0.326
+ ],
+ "angle": 0,
+ "content": "the sentence for each position utilizing a pointer network (Vinyals et al., 2015). In particular, we first employ multi-head attention to explore the relations between positions. Then the output distribution over the original sentence representation set for each position will be modelled by a pointer network. As mentioned in (Gu et al., 2018), naive non-autoregressive decoding suffers from severe problem of output repetition due to the complete conditional independence. To address this issue, we elaborate an exclusive loss constraining the exclusiveness between each sentence and position pair during training, and a greedy selective and removing strategy for inference. In summary, our main contributions are as follows:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.51,
+ 0.334,
+ 0.887,
+ 0.48
+ ],
+ "angle": 0,
+ "content": "- We propose a novel approach, non-autoregressive ordering network (NAON), for sentence ordering, which implements self-attention to explore the bilateral interactions between sentences, going beyond the unilateral dependencies of RNN-based step-by-step decoding. To the best of our knowledge, our approach is one of the first attempts using non-autoregressive decoders for the sentence ordering problem."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.51,
+ 0.489,
+ 0.887,
+ 0.603
+ ],
+ "angle": 0,
+ "content": "- To alleviate the repetition problem of vanilla non-autoregressive models, we design an exclusive loss under the implication of exclusive constraint between positions and sentences during training. For inference, we simply design a greedy selective and removing strategy to match each sentence-position pair."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.51,
+ 0.611,
+ 0.887,
+ 0.66
+ ],
+ "angle": 0,
+ "content": "- Extensive experiments are conducted on several common-used datasets, and the results demonstrate the effectiveness of the proposed approach."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.51,
+ 0.334,
+ 0.887,
+ 0.66
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.678,
+ 0.756,
+ 0.695
+ ],
+ "angle": 0,
+ "content": "2 The Proposed Approach"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.51,
+ 0.703,
+ 0.652,
+ 0.719
+ ],
+ "angle": 0,
+ "content": "2.1 Preliminary"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.724,
+ 0.886,
+ 0.853
+ ],
+ "angle": 0,
+ "content": "Sentence ordering aims to capture the coherence and recover the gold order of a set of unordered sentences. Specifically, given a sentence set \\( S = \\{s_1, s_2, \\ldots, s_N\\} \\) with \\( N \\) sentences, and each sentence \\( s_i = [w_{i1}, w_{i2}, \\ldots, w_{iL_i}] \\) contains a sequence of \\( L_i \\) words. The goal of sentence ordering is to find an order \\( O^* = [o_1^*, o_2^*, \\ldots, o_N^*] \\) conveying the coherent semantics, which can be modelled by:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.58,
+ 0.861,
+ 0.884,
+ 0.879
+ ],
+ "angle": 0,
+ "content": "\\[\nP (O ^ {*} | S) \\geq P (O | S), \\forall O \\in \\Omega , \\tag {1}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.509,
+ 0.888,
+ 0.884,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "where \\(P(O|S)\\) is the probability of order \\(O\\) given \\(S\\), and \\(\\Omega\\) denotes the exhaustive set of all potential"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.522,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "4199"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.175,
+ 0.082,
+ 0.833,
+ 0.301
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.113,
+ 0.316,
+ 0.884,
+ 0.346
+ ],
+ "angle": 0,
+ "content": "Figure 2: The overall flowchart of the proposed Non-Autoregressive Ordering Network (NAON), which consists of a basic sentence encoder, a contextual sentence encoder and a non-autoregressive decoder."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.37,
+ 0.49,
+ 0.451
+ ],
+ "angle": 0,
+ "content": "order with the size of \\( A_N^N \\). Existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019) generate the coherent paragraph sentence-by-sentence to recover the gold order, and optimize the model by maximizing:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.178,
+ 0.457,
+ 0.488,
+ 0.502
+ ],
+ "angle": 0,
+ "content": "\\[\n\\sum_ {j = 1} ^ {N} \\log P \\left(s _ {j} \\mid s _ {1}, s _ {2}, \\dots , s _ {j - 1}, S\\right). \\tag {2}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.507,
+ 0.49,
+ 0.556
+ ],
+ "angle": 0,
+ "content": "While our non-autoregressive ordering network implements the optimization in parallel by maximizing the logarithmic probabilities for all sentences:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.235,
+ 0.561,
+ 0.488,
+ 0.605
+ ],
+ "angle": 0,
+ "content": "\\[\n\\sum_ {j = 1} ^ {N} \\log P (s _ {j} | S). \\tag {3}\n\\]"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.612,
+ 0.402,
+ 0.627
+ ],
+ "angle": 0,
+ "content": "2.2 Basic Sentence Representation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.632,
+ 0.49,
+ 0.856
+ ],
+ "angle": 0,
+ "content": "To map discrete words of a sentence into compact representations, as shown on the left of Figure 2, a sequential encoder is first employed, e.g., an LSTM or BERT. Inspired by the success in language modeling of BERT (Devlin et al., 2019), we adopt a pre-trained BERT-base model for basic sentence representation. In particular, following previous work (Kumar et al., 2020), we concatenate the [CLS] token and word sequence as BERT encoder input and average the outputs of all the tokens to denote the entire sentence representation. To make the representation more compatible with sentence ordering task, we also fine-tune it in an end-to-end fashion on each dataset."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.867,
+ 0.446,
+ 0.882
+ ],
+ "angle": 0,
+ "content": "2.3 Contextual Sentence Representation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.888,
+ 0.49,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "Most existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019; Logeswaran et al.,"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.37,
+ 0.885,
+ 0.756
+ ],
+ "angle": 0,
+ "content": "2018) provide a \"paragraph representation\" to initialize the hidden state and memory cell of the decoder. A straightforward way is assuming that the given order is a \"pseudo order\" of the unordered sentences, and implementing an RNN-based encoder (Wang and Wan, 2019), e.g., LSTM or GRU, to map sentence representations to a dense feature vector for the paragraph. While such pseudo order assumption introduces incorrect sentence order, which may induce semantic incoherence and mislead the decoder in recovering the ground-truth order. To address this issue, Cui et al. (2018) proposed a deep attentive ordering network integrating a self-attention mechanism to derive the order invariant sentence representations, and adopted mean pooling across sentences to obtain the paragraph representation. Different from existing autoregressive approaches which need to initialize the decoder memory, our non-autoregressive decoder does not require such \"paragraph representation\". We design a contextual sentence encoder to interact each sentence with all the sentences in the set and injects all the contextual sentence representations into the decoder to find the correct order."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.759,
+ 0.885,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "Obviously, it is important to exploit semantic relations between sentences for recovering coherence. Towards this end, we design a contextual sentence encoder, shown in the middle of Figure 2, to interact each sentence with other sentences. Specifically, after obtaining the basic sentence representations \\( E^{b} = \\{e_{1}^{b}, e_{2}^{b}, \\dots, e_{N}^{b}\\} \\), we employ a Transformer-based architecture similar to (Cui et al., 2018), equipped with multi-head self-attention mechanism without the positional embedding, to exploit the"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.522,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "4200"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.085,
+ 0.373,
+ 0.1
+ ],
+ "angle": 0,
+ "content": "contextual information as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.13,
+ 0.106,
+ 0.49,
+ 0.143
+ ],
+ "angle": 0,
+ "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {4}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.163,
+ 0.15,
+ 0.488,
+ 0.168
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathrm {M H} (Q, K, V) = \\left[ H _ {1}, H _ {2}, \\dots , H _ {h} \\right] W, \\tag {5}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.143,
+ 0.171,
+ 0.489,
+ 0.192
+ ],
+ "angle": 0,
+ "content": "\\[\nH _ {i} = \\operatorname {A t t e n t i o n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right). \\tag {6}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.197,
+ 0.49,
+ 0.295
+ ],
+ "angle": 0,
+ "content": "To explore deep contextual interaction, we duplicate the multi-head self-attention block multiple times, e.g., 4 for NSF and arXiv datasets in this work. Finally, the contextual sentence representations \\( E^{c} = \\{e_{1}^{c}, e_{2}^{c}, \\ldots, e_{N}^{c}\\} \\) can be derived from the output of the last layer."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.304,
+ 0.402,
+ 0.319
+ ],
+ "angle": 0,
+ "content": "2.4 Non-Autoregressive Prediction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.324,
+ 0.49,
+ 0.759
+ ],
+ "angle": 0,
+ "content": "Almost all the previous generation approaches (Logeswaran et al., 2018; Cui et al., 2018; Basu Roy Chowdhury et al., 2021) adopt a recurrent-based architecture composing with a pointer network (Vinyals et al., 2015) to predict the ordered sentence step-by-step, which only leverages the unilateral dependencies with the past prediction, and cannot capture the coherent dependencies between all the sentences. Besides, as argued in (Bengio et al., 2015), the discrepancy between training and inference of RNN-based generation strategy may decrease the inference performance. Inspired by the potentials of the non-autoregressive transformer in neural machine translation, we devise a non-autoregressive decoder (NAD) for recovering the correct order of sentences, with the exploration of bilateral dependencies. Our NAD is designed based on a Non-Autoregressive Transformer (NAT) decoder in machine translation, which removes the autoregressive connections between steps and generates all the target words in parallel, rather than step-by-step (i.e., word-by-word). Beyond the aforementioned superiority of NAT in translation, NAD is especially suitable for sentence ordering from two aspects: (1) the length of the decoder is determined by the input set of sentences, and (2) the sentences and positions match exclusively."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.759,
+ 0.49,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "As suggested in (Gu et al., 2018), different inputs of non-autoregressive decoder may lead to quite different outputs resulting in a remarkable performance gap. Thus it is crucial to choose a proper information source for the input of NAD. Towards this end, we meticulously adopt positional information as the input of our non-autoregressive decoder. Because the sentence ordering problem could be interpreted as matching the unordered sentences to the right positions, and it is straightforward to"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.085,
+ 0.885,
+ 0.166
+ ],
+ "angle": 0,
+ "content": "take the position information as input and predict the correct sentence for each position in parallel. Specifically, we follow previous work (Vaswani et al., 2017) to project each discrete position into a compact representation \\( P \\) via:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.591,
+ 0.174,
+ 0.884,
+ 0.193
+ ],
+ "angle": 0,
+ "content": "\\[\np _ {i, 2 j} = \\sin \\left(i / 1 0 0 0 0 ^ {2 j / d _ {k}}\\right), \\tag {7}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.584,
+ 0.202,
+ 0.883,
+ 0.222
+ ],
+ "angle": 0,
+ "content": "\\[\np _ {i, 2 j + 1} = \\cos (i / 1 0 0 0 0 ^ {2 j / d _ {k}}), \\tag {8}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.227,
+ 0.885,
+ 0.291
+ ],
+ "angle": 0,
+ "content": "where \\(i\\) denotes the position and \\(j\\) is the \\(j\\)-th dimension in \\(p_i\\). We then exploit the interaction and relative information between positions implementing a multi-head self-attention as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.558,
+ 0.299,
+ 0.884,
+ 0.32
+ ],
+ "angle": 0,
+ "content": "\\[\n\\bar {P} = \\mathrm {M H A t t} \\left(P W ^ {Q}, P W ^ {K}, P W ^ {V}\\right), \\tag {9}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.329,
+ 0.884,
+ 0.472
+ ],
+ "angle": 0,
+ "content": "where MHAtt denotes multi-head self-attention block comprised of Equation 4-6. To connect the positions and the elements in the unordered context set, we further implement multi-head interattention between positions and sentences to exploit the correlations between them. Specifically, we query the contextual sentence representation set using \\(\\overline{P}\\), to obtain the attentive representations of sentences as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.581,
+ 0.479,
+ 0.884,
+ 0.515
+ ],
+ "angle": 0,
+ "content": "\\[\nE ^ {p} = \\operatorname {s o f t m a x} \\left(\\frac {\\bar {P} (E ^ {c}) ^ {T}}{\\sqrt {d _ {k}}}\\right) E ^ {c}, \\tag {10}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.522,
+ 0.884,
+ 0.635
+ ],
+ "angle": 0,
+ "content": "where \\( E^c \\) is the contextual sentence representation derived from the contextual encoder. With such an operation, the obtained \\( E^p \\) now is position sensitive representation, and could be used to select the most relevant sentence for each position. Specifically, the probability of the sentence \\( j \\) to be in the position \\( i \\) can be obtained by:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.576,
+ 0.644,
+ 0.883,
+ 0.664
+ ],
+ "angle": 0,
+ "content": "\\[\n\\omega_ {i j} = u ^ {T} \\tanh \\left(W _ {p} e _ {i} ^ {p} + W _ {b} e _ {j} ^ {b}\\right), \\tag {11}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.611,
+ 0.672,
+ 0.882,
+ 0.689
+ ],
+ "angle": 0,
+ "content": "\\[\nP t r _ {i} = \\text {s o f t m a x} (\\omega_ {i}), \\tag {12}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.695,
+ 0.885,
+ 0.776
+ ],
+ "angle": 0,
+ "content": "where \\( W_{*} \\) are learned parameters, and \\( u \\) is a column weight vector. \\( e_j^b \\) and \\( e_j^p \\) are the basic and positional sensitive representations of sentence \\( j \\). \\( P_{tr_i} \\) is the probabilistic distribution across all the sentences for \\( i \\)-th position."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.786,
+ 0.738,
+ 0.801
+ ],
+ "angle": 0,
+ "content": "2.5 Training and Inference"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.807,
+ 0.884,
+ 0.871
+ ],
+ "angle": 0,
+ "content": "Training: Following previous ordering works (Cui et al., 2018; Chen et al., 2016), the entire model could be trained to maximize the coherence probability by minimizing the cross entropy loss as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.578,
+ 0.88,
+ 0.883,
+ 0.922
+ ],
+ "angle": 0,
+ "content": "\\[\nL _ {c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log P \\left(o _ {i} \\mid p _ {i}; \\Theta\\right), \\tag {13}\n\\]"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.519,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "4201"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.182,
+ 0.089,
+ 0.429,
+ 0.288
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.114,
+ 0.301,
+ 0.49,
+ 0.346
+ ],
+ "angle": 0,
+ "content": "Figure 3: Illustration of our proposed exclusive loss, which simultaneously constrain the sentence selection and position matching."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.37,
+ 0.49,
+ 0.804
+ ],
+ "angle": 0,
+ "content": "where \\( o_i \\) is the ground-truth sentence for given \\( i \\)-th position \\( p_i \\), and \\( \\Theta \\) denotes all the learnable parameters in the model. As each sentence in the ordering problem is allowed to be selected only once, to avoid repeat selection, the RNN-based autoregressive model adopts a mask to remove selected sentences (Yin et al., 2020). However, the non-autoregressive model suffers from the repetition problem for its parallel decoding fashion (Gu et al., 2018). We have observed that the traditional pointer loss only penalizes the repetition for sentence choosing. Considering the characteristic of the ordering task, as depicted in Figure 3, every position should also choose one and only one sentence. Towards this end, we introduce an elegant exclusive loss to constrain the mutual exclusiveness of optimal matching between positions and elements. Specifically, we simultaneously calculate two pointers for sentence choosing and position choosing when we obtain the pointer map by Equation 11. Given the pointer map \\( \\omega \\), we implement softmax across column and row similar to Equation 12, where the column one interprets choosing sentence for each position and the row one assigns a position for each sentence. Then we optimize the entire model by minimizing the combined exclusive loss as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.129,
+ 0.815,
+ 0.488,
+ 0.87
+ ],
+ "angle": 0,
+ "content": "\\[\nL _ {e x} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\log P \\left(o _ {i} \\mid p _ {i}\\right) + \\log P \\left(p _ {i} \\mid o _ {i}\\right)\\right). \\tag {14}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.871,
+ 0.486,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "Through such an exclusive constraint, the repetition problem could be distinctly alleviated."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.904,
+ 0.486,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "Inference: Obviously, the proposed exclusive loss"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.085,
+ 0.885,
+ 0.327
+ ],
+ "angle": 0,
+ "content": "cannot completely avoid repetition problem during decoding. Similar with the masking strategy used in recurrent pointer networks, we introduce a simple yet effective greedy selective and removing strategy to ensure the exclusive matching for inference. Given the probability distribution array, as described in Algorithm 1, the proposed strategy greedily selects the largest probability and returns its index indicating the matching position and sentence. It then removes all the probability values of the corresponding column and row by setting them to zero. The process is repeated until all the values of the array are set to zero, which means that the matching process between positions and sentences is completely finished."
+ },
+ {
+ "type": "code_caption",
+ "bbox": [
+ 0.509,
+ 0.339,
+ 0.882,
+ 0.372
+ ],
+ "angle": 0,
+ "content": "Algorithm 1 Pseudo-code of greedy selective and removing in a PyTorch-like style."
+ },
+ {
+ "type": "algorithm",
+ "bbox": [
+ 0.526,
+ 0.374,
+ 0.884,
+ 0.496
+ ],
+ "angle": 0,
+ "content": "\\(\\# \\text{Ptr}\\) .. \\(N\\times N\\) output probability distribution array of NAD \nINPUTPtr \noutput \\(= []\\) \nwhile \\(\\mathrm{sum}(\\mathrm{Ptr}) > 0\\) . row_idx,col_idx \\(=\\) max_index(Ptr) #max_index(): get the index of the maximum value output.append((row_idx, col_idx)) Ptr[row_idx, :] \\(= 0\\) Ptr[:,col_idx] \\(= 0\\) \nRETURN output"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.51,
+ 0.525,
+ 0.656,
+ 0.541
+ ],
+ "angle": 0,
+ "content": "3 Experiments"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.551,
+ 0.884,
+ 0.631
+ ],
+ "angle": 0,
+ "content": "To verify the effectiveness of our NAON, we conduct extensive experiments on six datasets, and evaluate the performance with Acc, PMR, and \\(\\tau\\). More detailed experimental settings are described in the Appendix B."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.642,
+ 0.762,
+ 0.657
+ ],
+ "angle": 0,
+ "content": "3.1 Comparing with Baselines"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.662,
+ 0.885,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "As aforementioned, existing works lie in two categories: ranking-based and generation-based methods. To assess the proposed NAON, we compare it with the recent state-of-the-art methods, including Pairwise model via pairwise ranking (Chen et al., 2016), RankTxNet via deep attentive ranking (Kumar et al., 2020), and B-TSort searching correct order with topological sort (Prabhumoye et al., 2020) of ranking model, and the first architecture using pointer network LSTM-Ptr (Gong et al., 2016), SetLSTM designing an LSTM based attention to process the set encoding (Logeswaran et al., 2018), AON with a deep attentive order invariant paragraph encoder (Cui et al., 2018), HAN equipped with hierarchical attention (Wang and Wan, 2019), TGCM modeling topic-guide coherence (Oh et al.,"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "4202"
+ }
+ ],
+ [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.113,
+ 0.083,
+ 0.885,
+ 0.14
+ ],
+ "angle": 0,
+ "content": "Table 1: Comparison with baselines. The best and 2nd-best results are in bold and underlined, respectively. NAON-NE indicates no exclusive loss. Besides, we also take an attempt at relative order exploration and BART enhancement, namely NAON-RO and BART-NAON respectively, and compare them with BERSON and RE-BART, listed in the last two blocks in blue."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.115,
+ 0.15,
+ 0.887,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "| Unordered Sentences | Coherent Paragraph |
| S1. She poured herself a glass. | S2. Kelsi opened up a bottle of grape juice. |
| S2. Kelsi opened up a bottle of grape juice. | S1. She poured herself a glass. |
| S3. The carpet was now covered in juice. | S4. Accidentally she spilled some. |
| S4. Accidentally she spilled some. | S3. The carpet was now covered in juice. |
| S5. Kelsi just had to call the cleaners. | S5. Kelsi just had to call the cleaners. |
+
+Figure 1: An example of sentence ordering sourced from SIND dataset. The goal of sentence ordering is to understand the semantics and logic of a set of unordered sentences (the left box), and reorganize them to a coherent paragraph (the right box).
+
+sophisticated designs are costly in labor and time, expertise-required, and even cannot be well generalized to other scenarios. In the past couple of years, inspired by the great success of deep learning, dozens of deep neural methods for sentence ordering have been proposed and achieved great success (Cui et al., 2018; Kumar et al., 2020; Wang and Wan, 2019; Prabhumoye et al., 2020; Basu Roy Chowdhury et al., 2021).
+
+There exist two paradigms of neural sentence ordering: 1) ranking model via the scores of relative positions between paired sentences, and 2) generation model based on an encoder-decoder framework to predict the sentence order with sequence generation. The former one, ranking model (Prabhumoye et al., 2020; Chen et al., 2016), starts from calculating a score for each sentence pair to indicate their relative position in a coherent paragraph, then employs ranking or searching strategies to derive out the gold order. Since the score between paired sentences only measures the one-to-one relative position, which may fail to capture the one-to-many interactions on the global context level. Besides, the naive ranking process is somehow brute-force and time-consuming, and highly depends on the quality of sentence representation.
+
+The latter one, generation model regards sentence ordering as a sequence prediction problem to explore the relations between sentences and positions, which takes a set of sentences and generates
+
+the coherent sentence sequence. At the early stage, several encoder-decoder based approaches (Gong et al., 2016; Logeswaran et al., 2018) treat the unordered input sentences as a permutation sequence and encode them sequentially. Such sequential modeling encodes incorrect sentence order and semantic logic between sentences, which may mislead the decoder to predict an incoherent paragraph. Specifically, with sequential modeling, different permutations of the same paragraph may derive different paragraph representations and result in different output sentence order, which is not reasonable and inconsistent with the intuition of humans. To address this issue, Cui et al. (2018) first propose a deep attentive sentence ordering network equipped with self-attention to learn a reliable and consistent paragraph representation. Their model implements order-independent encoding with a non-positional variant of Transformer (Vaswani et al., 2017), but still predicts sentences via a pointer network as most previous works do. Nevertheless, the RNN-based pointer networks predict the sentence order step-by-step, which only explores the unilateral dependencies with past predictions and fails to leverage the comprehensive bilateral interactions between sentences for ordering.
+
+To overcome the above limitations, in this paper, we propose a novel Non-Autoregressive Ordering Network, dubbed NAON, which explores the bilateral dependencies between sentences and predicts the sentence for each position in parallel. Specifically, our NAON consists of a basic sentence encoder, an order-independent contextual sentence encoder and a non-autoregressive Transformer (NAT) (Gu et al., 2018) decoder. We first employ a sequence encoder, e.g., BERT in this work, to map discrete words in each sentence to a compact representation. The obtained sentence representations are injected to a Transformer (Vaswani et al., 2017) encoder to exploit the interaction and relation between sentences, and export contextual sentence representations. Note that we remove the positional embedding in the Transformer to achieve the order-invariant encoding for the sentences. In contrast to obtaining the paragraph representation via pooling to initialize the memory of auto-regressive decoder (Cui et al., 2018), we make full use of contextual sentence representations through multi-head attention in non-autoregressive decoder. We design a non-autoregressive decoder (NAD) that takes positions as input and predicts
+
+the sentence for each position utilizing a pointer network (Vinyals et al., 2015). In particular, we first employ multi-head attention to explore the relations between positions. Then the output distribution over the original sentence representation set for each position will be modelled by a pointer network. As mentioned in (Gu et al., 2018), naive non-autoregressive decoding suffers from severe problem of output repetition due to the complete conditional independence. To address this issue, we elaborate an exclusive loss constraining the exclusiveness between each sentence and position pair during training, and a greedy selective and removing strategy for inference. In summary, our main contributions are as follows:
+
+- We propose a novel approach, non-autoregressive ordering network (NAON), for sentence ordering, which implements self-attention to explore the bilateral interactions between sentences, going beyond the unilateral dependencies of RNN-based step-by-step decoding. To the best of our knowledge, our approach is one of the first attempts using non-autoregressive decoders for the sentence ordering problem.
+- To alleviate the repetition problem of vanilla non-autoregressive models, we design an exclusive loss under the implication of exclusive constraint between positions and sentences during training. For inference, we simply design a greedy selective and removing strategy to match each sentence-position pair.
+- Extensive experiments are conducted on several common-used datasets, and the results demonstrate the effectiveness of the proposed approach.
+
+# 2 The Proposed Approach
+
+# 2.1 Preliminary
+
+Sentence ordering aims to capture the coherence and recover the gold order of a set of unordered sentences. Specifically, given a sentence set $S = \{s_1, s_2, \ldots, s_N\}$ with $N$ sentences, and each sentence $s_i = [w_{i1}, w_{i2}, \ldots, w_{iL_i}]$ contains a sequence of $L_i$ words. The goal of sentence ordering is to find an order $O^* = [o_1^*, o_2^*, \ldots, o_N^*]$ conveying the coherent semantics, which can be modelled by:
+
+$$
+P (O ^ {*} | S) \geq P (O | S), \forall O \in \Omega , \tag {1}
+$$
+
+where $P(O|S)$ is the probability of order $O$ given $S$ , and $\Omega$ denotes the exhaustive set of all potential
+
+
+Figure 2: The overall flowchart of the proposed Non-Autoregressive Ordering Network (NAON), which consists of a basic sentence encoder, a contextual sentence encoder and a non-autoregressive decoder.
+
+order with the size of $A_N^N$ . Existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019) generate the coherent paragraph sentence-by-sentence to recover the gold order, and optimize the model by maximizing:
+
+$$
+\sum_ {j = 1} ^ {N} \log P \left(s _ {j} \mid s _ {1}, s _ {2}, \dots , s _ {j - 1}, S\right). \tag {2}
+$$
+
+While our non-autoregressive ordering network implements the optimization in parallel by maximizing the logarithmic probabilities for all sentences:
+
+$$
+\sum_ {j = 1} ^ {N} \log P (s _ {j} | S). \tag {3}
+$$
+
+# 2.2 Basic Sentence Representation
+
+To map discrete words of a sentence into compact representations, as shown on the left of Figure 2, a sequential encoder is first employed, e.g., an LSTM or BERT. Inspired by the success in language modeling of BERT (Devlin et al., 2019), we adopt a pre-trained BERT-base model for basic sentence representation. In particular, following previous work (Kumar et al., 2020), we concatenate the [CLS] token and word sequence as BERT encoder input and average the outputs of all the tokens to denote the entire sentence representation. To make the representation more compatible with sentence ordering task, we also fine-tune it in an end-to-end fashion on each dataset.
+
+# 2.3 Contextual Sentence Representation
+
+Most existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019; Logeswaran et al.,
+
+2018) provide a "paragraph representation" to initialize the hidden state and memory cell of the decoder. A straightforward way is assuming that the given order is a "pseudo order" of the unordered sentences, and implementing an RNN-based encoder (Wang and Wan, 2019), e.g., LSTM or GRU, to map sentence representations to a dense feature vector for the paragraph. While such pseudo order assumption introduces incorrect sentence order, which may induce semantic incoherence and mislead the decoder in recovering the ground-truth order. To address this issue, Cui et al. (2018) proposed a deep attentive ordering network integrating a self-attention mechanism to derive the order invariant sentence representations, and adopted mean pooling across sentences to obtain the paragraph representation. Different from existing autoregressive approaches which need to initialize the decoder memory, our non-autoregressive decoder does not require such "paragraph representation". We design a contextual sentence encoder to interact each sentence with all the sentences in the set and injects all the contextual sentence representations into the decoder to find the correct order.
+
+Obviously, it is important to exploit semantic relations between sentences for recovering coherence. Towards this end, we design a contextual sentence encoder, shown in the middle of Figure 2, to interact each sentence with other sentences. Specifically, after obtaining the basic sentence representations $E^{b} = \{e_{1}^{b}, e_{2}^{b}, \dots, e_{N}^{b}\}$ , we employ a Transformer-based architecture similar to (Cui et al., 2018), equipped with multi-head self-attention mechanism without the positional embedding, to exploit the
+
+contextual information as follows:
+
+$$
+\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V, \tag {4}
+$$
+
+$$
+\mathrm {M H} (Q, K, V) = \left[ H _ {1}, H _ {2}, \dots , H _ {h} \right] W, \tag {5}
+$$
+
+$$
+H _ {i} = \operatorname {A t t e n t i o n} \left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\right). \tag {6}
+$$
+
+To explore deep contextual interaction, we duplicate the multi-head self-attention block multiple times, e.g., 4 for NSF and arXiv datasets in this work. Finally, the contextual sentence representations $E^{c} = \{e_{1}^{c}, e_{2}^{c}, \ldots, e_{N}^{c}\}$ can be derived from the output of the last layer.
+
+# 2.4 Non-Autoregressive Prediction
+
+Almost all the previous generation approaches (Logeswaran et al., 2018; Cui et al., 2018; Basu Roy Chowdhury et al., 2021) adopt a recurrent-based architecture composing with a pointer network (Vinyals et al., 2015) to predict the ordered sentence step-by-step, which only leverages the unilateral dependencies with the past prediction, and cannot capture the coherent dependencies between all the sentences. Besides, as argued in (Bengio et al., 2015), the discrepancy between training and inference of RNN-based generation strategy may decrease the inference performance. Inspired by the potentials of the non-autoregressive transformer in neural machine translation, we devise a non-autoregressive decoder (NAD) for recovering the correct order of sentences, with the exploration of bilateral dependencies. Our NAD is designed based on a Non-Autoregressive Transformer (NAT) decoder in machine translation, which removes the autoregressive connections between steps and generates all the target words in parallel, rather than step-by-step (i.e., word-by-word). Beyond the aforementioned superiority of NAT in translation, NAD is especially suitable for sentence ordering from two aspects: (1) the length of the decoder is determined by the input set of sentences, and (2) the sentences and positions match exclusively.
+
+As suggested in (Gu et al., 2018), different inputs of non-autoregressive decoder may lead to quite different outputs resulting in a remarkable performance gap. Thus it is crucial to choose a proper information source for the input of NAD. Towards this end, we meticulously adopt positional information as the input of our non-autoregressive decoder. Because the sentence ordering problem could be interpreted as matching the unordered sentences to the right positions, and it is straightforward to
+
+take the position information as input and predict the correct sentence for each position in parallel. Specifically, we follow previous work (Vaswani et al., 2017) to project each discrete position into a compact representation $P$ via:
+
+$$
+p _ {i, 2 j} = \sin \left(i / 1 0 0 0 0 ^ {2 j / d _ {k}}\right), \tag {7}
+$$
+
+$$
+p _ {i, 2 j + 1} = \cos (i / 1 0 0 0 0 ^ {2 j / d _ {k}}), \tag {8}
+$$
+
+where $i$ denotes the position and $j$ is the $j$ -th dimension in $p_i$ . We then exploit the interaction and relative information between positions implementing a multi-head self-attention as:
+
+$$
+\bar {P} = \mathrm {M H A t t} \left(P W ^ {Q}, P W ^ {K}, P W ^ {V}\right), \tag {9}
+$$
+
+where MHAtt denotes multi-head self-attention block comprised of Equation 4-6. To connect the positions and the elements in the unordered context set, we further implement multi-head interattention between positions and sentences to exploit the correlations between them. Specifically, we query the contextual sentence representation set using $\overline{P}$ , to obtain the attentive representations of sentences as:
+
+$$
+E ^ {p} = \operatorname {s o f t m a x} \left(\frac {\bar {P} (E ^ {c}) ^ {T}}{\sqrt {d _ {k}}}\right) E ^ {c}, \tag {10}
+$$
+
+where $E^c$ is the contextual sentence representation derived from the contextual encoder. With such an operation, the obtained $E^p$ now is position sensitive representation, and could be used to select the most relevant sentence for each position. Specifically, the probability of the sentence $j$ to be in the position $i$ can be obtained by:
+
+$$
+\omega_ {i j} = u ^ {T} \tanh \left(W _ {p} e _ {i} ^ {p} + W _ {b} e _ {j} ^ {b}\right), \tag {11}
+$$
+
+$$
+P t r _ {i} = \text {s o f t m a x} (\omega_ {i}), \tag {12}
+$$
+
+where $W_{*}$ are learned parameters, and $u$ is a column weight vector. $e_j^b$ and $e_j^p$ are the basic and positional sensitive representations of sentence $j$ . $P_{tr_i}$ is the probabilistic distribution across all the sentences for $i$ -th position.
+
+# 2.5 Training and Inference
+
+Training: Following previous ordering works (Cui et al., 2018; Chen et al., 2016), the entire model could be trained to maximize the coherence probability by minimizing the cross entropy loss as:
+
+$$
+L _ {c} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log P \left(o _ {i} \mid p _ {i}; \Theta\right), \tag {13}
+$$
+
+
+Figure 3: Illustration of our proposed exclusive loss, which simultaneously constrain the sentence selection and position matching.
+
+where $o_i$ is the ground-truth sentence for given $i$ -th position $p_i$ , and $\Theta$ denotes all the learnable parameters in the model. As each sentence in the ordering problem is allowed to be selected only once, to avoid repeat selection, the RNN-based autoregressive model adopts a mask to remove selected sentences (Yin et al., 2020). However, the non-autoregressive model suffers from the repetition problem for its parallel decoding fashion (Gu et al., 2018). We have observed that the traditional pointer loss only penalizes the repetition for sentence choosing. Considering the characteristic of the ordering task, as depicted in Figure 3, every position should also choose one and only one sentence. Towards this end, we introduce an elegant exclusive loss to constrain the mutual exclusiveness of optimal matching between positions and elements. Specifically, we simultaneously calculate two pointers for sentence choosing and position choosing when we obtain the pointer map by Equation 11. Given the pointer map $\omega$ , we implement softmax across column and row similar to Equation 12, where the column one interprets choosing sentence for each position and the row one assigns a position for each sentence. Then we optimize the entire model by minimizing the combined exclusive loss as follows:
+
+$$
+L _ {e x} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\log P \left(o _ {i} \mid p _ {i}\right) + \log P \left(p _ {i} \mid o _ {i}\right)\right). \tag {14}
+$$
+
+Through such an exclusive constraint, the repetition problem could be distinctly alleviated.
+
+Inference: Obviously, the proposed exclusive loss
+
+cannot completely avoid repetition problem during decoding. Similar with the masking strategy used in recurrent pointer networks, we introduce a simple yet effective greedy selective and removing strategy to ensure the exclusive matching for inference. Given the probability distribution array, as described in Algorithm 1, the proposed strategy greedily selects the largest probability and returns its index indicating the matching position and sentence. It then removes all the probability values of the corresponding column and row by setting them to zero. The process is repeated until all the values of the array are set to zero, which means that the matching process between positions and sentences is completely finished.
+
+Algorithm 1 Pseudo-code of greedy selective and removing in a PyTorch-like style.
+$\# \text{Ptr}$ .. $N\times N$ output probability distribution array of NAD
+INPUTPtr
+output $= []$
+while $\mathrm{sum}(\mathrm{Ptr}) > 0$ . row_idx,col_idx $=$ max_index(Ptr) #max_index(): get the index of the maximum value output.append((row_idx, col_idx)) Ptr[row_idx, :] $= 0$ Ptr[:,col_idx] $= 0$
+RETURN output
+
+# 3 Experiments
+
+To verify the effectiveness of our NAON, we conduct extensive experiments on six datasets, and evaluate the performance with Acc, PMR, and $\tau$ . More detailed experimental settings are described in the Appendix B.
+
+# 3.1 Comparing with Baselines
+
+As aforementioned, existing works lie in two categories: ranking-based and generation-based methods. To assess the proposed NAON, we compare it with the recent state-of-the-art methods, including Pairwise model via pairwise ranking (Chen et al., 2016), RankTxNet via deep attentive ranking (Kumar et al., 2020), and B-TSort searching correct order with topological sort (Prabhumoye et al., 2020) of ranking model, and the first architecture using pointer network LSTM-Ptr (Gong et al., 2016), SetLSTM designing an LSTM based attention to process the set encoding (Logeswaran et al., 2018), AON with a deep attentive order invariant paragraph encoder (Cui et al., 2018), HAN equipped with hierarchical attention (Wang and Wan, 2019), TGCM modeling topic-guide coherence (Oh et al.,
+
+Table 1: Comparison with baselines. The best and 2nd-best results are in bold and underlined, respectively. NAON-NE indicates no exclusive loss. Besides, we also take an attempt at relative order exploration and BART enhancement, namely NAON-RO and BART-NAON respectively, and compare them with BERSON and RE-BART, listed in the last two blocks in blue.
+
+| Unordered Sentences | Coherent Paragraph |
| S1. She poured herself a glass. | S2. Kelsi opened up a bottle of grape juice. |
| S2. Kelsi opened up a bottle of grape juice. | S1. She poured herself a glass. |
| S3. The carpet was now covered in juice. | S4. Accidentally she spilled some. |
| S4. Accidentally she spilled some. | S3. The carpet was now covered in juice. |
| S5. Kelsi just had to call the cleaners. | S5. Kelsi just had to call the cleaners. |
",
+ "image_path": "84f73ce508bd482ef412c2c0c1b6a87233982b8083cecaef98301d6d9c4e0fc0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 285,
+ 526,
+ 346
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 285,
+ 526,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 285,
+ 526,
+ 346
+ ],
+ "type": "text",
+ "content": "Figure 1: An example of sentence ordering sourced from SIND dataset. The goal of sentence ordering is to understand the semantics and logic of a set of unordered sentences (the left box), and reorganize them to a coherent paragraph (the right box)."
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 366,
+ 525,
+ 488
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 366,
+ 525,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 366,
+ 525,
+ 488
+ ],
+ "type": "text",
+ "content": "sophisticated designs are costly in labor and time, expertise-required, and even cannot be well generalized to other scenarios. In the past couple of years, inspired by the great success of deep learning, dozens of deep neural methods for sentence ordering have been proposed and achieved great success (Cui et al., 2018; Kumar et al., 2020; Wang and Wan, 2019; Prabhumoye et al., 2020; Basu Roy Chowdhury et al., 2021)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 302,
+ 489,
+ 525,
+ 718
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 489,
+ 525,
+ 718
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 489,
+ 525,
+ 718
+ ],
+ "type": "text",
+ "content": "There exist two paradigms of neural sentence ordering: 1) ranking model via the scores of relative positions between paired sentences, and 2) generation model based on an encoder-decoder framework to predict the sentence order with sequence generation. The former one, ranking model (Prabhumoye et al., 2020; Chen et al., 2016), starts from calculating a score for each sentence pair to indicate their relative position in a coherent paragraph, then employs ranking or searching strategies to derive out the gold order. Since the score between paired sentences only measures the one-to-one relative position, which may fail to capture the one-to-many interactions on the global context level. Besides, the naive ranking process is somehow brute-force and time-consuming, and highly depends on the quality of sentence representation."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 719,
+ 526,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 719,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "content": "The latter one, generation model regards sentence ordering as a sequence prediction problem to explore the relations between sentences and positions, which takes a set of sentences and generates"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "4198"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 805
+ ],
+ "type": "footer",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 805
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 805
+ ],
+ "type": "text",
+ "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4198-4214"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 165,
+ 805,
+ 428,
+ 817
+ ],
+ "type": "footer",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 165,
+ 805,
+ 428,
+ 817
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 165,
+ 805,
+ 428,
+ 817
+ ],
+ "type": "text",
+ "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 293,
+ 423
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 293,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 293,
+ 423
+ ],
+ "type": "text",
+ "content": "the coherent sentence sequence. At the early stage, several encoder-decoder based approaches (Gong et al., 2016; Logeswaran et al., 2018) treat the unordered input sentences as a permutation sequence and encode them sequentially. Such sequential modeling encodes incorrect sentence order and semantic logic between sentences, which may mislead the decoder to predict an incoherent paragraph. Specifically, with sequential modeling, different permutations of the same paragraph may derive different paragraph representations and result in different output sentence order, which is not reasonable and inconsistent with the intuition of humans. To address this issue, Cui et al. (2018) first propose a deep attentive sentence ordering network equipped with self-attention to learn a reliable and consistent paragraph representation. Their model implements order-independent encoding with a non-positional variant of Transformer (Vaswani et al., 2017), but still predicts sentences via a pointer network as most previous works do. Nevertheless, the RNN-based pointer networks predict the sentence order step-by-step, which only explores the unilateral dependencies with past predictions and fails to leverage the comprehensive bilateral interactions between sentences for ordering."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 435,
+ 292,
+ 775
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 435,
+ 292,
+ 775
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 435,
+ 292,
+ 775
+ ],
+ "type": "text",
+ "content": "To overcome the above limitations, in this paper, we propose a novel Non-Autoregressive Ordering Network, dubbed NAON, which explores the bilateral dependencies between sentences and predicts the sentence for each position in parallel. Specifically, our NAON consists of a basic sentence encoder, an order-independent contextual sentence encoder and a non-autoregressive Transformer (NAT) (Gu et al., 2018) decoder. We first employ a sequence encoder, e.g., BERT in this work, to map discrete words in each sentence to a compact representation. The obtained sentence representations are injected to a Transformer (Vaswani et al., 2017) encoder to exploit the interaction and relation between sentences, and export contextual sentence representations. Note that we remove the positional embedding in the Transformer to achieve the order-invariant encoding for the sentences. In contrast to obtaining the paragraph representation via pooling to initialize the memory of auto-regressive decoder (Cui et al., 2018), we make full use of contextual sentence representations through multi-head attention in non-autoregressive decoder. We design a non-autoregressive decoder (NAD) that takes positions as input and predicts"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 302,
+ 71,
+ 527,
+ 274
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 527,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 527,
+ 274
+ ],
+ "type": "text",
+ "content": "the sentence for each position utilizing a pointer network (Vinyals et al., 2015). In particular, we first employ multi-head attention to explore the relations between positions. Then the output distribution over the original sentence representation set for each position will be modelled by a pointer network. As mentioned in (Gu et al., 2018), naive non-autoregressive decoding suffers from severe problem of output repetition due to the complete conditional independence. To address this issue, we elaborate an exclusive loss constraining the exclusiveness between each sentence and position pair during training, and a greedy selective and removing strategy for inference. In summary, our main contributions are as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 303,
+ 280,
+ 527,
+ 555
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 6,
+ "blocks": [
+ {
+ "bbox": [
+ 303,
+ 280,
+ 527,
+ 403
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 303,
+ 280,
+ 527,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 280,
+ 527,
+ 403
+ ],
+ "type": "text",
+ "content": "- We propose a novel approach, non-autoregressive ordering network (NAON), for sentence ordering, which implements self-attention to explore the bilateral interactions between sentences, going beyond the unilateral dependencies of RNN-based step-by-step decoding. To the best of our knowledge, our approach is one of the first attempts using non-autoregressive decoders for the sentence ordering problem."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 303,
+ 411,
+ 527,
+ 507
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 303,
+ 411,
+ 527,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 411,
+ 527,
+ 507
+ ],
+ "type": "text",
+ "content": "- To alleviate the repetition problem of vanilla non-autoregressive models, we design an exclusive loss under the implication of exclusive constraint between positions and sentences during training. For inference, we simply design a greedy selective and removing strategy to match each sentence-position pair."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 303,
+ 513,
+ 527,
+ 555
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 303,
+ 513,
+ 527,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 513,
+ 527,
+ 555
+ ],
+ "type": "text",
+ "content": "- Extensive experiments are conducted on several common-used datasets, and the results demonstrate the effectiveness of the proposed approach."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 570,
+ 449,
+ 584
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 570,
+ 449,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 570,
+ 449,
+ 584
+ ],
+ "type": "text",
+ "content": "2 The Proposed Approach"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 303,
+ 591,
+ 387,
+ 604
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 303,
+ 591,
+ 387,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 591,
+ 387,
+ 604
+ ],
+ "type": "text",
+ "content": "2.1 Preliminary"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "text",
+ "content": "Sentence ordering aims to capture the coherence and recover the gold order of a set of unordered sentences. Specifically, given a sentence set "
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "inline_equation",
+ "content": "S = \\{s_1, s_2, \\ldots, s_N\\}"
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "text",
+ "content": " with "
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "inline_equation",
+ "content": "N"
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "text",
+ "content": " sentences, and each sentence "
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "inline_equation",
+ "content": "s_i = [w_{i1}, w_{i2}, \\ldots, w_{iL_i}]"
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "text",
+ "content": " contains a sequence of "
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "inline_equation",
+ "content": "L_i"
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "text",
+ "content": " words. The goal of sentence ordering is to find an order "
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "inline_equation",
+ "content": "O^* = [o_1^*, o_2^*, \\ldots, o_N^*]"
+ },
+ {
+ "bbox": [
+ 302,
+ 608,
+ 527,
+ 717
+ ],
+ "type": "text",
+ "content": " conveying the coherent semantics, which can be modelled by:"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 345,
+ 724,
+ 525,
+ 739
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 345,
+ 724,
+ 525,
+ 739
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 724,
+ 525,
+ 739
+ ],
+ "type": "interline_equation",
+ "content": "P (O ^ {*} | S) \\geq P (O | S), \\forall O \\in \\Omega , \\tag {1}",
+ "image_path": "d16da8411d4c5dee517274dc790bcebd043179ffe89af733a270c0ea9c873951.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "P(O|S)"
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "text",
+ "content": " is the probability of order "
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "O"
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "text",
+ "content": " given "
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "S"
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "text",
+ "content": ", and "
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "\\Omega"
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 773
+ ],
+ "type": "text",
+ "content": " denotes the exhaustive set of all potential"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "text",
+ "content": "4199"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 104,
+ 68,
+ 495,
+ 253
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 104,
+ 68,
+ 495,
+ 253
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 68,
+ 495,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 68,
+ 495,
+ 253
+ ],
+ "type": "image",
+ "image_path": "3c0cce24d4272f86916a212ca75dd8e54d094ff552edd61d96ac65d379f59edc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 67,
+ 265,
+ 525,
+ 290
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 265,
+ 525,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 265,
+ 525,
+ 290
+ ],
+ "type": "text",
+ "content": "Figure 2: The overall flowchart of the proposed Non-Autoregressive Ordering Network (NAON), which consists of a basic sentence encoder, a contextual sentence encoder and a non-autoregressive decoder."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 379
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 379
+ ],
+ "type": "text",
+ "content": "order with the size of "
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 379
+ ],
+ "type": "inline_equation",
+ "content": "A_N^N"
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 379
+ ],
+ "type": "text",
+ "content": ". Existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019) generate the coherent paragraph sentence-by-sentence to recover the gold order, and optimize the model by maximizing:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 290,
+ 422
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 290,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 290,
+ 422
+ ],
+ "type": "interline_equation",
+ "content": "\\sum_ {j = 1} ^ {N} \\log P \\left(s _ {j} \\mid s _ {1}, s _ {2}, \\dots , s _ {j - 1}, S\\right). \\tag {2}",
+ "image_path": "5df61a12b162b55efa9c4b340c0974c53e37b4497d5747c485370ebb0fb19edf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 426,
+ 291,
+ 467
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 426,
+ 291,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 426,
+ 291,
+ 467
+ ],
+ "type": "text",
+ "content": "While our non-autoregressive ordering network implements the optimization in parallel by maximizing the logarithmic probabilities for all sentences:"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 139,
+ 471,
+ 290,
+ 508
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 471,
+ 290,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 471,
+ 290,
+ 508
+ ],
+ "type": "interline_equation",
+ "content": "\\sum_ {j = 1} ^ {N} \\log P (s _ {j} | S). \\tag {3}",
+ "image_path": "5bf9148912a722906e08373682b870f031df0640c5ab20280ce9f511960e93e1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 67,
+ 514,
+ 239,
+ 527
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 514,
+ 239,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 514,
+ 239,
+ 527
+ ],
+ "type": "text",
+ "content": "2.2 Basic Sentence Representation"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 67,
+ 531,
+ 291,
+ 719
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 531,
+ 291,
+ 719
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 531,
+ 291,
+ 719
+ ],
+ "type": "text",
+ "content": "To map discrete words of a sentence into compact representations, as shown on the left of Figure 2, a sequential encoder is first employed, e.g., an LSTM or BERT. Inspired by the success in language modeling of BERT (Devlin et al., 2019), we adopt a pre-trained BERT-base model for basic sentence representation. In particular, following previous work (Kumar et al., 2020), we concatenate the [CLS] token and word sequence as BERT encoder input and average the outputs of all the tokens to denote the entire sentence representation. To make the representation more compatible with sentence ordering task, we also fine-tune it in an end-to-end fashion on each dataset."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 67,
+ 729,
+ 265,
+ 741
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 729,
+ 265,
+ 741
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 729,
+ 265,
+ 741
+ ],
+ "type": "text",
+ "content": "2.3 Contextual Sentence Representation"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 67,
+ 746,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 746,
+ 291,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 746,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": "Most existing autoregressive approaches (Cui et al., 2018; Wang and Wan, 2019; Logeswaran et al.,"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 311,
+ 526,
+ 635
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 311,
+ 526,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 311,
+ 526,
+ 635
+ ],
+ "type": "text",
+ "content": "2018) provide a \"paragraph representation\" to initialize the hidden state and memory cell of the decoder. A straightforward way is assuming that the given order is a \"pseudo order\" of the unordered sentences, and implementing an RNN-based encoder (Wang and Wan, 2019), e.g., LSTM or GRU, to map sentence representations to a dense feature vector for the paragraph. While such pseudo order assumption introduces incorrect sentence order, which may induce semantic incoherence and mislead the decoder in recovering the ground-truth order. To address this issue, Cui et al. (2018) proposed a deep attentive ordering network integrating a self-attention mechanism to derive the order invariant sentence representations, and adopted mean pooling across sentences to obtain the paragraph representation. Different from existing autoregressive approaches which need to initialize the decoder memory, our non-autoregressive decoder does not require such \"paragraph representation\". We design a contextual sentence encoder to interact each sentence with all the sentences in the set and injects all the contextual sentence representations into the decoder to find the correct order."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "content": "Obviously, it is important to exploit semantic relations between sentences for recovering coherence. Towards this end, we design a contextual sentence encoder, shown in the middle of Figure 2, to interact each sentence with other sentences. Specifically, after obtaining the basic sentence representations "
+ },
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "E^{b} = \\{e_{1}^{b}, e_{2}^{b}, \\dots, e_{N}^{b}\\}"
+ },
+ {
+ "bbox": [
+ 302,
+ 638,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "content": ", we employ a Transformer-based architecture similar to (Cui et al., 2018), equipped with multi-head self-attention mechanism without the positional embedding, to exploit the"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "text",
+ "content": "4200"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 221,
+ 84
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 221,
+ 84
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 221,
+ 84
+ ],
+ "type": "text",
+ "content": "contextual information as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 77,
+ 89,
+ 291,
+ 120
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 89,
+ 291,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 89,
+ 291,
+ 120
+ ],
+ "type": "interline_equation",
+ "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {4}",
+ "image_path": "f67254d116c662b959c4ec6824f4ba644a6bb006f95f71c6bdc974db56e28888.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 96,
+ 126,
+ 290,
+ 141
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 96,
+ 126,
+ 290,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 96,
+ 126,
+ 290,
+ 141
+ ],
+ "type": "interline_equation",
+ "content": "\\mathrm {M H} (Q, K, V) = \\left[ H _ {1}, H _ {2}, \\dots , H _ {h} \\right] W, \\tag {5}",
+ "image_path": "39a49c3ac14feb3cf936a6c7d3a5880ecb1f6214d53db169fe6205a2ea442336.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 85,
+ 143,
+ 290,
+ 161
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 85,
+ 143,
+ 290,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 85,
+ 143,
+ 290,
+ 161
+ ],
+ "type": "interline_equation",
+ "content": "H _ {i} = \\operatorname {A t t e n t i o n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right). \\tag {6}",
+ "image_path": "12668fd5d3e3ef605787f0d0abdc939402b0d20a25bc41091a1e4acb01e4dcd5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 165,
+ 291,
+ 248
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 165,
+ 291,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 165,
+ 291,
+ 248
+ ],
+ "type": "text",
+ "content": "To explore deep contextual interaction, we duplicate the multi-head self-attention block multiple times, e.g., 4 for NSF and arXiv datasets in this work. Finally, the contextual sentence representations "
+ },
+ {
+ "bbox": [
+ 67,
+ 165,
+ 291,
+ 248
+ ],
+ "type": "inline_equation",
+ "content": "E^{c} = \\{e_{1}^{c}, e_{2}^{c}, \\ldots, e_{N}^{c}\\}"
+ },
+ {
+ "bbox": [
+ 67,
+ 165,
+ 291,
+ 248
+ ],
+ "type": "text",
+ "content": " can be derived from the output of the last layer."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 255,
+ 239,
+ 268
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 255,
+ 239,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 255,
+ 239,
+ 268
+ ],
+ "type": "text",
+ "content": "2.4 Non-Autoregressive Prediction"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 67,
+ 272,
+ 291,
+ 638
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 272,
+ 291,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 272,
+ 291,
+ 638
+ ],
+ "type": "text",
+ "content": "Almost all the previous generation approaches (Logeswaran et al., 2018; Cui et al., 2018; Basu Roy Chowdhury et al., 2021) adopt a recurrent-based architecture composing with a pointer network (Vinyals et al., 2015) to predict the ordered sentence step-by-step, which only leverages the unilateral dependencies with the past prediction, and cannot capture the coherent dependencies between all the sentences. Besides, as argued in (Bengio et al., 2015), the discrepancy between training and inference of RNN-based generation strategy may decrease the inference performance. Inspired by the potentials of the non-autoregressive transformer in neural machine translation, we devise a non-autoregressive decoder (NAD) for recovering the correct order of sentences, with the exploration of bilateral dependencies. Our NAD is designed based on a Non-Autoregressive Transformer (NAT) decoder in machine translation, which removes the autoregressive connections between steps and generates all the target words in parallel, rather than step-by-step (i.e., word-by-word). Beyond the aforementioned superiority of NAT in translation, NAD is especially suitable for sentence ordering from two aspects: (1) the length of the decoder is determined by the input set of sentences, and (2) the sentences and positions match exclusively."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 67,
+ 638,
+ 291,
+ 773
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 638,
+ 291,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 638,
+ 291,
+ 773
+ ],
+ "type": "text",
+ "content": "As suggested in (Gu et al., 2018), different inputs of non-autoregressive decoder may lead to quite different outputs resulting in a remarkable performance gap. Thus it is crucial to choose a proper information source for the input of NAD. Towards this end, we meticulously adopt positional information as the input of our non-autoregressive decoder. Because the sentence ordering problem could be interpreted as matching the unordered sentences to the right positions, and it is straightforward to"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 139
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 139
+ ],
+ "type": "text",
+ "content": "take the position information as input and predict the correct sentence for each position in parallel. Specifically, we follow previous work (Vaswani et al., 2017) to project each discrete position into a compact representation "
+ },
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 139
+ ],
+ "type": "inline_equation",
+ "content": "P"
+ },
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 139
+ ],
+ "type": "text",
+ "content": " via:"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 351,
+ 146,
+ 525,
+ 162
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 351,
+ 146,
+ 525,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 351,
+ 146,
+ 525,
+ 162
+ ],
+ "type": "interline_equation",
+ "content": "p _ {i, 2 j} = \\sin \\left(i / 1 0 0 0 0 ^ {2 j / d _ {k}}\\right), \\tag {7}",
+ "image_path": "3e3370a777f950482c6bf6a0a6475269024574b4db1b5d400988e62544a3e001.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 347,
+ 169,
+ 525,
+ 186
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 347,
+ 169,
+ 525,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 347,
+ 169,
+ 525,
+ 186
+ ],
+ "type": "interline_equation",
+ "content": "p _ {i, 2 j + 1} = \\cos (i / 1 0 0 0 0 ^ {2 j / d _ {k}}), \\tag {8}",
+ "image_path": "94a96c6dad79b4107f301794f8e7682d55118e67056d884a2fe473b4b1c43f31.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "inline_equation",
+ "content": "i"
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "text",
+ "content": " denotes the position and "
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "inline_equation",
+ "content": "j"
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "text",
+ "content": " is the "
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "inline_equation",
+ "content": "j"
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "text",
+ "content": "-th dimension in "
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "inline_equation",
+ "content": "p_i"
+ },
+ {
+ "bbox": [
+ 302,
+ 190,
+ 526,
+ 244
+ ],
+ "type": "text",
+ "content": ". We then exploit the interaction and relative information between positions implementing a multi-head self-attention as:"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 332,
+ 251,
+ 525,
+ 269
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 332,
+ 251,
+ 525,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 332,
+ 251,
+ 525,
+ 269
+ ],
+ "type": "interline_equation",
+ "content": "\\bar {P} = \\mathrm {M H A t t} \\left(P W ^ {Q}, P W ^ {K}, P W ^ {V}\\right), \\tag {9}",
+ "image_path": "61a827bf3e74169d7a8381e6c64090c2f35b92d5eae200af4844cf8e57e33d8f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 302,
+ 276,
+ 525,
+ 396
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 276,
+ 525,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 276,
+ 525,
+ 396
+ ],
+ "type": "text",
+ "content": "where MHAtt denotes multi-head self-attention block comprised of Equation 4-6. To connect the positions and the elements in the unordered context set, we further implement multi-head interattention between positions and sentences to exploit the correlations between them. Specifically, we query the contextual sentence representation set using "
+ },
+ {
+ "bbox": [
+ 302,
+ 276,
+ 525,
+ 396
+ ],
+ "type": "inline_equation",
+ "content": "\\overline{P}"
+ },
+ {
+ "bbox": [
+ 302,
+ 276,
+ 525,
+ 396
+ ],
+ "type": "text",
+ "content": ", to obtain the attentive representations of sentences as:"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 345,
+ 402,
+ 525,
+ 433
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 345,
+ 402,
+ 525,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 402,
+ 525,
+ 433
+ ],
+ "type": "interline_equation",
+ "content": "E ^ {p} = \\operatorname {s o f t m a x} \\left(\\frac {\\bar {P} (E ^ {c}) ^ {T}}{\\sqrt {d _ {k}}}\\right) E ^ {c}, \\tag {10}",
+ "image_path": "f8ad0f480372adbb1553b8acf0f539e19c6bda12bb0288fe66152038ee3108fc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "inline_equation",
+ "content": "E^c"
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "text",
+ "content": " is the contextual sentence representation derived from the contextual encoder. With such an operation, the obtained "
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "inline_equation",
+ "content": "E^p"
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "text",
+ "content": " now is position sensitive representation, and could be used to select the most relevant sentence for each position. Specifically, the probability of the sentence "
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "inline_equation",
+ "content": "j"
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "text",
+ "content": " to be in the position "
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "inline_equation",
+ "content": "i"
+ },
+ {
+ "bbox": [
+ 302,
+ 439,
+ 525,
+ 534
+ ],
+ "type": "text",
+ "content": " can be obtained by:"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 342,
+ 541,
+ 525,
+ 558
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 342,
+ 541,
+ 525,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 342,
+ 541,
+ 525,
+ 558
+ ],
+ "type": "interline_equation",
+ "content": "\\omega_ {i j} = u ^ {T} \\tanh \\left(W _ {p} e _ {i} ^ {p} + W _ {b} e _ {j} ^ {b}\\right), \\tag {11}",
+ "image_path": "43bd2ac2fa0d35b92d232a358f7efe69d0e2413276d2cb8f7347dc8df1184653.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 363,
+ 565,
+ 524,
+ 579
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 363,
+ 565,
+ 524,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 565,
+ 524,
+ 579
+ ],
+ "type": "interline_equation",
+ "content": "P t r _ {i} = \\text {s o f t m a x} (\\omega_ {i}), \\tag {12}",
+ "image_path": "58d492362141d6ea8d7eb66baf33331f59e35e928eb487a12092574ca2308eb3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "inline_equation",
+ "content": "W_{*}"
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": " are learned parameters, and "
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "inline_equation",
+ "content": "u"
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": " is a column weight vector. "
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "inline_equation",
+ "content": "e_j^b"
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "inline_equation",
+ "content": "e_j^p"
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": " are the basic and positional sensitive representations of sentence "
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "inline_equation",
+ "content": "j"
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": ". "
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "inline_equation",
+ "content": "P_{tr_i}"
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": " is the probabilistic distribution across all the sentences for "
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "inline_equation",
+ "content": "i"
+ },
+ {
+ "bbox": [
+ 302,
+ 584,
+ 526,
+ 652
+ ],
+ "type": "text",
+ "content": "-th position."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 302,
+ 661,
+ 439,
+ 673
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 661,
+ 439,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 661,
+ 439,
+ 673
+ ],
+ "type": "text",
+ "content": "2.5 Training and Inference"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 302,
+ 678,
+ 525,
+ 732
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 678,
+ 525,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 678,
+ 525,
+ 732
+ ],
+ "type": "text",
+ "content": "Training: Following previous ordering works (Cui et al., 2018; Chen et al., 2016), the entire model could be trained to maximize the coherence probability by minimizing the cross entropy loss as:"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 343,
+ 740,
+ 525,
+ 775
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 343,
+ 740,
+ 525,
+ 775
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 343,
+ 740,
+ 525,
+ 775
+ ],
+ "type": "interline_equation",
+ "content": "L _ {c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log P \\left(o _ {i} \\mid p _ {i}; \\Theta\\right), \\tag {13}",
+ "image_path": "47d61a1f348512da584a6216f5e6052ff70a34b7697cd0597713d513b35dfdd6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 308,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 308,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 308,
+ 791
+ ],
+ "type": "text",
+ "content": "4201"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 74,
+ 255,
+ 242
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 108,
+ 74,
+ 255,
+ 242
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 74,
+ 255,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 74,
+ 255,
+ 242
+ ],
+ "type": "image",
+ "image_path": "627068c9fcaac382cb196d0322e9dcc66362e23883e10f71487297c90d47ab52.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 67,
+ 253,
+ 291,
+ 290
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 253,
+ 291,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 253,
+ 291,
+ 290
+ ],
+ "type": "text",
+ "content": "Figure 3: Illustration of our proposed exclusive loss, which simultaneously constrain the sentence selection and position matching."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "inline_equation",
+ "content": "o_i"
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "text",
+ "content": " is the ground-truth sentence for given "
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "inline_equation",
+ "content": "i"
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "text",
+ "content": "-th position "
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "inline_equation",
+ "content": "p_i"
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "text",
+ "content": ", and "
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "inline_equation",
+ "content": "\\Theta"
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "text",
+ "content": " denotes all the learnable parameters in the model. As each sentence in the ordering problem is allowed to be selected only once, to avoid repeat selection, the RNN-based autoregressive model adopts a mask to remove selected sentences (Yin et al., 2020). However, the non-autoregressive model suffers from the repetition problem for its parallel decoding fashion (Gu et al., 2018). We have observed that the traditional pointer loss only penalizes the repetition for sentence choosing. Considering the characteristic of the ordering task, as depicted in Figure 3, every position should also choose one and only one sentence. Towards this end, we introduce an elegant exclusive loss to constrain the mutual exclusiveness of optimal matching between positions and elements. Specifically, we simultaneously calculate two pointers for sentence choosing and position choosing when we obtain the pointer map by Equation 11. Given the pointer map "
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "inline_equation",
+ "content": "\\omega"
+ },
+ {
+ "bbox": [
+ 67,
+ 311,
+ 291,
+ 676
+ ],
+ "type": "text",
+ "content": ", we implement softmax across column and row similar to Equation 12, where the column one interprets choosing sentence for each position and the row one assigns a position for each sentence. Then we optimize the entire model by minimizing the combined exclusive loss as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 76,
+ 685,
+ 290,
+ 731
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 76,
+ 685,
+ 290,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 76,
+ 685,
+ 290,
+ 731
+ ],
+ "type": "interline_equation",
+ "content": "L _ {e x} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\log P \\left(o _ {i} \\mid p _ {i}\\right) + \\log P \\left(p _ {i} \\mid o _ {i}\\right)\\right). \\tag {14}",
+ "image_path": "167dbfff3b12076afe93c78e4c225ef43278d2aa070404a8921e9074c3cc1783.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 732,
+ 289,
+ 758
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 732,
+ 289,
+ 758
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 732,
+ 289,
+ 758
+ ],
+ "type": "text",
+ "content": "Through such an exclusive constraint, the repetition problem could be distinctly alleviated."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 760,
+ 289,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 760,
+ 289,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 760,
+ 289,
+ 772
+ ],
+ "type": "text",
+ "content": "Inference: Obviously, the proposed exclusive loss"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 275
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 71,
+ 526,
+ 275
+ ],
+ "type": "text",
+ "content": "cannot completely avoid repetition problem during decoding. Similar with the masking strategy used in recurrent pointer networks, we introduce a simple yet effective greedy selective and removing strategy to ensure the exclusive matching for inference. Given the probability distribution array, as described in Algorithm 1, the proposed strategy greedily selects the largest probability and returns its index indicating the matching position and sentence. It then removes all the probability values of the corresponding column and row by setting them to zero. The process is repeated until all the values of the array are set to zero, which means that the matching process between positions and sentences is completely finished."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "code",
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 302,
+ 285,
+ 524,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 285,
+ 524,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 285,
+ 524,
+ 312
+ ],
+ "type": "text",
+ "content": "Algorithm 1 Pseudo-code of greedy selective and removing in a PyTorch-like style."
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "code_caption"
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "\\# \\text{Ptr}"
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "text",
+ "content": " .. "
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "N\\times N"
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "text",
+ "content": " output probability distribution array of NAD \nINPUTPtr \noutput "
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "= []"
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "text",
+ "content": " \nwhile "
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{sum}(\\mathrm{Ptr}) > 0"
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "text",
+ "content": " . row_idx,col_idx "
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "="
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "text",
+ "content": " max_index(Ptr) #max_index(): get the index of the maximum value output.append((row_idx, col_idx)) Ptr[row_idx, :] "
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "= 0"
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "text",
+ "content": " Ptr[:,col_idx] "
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "= 0"
+ },
+ {
+ "bbox": [
+ 312,
+ 314,
+ 525,
+ 417
+ ],
+ "type": "text",
+ "content": " \nRETURN output"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "code_body"
+ }
+ ],
+ "index": 8,
+ "sub_type": "algorithm"
+ },
+ {
+ "bbox": [
+ 303,
+ 441,
+ 390,
+ 454
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 303,
+ 441,
+ 390,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 441,
+ 390,
+ 454
+ ],
+ "type": "text",
+ "content": "3 Experiments"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 463,
+ 525,
+ 530
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 463,
+ 525,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 463,
+ 525,
+ 530
+ ],
+ "type": "text",
+ "content": "To verify the effectiveness of our NAON, we conduct extensive experiments on six datasets, and evaluate the performance with Acc, PMR, and "
+ },
+ {
+ "bbox": [
+ 302,
+ 463,
+ 525,
+ 530
+ ],
+ "type": "inline_equation",
+ "content": "\\tau"
+ },
+ {
+ "bbox": [
+ 302,
+ 463,
+ 525,
+ 530
+ ],
+ "type": "text",
+ "content": ". More detailed experimental settings are described in the Appendix B."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 539,
+ 453,
+ 552
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 539,
+ 453,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 539,
+ 453,
+ 552
+ ],
+ "type": "text",
+ "content": "3.1 Comparing with Baselines"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 302,
+ 556,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 556,
+ 526,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 556,
+ 526,
+ 772
+ ],
+ "type": "text",
+ "content": "As aforementioned, existing works lie in two categories: ranking-based and generation-based methods. To assess the proposed NAON, we compare it with the recent state-of-the-art methods, including Pairwise model via pairwise ranking (Chen et al., 2016), RankTxNet via deep attentive ranking (Kumar et al., 2020), and B-TSort searching correct order with topological sort (Prabhumoye et al., 2020) of ranking model, and the first architecture using pointer network LSTM-Ptr (Gong et al., 2016), SetLSTM designing an LSTM based attention to process the set encoding (Logeswaran et al., 2018), AON with a deep attentive order invariant paragraph encoder (Cui et al., 2018), HAN equipped with hierarchical attention (Wang and Wan, 2019), TGCM modeling topic-guide coherence (Oh et al.,"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "4202"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 68,
+ 126,
+ 527,
+ 314
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 67,
+ 69,
+ 526,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 69,
+ 526,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 69,
+ 526,
+ 117
+ ],
+ "type": "text",
+ "content": "Table 1: Comparison with baselines. The best and 2nd-best results are in bold and underlined, respectively. NAON-NE indicates no exclusive loss. Besides, we also take an attempt at relative order exploration and BART enhancement, namely NAON-RO and BART-NAON respectively, and compare them with BERSON and RE-BART, listed in the last two blocks in blue."
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 68,
+ 126,
+ 527,
+ 314
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 126,
+ 527,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 126,
+ 527,
+ 314
+ ],
+ "type": "table",
+ "html": "| Fig. | Phrase A | Phrase B | Rating | Sent. H | Sent. R-L |
| ✓ | a nearly terminal case | of the cutes | 4.11 | 5.33 | 2.00 |
| the franchise's best years | are long past | -4.11 | 0.33 | 1.00 |
| ✓ | all the excitement | of eating oatmeal | -3.00 | 2.00 | 2.33 |
| just a simple fable | done in an artless style | -2.56 | 1.00 | 1.33 |
| ✓ | a pressure cooker | of horrified awe | -2.56 | 1.00 | 3.67 |
| something creepy and vague | is in the works | 2.33 | 3.67 | 2.00 |
| a no-surprise series | of explosions and violence | -2.22 | 0.67 | 2.00 |
| the numerous scenes | of gory mayhem | -1.89 | 1.67 | 3.00 |
| the momentary joys | of pretty and weightless intellectual entertainment | -1.89 | 2.00 | 4.00 |
| flourishes - artsy fantasy sequences - | that simply feel wrong | -1.78 | 1.00 | 1.67 |
| the overall feel of the film | is pretty cheesy | -1.67 | 0.67 | 1.00 |
| the contrived nature | of its provocative conclusion | -1.67 | 1.33 | 2.00 |
| ✓ | fans of the animated wildlife adventure show | will be in warthog heaven | 1.56 | 5.67 | 5.00 |
| a pointed little chiller | about the frightening seductiveness of new technology | -1.56 | 2.67 | 4.00 |
| ✓ | a real human soul | buried beneath a spellbinding serpent's smirk | 1.56 | 4.67 | 5.00 |
| the comic heights | it obviously desired | -1.56 | 2.67 | 3.67 |
| the belly laughs | of lowbrow comedy | 1.56 | 4.67 | 4.00 |
| the intellectual and emotional pedigree | of your date | -1.56 | 2.67 | 3.33 |
| the experience of going to a film festival | is a rewarding one | 1.56 | 6.00 | 5.00 |
| a serious exploration | of nuclear terrorism | 1.56 | 4.33 | 4.00 |
| a simple tale | of an unlikely friendship | 1.56 | 5.33 | 4.00 |
| the real charm | of this trifle | -1.44 | 3.67 | 4.00 |
| the human story | is pushed to one side | -1.44 | 2.00 | 2.00 |
| a pale imitation | of the real deal | -1.44 | 0.67 | 1.00 |
| a main character | who sometimes defies sympathy | -1.44 | 1.67 | 2.33 |
| ✓ | the obligatory moments | of sentimental ooze | -1.44 | 2.00 | 3.00 |
| the actresses in the lead roles | are all more than competent | 1.44 | 5.33 | 5.00 |
| the debate it joins | is a necessary and timely one | 1.44 | 4.67 | 5.00 |
| an audacious tour | of the past | -1.33 | 2.67 | 3.67 |
| only a document | of the worst possibilities of mankind | 1.33 | 2.00 | 1.00 |
| the life experiences | of a particular theatrical family | 1.33 | 4.67 | 3.00 |
| the entire point | of a shaggy dog story | 1.33 | 4.00 | 3.00 |
| an inexpressible and drab wannabe | looking for that exact niche | 1.33 | 2.00 | 1.00 |
| everyone involved with moviemaking | is a con artist and a liar | -1.33 | 0.00 | 1.00 |
| a very original artist | in his medium | 1.33 | 5.33 | 5.00 |
| the visuals and eccentricities | of many of the characters | 1.33 | 4.33 | 3.00 |
| ✓ | a violent initiation rite | for the audience | -1.22 | 2.00 | 2.67 |
| these curious owners | of architectural oddities | -1.22 | 2.67 | 3.00 |
| this engaging mix | of love and bloodletting | 1.22 | 5.00 | 4.67 |
| a sudden lunch rush | at the diner | -1.22 | 2.67 | 3.00 |
| the true potential | of the medium | -1.22 | 3.00 | 4.00 |
| ✓ | a therapeutic zap | of shock treatment | -1.22 | 3.00 | 3.33 |
| the exotic world | of belly dancing | -1.22 | 3.00 | 3.67 |
| ✓ | the real story | starts just around the corner | 1.22 | 3.67 | 3.00 |
| the most offensive thing | about the movie | -1.22 | 1.00 | 1.00 |
| the issue of faith | is not explored very deeply | -1.22 | 1.00 | 2.00 |
| a big box | of consolation candy | 1.22 | 3.67 | 3.00 |
| the playful paranoia | of the film's past | 1.22 | 4.00 | 3.67 |
| the overlooked pitfalls | of such an endeavour | -1.22 | 2.00 | 3.00 |
| ✓ | the modern master of the chase sequence | returns with a chase to end all chases | 1.11 | 6.00 | 5.00 |
| the sheer beauty | of his images | 1.11 | 5.67 | 5.00 |
| a compelling slice | of awkward emotions | 1.11 | 4.67 | 5.00 |
| a great actress | tearing into a landmark role | -1.11 | 4.67 | 5.33 |
| much of the writing | is genuinely witty | 1.11 | 5.00 | 5.00 |
| an especially poignant portrait | of her friendship | -1.11 | 3.67 | 5.00 |
| the insight and honesty | of this disarming indie | -1.11 | 3.67 | 5.00 |
| the emotional arc | of its raw blues soundtrack | 1.11 | 4.33 | 3.67 |
| the kind of art shots | that fill gallery shows | 1.11 | 4.67 | 4.00 |
| a new software program | spit out the screenplay | 1.11 | 2.67 | 3.00 |
| a very good time | at the cinema | 1.11 | 5.67 | 5.67 |
",
+ "bbox": [
+ 115,
+ 123,
+ 882,
+ 821
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Table 3: 60 phrases with a rating that exceeds an absolute value of 1. We indicate the rating from the MAX variant of our resource. We indicate whether we believe the phrase to represent a figurative phrase, and also provide the sentiment assigned by the annotators (Sent. H.) and by ROBERTA-LARGE (Sent. R-L)",
+ "bbox": [
+ 112,
+ 829,
+ 884,
+ 873
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "page_number",
+ "text": "5161",
+ "bbox": [
+ 480,
+ 928,
+ 519,
+ 940
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "D Models & model evaluation",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 84,
+ 389,
+ 98
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Experimental setup We fine-tune the following models, all obtained from the HuggingFace model hub:",
+ "bbox": [
+ 112,
+ 109,
+ 882,
+ 124
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- ROBERTA-BASE https://huggingface.co/roberta-base",
+ "- ROBERTA-LARGE https://huggingface.co/roberta-large",
+ "- TIMELM (Loureiro et al., 2022):"
+ ],
+ "bbox": [
+ 134,
+ 136,
+ 626,
+ 183
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "https://huggingface.co/cardiffnlp/twitter-roberta-base",
+ "bbox": [
+ 149,
+ 185,
+ 652,
+ 199
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- BERTTweet (Nguyen et al., 2020): https://huggingface.co/vinai/bertweet-base",
+ "- Camacho-collados et al. (2022):"
+ ],
+ "bbox": [
+ 134,
+ 200,
+ 811,
+ 231
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "https://huggingface.co/cardiffnlp/twitter-rosberta-base-sentimentlatest",
+ "Pérez et al. (2021):"
+ ],
+ "bbox": [
+ 136,
+ 233,
+ 808,
+ 263
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "https://huggingface.co/finiteautomata/bertweet-base-sentiment-analysis",
+ "bbox": [
+ 149,
+ 265,
+ 800,
+ 279
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "Hartmann et al. (2021):",
+ "https://huggingface.co/j-hartmann/sentiment-roberta-large-english-3-classes",
+ "- IMDB ROBERTA-BASE: https://huggingface.co/textattack/roberta-base-imdb"
+ ],
+ "bbox": [
+ 134,
+ 281,
+ 847,
+ 328
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "The experimental setup used to fine-tune them on SST is as follows: we use a batch size of 32, learning rate $5e - 6$ , the Adam optimiser with a cosine-based warmup scheduler (warmup is $20\\%$ ). We train for 5 epochs, selecting the best model based on the validation data and evaluate that on the test sets. We do not experiment with these hyperparameters extensively, as they are within the range of recommended settings and all models use the same base models (ROBERTA-BASE or ROBERTA-LARGE).",
+ "bbox": [
+ 112,
+ 340,
+ 880,
+ 418
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "We train using NVIDIA A100-SXM-80GB GPUs on which training one model for one seed takes up to 30 minutes for the BASE model, and 50 minutes for the LARGE model.",
+ "bbox": [
+ 112,
+ 420,
+ 882,
+ 451
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "**Characterising our experiments** Through our experiments, we contribute to generalisation evaluation in NLP: evaluating models' generalisation capabilities beyond standard i.i.d. testing, in particular through the evaluation on the 67 most non-compositional phrases. We characterise our experiments using the taxonomy of Hupkes et al. (2023): our experiments are cognitively motivated and focus on (non-)compositional generalisation. By separating inputs based on compositionality, we create a covariate shift between train and test data. That shift has a partitioned natural source, since the stimuli are from a human-written corpus, but the split we create is curated. In our experiments, we examine the influence of fine-tuning on SST-7 and evaluating on non-compositional examples (fine-tune train-test locus) but also examine the effect of a data shift in the pretrain-train locus, by considering models with different pretraining corpora.",
+ "bbox": [
+ 112,
+ 461,
+ 882,
+ 621
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "image",
+ "img_path": "images/4e252e5c760fd260c87bb128ddf8bd69f9524d15e7068c1361ee328e69f8d333.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 233,
+ 644,
+ 768,
+ 800
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "page_number",
+ "text": "5162",
+ "bbox": [
+ 480,
+ 928,
+ 519,
+ 940
+ ],
+ "page_idx": 12
+ }
+]
\ No newline at end of file
diff --git a/2023/Non-Compositionality in Sentiment_ New Data and Analyses/3184a1e0-ef47-41f1-9868-1f2f8500adaa_model.json b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/3184a1e0-ef47-41f1-9868-1f2f8500adaa_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..0c01204e0420d8a41f6ea0632b9c7fefb3c7f5d6
--- /dev/null
+++ b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/3184a1e0-ef47-41f1-9868-1f2f8500adaa_model.json
@@ -0,0 +1,2173 @@
+[
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.186,
+ 0.091,
+ 0.817,
+ 0.113
+ ],
+ "angle": 0,
+ "content": "Non-Compositionality in Sentiment: New Data and Analyses"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.316,
+ 0.143,
+ 0.686,
+ 0.159
+ ],
+ "angle": 0,
+ "content": "Verna Dankers and Christopher G. Lucas"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.294,
+ 0.16,
+ 0.71,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "Institute for Language, Cognition and Computation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.402,
+ 0.178,
+ 0.6,
+ 0.193
+ ],
+ "angle": 0,
+ "content": "University of Edinburgh"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.304,
+ 0.194,
+ 0.7,
+ 0.21
+ ],
+ "angle": 0,
+ "content": "vernadankers@gmail.com, c.lucas@ed.ac.uk"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.261,
+ 0.269,
+ 0.342,
+ 0.285
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.142,
+ 0.292,
+ 0.461,
+ 0.533
+ ],
+ "angle": 0,
+ "content": "When natural language phrases are combined, their meaning is often more than the sum of their parts. In the context of NLP tasks such as sentiment analysis, where the meaning of a phrase is its sentiment, that still applies. Many NLP studies on sentiment analysis, however, focus on the fact that sentiment computations are largely compositional. We, instead, set out to obtain non-compositionality ratings for phrases with respect to their sentiment. Our contributions are as follows: a) a methodology for obtaining those non-compositionality ratings, b) a resource of ratings for 259 phrases - NON-COMPSST - along with an analysis of that resource, and c) an evaluation of computational models for sentiment analysis using this new resource."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.541,
+ 0.262,
+ 0.557
+ ],
+ "angle": 0,
+ "content": "1 Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.566,
+ 0.49,
+ 0.886
+ ],
+ "angle": 0,
+ "content": "In NLP, the topics of the compositionality of language and neural models' capabilities to compute meaning compositionally have gained substantial interest in recent years. Yet, the meaning of linguistic utterances often does not adhere to strict patterns and can be surprising when looking at the individual words involved. This affects how those utterances behave in downstream tasks, such as sentiment analysis. Given a phrase or sentence, that task involves predicting the polarity as positive, negative or neutral. Sentiment largely adheres to compositional principles (Moilanen and Pulman, 2007, p.1): \"If the meaning of a sentence is a function of the meanings of its parts then the global polarity of a sentence is a function of the polarities of its parts.\" Modelling sentiment as a compositional process is, therefore, often mentioned as a design principle for computational sentiment models (e.g. by Socher et al., 2013; Sutherland et al., 2020; Yin et al., 2020)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.888,
+ 0.489,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "Nonetheless, one can think of examples where the sentiment of a phrase is unexpected given the"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.512,
+ 0.266,
+ 0.688,
+ 0.345
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.536,
+ 0.35,
+ 0.664,
+ 0.372
+ ],
+ "angle": 0,
+ "content": "all the excitement + of eating oatmeal"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.706,
+ 0.266,
+ 0.883,
+ 0.345
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.719,
+ 0.35,
+ 0.871,
+ 0.372
+ ],
+ "angle": 0,
+ "content": "a nearly terminal case + of the cutes"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.509,
+ 0.386,
+ 0.886,
+ 0.457
+ ],
+ "angle": 0,
+ "content": "Figure 1: Illustration of how the observed sentiment deviates from the expected sentiment when viewing polarity as a function of the polarity of the subphrases. These examples were obtained from our newly annotated stimuli."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.484,
+ 0.885,
+ 0.709
+ ],
+ "angle": 0,
+ "content": "sentiment of the individual parts (e.g. see Zhu et al., 2015; Hwang and Hidey, 2019; Barnes et al., 2019; Tahayna et al., 2022, for work on non-compositional sentiment). These include the case of sarcasm (\"life is good, you should get one\"), opposing sentiments (\"terribrly fascinating\"), idiomatic expressions (\"break a leg\") and neutral terms that, when composed, suddenly convey sentiment (\"yeah right\"). Adequately capturing sentiment computationally requires both learning compositional rules and understanding when such exceptions exist, where most contemporary sentiment models are expected to learn that via mere end-to-end training on examples."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.71,
+ 0.886,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "How can we identify whether the sentiment of a phrase is non-compositional? We design a protocol to obtain such non-compositionality judgments based on human-annotated sentiment. Our methodology (elaborated on in §3) utilises phrases from the Stanford Sentiment Treebank (SST) (Socher et al., 2013) and contrasts the sentiment of a phrase with control stimuli, in which one of the two subphrases has been replaced. Phrases whose annotated sentiment deviates from what is expected based on the controls are considered less compositional, as is illustrated in Figure 1. We analyse the resulting non-compositionality ranking of phrases"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.48,
+ 0.928,
+ 0.522,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "5150"
+ },
+ {
+ "type": "footer",
+ "bbox": [
+ 0.218,
+ 0.946,
+ 0.78,
+ 0.959
+ ],
+ "angle": 0,
+ "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5150-5162"
+ },
+ {
+ "type": "footer",
+ "bbox": [
+ 0.278,
+ 0.959,
+ 0.72,
+ 0.973
+ ],
+ "angle": 0,
+ "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.085,
+ 0.491,
+ 0.198
+ ],
+ "angle": 0,
+ "content": "\\((\\S 4)\\) and show how the constructed resource can be used to evaluate sentiment models \\((\\S 5)\\). Our new resource (NONCOMPSST) can further improve the understanding of what underlies non-compositionality in sentiment analysis, and can complement existing evaluation protocols for sentiment analysis models."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.209,
+ 0.266,
+ 0.224
+ ],
+ "angle": 0,
+ "content": "2 Related work"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.234,
+ 0.49,
+ 0.475
+ ],
+ "angle": 0,
+ "content": "Over the course of years, sentiment analysis systems went from using rule-based models and sentiment lexicons (e.g. Moilanen and Pulman, 2007; Taboada et al., 2011) to using recursive neural networks (Socher et al., 2013; Tai et al., 2015; Zhu et al., 2015), to abandoning the use of structure altogether by finetuning pretrained large language models (e.g. Pérez et al., 2021; Camacho-collados et al., 2022; Hartmann et al., 2023), and recently, to abandoning training, via zero-shot generalisation (Wang et al., 2023). Crucial to the development of these systems has been the introduction of benchmarks, such as SST (Socher et al., 2013) and SemEval's Twitter benchmarks (e.g. Rosenthal et al., 2015; Nakov et al., 2016; Rosenthal et al., 2017)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.476,
+ 0.49,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "Although the vast majority of related work focused on simply improving on benchmarks, there have been studies more closely related to ours, asking questions such as: How do phrases with opposing sentiment affect each other (Kiritchenko and Mohammad, 2016a,b)? What is the role of negations (Zhu et al., 2014), modals and adverbs (Kiritchenko and Mohammad, 2016c)? Do idioms have non-compositional sentiment (Hwang and Hidey, 2019)? Which linguistic phenomena are still problematic for SOTA sentiment systems (Barnes et al., 2019)? Can we incorporate compositional and non-compositional processing in one system (Zhu et al., 2015)? And can we computationally rank sentences according to their sentiment compositionality (Dankers and Titov, 2022)? Gaining a better understanding of the contexts in which sentiment functions non-compositionally and is challenging to predict is crucial for the evaluation of sentiment models in an age where sentiment benchmarks may appear saturated (Barnes et al., 2019).| Fig. | Phrase A | Phrase B | Rating | Sent. H | Sent. R-L |
| ✓ | a nearly terminal case | of the cutes | 4.11 | 5.33 | 2.00 |
| the franchise's best years | are long past | -4.11 | 0.33 | 1.00 |
| ✓ | all the excitement | of eating oatmeal | -3.00 | 2.00 | 2.33 |
| just a simple fable | done in an artless style | -2.56 | 1.00 | 1.33 |
| ✓ | a pressure cooker | of horrified awe | -2.56 | 1.00 | 3.67 |
| something creepy and vague | is in the works | 2.33 | 3.67 | 2.00 |
| a no-surprise series | of explosions and violence | -2.22 | 0.67 | 2.00 |
| the numerous scenes | of gory mayhem | -1.89 | 1.67 | 3.00 |
| the momentary joys | of pretty and weightless intellectual entertainment | -1.89 | 2.00 | 4.00 |
| flourishes - artsy fantasy sequences - | that simply feel wrong | -1.78 | 1.00 | 1.67 |
| the overall feel of the film | is pretty cheesy | -1.67 | 0.67 | 1.00 |
| the contrived nature | of its provocative conclusion | -1.67 | 1.33 | 2.00 |
| ✓ | fans of the animated wildlife adventure show | will be in warthog heaven | 1.56 | 5.67 | 5.00 |
| a pointed little chiller | about the frightening seductiveness of new technology | -1.56 | 2.67 | 4.00 |
| ✓ | a real human soul | buried beneath a spellbinding serpent's smirk | 1.56 | 4.67 | 5.00 |
| the comic heights | it obviously desired | -1.56 | 2.67 | 3.67 |
| the belly laughs | of lowbrow comedy | 1.56 | 4.67 | 4.00 |
| the intellectual and emotional pedigree | of your date | -1.56 | 2.67 | 3.33 |
| the experience of going to a film festival | is a rewarding one | 1.56 | 6.00 | 5.00 |
| a serious exploration | of nuclear terrorism | 1.56 | 4.33 | 4.00 |
| a simple tale | of an unlikely friendship | 1.56 | 5.33 | 4.00 |
| the real charm | of this trifle | -1.44 | 3.67 | 4.00 |
| the human story | is pushed to one side | -1.44 | 2.00 | 2.00 |
| a pale imitation | of the real deal | -1.44 | 0.67 | 1.00 |
| a main character | who sometimes defies sympathy | -1.44 | 1.67 | 2.33 |
| ✓ | the obligatory moments | of sentimental ooze | -1.44 | 2.00 | 3.00 |
| the actresses in the lead roles | are all more than competent | 1.44 | 5.33 | 5.00 |
| the debate it joins | is a necessary and timely one | 1.44 | 4.67 | 5.00 |
| an audacious tour | of the past | -1.33 | 2.67 | 3.67 |
| only a document | of the worst possibilities of mankind | 1.33 | 2.00 | 1.00 |
| the life experiences | of a particular theatrical family | 1.33 | 4.67 | 3.00 |
| the entire point | of a shaggy dog story | 1.33 | 4.00 | 3.00 |
| an inexpressible and drab wannabe | looking for that exact niche | 1.33 | 2.00 | 1.00 |
| everyone involved with moviemaking | is a con artist and a liar | -1.33 | 0.00 | 1.00 |
| a very original artist | in his medium | 1.33 | 5.33 | 5.00 |
| the visuals and eccentricities | of many of the characters | 1.33 | 4.33 | 3.00 |
| ✓ | a violent initiation rite | for the audience | -1.22 | 2.00 | 2.67 |
| these curious owners | of architectural oddities | -1.22 | 2.67 | 3.00 |
| this engaging mix | of love and bloodletting | 1.22 | 5.00 | 4.67 |
| a sudden lunch rush | at the diner | -1.22 | 2.67 | 3.00 |
| the true potential | of the medium | -1.22 | 3.00 | 4.00 |
| ✓ | a therapeutic zap | of shock treatment | -1.22 | 3.00 | 3.33 |
| the exotic world | of belly dancing | -1.22 | 3.00 | 3.67 |
| ✓ | the real story | starts just around the corner | 1.22 | 3.67 | 3.00 |
| the most offensive thing | about the movie | -1.22 | 1.00 | 1.00 |
| the issue of faith | is not explored very deeply | -1.22 | 1.00 | 2.00 |
| a big box | of consolation candy | 1.22 | 3.67 | 3.00 |
| the playful paranoia | of the film's past | 1.22 | 4.00 | 3.67 |
| the overlooked pitfalls | of such an endeavour | -1.22 | 2.00 | 3.00 |
| ✓ | the modern master of the chase sequence | returns with a chase to end all chases | 1.11 | 6.00 | 5.00 |
| the sheer beauty | of his images | 1.11 | 5.67 | 5.00 |
| a compelling slice | of awkward emotions | 1.11 | 4.67 | 5.00 |
| a great actress | tearing into a landmark role | -1.11 | 4.67 | 5.33 |
| much of the writing | is genuinely witty | 1.11 | 5.00 | 5.00 |
| an especially poignant portrait | of her friendship | -1.11 | 3.67 | 5.00 |
| the insight and honesty | of this disarming indie | -1.11 | 3.67 | 5.00 |
| the emotional arc | of its raw blues soundtrack | 1.11 | 4.33 | 3.67 |
| the kind of art shots | that fill gallery shows | 1.11 | 4.67 | 4.00 |
| a new software program | spit out the screenplay | 1.11 | 2.67 | 3.00 |
| a very good time | at the cinema | 1.11 | 5.67 | 5.67 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.113,
+ 0.83,
+ 0.885,
+ 0.874
+ ],
+ "angle": 0,
+ "content": "Table 3: 60 phrases with a rating that exceeds an absolute value of 1. We indicate the rating from the MAX variant of our resource. We indicate whether we believe the phrase to represent a figurative phrase, and also provide the sentiment assigned by the annotators (Sent. H.) and by ROBERTA-LARGE (Sent. R-L)"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.929,
+ 0.52,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "5161"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.085,
+ 0.391,
+ 0.099
+ ],
+ "angle": 0,
+ "content": "D Models & model evaluation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.11,
+ 0.884,
+ 0.126
+ ],
+ "angle": 0,
+ "content": "Experimental setup We fine-tune the following models, all obtained from the HuggingFace model hub:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.136,
+ 0.137,
+ 0.608,
+ 0.152
+ ],
+ "angle": 0,
+ "content": "- ROBERTA-BASE https://huggingface.co/roberta-base"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.136,
+ 0.153,
+ 0.627,
+ 0.169
+ ],
+ "angle": 0,
+ "content": "- ROBERTA-LARGE https://huggingface.co/roberta-large"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.137,
+ 0.17,
+ 0.4,
+ 0.184
+ ],
+ "angle": 0,
+ "content": "- TIMELM (Loureiro et al., 2022):"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.136,
+ 0.137,
+ 0.627,
+ 0.184
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.15,
+ 0.186,
+ 0.653,
+ 0.2
+ ],
+ "angle": 0,
+ "content": "https://huggingface.co/cardiffnlp/twitter-roberta-base"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.136,
+ 0.202,
+ 0.812,
+ 0.216
+ ],
+ "angle": 0,
+ "content": "- BERTTweet (Nguyen et al., 2020): https://huggingface.co/vinai/bertweet-base"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.137,
+ 0.218,
+ 0.392,
+ 0.232
+ ],
+ "angle": 0,
+ "content": "- Camacho-collados et al. (2022):"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.136,
+ 0.202,
+ 0.812,
+ 0.232
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.15,
+ 0.234,
+ 0.81,
+ 0.248
+ ],
+ "angle": 0,
+ "content": "https://huggingface.co/cardiffnlp/twitter-rosberta-base-sentimentlatest"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.137,
+ 0.25,
+ 0.296,
+ 0.264
+ ],
+ "angle": 0,
+ "content": "Pérez et al. (2021):"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.137,
+ 0.234,
+ 0.81,
+ 0.264
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.15,
+ 0.266,
+ 0.8,
+ 0.28
+ ],
+ "angle": 0,
+ "content": "https://huggingface.co/finiteautomata/bertweet-base-sentiment-analysis"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.137,
+ 0.282,
+ 0.329,
+ 0.297
+ ],
+ "angle": 0,
+ "content": "Hartmann et al. (2021):"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.15,
+ 0.298,
+ 0.848,
+ 0.312
+ ],
+ "angle": 0,
+ "content": "https://huggingface.co/j-hartmann/sentiment-roberta-large-english-3-classes"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.136,
+ 0.314,
+ 0.817,
+ 0.329
+ ],
+ "angle": 0,
+ "content": "- IMDB ROBERTA-BASE: https://huggingface.co/textattack/roberta-base-imdb"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.136,
+ 0.282,
+ 0.848,
+ 0.329
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.341,
+ 0.882,
+ 0.419
+ ],
+ "angle": 0,
+ "content": "The experimental setup used to fine-tune them on SST is as follows: we use a batch size of 32, learning rate \\(5e - 6\\), the Adam optimiser with a cosine-based warmup scheduler (warmup is \\(20\\%\\)). We train for 5 epochs, selecting the best model based on the validation data and evaluate that on the test sets. We do not experiment with these hyperparameters extensively, as they are within the range of recommended settings and all models use the same base models (ROBERTA-BASE or ROBERTA-LARGE)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.421,
+ 0.884,
+ 0.452
+ ],
+ "angle": 0,
+ "content": "We train using NVIDIA A100-SXM-80GB GPUs on which training one model for one seed takes up to 30 minutes for the BASE model, and 50 minutes for the LARGE model."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.462,
+ 0.884,
+ 0.623
+ ],
+ "angle": 0,
+ "content": "**Characterising our experiments** Through our experiments, we contribute to generalisation evaluation in NLP: evaluating models' generalisation capabilities beyond standard i.i.d. testing, in particular through the evaluation on the 67 most non-compositional phrases. We characterise our experiments using the taxonomy of Hupkes et al. (2023): our experiments are cognitively motivated and focus on (non-)compositional generalisation. By separating inputs based on compositionality, we create a covariate shift between train and test data. That shift has a partitioned natural source, since the stimuli are from a human-written corpus, but the split we create is curated. In our experiments, we examine the influence of fine-tuning on SST-7 and evaluating on non-compositional examples (fine-tune train-test locus) but also examine the effect of a data shift in the pretrain-train locus, by considering models with different pretraining corpora."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.234,
+ 0.645,
+ 0.769,
+ 0.801
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.929,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "5162"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/2023/Non-Compositionality in Sentiment_ New Data and Analyses/3184a1e0-ef47-41f1-9868-1f2f8500adaa_origin.pdf b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/3184a1e0-ef47-41f1-9868-1f2f8500adaa_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..da266116ede77ac88604d62398fd1a39396268bc
--- /dev/null
+++ b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/3184a1e0-ef47-41f1-9868-1f2f8500adaa_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:882a354eca43b0e82fab1e0c90d944bd88c3ad89cd41cb1ba36df01db8e3032b
+size 684193
diff --git a/2023/Non-Compositionality in Sentiment_ New Data and Analyses/full.md b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..62571abced52aa95124ac5b9f5d5efaaa45bc2bf
--- /dev/null
+++ b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/full.md
@@ -0,0 +1,280 @@
+# Non-Compositionality in Sentiment: New Data and Analyses
+
+Verna Dankers and Christopher G. Lucas
+
+Institute for Language, Cognition and Computation
+
+University of Edinburgh
+
+vernadankers@gmail.com, c.lucas@ed.ac.uk
+
+# Abstract
+
+When natural language phrases are combined, their meaning is often more than the sum of their parts. In the context of NLP tasks such as sentiment analysis, where the meaning of a phrase is its sentiment, that still applies. Many NLP studies on sentiment analysis, however, focus on the fact that sentiment computations are largely compositional. We, instead, set out to obtain non-compositionality ratings for phrases with respect to their sentiment. Our contributions are as follows: a) a methodology for obtaining those non-compositionality ratings, b) a resource of ratings for 259 phrases - NON-COMPSST - along with an analysis of that resource, and c) an evaluation of computational models for sentiment analysis using this new resource.
+
+# 1 Introduction
+
+In NLP, the topics of the compositionality of language and neural models' capabilities to compute meaning compositionally have gained substantial interest in recent years. Yet, the meaning of linguistic utterances often does not adhere to strict patterns and can be surprising when looking at the individual words involved. This affects how those utterances behave in downstream tasks, such as sentiment analysis. Given a phrase or sentence, that task involves predicting the polarity as positive, negative or neutral. Sentiment largely adheres to compositional principles (Moilanen and Pulman, 2007, p.1): "If the meaning of a sentence is a function of the meanings of its parts then the global polarity of a sentence is a function of the polarities of its parts." Modelling sentiment as a compositional process is, therefore, often mentioned as a design principle for computational sentiment models (e.g. by Socher et al., 2013; Sutherland et al., 2020; Yin et al., 2020).
+
+Nonetheless, one can think of examples where the sentiment of a phrase is unexpected given the
+
+
+all the excitement + of eating oatmeal
+
+
+a nearly terminal case + of the cutes
+Figure 1: Illustration of how the observed sentiment deviates from the expected sentiment when viewing polarity as a function of the polarity of the subphrases. These examples were obtained from our newly annotated stimuli.
+
+sentiment of the individual parts (e.g. see Zhu et al., 2015; Hwang and Hidey, 2019; Barnes et al., 2019; Tahayna et al., 2022, for work on non-compositional sentiment). These include the case of sarcasm ("life is good, you should get one"), opposing sentiments ("terribrly fascinating"), idiomatic expressions ("break a leg") and neutral terms that, when composed, suddenly convey sentiment ("yeah right"). Adequately capturing sentiment computationally requires both learning compositional rules and understanding when such exceptions exist, where most contemporary sentiment models are expected to learn that via mere end-to-end training on examples.
+
+How can we identify whether the sentiment of a phrase is non-compositional? We design a protocol to obtain such non-compositionality judgments based on human-annotated sentiment. Our methodology (elaborated on in §3) utilises phrases from the Stanford Sentiment Treebank (SST) (Socher et al., 2013) and contrasts the sentiment of a phrase with control stimuli, in which one of the two subphrases has been replaced. Phrases whose annotated sentiment deviates from what is expected based on the controls are considered less compositional, as is illustrated in Figure 1. We analyse the resulting non-compositionality ranking of phrases
+
+$(\S 4)$ and show how the constructed resource can be used to evaluate sentiment models $(\S 5)$ . Our new resource (NONCOMPSST) can further improve the understanding of what underlies non-compositionality in sentiment analysis, and can complement existing evaluation protocols for sentiment analysis models.
+
+# 2 Related work
+
+Over the course of years, sentiment analysis systems went from using rule-based models and sentiment lexicons (e.g. Moilanen and Pulman, 2007; Taboada et al., 2011) to using recursive neural networks (Socher et al., 2013; Tai et al., 2015; Zhu et al., 2015), to abandoning the use of structure altogether by finetuning pretrained large language models (e.g. Pérez et al., 2021; Camacho-collados et al., 2022; Hartmann et al., 2023), and recently, to abandoning training, via zero-shot generalisation (Wang et al., 2023). Crucial to the development of these systems has been the introduction of benchmarks, such as SST (Socher et al., 2013) and SemEval's Twitter benchmarks (e.g. Rosenthal et al., 2015; Nakov et al., 2016; Rosenthal et al., 2017).
+
+Although the vast majority of related work focused on simply improving on benchmarks, there have been studies more closely related to ours, asking questions such as: How do phrases with opposing sentiment affect each other (Kiritchenko and Mohammad, 2016a,b)? What is the role of negations (Zhu et al., 2014), modals and adverbs (Kiritchenko and Mohammad, 2016c)? Do idioms have non-compositional sentiment (Hwang and Hidey, 2019)? Which linguistic phenomena are still problematic for SOTA sentiment systems (Barnes et al., 2019)? Can we incorporate compositional and non-compositional processing in one system (Zhu et al., 2015)? And can we computationally rank sentences according to their sentiment compositionality (Dankers and Titov, 2022)? Gaining a better understanding of the contexts in which sentiment functions non-compositionally and is challenging to predict is crucial for the evaluation of sentiment models in an age where sentiment benchmarks may appear saturated (Barnes et al., 2019).| Fig. | Phrase A | Phrase B | Rating | Sent. H | Sent. R-L |
| ✓ | a nearly terminal case | of the cutes | 4.11 | 5.33 | 2.00 |
| the franchise's best years | are long past | -4.11 | 0.33 | 1.00 |
| ✓ | all the excitement | of eating oatmeal | -3.00 | 2.00 | 2.33 |
| just a simple fable | done in an artless style | -2.56 | 1.00 | 1.33 |
| ✓ | a pressure cooker | of horrified awe | -2.56 | 1.00 | 3.67 |
| something creepy and vague | is in the works | 2.33 | 3.67 | 2.00 |
| a no-surprise series | of explosions and violence | -2.22 | 0.67 | 2.00 |
| the numerous scenes | of gory mayhem | -1.89 | 1.67 | 3.00 |
| the momentary joys | of pretty and weightless intellectual entertainment | -1.89 | 2.00 | 4.00 |
| flourishes - artsy fantasy sequences - | that simply feel wrong | -1.78 | 1.00 | 1.67 |
| the overall feel of the film | is pretty cheesy | -1.67 | 0.67 | 1.00 |
| the contrived nature | of its provocative conclusion | -1.67 | 1.33 | 2.00 |
| ✓ | fans of the animated wildlife adventure show | will be in warthog heaven | 1.56 | 5.67 | 5.00 |
| a pointed little chiller | about the frightening seductiveness of new technology | -1.56 | 2.67 | 4.00 |
| ✓ | a real human soul | buried beneath a spellbinding serpent's smirk | 1.56 | 4.67 | 5.00 |
| the comic heights | it obviously desired | -1.56 | 2.67 | 3.67 |
| the belly laughs | of lowbrow comedy | 1.56 | 4.67 | 4.00 |
| the intellectual and emotional pedigree | of your date | -1.56 | 2.67 | 3.33 |
| the experience of going to a film festival | is a rewarding one | 1.56 | 6.00 | 5.00 |
| a serious exploration | of nuclear terrorism | 1.56 | 4.33 | 4.00 |
| a simple tale | of an unlikely friendship | 1.56 | 5.33 | 4.00 |
| the real charm | of this trifle | -1.44 | 3.67 | 4.00 |
| the human story | is pushed to one side | -1.44 | 2.00 | 2.00 |
| a pale imitation | of the real deal | -1.44 | 0.67 | 1.00 |
| a main character | who sometimes defies sympathy | -1.44 | 1.67 | 2.33 |
| ✓ | the obligatory moments | of sentimental ooze | -1.44 | 2.00 | 3.00 |
| the actresses in the lead roles | are all more than competent | 1.44 | 5.33 | 5.00 |
| the debate it joins | is a necessary and timely one | 1.44 | 4.67 | 5.00 |
| an audacious tour | of the past | -1.33 | 2.67 | 3.67 |
| only a document | of the worst possibilities of mankind | 1.33 | 2.00 | 1.00 |
| the life experiences | of a particular theatrical family | 1.33 | 4.67 | 3.00 |
| the entire point | of a shaggy dog story | 1.33 | 4.00 | 3.00 |
| an inexpressible and drab wannabe | looking for that exact niche | 1.33 | 2.00 | 1.00 |
| everyone involved with moviemaking | is a con artist and a liar | -1.33 | 0.00 | 1.00 |
| a very original artist | in his medium | 1.33 | 5.33 | 5.00 |
| the visuals and eccentricities | of many of the characters | 1.33 | 4.33 | 3.00 |
| ✓ | a violent initiation rite | for the audience | -1.22 | 2.00 | 2.67 |
| these curious owners | of architectural oddities | -1.22 | 2.67 | 3.00 |
| this engaging mix | of love and bloodletting | 1.22 | 5.00 | 4.67 |
| a sudden lunch rush | at the diner | -1.22 | 2.67 | 3.00 |
| the true potential | of the medium | -1.22 | 3.00 | 4.00 |
| ✓ | a therapeutic zap | of shock treatment | -1.22 | 3.00 | 3.33 |
| the exotic world | of belly dancing | -1.22 | 3.00 | 3.67 |
| ✓ | the real story | starts just around the corner | 1.22 | 3.67 | 3.00 |
| the most offensive thing | about the movie | -1.22 | 1.00 | 1.00 |
| the issue of faith | is not explored very deeply | -1.22 | 1.00 | 2.00 |
| a big box | of consolation candy | 1.22 | 3.67 | 3.00 |
| the playful paranoia | of the film's past | 1.22 | 4.00 | 3.67 |
| the overlooked pitfalls | of such an endeavour | -1.22 | 2.00 | 3.00 |
| ✓ | the modern master of the chase sequence | returns with a chase to end all chases | 1.11 | 6.00 | 5.00 |
| the sheer beauty | of his images | 1.11 | 5.67 | 5.00 |
| a compelling slice | of awkward emotions | 1.11 | 4.67 | 5.00 |
| a great actress | tearing into a landmark role | -1.11 | 4.67 | 5.33 |
| much of the writing | is genuinely witty | 1.11 | 5.00 | 5.00 |
| an especially poignant portrait | of her friendship | -1.11 | 3.67 | 5.00 |
| the insight and honesty | of this disarming indie | -1.11 | 3.67 | 5.00 |
| the emotional arc | of its raw blues soundtrack | 1.11 | 4.33 | 3.67 |
| the kind of art shots | that fill gallery shows | 1.11 | 4.67 | 4.00 |
| a new software program | spit out the screenplay | 1.11 | 2.67 | 3.00 |
| a very good time | at the cinema | 1.11 | 5.67 | 5.67 |
+
+Table 3: 60 phrases with a rating that exceeds an absolute value of 1. We indicate the rating from the MAX variant of our resource. We indicate whether we believe the phrase to represent a figurative phrase, and also provide the sentiment assigned by the annotators (Sent. H.) and by ROBERTA-LARGE (Sent. R-L)
+
+# D Models & model evaluation
+
+Experimental setup We fine-tune the following models, all obtained from the HuggingFace model hub:
+
+- ROBERTA-BASE https://huggingface.co/roberta-base
+- ROBERTA-LARGE https://huggingface.co/roberta-large
+- TIMELM (Loureiro et al., 2022):
+
+https://huggingface.co/cardiffnlp/twitter-roberta-base
+
+- BERTTweet (Nguyen et al., 2020): https://huggingface.co/vinai/bertweet-base
+- Camacho-collados et al. (2022):
+
+https://huggingface.co/cardiffnlp/twitter-rosberta-base-sentimentlatest
+Pérez et al. (2021):
+
+https://huggingface.co/finiteautomata/bertweet-base-sentiment-analysis
+
+Hartmann et al. (2021):
+https://huggingface.co/j-hartmann/sentiment-roberta-large-english-3-classes
+- IMDB ROBERTA-BASE: https://huggingface.co/textattack/roberta-base-imdb
+
+The experimental setup used to fine-tune them on SST is as follows: we use a batch size of 32, learning rate $5e - 6$ , the Adam optimiser with a cosine-based warmup scheduler (warmup is $20\%$ ). We train for 5 epochs, selecting the best model based on the validation data and evaluate that on the test sets. We do not experiment with these hyperparameters extensively, as they are within the range of recommended settings and all models use the same base models (ROBERTA-BASE or ROBERTA-LARGE).
+
+We train using NVIDIA A100-SXM-80GB GPUs on which training one model for one seed takes up to 30 minutes for the BASE model, and 50 minutes for the LARGE model.
+
+**Characterising our experiments** Through our experiments, we contribute to generalisation evaluation in NLP: evaluating models' generalisation capabilities beyond standard i.i.d. testing, in particular through the evaluation on the 67 most non-compositional phrases. We characterise our experiments using the taxonomy of Hupkes et al. (2023): our experiments are cognitively motivated and focus on (non-)compositional generalisation. By separating inputs based on compositionality, we create a covariate shift between train and test data. That shift has a partitioned natural source, since the stimuli are from a human-written corpus, but the split we create is curated. In our experiments, we examine the influence of fine-tuning on SST-7 and evaluating on non-compositional examples (fine-tune train-test locus) but also examine the effect of a data shift in the pretrain-train locus, by considering models with different pretraining corpora.
+
+
\ No newline at end of file
diff --git a/2023/Non-Compositionality in Sentiment_ New Data and Analyses/images.zip b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/images.zip
new file mode 100644
index 0000000000000000000000000000000000000000..0a087c74e5b4c81bde5f9c8f1c08d5f140609bf6
--- /dev/null
+++ b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/images.zip
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33a5b3d784789de50b2451d198e6531c93cf0e1d022f9273856291626da9a70c
+size 731460
diff --git a/2023/Non-Compositionality in Sentiment_ New Data and Analyses/layout.json b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..199218bdc27cad7dd2326b6cb6f539402d81e8ea
--- /dev/null
+++ b/2023/Non-Compositionality in Sentiment_ New Data and Analyses/layout.json
@@ -0,0 +1,7972 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 110,
+ 76,
+ 486,
+ 95
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 76,
+ 486,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 76,
+ 486,
+ 95
+ ],
+ "type": "text",
+ "content": "Non-Compositionality in Sentiment: New Data and Analyses"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 188,
+ 120,
+ 408,
+ 133
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 188,
+ 120,
+ 408,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 188,
+ 120,
+ 408,
+ 133
+ ],
+ "type": "text",
+ "content": "Verna Dankers and Christopher G. Lucas"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 174,
+ 134,
+ 422,
+ 148
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 174,
+ 134,
+ 422,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 174,
+ 134,
+ 422,
+ 148
+ ],
+ "type": "text",
+ "content": "Institute for Language, Cognition and Computation"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 239,
+ 149,
+ 357,
+ 162
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 239,
+ 149,
+ 357,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 239,
+ 149,
+ 357,
+ 162
+ ],
+ "type": "text",
+ "content": "University of Edinburgh"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 180,
+ 163,
+ 416,
+ 176
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 180,
+ 163,
+ 416,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 180,
+ 163,
+ 416,
+ 176
+ ],
+ "type": "text",
+ "content": "vernadankers@gmail.com, c.lucas@ed.ac.uk"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 155,
+ 226,
+ 203,
+ 239
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 155,
+ 226,
+ 203,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 155,
+ 226,
+ 203,
+ 239
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 84,
+ 245,
+ 274,
+ 448
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 84,
+ 245,
+ 274,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 84,
+ 245,
+ 274,
+ 448
+ ],
+ "type": "text",
+ "content": "When natural language phrases are combined, their meaning is often more than the sum of their parts. In the context of NLP tasks such as sentiment analysis, where the meaning of a phrase is its sentiment, that still applies. Many NLP studies on sentiment analysis, however, focus on the fact that sentiment computations are largely compositional. We, instead, set out to obtain non-compositionality ratings for phrases with respect to their sentiment. Our contributions are as follows: a) a methodology for obtaining those non-compositionality ratings, b) a resource of ratings for 259 phrases - NON-COMPSST - along with an analysis of that resource, and c) an evaluation of computational models for sentiment analysis using this new resource."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 68,
+ 454,
+ 155,
+ 468
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 454,
+ 155,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 454,
+ 155,
+ 468
+ ],
+ "type": "text",
+ "content": "1 Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 67,
+ 476,
+ 291,
+ 745
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 476,
+ 291,
+ 745
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 476,
+ 291,
+ 745
+ ],
+ "type": "text",
+ "content": "In NLP, the topics of the compositionality of language and neural models' capabilities to compute meaning compositionally have gained substantial interest in recent years. Yet, the meaning of linguistic utterances often does not adhere to strict patterns and can be surprising when looking at the individual words involved. This affects how those utterances behave in downstream tasks, such as sentiment analysis. Given a phrase or sentence, that task involves predicting the polarity as positive, negative or neutral. Sentiment largely adheres to compositional principles (Moilanen and Pulman, 2007, p.1): \"If the meaning of a sentence is a function of the meanings of its parts then the global polarity of a sentence is a function of the polarities of its parts.\" Modelling sentiment as a compositional process is, therefore, often mentioned as a design principle for computational sentiment models (e.g. by Socher et al., 2013; Sutherland et al., 2020; Yin et al., 2020)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 67,
+ 746,
+ 290,
+ 773
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 746,
+ 290,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 746,
+ 290,
+ 773
+ ],
+ "type": "text",
+ "content": "Nonetheless, one can think of examples where the sentiment of a phrase is unexpected given the"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 304,
+ 223,
+ 409,
+ 290
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 304,
+ 223,
+ 409,
+ 290
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 223,
+ 409,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 223,
+ 409,
+ 290
+ ],
+ "type": "image",
+ "image_path": "19471c6fd68a4682d1df5121ef5ade14ce461414766343238993cde2df0ce6eb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 318,
+ 294,
+ 395,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 294,
+ 395,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 294,
+ 395,
+ 312
+ ],
+ "type": "text",
+ "content": "all the excitement + of eating oatmeal"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 420,
+ 223,
+ 525,
+ 290
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 420,
+ 223,
+ 525,
+ 290
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 420,
+ 223,
+ 525,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 420,
+ 223,
+ 525,
+ 290
+ ],
+ "type": "image",
+ "image_path": "db52c7c7a769f6b3b7ecb819becdd5615af2d0ae65545dc97b73fe501fbee6e2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 427,
+ 294,
+ 518,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 427,
+ 294,
+ 518,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 427,
+ 294,
+ 518,
+ 312
+ ],
+ "type": "text",
+ "content": "a nearly terminal case + of the cutes"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 302,
+ 324,
+ 527,
+ 384
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 324,
+ 527,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 324,
+ 527,
+ 384
+ ],
+ "type": "text",
+ "content": "Figure 1: Illustration of how the observed sentiment deviates from the expected sentiment when viewing polarity as a function of the polarity of the subphrases. These examples were obtained from our newly annotated stimuli."
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 302,
+ 407,
+ 526,
+ 596
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 407,
+ 526,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 407,
+ 526,
+ 596
+ ],
+ "type": "text",
+ "content": "sentiment of the individual parts (e.g. see Zhu et al., 2015; Hwang and Hidey, 2019; Barnes et al., 2019; Tahayna et al., 2022, for work on non-compositional sentiment). These include the case of sarcasm (\"life is good, you should get one\"), opposing sentiments (\"terribrly fascinating\"), idiomatic expressions (\"break a leg\") and neutral terms that, when composed, suddenly convey sentiment (\"yeah right\"). Adequately capturing sentiment computationally requires both learning compositional rules and understanding when such exceptions exist, where most contemporary sentiment models are expected to learn that via mere end-to-end training on examples."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 302,
+ 597,
+ 527,
+ 773
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 597,
+ 527,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 597,
+ 527,
+ 773
+ ],
+ "type": "text",
+ "content": "How can we identify whether the sentiment of a phrase is non-compositional? We design a protocol to obtain such non-compositionality judgments based on human-annotated sentiment. Our methodology (elaborated on in §3) utilises phrases from the Stanford Sentiment Treebank (SST) (Socher et al., 2013) and contrasts the sentiment of a phrase with control stimuli, in which one of the two subphrases has been replaced. Phrases whose annotated sentiment deviates from what is expected based on the controls are considered less compositional, as is illustrated in Figure 1. We analyse the resulting non-compositionality ranking of phrases"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 310,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "text",
+ "content": "5150"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 806
+ ],
+ "type": "footer",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 806
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 806
+ ],
+ "type": "text",
+ "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5150-5162"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 165,
+ 806,
+ 428,
+ 818
+ ],
+ "type": "footer",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 165,
+ 806,
+ 428,
+ 818
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 165,
+ 806,
+ 428,
+ 818
+ ],
+ "type": "text",
+ "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 292,
+ 166
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 292,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 292,
+ 166
+ ],
+ "type": "inline_equation",
+ "content": "(\\S 4)"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 292,
+ 166
+ ],
+ "type": "text",
+ "content": " and show how the constructed resource can be used to evaluate sentiment models "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 292,
+ 166
+ ],
+ "type": "inline_equation",
+ "content": "(\\S 5)"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 292,
+ 166
+ ],
+ "type": "text",
+ "content": ". Our new resource (NONCOMPSST) can further improve the understanding of what underlies non-compositionality in sentiment analysis, and can complement existing evaluation protocols for sentiment analysis models."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 175,
+ 158,
+ 188
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 175,
+ 158,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 175,
+ 158,
+ 188
+ ],
+ "type": "text",
+ "content": "2 Related work"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 67,
+ 196,
+ 291,
+ 399
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 196,
+ 291,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 196,
+ 291,
+ 399
+ ],
+ "type": "text",
+ "content": "Over the course of years, sentiment analysis systems went from using rule-based models and sentiment lexicons (e.g. Moilanen and Pulman, 2007; Taboada et al., 2011) to using recursive neural networks (Socher et al., 2013; Tai et al., 2015; Zhu et al., 2015), to abandoning the use of structure altogether by finetuning pretrained large language models (e.g. Pérez et al., 2021; Camacho-collados et al., 2022; Hartmann et al., 2023), and recently, to abandoning training, via zero-shot generalisation (Wang et al., 2023). Crucial to the development of these systems has been the introduction of benchmarks, such as SST (Socher et al., 2013) and SemEval's Twitter benchmarks (e.g. Rosenthal et al., 2015; Nakov et al., 2016; Rosenthal et al., 2017)."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 400,
+ 291,
+ 684
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 400,
+ 291,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 400,
+ 291,
+ 684
+ ],
+ "type": "text",
+ "content": "Although the vast majority of related work focused on simply improving on benchmarks, there have been studies more closely related to ours, asking questions such as: How do phrases with opposing sentiment affect each other (Kiritchenko and Mohammad, 2016a,b)? What is the role of negations (Zhu et al., 2014), modals and adverbs (Kiritchenko and Mohammad, 2016c)? Do idioms have non-compositional sentiment (Hwang and Hidey, 2019)? Which linguistic phenomena are still problematic for SOTA sentiment systems (Barnes et al., 2019)? Can we incorporate compositional and non-compositional processing in one system (Zhu et al., 2015)? And can we computationally rank sentences according to their sentiment compositionality (Dankers and Titov, 2022)? Gaining a better understanding of the contexts in which sentiment functions non-compositionally and is challenging to predict is crucial for the evaluation of sentiment models in an age where sentiment benchmarks may appear saturated (Barnes et al., 2019).| Fig. | Phrase A | Phrase B | Rating | Sent. H | Sent. R-L |
| ✓ | a nearly terminal case | of the cutes | 4.11 | 5.33 | 2.00 |
| the franchise's best years | are long past | -4.11 | 0.33 | 1.00 |
| ✓ | all the excitement | of eating oatmeal | -3.00 | 2.00 | 2.33 |
| just a simple fable | done in an artless style | -2.56 | 1.00 | 1.33 |
| ✓ | a pressure cooker | of horrified awe | -2.56 | 1.00 | 3.67 |
| something creepy and vague | is in the works | 2.33 | 3.67 | 2.00 |
| a no-surprise series | of explosions and violence | -2.22 | 0.67 | 2.00 |
| the numerous scenes | of gory mayhem | -1.89 | 1.67 | 3.00 |
| the momentary joys | of pretty and weightless intellectual entertainment | -1.89 | 2.00 | 4.00 |
| flourishes - artsy fantasy sequences - | that simply feel wrong | -1.78 | 1.00 | 1.67 |
| the overall feel of the film | is pretty cheesy | -1.67 | 0.67 | 1.00 |
| the contrived nature | of its provocative conclusion | -1.67 | 1.33 | 2.00 |
| ✓ | fans of the animated wildlife adventure show | will be in warthog heaven | 1.56 | 5.67 | 5.00 |
| a pointed little chiller | about the frightening seductiveness of new technology | -1.56 | 2.67 | 4.00 |
| ✓ | a real human soul | buried beneath a spellbinding serpent's smirk | 1.56 | 4.67 | 5.00 |
| the comic heights | it obviously desired | -1.56 | 2.67 | 3.67 |
| the belly laughs | of lowbrow comedy | 1.56 | 4.67 | 4.00 |
| the intellectual and emotional pedigree | of your date | -1.56 | 2.67 | 3.33 |
| the experience of going to a film festival | is a rewarding one | 1.56 | 6.00 | 5.00 |
| a serious exploration | of nuclear terrorism | 1.56 | 4.33 | 4.00 |
| a simple tale | of an unlikely friendship | 1.56 | 5.33 | 4.00 |
| the real charm | of this trifle | -1.44 | 3.67 | 4.00 |
| the human story | is pushed to one side | -1.44 | 2.00 | 2.00 |
| a pale imitation | of the real deal | -1.44 | 0.67 | 1.00 |
| a main character | who sometimes defies sympathy | -1.44 | 1.67 | 2.33 |
| ✓ | the obligatory moments | of sentimental ooze | -1.44 | 2.00 | 3.00 |
| the actresses in the lead roles | are all more than competent | 1.44 | 5.33 | 5.00 |
| the debate it joins | is a necessary and timely one | 1.44 | 4.67 | 5.00 |
| an audacious tour | of the past | -1.33 | 2.67 | 3.67 |
| only a document | of the worst possibilities of mankind | 1.33 | 2.00 | 1.00 |
| the life experiences | of a particular theatrical family | 1.33 | 4.67 | 3.00 |
| the entire point | of a shaggy dog story | 1.33 | 4.00 | 3.00 |
| an inexpressible and drab wannabe | looking for that exact niche | 1.33 | 2.00 | 1.00 |
| everyone involved with moviemaking | is a con artist and a liar | -1.33 | 0.00 | 1.00 |
| a very original artist | in his medium | 1.33 | 5.33 | 5.00 |
| the visuals and eccentricities | of many of the characters | 1.33 | 4.33 | 3.00 |
| ✓ | a violent initiation rite | for the audience | -1.22 | 2.00 | 2.67 |
| these curious owners | of architectural oddities | -1.22 | 2.67 | 3.00 |
| this engaging mix | of love and bloodletting | 1.22 | 5.00 | 4.67 |
| a sudden lunch rush | at the diner | -1.22 | 2.67 | 3.00 |
| the true potential | of the medium | -1.22 | 3.00 | 4.00 |
| ✓ | a therapeutic zap | of shock treatment | -1.22 | 3.00 | 3.33 |
| the exotic world | of belly dancing | -1.22 | 3.00 | 3.67 |
| ✓ | the real story | starts just around the corner | 1.22 | 3.67 | 3.00 |
| the most offensive thing | about the movie | -1.22 | 1.00 | 1.00 |
| the issue of faith | is not explored very deeply | -1.22 | 1.00 | 2.00 |
| a big box | of consolation candy | 1.22 | 3.67 | 3.00 |
| the playful paranoia | of the film's past | 1.22 | 4.00 | 3.67 |
| the overlooked pitfalls | of such an endeavour | -1.22 | 2.00 | 3.00 |
| ✓ | the modern master of the chase sequence | returns with a chase to end all chases | 1.11 | 6.00 | 5.00 |
| the sheer beauty | of his images | 1.11 | 5.67 | 5.00 |
| a compelling slice | of awkward emotions | 1.11 | 4.67 | 5.00 |
| a great actress | tearing into a landmark role | -1.11 | 4.67 | 5.33 |
| much of the writing | is genuinely witty | 1.11 | 5.00 | 5.00 |
| an especially poignant portrait | of her friendship | -1.11 | 3.67 | 5.00 |
| the insight and honesty | of this disarming indie | -1.11 | 3.67 | 5.00 |
| the emotional arc | of its raw blues soundtrack | 1.11 | 4.33 | 3.67 |
| the kind of art shots | that fill gallery shows | 1.11 | 4.67 | 4.00 |
| a new software program | spit out the screenplay | 1.11 | 2.67 | 3.00 |
| a very good time | at the cinema | 1.11 | 5.67 | 5.67 |
",
+ "image_path": "76413e1f1e61644b7aea12a6ac230b89ac978bf418a20b1b9929c514bf2c864f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 698,
+ 526,
+ 735
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 698,
+ 526,
+ 735
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 698,
+ 526,
+ 735
+ ],
+ "type": "text",
+ "content": "Table 3: 60 phrases with a rating that exceeds an absolute value of 1. We indicate the rating from the MAX variant of our resource. We indicate whether we believe the phrase to represent a figurative phrase, and also provide the sentiment assigned by the annotators (Sent. H.) and by ROBERTA-LARGE (Sent. R-L)"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "5161"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 11
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 68,
+ 71,
+ 232,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 71,
+ 232,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 71,
+ 232,
+ 83
+ ],
+ "type": "text",
+ "content": "D Models & model evaluation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 92,
+ 525,
+ 105
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 92,
+ 525,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 92,
+ 525,
+ 105
+ ],
+ "type": "text",
+ "content": "Experimental setup We fine-tune the following models, all obtained from the HuggingFace model hub:"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 80,
+ 115,
+ 373,
+ 154
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 5,
+ "blocks": [
+ {
+ "bbox": [
+ 80,
+ 115,
+ 361,
+ 127
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 80,
+ 115,
+ 361,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 80,
+ 115,
+ 361,
+ 127
+ ],
+ "type": "text",
+ "content": "- ROBERTA-BASE https://huggingface.co/roberta-base"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 80,
+ 128,
+ 373,
+ 142
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 80,
+ 128,
+ 373,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 80,
+ 128,
+ 373,
+ 142
+ ],
+ "type": "text",
+ "content": "- ROBERTA-LARGE https://huggingface.co/roberta-large"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 81,
+ 142,
+ 238,
+ 154
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 81,
+ 142,
+ 238,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 81,
+ 142,
+ 238,
+ 154
+ ],
+ "type": "text",
+ "content": "- TIMELM (Loureiro et al., 2022):"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 89,
+ 156,
+ 388,
+ 168
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 89,
+ 156,
+ 388,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 89,
+ 156,
+ 388,
+ 168
+ ],
+ "type": "text",
+ "content": "https://huggingface.co/cardiffnlp/twitter-roberta-base"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 80,
+ 169,
+ 483,
+ 195
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 9,
+ "blocks": [
+ {
+ "bbox": [
+ 80,
+ 169,
+ 483,
+ 181
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 80,
+ 169,
+ 483,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 80,
+ 169,
+ 483,
+ 181
+ ],
+ "type": "text",
+ "content": "- BERTTweet (Nguyen et al., 2020): https://huggingface.co/vinai/bertweet-base"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 81,
+ 183,
+ 233,
+ 195
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 81,
+ 183,
+ 233,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 81,
+ 183,
+ 233,
+ 195
+ ],
+ "type": "text",
+ "content": "- Camacho-collados et al. (2022):"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 81,
+ 196,
+ 481,
+ 222
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 12,
+ "blocks": [
+ {
+ "bbox": [
+ 89,
+ 196,
+ 481,
+ 208
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 89,
+ 196,
+ 481,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 89,
+ 196,
+ 481,
+ 208
+ ],
+ "type": "text",
+ "content": "https://huggingface.co/cardiffnlp/twitter-rosberta-base-sentimentlatest"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 81,
+ 210,
+ 176,
+ 222
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 81,
+ 210,
+ 176,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 81,
+ 210,
+ 176,
+ 222
+ ],
+ "type": "text",
+ "content": "Pérez et al. (2021):"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 89,
+ 223,
+ 476,
+ 235
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 89,
+ 223,
+ 476,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 89,
+ 223,
+ 476,
+ 235
+ ],
+ "type": "text",
+ "content": "https://huggingface.co/finiteautomata/bertweet-base-sentiment-analysis"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 80,
+ 237,
+ 504,
+ 276
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 17,
+ "blocks": [
+ {
+ "bbox": [
+ 81,
+ 237,
+ 195,
+ 249
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 81,
+ 237,
+ 195,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 81,
+ 237,
+ 195,
+ 249
+ ],
+ "type": "text",
+ "content": "Hartmann et al. (2021):"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 89,
+ 250,
+ 504,
+ 262
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 89,
+ 250,
+ 504,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 89,
+ 250,
+ 504,
+ 262
+ ],
+ "type": "text",
+ "content": "https://huggingface.co/j-hartmann/sentiment-roberta-large-english-3-classes"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 80,
+ 264,
+ 486,
+ 276
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 80,
+ 264,
+ 486,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 80,
+ 264,
+ 486,
+ 276
+ ],
+ "type": "text",
+ "content": "- IMDB ROBERTA-BASE: https://huggingface.co/textattack/roberta-base-imdb"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 67,
+ 286,
+ 524,
+ 352
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 286,
+ 524,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 286,
+ 524,
+ 352
+ ],
+ "type": "text",
+ "content": "The experimental setup used to fine-tune them on SST is as follows: we use a batch size of 32, learning rate "
+ },
+ {
+ "bbox": [
+ 67,
+ 286,
+ 524,
+ 352
+ ],
+ "type": "inline_equation",
+ "content": "5e - 6"
+ },
+ {
+ "bbox": [
+ 67,
+ 286,
+ 524,
+ 352
+ ],
+ "type": "text",
+ "content": ", the Adam optimiser with a cosine-based warmup scheduler (warmup is "
+ },
+ {
+ "bbox": [
+ 67,
+ 286,
+ 524,
+ 352
+ ],
+ "type": "inline_equation",
+ "content": "20\\%"
+ },
+ {
+ "bbox": [
+ 67,
+ 286,
+ 524,
+ 352
+ ],
+ "type": "text",
+ "content": "). We train for 5 epochs, selecting the best model based on the validation data and evaluate that on the test sets. We do not experiment with these hyperparameters extensively, as they are within the range of recommended settings and all models use the same base models (ROBERTA-BASE or ROBERTA-LARGE)."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 67,
+ 354,
+ 525,
+ 380
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 354,
+ 525,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 354,
+ 525,
+ 380
+ ],
+ "type": "text",
+ "content": "We train using NVIDIA A100-SXM-80GB GPUs on which training one model for one seed takes up to 30 minutes for the BASE model, and 50 minutes for the LARGE model."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 67,
+ 388,
+ 525,
+ 523
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 388,
+ 525,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 388,
+ 525,
+ 523
+ ],
+ "type": "text",
+ "content": "**Characterising our experiments** Through our experiments, we contribute to generalisation evaluation in NLP: evaluating models' generalisation capabilities beyond standard i.i.d. testing, in particular through the evaluation on the 67 most non-compositional phrases. We characterise our experiments using the taxonomy of Hupkes et al. (2023): our experiments are cognitively motivated and focus on (non-)compositional generalisation. By separating inputs based on compositionality, we create a covariate shift between train and test data. That shift has a partitioned natural source, since the stimuli are from a human-written corpus, but the split we create is curated. In our experiments, we examine the influence of fine-tuning on SST-7 and evaluating on non-compositional examples (fine-tune train-test locus) but also examine the effect of a data shift in the pretrain-train locus, by considering models with different pretraining corpora."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 139,
+ 542,
+ 457,
+ 673
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 139,
+ 542,
+ 457,
+ 673
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 542,
+ 457,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 542,
+ 457,
+ 673
+ ],
+ "type": "image",
+ "image_path": "4e252e5c760fd260c87bb128ddf8bd69f9524d15e7068c1361ee328e69f8d333.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "5162"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 12
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_content_list.json b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..1c0ef3a22df1618df53cf95c9b3e5ceebe19935c
--- /dev/null
+++ b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_content_list.json
@@ -0,0 +1,2055 @@
+[
+ {
+ "type": "text",
+ "text": "Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning",
+ "text_level": 1,
+ "bbox": [
+ 115,
+ 87,
+ 880,
+ 127
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Jianing Zhou $^{1}$ , Ziheng Zeng $^{1}$ , Hongyu Gong $^{2}$ , Suma Bhat $^{1}$",
+ "bbox": [
+ 248,
+ 140,
+ 749,
+ 158
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "| Easy | Input Sentence | you would n't get anything from anywhere else because they 've cut it [MASK] . |
| Target Sentence | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Vanilla | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| Competence + SL | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Competence + WR | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Norm-based | you would n't get anything from anywhere else because they 've cut it to the bone . |
| SGCL + fixed | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| SGCL + dynamic | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Ours | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Medium | Input Sentence | these were some of the qualities needed to [MASK] as a sportsmen even at the more modest levels . |
| Target Sentence | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Vanilla | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Competence + SL | these were some of the qualities needed to stand up and be counted as a sportsmen even at the more modest levels . |
| Competence + WR | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| Norm-based | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| SGCL + fixed | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| SGCL + dynamic | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Ours | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Hard | Input Sentence | Are the Americans going [MASK] again , or is this an indictment which we should place on trial ? |
| Target Sentence | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
| Vanilla | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + SL | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + WR | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Norm-based | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| SGCL + fixed | Are the Americans going through the motions again , or is this an indictment which we should place on trial ? |
| SGCL + dynamic | Are the Americans going behind our backs again , or is this an indictment which we should place on trial ? |
| Ours | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
",
+ "bbox": [
+ 117,
+ 438,
+ 878,
+ 736
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Table 9: A sample of the generated sentences on MAGPIE highlighting the correct idioms, and the wrong idioms. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels.",
+ "bbox": [
+ 112,
+ 745,
+ 884,
+ 816
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "page_number",
+ "text": "4334",
+ "bbox": [
+ 480,
+ 928,
+ 519,
+ 940
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "table",
+ "img_path": "images/a7d181822d0252eb25411974c0d0b5e9a1e9ae70d94bafe9c624194ca36118ca.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Easy | Input Sentence | Whose wrath <V> loomed <V> heavy o'er the trojan towers : |
| Target Sentence | Whose wrath hung heavy o'er the trojan towers : |
| Vanilla | Whose wrath hangs heavy o'er the trojan towers : |
| Competence + SL | Whose wrath rises heavy o'er the trojan towers : |
| Competence + WR | Whose wrath hangs heavy o'er the trojan towers : |
| Norm-based | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + fixed | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + dynamic | Whose wrath hangs heavy o'er the trojan towers : |
| Ours | Whose wrath hung heavy o'er the trojan towers : |
| Medium | Input Sentence | To that which never can us comfort <V> say <V> . |
| Target Sentence | To that which never can us comfort bring . |
| Vanilla | To that which never can us comfort give . |
| Competence + SL | To that which never can us comfort bring . |
| Competence + WR | To that which never can us comfort find . |
| Norm-based | To that which never can us comfort find . |
| SGCL + fixed | To that which never can us comfort make . |
| SGCL + dynamic | To that which never can us comfort give . |
| Ours | To that which never can us comfort bring . |
| Hard | Input Sentence | He hardly knew when waves he <V> used <V> on |
| Target Sentence | He hardly knew when waves he tossed on |
| Vanilla | He hardly knew when waves he plowed on |
| Competence + SL | He hardly knew when waves he went on |
| Competence + WR | He hardly knew when waves he put on |
| Norm-based | He hardly knew when waves he put on |
| SGCL + fixed | He hardly knew when waves he rolled on |
| SGCL + dynamic | He hardly knew when waves he put on |
| Ours | He hardly knew when waves he tossed on |
",
+ "bbox": [
+ 233,
+ 296,
+ 764,
+ 618
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Table 10: A sample of the generated sentences on MERMAID highlighting the correct metaphor, and the wrong metaphor. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels.",
+ "bbox": [
+ 112,
+ 627,
+ 884,
+ 700
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "page_number",
+ "text": "4335",
+ "bbox": [
+ 480,
+ 928,
+ 519,
+ 940
+ ],
+ "page_idx": 15
+ }
+]
\ No newline at end of file
diff --git a/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_model.json b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..9d8dd94f5a732fd666224f939975b06266c32c6e
--- /dev/null
+++ b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_model.json
@@ -0,0 +1,2806 @@
+[
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.117,
+ 0.089,
+ 0.882,
+ 0.128
+ ],
+ "angle": 0,
+ "content": "Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.249,
+ 0.141,
+ 0.751,
+ 0.159
+ ],
+ "angle": 0,
+ "content": "Jianing Zhou\\(^{1}\\), Ziheng Zeng\\(^{1}\\), Hongyu Gong\\(^{2}\\), Suma Bhat\\(^{1}\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.332,
+ 0.16,
+ 0.671,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "| Easy | Input Sentence | you would n't get anything from anywhere else because they 've cut it [MASK] . |
| Target Sentence | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Vanilla | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| Competence + SL | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Competence + WR | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Norm-based | you would n't get anything from anywhere else because they 've cut it to the bone . |
| SGCL + fixed | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| SGCL + dynamic | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Ours | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Medium | Input Sentence | these were some of the qualities needed to [MASK] as a sportsmen even at the more modest levels . |
| Target Sentence | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Vanilla | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Competence + SL | these were some of the qualities needed to stand up and be counted as a sportsmen even at the more modest levels . |
| Competence + WR | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| Norm-based | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| SGCL + fixed | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| SGCL + dynamic | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Ours | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Hard | Input Sentence | Are the Americans going [MASK] again , or is this an indictment which we should place on trial ? |
| Target Sentence | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
| Vanilla | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + SL | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + WR | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Norm-based | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| SGCL + fixed | Are the Americans going through the motions again , or is this an indictment which we should place on trial ? |
| SGCL + dynamic | Are the Americans going behind our backs again , or is this an indictment which we should place on trial ? |
| Ours | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.113,
+ 0.746,
+ 0.885,
+ 0.818
+ ],
+ "angle": 0,
+ "content": "Table 9: A sample of the generated sentences on MAGPIE highlighting the correct idioms, and the wrong idioms. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.929,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "4334"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.234,
+ 0.297,
+ 0.766,
+ 0.619
+ ],
+ "angle": 0,
+ "content": "| Easy | Input Sentence | Whose wrath <V> loomed <V> heavy o'er the trojan towers : |
| Target Sentence | Whose wrath hung heavy o'er the trojan towers : |
| Vanilla | Whose wrath hangs heavy o'er the trojan towers : |
| Competence + SL | Whose wrath rises heavy o'er the trojan towers : |
| Competence + WR | Whose wrath hangs heavy o'er the trojan towers : |
| Norm-based | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + fixed | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + dynamic | Whose wrath hangs heavy o'er the trojan towers : |
| Ours | Whose wrath hung heavy o'er the trojan towers : |
| Medium | Input Sentence | To that which never can us comfort <V> say <V> . |
| Target Sentence | To that which never can us comfort bring . |
| Vanilla | To that which never can us comfort give . |
| Competence + SL | To that which never can us comfort bring . |
| Competence + WR | To that which never can us comfort find . |
| Norm-based | To that which never can us comfort find . |
| SGCL + fixed | To that which never can us comfort make . |
| SGCL + dynamic | To that which never can us comfort give . |
| Ours | To that which never can us comfort bring . |
| Hard | Input Sentence | He hardly knew when waves he <V> used <V> on |
| Target Sentence | He hardly knew when waves he tossed on |
| Vanilla | He hardly knew when waves he plowed on |
| Competence + SL | He hardly knew when waves he went on |
| Competence + WR | He hardly knew when waves he put on |
| Norm-based | He hardly knew when waves he put on |
| SGCL + fixed | He hardly knew when waves he rolled on |
| SGCL + dynamic | He hardly knew when waves he put on |
| Ours | He hardly knew when waves he tossed on |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.113,
+ 0.628,
+ 0.885,
+ 0.701
+ ],
+ "angle": 0,
+ "content": "Table 10: A sample of the generated sentences on MERMAID highlighting the correct metaphor, and the wrong metaphor. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.929,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "4335"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_origin.pdf b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..433911d8c1de1fb55be679447ab7c9abcd34eaa5
--- /dev/null
+++ b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ef4dc0fc9ace187439f179a77039f7946c1f9a02e3b47f48a42df41ca6a8b17
+size 558829
diff --git a/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/full.md b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..7cc97f16ddb577ef1dbc0c0ed2b3e97667bb0abd
--- /dev/null
+++ b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/full.md
@@ -0,0 +1,427 @@
+# Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning
+
+Jianing Zhou $^{1}$ , Ziheng Zeng $^{1}$ , Hongyu Gong $^{2}$ , Suma Bhat $^{1}$
+
+| Easy | Input Sentence | you would n't get anything from anywhere else because they 've cut it [MASK] . |
| Target Sentence | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Vanilla | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| Competence + SL | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Competence + WR | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Norm-based | you would n't get anything from anywhere else because they 've cut it to the bone . |
| SGCL + fixed | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| SGCL + dynamic | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Ours | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Medium | Input Sentence | these were some of the qualities needed to [MASK] as a sportsmen even at the more modest levels . |
| Target Sentence | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Vanilla | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Competence + SL | these were some of the qualities needed to stand up and be counted as a sportsmen even at the more modest levels . |
| Competence + WR | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| Norm-based | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| SGCL + fixed | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| SGCL + dynamic | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Ours | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Hard | Input Sentence | Are the Americans going [MASK] again , or is this an indictment which we should place on trial ? |
| Target Sentence | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
| Vanilla | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + SL | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + WR | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Norm-based | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| SGCL + fixed | Are the Americans going through the motions again , or is this an indictment which we should place on trial ? |
| SGCL + dynamic | Are the Americans going behind our backs again , or is this an indictment which we should place on trial ? |
| Ours | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
+
+Table 9: A sample of the generated sentences on MAGPIE highlighting the correct idioms, and the wrong idioms. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels.
+
+| Easy | Input Sentence | Whose wrath <V> loomed <V> heavy o'er the trojan towers : |
| Target Sentence | Whose wrath hung heavy o'er the trojan towers : |
| Vanilla | Whose wrath hangs heavy o'er the trojan towers : |
| Competence + SL | Whose wrath rises heavy o'er the trojan towers : |
| Competence + WR | Whose wrath hangs heavy o'er the trojan towers : |
| Norm-based | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + fixed | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + dynamic | Whose wrath hangs heavy o'er the trojan towers : |
| Ours | Whose wrath hung heavy o'er the trojan towers : |
| Medium | Input Sentence | To that which never can us comfort <V> say <V> . |
| Target Sentence | To that which never can us comfort bring . |
| Vanilla | To that which never can us comfort give . |
| Competence + SL | To that which never can us comfort bring . |
| Competence + WR | To that which never can us comfort find . |
| Norm-based | To that which never can us comfort find . |
| SGCL + fixed | To that which never can us comfort make . |
| SGCL + dynamic | To that which never can us comfort give . |
| Ours | To that which never can us comfort bring . |
| Hard | Input Sentence | He hardly knew when waves he <V> used <V> on |
| Target Sentence | He hardly knew when waves he tossed on |
| Vanilla | He hardly knew when waves he plowed on |
| Competence + SL | He hardly knew when waves he went on |
| Competence + WR | He hardly knew when waves he put on |
| Norm-based | He hardly knew when waves he put on |
| SGCL + fixed | He hardly knew when waves he rolled on |
| SGCL + dynamic | He hardly knew when waves he put on |
| Ours | He hardly knew when waves he tossed on |
+
+Table 10: A sample of the generated sentences on MERMAID highlighting the correct metaphor, and the wrong metaphor. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels.
\ No newline at end of file
diff --git a/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/images.zip b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/images.zip
new file mode 100644
index 0000000000000000000000000000000000000000..9ba3f8ca24760c046c4ca91607f550ccc5a7a81c
--- /dev/null
+++ b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/images.zip
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c05e42bee8f7daef24efe627dfbbd4b4d77be5884956a89a8fc13dc6020538c1
+size 933668
diff --git a/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/layout.json b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..bc21bed4c4e2c0be519be8ccc4c4423af198b128
--- /dev/null
+++ b/2023/Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning/layout.json
@@ -0,0 +1,10054 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 69,
+ 74,
+ 524,
+ 107
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 74,
+ 524,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 74,
+ 524,
+ 107
+ ],
+ "type": "text",
+ "content": "Non-compositional Expression Generation Based on Curriculum Learning and Continual Learning"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "text",
+ "content": "Jianing Zhou"
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "text",
+ "content": ", Ziheng Zeng"
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "text",
+ "content": ", Hongyu Gong"
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "text",
+ "content": ", Suma Bhat"
+ },
+ {
+ "bbox": [
+ 148,
+ 118,
+ 446,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 197,
+ 134,
+ 399,
+ 148
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 197,
+ 134,
+ 399,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 197,
+ 134,
+ 399,
+ 148
+ ],
+ "type": "text",
+ "content": "| Easy | Input Sentence | you would n't get anything from anywhere else because they 've cut it [MASK] . |
| Target Sentence | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Vanilla | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| Competence + SL | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Competence + WR | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Norm-based | you would n't get anything from anywhere else because they 've cut it to the bone . |
| SGCL + fixed | you wouldn't get anything from anywhere else because they don't cut it at the end of the day . |
| SGCL + dynamic | you wouldn't get anything from anywhere else because they've cut it by the back door . |
| Ours | you would n't get anything from anywhere else because they 've cut it to the bone . |
| Medium | Input Sentence | these were some of the qualities needed to [MASK] as a sportsmen even at the more modest levels . |
| Target Sentence | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Vanilla | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Competence + SL | these were some of the qualities needed to stand up and be counted as a sportsmen even at the more modest levels . |
| Competence + WR | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| Norm-based | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| SGCL + fixed | these were some of the qualities needed to make the killing as a sportsmen even at the more modest levels . |
| SGCL + dynamic | these were some of the qualities needed to rise to the occasion as a sportsmen even at the more modest levels . |
| Ours | these were some of the qualities needed to make the grade as a sportsmen even at the more modest levels . |
| Hard | Input Sentence | Are the Americans going [MASK] again , or is this an indictment which we should place on trial ? |
| Target Sentence | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
| Vanilla | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + SL | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Competence + WR | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| Norm-based | Are the Americans going behind the scenes again , or is this an indictment which we should place on trial ? |
| SGCL + fixed | Are the Americans going through the motions again , or is this an indictment which we should place on trial ? |
| SGCL + dynamic | Are the Americans going behind our backs again , or is this an indictment which we should place on trial ? |
| Ours | Are the Americans going over the top again , or is this an indictment which we should place on trial ? |
",
+ "image_path": "40bd8c25d0375cba986cb60b1fc3d055f79812a8be8b5717c5312a42ce14fdb9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 627,
+ 526,
+ 687
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 627,
+ 526,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 627,
+ 526,
+ 687
+ ],
+ "type": "text",
+ "content": "Table 9: A sample of the generated sentences on MAGPIE highlighting the correct idioms, and the wrong idioms. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "4334"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 14
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 139,
+ 249,
+ 455,
+ 520
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 139,
+ 249,
+ 455,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 249,
+ 455,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 249,
+ 455,
+ 520
+ ],
+ "type": "table",
+ "html": "| Easy | Input Sentence | Whose wrath <V> loomed <V> heavy o'er the trojan towers : |
| Target Sentence | Whose wrath hung heavy o'er the trojan towers : |
| Vanilla | Whose wrath hangs heavy o'er the trojan towers : |
| Competence + SL | Whose wrath rises heavy o'er the trojan towers : |
| Competence + WR | Whose wrath hangs heavy o'er the trojan towers : |
| Norm-based | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + fixed | Whose wrath hangs heavy o'er the trojan towers : |
| SGCL + dynamic | Whose wrath hangs heavy o'er the trojan towers : |
| Ours | Whose wrath hung heavy o'er the trojan towers : |
| Medium | Input Sentence | To that which never can us comfort <V> say <V> . |
| Target Sentence | To that which never can us comfort bring . |
| Vanilla | To that which never can us comfort give . |
| Competence + SL | To that which never can us comfort bring . |
| Competence + WR | To that which never can us comfort find . |
| Norm-based | To that which never can us comfort find . |
| SGCL + fixed | To that which never can us comfort make . |
| SGCL + dynamic | To that which never can us comfort give . |
| Ours | To that which never can us comfort bring . |
| Hard | Input Sentence | He hardly knew when waves he <V> used <V> on |
| Target Sentence | He hardly knew when waves he tossed on |
| Vanilla | He hardly knew when waves he plowed on |
| Competence + SL | He hardly knew when waves he went on |
| Competence + WR | He hardly knew when waves he put on |
| Norm-based | He hardly knew when waves he put on |
| SGCL + fixed | He hardly knew when waves he rolled on |
| SGCL + dynamic | He hardly knew when waves he put on |
| Ours | He hardly knew when waves he tossed on |
",
+ "image_path": "a7d181822d0252eb25411974c0d0b5e9a1e9ae70d94bafe9c624194ca36118ca.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 528,
+ 526,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 528,
+ 526,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 528,
+ 526,
+ 589
+ ],
+ "type": "text",
+ "content": "Table 10: A sample of the generated sentences on MERMAID highlighting the correct metaphor, and the wrong metaphor. Easy represents the easy example randomly selected from the examples in the start after ranking based on difficulty levels. Medium represents the example randomly selected from the examples in the middle after ranking based on difficulty levels. Hard represents the example randomly selected from the examples in the final after ranking based on difficulty levels."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 781,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "4335"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 15
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_content_list.json b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..982414068b03055ab8eee2d986602ac28ba236f8
--- /dev/null
+++ b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_content_list.json
@@ -0,0 +1,1461 @@
+[
+ {
+ "type": "text",
+ "text": "Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modeling",
+ "text_level": 1,
+ "bbox": [
+ 159,
+ 78,
+ 836,
+ 118
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Linqin Wang $^{1,2}$ , Zhengtao Yu $^{1,2}$ , *Yuanzhang Yang $^{1,2}$ , Shengxiang Gao $^{1,2}$ , Cunli Mao $^{1,2}$ , Yuxin Huang $^{1,2}$",
+ "bbox": [
+ 270,
+ 124,
+ 729,
+ 158
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "$^{1}$ Faculty of Information Engineering and Automation,",
+ "bbox": [
+ 280,
+ 159,
+ 719,
+ 175
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Kunming University of Science and Technology, Kunming, China",
+ "bbox": [
+ 233,
+ 175,
+ 766,
+ 192
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "2 Yunnan Key Laboratory of Artificial Intelligence, Kunming, China {linqinwang7767}@163.com, {ztyu,gaoshengxiang.yn}@ hotmail.com, {maocunli, yyz845935161, huangyuxin2004}@163.com",
+ "bbox": [
+ 196,
+ 192,
+ 806,
+ 241
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 260,
+ 252,
+ 339,
+ 266
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Existing accent transfer works rely on parallel data or speech recognition models. This paper focuses on the practical application of accent transfer and aims to implement accent transfer using non-parallel datasets. The study has encountered the challenge of speech representation disentanglement and modeling accents. In our accent modeling transfer framework, we manage to solve these problems by two proposed methods. First, we learn the suprasegmental information associated with tone to finely model the accents in terms of tone and rhythm. Second, we propose to use mutual information learning to disentangle the accent features and control the accent of the generated speech during the inference time. Experiments show that the proposed framework attains superior performance to the baseline models in terms of accentedness and audio quality.",
+ "bbox": [
+ 141,
+ 275,
+ 460,
+ 545
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 Introduction",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 552,
+ 258,
+ 567
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "The accent transfer task refers to the synthesis of speech with an accent, such as British English for North Americans. Accent pronunciation is a distinctive form of expression influenced by the native language, the speaker's social group or speaking in a particular region (Loots and Niesler, 2011). In general, people find it easier to talk with others in their own accent group. The use of speech is now widely adopted, for example in the field of chatbots and film dubbing requires research on accent transfer of speech.",
+ "bbox": [
+ 112,
+ 577,
+ 489,
+ 753
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "At present, the accent transfer task for parallel data has achieved sound research results and increasing performance. Divided according to training data, the methods are specifically divided into: (1) Parallel corpus of different accents of the same speaker using source and target speech content and time alignment (Finkelstein et al., 2022; Liu et al., 2022; Hida et al., 2022; Toda et al., 2007; Oyamada et al., 2017). (2) Non-parallel corpus of",
+ "bbox": [
+ 112,
+ 755,
+ 489,
+ 900
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "multiple speakers with multiple accents using inconsistent source and target speech content (Wang et al., 2021; Zhao et al., 2018, 2019; Kaneko and Kameoka, 2017; Kaneko et al., 2019, 2020a, 2021; Finkelstein et al., 2022) used a multi-stage trained tts model to achieve transfer of North American accents, Australian accents, and British accents, and used a CHiVE-BERT pre-training model to enhance the audio effect of accent generation. Liu et al. (2022) added an accent variance adaptor to model the rhythmicity of accent variance, and also enhanced the accent generation audio by using a consistency constraint module. The use of phonetic posteriorgrams (PPG) is an essential idea in the application of non-parallel data (Wang et al., 2021; Zhao et al., 2018, 2019). Wang et al. (2021) extracted PPG from a Chinese pre-trained speech recognition model and then used them in an end-to-end speech conversion model based on adversarial learning disentangling. This approach achieved accent transfer from Chinese Mandarin to Tianjin and obtained decent results.",
+ "bbox": [
+ 507,
+ 253,
+ 884,
+ 605
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "However, existing accent transfer works highly rely on a large amount of labelled parallel data or advanced speech recognition models. Working with an enormous amount of labeling data is always hectic, labor-intensive, and time-consuming, which is more severe for low-resource languages. This limitation hinders the wider application of accent transfer in low-resource scenarios. Hence, it is a timely question: Is it feasible to do non-parallel accent transfer task under a unified framework? It is challenging because the speech representation containing various components, including speaker timbre, accent characteristics, and linguistic content, which are difficult to disentangle, especially for non-parallel accent transfer task.",
+ "bbox": [
+ 507,
+ 609,
+ 882,
+ 851
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "The dataset for the task in this paper can be represented as $\\{S_a(A_a), S_b(A_b)\\}$ , where the $a$ speaker $S_a$ can only speak the accent $A_a$ , and the $b$ speaker $S_b$ can only speak the accent $A_b$ . The objec",
+ "bbox": [
+ 507,
+ 854,
+ 882,
+ 917
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_footnote",
+ "text": "* Corresponding author",
+ "bbox": [
+ 141,
+ 904,
+ 294,
+ 917
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_number",
+ "text": "9288",
+ "bbox": [
+ 478,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "footer",
+ "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9288-9298 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+ "bbox": [
+ 216,
+ 945,
+ 779,
+ 972
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "tives of this paper are to achieve respectively: (1) $S_{a}(A_{a}) \\to S_{b}(A_{a})$ , where the $A_{a}$ accent transfer to the $S_{b}$ speaker without changing the linguistic content of the speech itself. (2) $S_{b}(A_{b}) \\to S_{a}(A_{b})$ , where the $A_{b}$ accent transfer to the $S_{a}$ speaker without changing the linguistic content of the speech itself. (3) $S_{a}(A_{a}) \\Longleftrightarrow S_{b}(A_{b})$ , two-way speaker timbre and accent transfer between the $S_{a}$ speaker and the $S_{b}$ speaker.",
+ "bbox": [
+ 112,
+ 84,
+ 489,
+ 228
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Following the success of mutual information learning to disentangle speaker information in the One-shot voice conversion (VC) task (Yang et al., 2022), this paper applies the non-parallel databased voice conversion model MaskCycleGANVC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. The most challenging task is to achieve effective disentangling of accent features, linguistic features, speaker timbre features and fine-grained embodied modeling of accents in a unified model architecture, and to achieve controlled and effective speaker timbre and accent transfer in the prediction phase of the model. The accurate modeling of phonetic pronunciation tones in the task of accent transfer is crucial. The contributions of this paper are as follows.",
+ "bbox": [
+ 115,
+ 231,
+ 489,
+ 537
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(1) For accents being difficult to model fine-grained concretely, this paper fine-grained concretely models accents in terms of phonetic intonation, rhythmic pauses and other pronunciation features. Then, an accent feature encoder is proposed, which can effectively extract accent features in the inference stage and realise accent controllability modeling.",
+ "(2) To address the problem of difficult speaker information disentangling in non-parallel data sets with different speakers with different accents, this paper proposes mutual information learning to maximize the mutual information upper bound of accent features and speech features. It can effectively disentangle the speaker features, accent features and phonetic features of speech.",
+ "(3) Experimental results show that method converts the speech up to a MOS score of 4.12. Achieving optimal results compared to baselines on the English accent transfer task for the public VCTK dataset and on the Lao accent transfer for the self-constructed Lao dataset. It significantly improves the accentedness and audio quality."
+ ],
+ "bbox": [
+ 112,
+ 542,
+ 490,
+ 917
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 Method",
+ "text_level": 1,
+ "bbox": [
+ 509,
+ 83,
+ 613,
+ 98
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In this section, we first describe our model architecture. Then we introduce the fine-grained accent modeling and adversarial mutual information learning and show how accent transfer between speaker $S_{a}$ and speaker $S_{b}$ on non-parallel datasets with different speakers and accents is achieved.",
+ "bbox": [
+ 507,
+ 110,
+ 884,
+ 206
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2.1 Architecture of the proposed model",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 219,
+ 833,
+ 234
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The generator structure of the model is shown in Figure. 1. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The generator part is composed of five parts: an accent encoder $E_{ac}$ , a speaker encoder $E_{s}$ , a speech content encoder $E_{c}$ , a speech generator $G$ , the feature disentanglements $C1$ and $C2$ . The model training strategy adopts a non-parallel voice conversion approach. Given a non-parallel corpus $D(x,y)$ , the training mechanism involves mapping source speech $x$ to converted speech $y$ and then back to $x'$ , with the primary training objective being the minimization of the mean square error between $x$ and $x'$ . The baseline models training details and our model parameters please refer to Appendix 5.",
+ "bbox": [
+ 507,
+ 240,
+ 884,
+ 498
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2.2 Encoder and accent modeling",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 510,
+ 791,
+ 525
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Encoder: The accent encoder $E_{ac}$ takes the mel-spectrogram $S$ and normalized pitch contour $P$ of the speech as inputs. The accent encoder provides global speech accent features to control the speaker's accent. We use a vector quantized variational autoencoder (VQ-VAE) (Van Den Oord et al., 2017; Polyak et al., 2021) model to learn suprasegmental representations related to tone. The speaker encoder is the same structure in (Chen et al., 2021), using Conv1D as the main structure to extract speaker features. The speech content encoder is the downsampling module in MaskCycleGAN (Kaneko et al., 2021), the rhythm encoder $E_r$ extract speech rhythm features, we have:",
+ "bbox": [
+ 507,
+ 531,
+ 884,
+ 756
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\begin{array}{l} Z _ {a c} = E _ {a c} \\left(E _ {r} (S), V Q - V A E (P)\\right), \\\\ Z _ {c} = E _ {c} (S), \\tag {1} \\\\ Z _ {t} = E _ {s} (S), \\\\ \\end{array}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 552,
+ 769,
+ 880,
+ 826
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "where $Z_{ac}$ represents the accent features, $Z_{c}$ represents the speech content features, $Z_{t}$ represents the speaker features.",
+ "bbox": [
+ 507,
+ 838,
+ 882,
+ 885
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Accent modeling: Tones alteration stands out as one of the primary distinguishing features of",
+ "bbox": [
+ 507,
+ 887,
+ 882,
+ 917
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "9289",
+ "bbox": [
+ 480,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/f2920a1775a0213513da8d4a3440e392413387912e4c4eb3378654955912ce50.jpg",
+ "image_caption": [
+ "Figure 1: Framework of proposed model. The $Z_{ac}$ is the output hidden states of accent encoder $E_{ac}$ , $Z_{t}$ is the output hidden states of speaker encoder $E_{s}$ , $Z_{c}$ is the output hidden states of content encoder $E_{c}$ , $Z_{r}$ is the output hidden states of rhythm encoder $E_{r}$ . $C1$ is speaker identity classify with linear, $C2$ is accent classify with gradient reverse linear."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 127,
+ 85,
+ 884,
+ 451
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "accents, and these different tones are reflected in the spectrum by different fundamental frequencies, so tone features can be represented by the discrete speech feature F0. To better model the tone at fine-grained degree, we adopt the VQ-VAE framework to train a pre-training model for speech tone feature F0. The yet another algorithm for pitch tracking (YAAPT) (Kasi and Zahorian, 2002) algorithm is used to extract the F0 from the input signal, $x$ , generating $P = (P_{1},\\dots,P_{T^{\\prime}})$ , we have:",
+ "bbox": [
+ 112,
+ 542,
+ 489,
+ 705
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nz _ {1: L ^ {\\prime}} ^ {F} = \\mathrm {E n c o d e r} _ {\\Phi_ {1} F} (P _ {1: T ^ {\\prime}}),\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 226,
+ 714,
+ 442,
+ 734
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\ne _ {1: L ^ {\\prime}} ^ {F} = \\operatorname {V e c t o r Q u a n t i z a t i o n} _ {\\Phi_ {2} F} \\left(z _ {1: L ^ {\\prime}}\\right), \\quad (2)\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 139,
+ 736,
+ 487,
+ 755
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {P} _ {1: L ^ {\\prime}} = \\mathrm {D e c o d e r} _ {\\Phi_ {3}} (e _ {1: L ^ {\\prime}} ^ {F}),\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 238,
+ 758,
+ 442,
+ 778
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Each element in $z_{1:L'}^F$ is an integer $z_s \\in \\{0,1,\\dots,K'\\}$ , where $K'$ is the encoder dictionary size. The VectorQuantization $\\Phi_{2F}$ , a bottleneck with a learned codebook $C = (e_1,\\dots,e_{K'})$ , where each item in $C$ is a 128-dimensional vector. The encoder extracts a sequence of latent vectors $\\mathrm{Encoder}_{\\Phi_1F}(P) = (h_1,\\dots,h_{L'})$ from the raw audio, where $h_i \\in \\mathbb{R}^{128}$ , for all $1 \\leq i \\leq L'$ .",
+ "bbox": [
+ 112,
+ 789,
+ 490,
+ 917
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Then, the bottleneck maps each latent vector to its nearest vector in the codebook $C$ . The embedded latent vectors are then being fed into the decoder $\\mathrm{Decoder}_{\\Phi_3}(e_{1:L'}^F) = \\hat{P}$ which reconstructs the original F0 signal. Similar to (Dhariwal et al., 2020), we use Exponential Moving Average updates to learn the codebook and employ random restarts for unused embeddings, we use the indices of the mapped latent vectors to generate $Z_p$ . The rhythm of the speech is extracted using the same structure as $E_r$ in SpeechSplit (Qian et al., 2020) to obtain rhythm features output $Z_r$ . At last, the $Z_p \\oplus Z_r$ as a representation of accent $Z_{ac}$ , it is incorporated into the speech generation model. We chose $E_p$ and $E_r$ as components of the accent encoder $E_{ac}$ because pitch and rhythm represent the most significant aspects of accent variation. Additionally, recent works (Qian et al., 2020; Dhariwal et al., 2020) have demonstrated their effectiveness in extracting pitch and rhythm features. We hope that the discrete representations learned from F0 capture pitch patterns and/or other suprasegmental information. The proposed extension is straightforward",
+ "bbox": [
+ 507,
+ 542,
+ 884,
+ 912
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "9290",
+ "bbox": [
+ 480,
+ 927,
+ 521,
+ 940
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "ward, but we observe that it results in impressive improvements for accent modeling, especially F0 of reconstructed speech waveforms in Lao.",
+ "bbox": [
+ 112,
+ 84,
+ 487,
+ 131
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.3 Generator and accent transfer",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 143,
+ 400,
+ 156
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Generator: The speech generator $G$ is based on the architecture of MaskCycleGAN-VC (Kaneko et al., 2021) and uses the Filling in Frames (FIF) strategy during training, where a random part of the spectrogram is masked. The mel-spectrogram is downsampled and mapped from high dimension to low dimension $Z_{c}$ . Then, the upsampled $Z_{t}$ (speaker features) and $Z_{ac}$ (accent features) are combined to map these features into the mel-spectrogram of the target speaker. The discriminator has the same structure as MaskCycleGAN-VC and takes as input the mel-spectrogram generated by the generator for the target speaker. We have:",
+ "bbox": [
+ 112,
+ 164,
+ 489,
+ 373
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {S} = G \\left(Z _ {a c}, Z _ {c}, Z _ {t}\\right), \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 220,
+ 387,
+ 487,
+ 406
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $\\hat{S}$ represents the converted speech.",
+ "bbox": [
+ 112,
+ 418,
+ 421,
+ 435
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Controllable accent transfer: During the model inference stage, different target accents can be achieved by controlling the accent feature $Z_{ac}$ .",
+ "bbox": [
+ 112,
+ 436,
+ 485,
+ 483
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\begin{array}{l} \\hat {S} _ {S _ {b} \\left(A _ {a}\\right)} = G \\left(Z _ {a c} \\left(A _ {a}\\right), Z _ {c} \\left(S _ {a}\\right), Z _ {t} \\left(S _ {b}\\right)\\right), \\tag {4} \\\\ \\hat {S} _ {S _ {a} (A _ {b})} = G \\left(Z _ {a c} (A _ {b}), Z _ {c} (S _ {b}), Z _ {t} (S _ {a})\\right), \\\\ \\end{array}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 137,
+ 508,
+ 487,
+ 550
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $\\hat{S}_{S_b(A_a)}$ represents the speaker $S_{b}$ with accent $A_{a}$ , which is not present in training data (only present the speaker $S_{b}$ with accent $A_{b}$ ).",
+ "bbox": [
+ 112,
+ 564,
+ 487,
+ 613
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.4 Disentanglement and loss",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 625,
+ 361,
+ 639
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Speaker information disentanglement: As shown in Fig. 1, we use common classifier $C_1$ and adversarial speaker classifier $C_2$ with gradient reverse linear (GRL) (Ganin et al., 2016) to recognize the identity of speaker, The Fig. 2 illustrates the model architecture of the two classifiers. Both classifiers, $C_1$ and $C_2$ , utilize speaker IDs as supervisory labels during the training process. The objective of $C_1$ is to accurately classify the speaker ID associated with $Z_{t}$ as the training progresses. On the contrary, as the training progresses, $C_2$ is designed to gradually struggle in correctly classifying the speaker ID for $Z_{ac}$ and $Z_{c}$ .",
+ "bbox": [
+ 112,
+ 646,
+ 489,
+ 853
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The variational contrastive log-ratio upper bound (vCLUB) (Cheng et al., 2020) is used to compute the upper bound of mutual information (MI) for irrelevant information of the speaker, decreasing",
+ "bbox": [
+ 112,
+ 854,
+ 487,
+ 917
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "image",
+ "img_path": "images/aaf86af8ca179b4ece27e6cb1958a38f80f64ef4712e600391877753dbd2a458.jpg",
+ "image_caption": [
+ "Figure 2: Framework of common classifier $C1$ and adversarial speaker classifier $C2$ ."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 526,
+ 82,
+ 882,
+ 227
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "the correlation among different speaker-irrelevant speech representations:",
+ "bbox": [
+ 507,
+ 291,
+ 880,
+ 323
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\begin{array}{l} \\hat {\\mathcal {I}} (Z _ {a c}, Z _ {c}) _ {m i n} \\\\ = \\frac {1}{N ^ {2}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left[ \\log q _ {\\theta} \\left(Z _ {a c _ {i}} \\mid Z _ {c _ {i}}\\right) \\right. \\tag {5} \\\\ - \\log q _ {\\theta} (Z _ {a c _ {j}} | Z _ {c _ {i}}) ], \\\\ \\end{array}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 579,
+ 336,
+ 880,
+ 423
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $q_{\\theta}(Z_{ac}|Z_c)$ is a variational distribution with parameter $\\theta$ to approximate $p(Z_{ac}|Z_c)$ . $\\hat{\\mathcal{I}}$ is the unbiased estimator for vCLUB with samples $\\{Z_{ac_i}, Z_{c_i}\\}$ . The indexes $i$ and $j$ are the samples of $Z_{ac}$ and $Z_c$ . The MI loss is:",
+ "bbox": [
+ 507,
+ 434,
+ 880,
+ 514
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} _ {M I} = \\hat {\\mathcal {I}} \\left(Z _ {a c}, Z _ {c}\\right). \\tag {6}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 620,
+ 529,
+ 880,
+ 546
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Loss: The final training objective of the proposed model is to train the generator $G_{(S_a(A_a) \\longleftrightarrow (S_b(A_b))}$ , achieving bidirectional conversion of speech and accent between speakers $S_a, S_b$ and accents $A_a, A_b$ . A full objective $\\mathcal{L}_{full}$ is written as follows:",
+ "bbox": [
+ 507,
+ 562,
+ 882,
+ 643
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} _ {f u l l} = \\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})} ^ {m a s k - a d v}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 658,
+ 656,
+ 838,
+ 677
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n+ \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {\\text {m a s k - a d v}}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 709,
+ 680,
+ 838,
+ 701
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n+ \\lambda_ {c y c} \\left(\\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {c y c}\\right)\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 611,
+ 703,
+ 838,
+ 724
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\left. + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})}\\right) \\tag {7}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 645,
+ 727,
+ 880,
+ 752
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\begin{array}{l} + \\lambda_ {i d} \\left(\\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})} ^ {i d} + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {i d}\\right) \\\\ + \\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {a d v 2} \\\\ \\left. + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})}\\right) \\\\ + \\alpha \\mathcal {L} _ {\\text {c o m} - C _ {1}} + \\beta \\mathcal {L} _ {\\text {a d v} - C _ {2}} + \\gamma \\mathcal {L} _ {M I}, \\\\ \\end{array}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 532,
+ 750,
+ 840,
+ 835
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathcal{L}^{mask - adv}$ , $\\mathcal{L}^{cyc}$ , $\\mathcal{L}^{id}$ and $\\mathcal{L}^{adv2}$ are loss function defined in MaskCycleGAN-VC (Kaneko et al., 2021). $\\mathcal{L}_{com - C_1}$ and $\\mathcal{L}_{adv - C_2}$ are the cross-entropy loss of the classifiers $C_1$ and $C_2$ . The",
+ "bbox": [
+ 507,
+ 852,
+ 882,
+ 917
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "9291",
+ "bbox": [
+ 480,
+ 927,
+ 517,
+ 940
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "$\\mathcal{L}_{MI}$ loss corresponds to minimizing mutual information of $Z_{ac}$ and $Z_{c}$ . $\\lambda_{cyc}, \\lambda_{id}, \\alpha, \\beta$ and $\\gamma$ are the hyperparameters.",
+ "bbox": [
+ 112,
+ 84,
+ 489,
+ 134
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3 Experiments",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 143,
+ 260,
+ 161
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.1 Data",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 170,
+ 200,
+ 184
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The experiments are conducted on the VCTK (Veaux et al., 2016) corpus. For English accent transfer, we conducted speaker selection involving individuals with diverse accents to validate our approach. In addition, we conducted comprehensive experiments on accent transfer in the low-resource language of Laotian. We utilize a total of 1000 Lao Vientiane accent and 1000 HuaPhan accent utterances, with 100 samples in the validation and testing set, totaling about 1 hour. All audio data used in the experiment have a sampling rate of $22.05\\mathrm{kHz}$ .",
+ "bbox": [
+ 112,
+ 191,
+ 490,
+ 384
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.2 Model and training setup",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 395,
+ 359,
+ 411
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "In the experiments, $a$ accent $S_{a}(A_{a})$ is used as the source speech and $b$ accent $S_{b}(A_{b})$ is used as the target speech for training. We compare our proposed method with the current best non-parallel speech conversion models:",
+ "bbox": [
+ 112,
+ 414,
+ 487,
+ 495
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "CycleGAN-VC2 (Kaneko et al., 2019). A GAN-based speech conversion model that uses mel cepstrum as input and output.",
+ "bbox": [
+ 112,
+ 501,
+ 489,
+ 549
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "CycleGAN-VC3 (Kaneko et al., 2020a). This model uses mel spectrogram as input and output instead of mel cepstrum, and incorporates a time-frequency adaptive normalization (TFAN) module on the basis of CycleGAN-VC2.",
+ "bbox": [
+ 112,
+ 554,
+ 489,
+ 634
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "MaskCycleGAN-VC (Kaneko et al., 2021). This model adds a mask mechanism on the basis of CycleGAN-VC2.",
+ "bbox": [
+ 112,
+ 640,
+ 487,
+ 688
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "SRD (Yang et al., 2022). A method that disentangles speaker information based on mutual information learning.",
+ "bbox": [
+ 112,
+ 694,
+ 489,
+ 741
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "In the experimental application, the feature of the speech is an 80-dimensional Mel-spectrogram, and the tone feature is represented F0. The F0 is extracted from the raw audio using a window size of $20ms$ and a $5ms$ hop. The VQ-VAE quantization described at Sec. 2.2, is applied using an F0 codebook of $K0 = 20$ tokens and an encoder that downsamples the signal by $\\times 16$ . The configurations of MaskCycleGAN-VC followed its original paper (Kaneko et al., 2021). We chose the model checkpoint with the lowest loss on the",
+ "bbox": [
+ 112,
+ 741,
+ 489,
+ 917
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Table 1: Evaluation results of different models. Average MCD, RMSE, SSIM, MOS with $95\\%$ confidence between the converted speech and the ground truth reference. “*” denotes the proposed model. “ $w/o$ ” is short for “without” in ablation study. The “ $A_{p243}$ ”, “ $A_{p329}$ ”, “ $A_{p248}$ ” and “ $A_{p244}$ ” represent the London accent, American accent, Indian accent and Manchester accent of English, respectively. $\\Delta$ : the difference value between our model and the best baseline model, $\\uparrow$ : improved performance compared to the best baseline model, Bold: the best performance under each category, underline: the second best performance, “-”: results are not available.",
+ "bbox": [
+ 507,
+ 82,
+ 884,
+ 267
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "table",
+ "img_path": "images/8fc33f95199a8797bfd4445ae1acc7f87498937e8303191db34b0c09fa07b0de.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Symbol | Definition |
| x | Original data of speech a |
| y | Original data of speech b |
| x' | Generate new sample of speech a |
| y' | Generate new sample of speech b |
| Gθ1X→Y | Forward conversion from speech a to speech b with parameters θ1 |
| Gθ2Y→X | Inverse conversion from speech b to speech a with parameters θ2 |
| S | The mel-spectrogram |
| P | The normalized pitch contour |
| Sa | The speaker identity of speech a |
| Sb | The speaker identity of speech b |
| Aa | The accent identity of speech a |
| Ab | The accent identity of speech b |
| Eac | An accent encoder |
| Es | A speaker encoder |
| Ec | A speech content encoder |
| C1 | A speaker identity classifier with linear |
| C2 | A speaker identity classifier with gradient reverse linear |
| Zac | The accent feature from Eac |
| Zt | The speaker feature from Es |
| Zc | The speaker feature from Ec |
| Zac | The accent feature from Ac |
| I | The unbiased estimator for vCLUB (Cheng et al., 2020) |
",
+ "bbox": [
+ 211,
+ 107,
+ 784,
+ 483
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "the input mel-spectrogram and encourage the converter to fill in missing frames based on surrounding frames. These adjustments add some structure to the text and make it even more reader-friendly.",
+ "bbox": [
+ 112,
+ 508,
+ 489,
+ 571
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "This paper applies the non-parallel database-based voice conversion model MaskCycleGAN-VC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The comprehensive list of the primary symbols used throughout this paper is presented in Table 3.",
+ "bbox": [
+ 112,
+ 571,
+ 489,
+ 750
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "page_number",
+ "text": "9298",
+ "bbox": [
+ 480,
+ 927,
+ 519,
+ 940
+ ],
+ "page_idx": 10
+ }
+]
\ No newline at end of file
diff --git a/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_model.json b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..94caba7c1a88812147514b756893e256e737a1c2
--- /dev/null
+++ b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_model.json
@@ -0,0 +1,1795 @@
+[
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.161,
+ 0.079,
+ 0.838,
+ 0.12
+ ],
+ "angle": 0,
+ "content": "Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modeling"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.272,
+ 0.125,
+ 0.731,
+ 0.159
+ ],
+ "angle": 0,
+ "content": "Linqin Wang\\(^{1,2}\\), Zhengtao Yu\\(^{1,2}\\), *Yuanzhang Yang\\(^{1,2}\\), Shengxiang Gao\\(^{1,2}\\), Cunli Mao\\(^{1,2}\\), Yuxin Huang\\(^{1,2}\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.282,
+ 0.16,
+ 0.721,
+ 0.176
+ ],
+ "angle": 0,
+ "content": "\\(^{1}\\) Faculty of Information Engineering and Automation,"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.235,
+ 0.177,
+ 0.768,
+ 0.193
+ ],
+ "angle": 0,
+ "content": "Kunming University of Science and Technology, Kunming, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.198,
+ 0.193,
+ 0.807,
+ 0.242
+ ],
+ "angle": 0,
+ "content": "2 Yunnan Key Laboratory of Artificial Intelligence, Kunming, China {linqinwang7767}@163.com, {ztyu,gaoshengxiang.yn}@ hotmail.com, {maocunli, yyz845935161, huangyuxin2004}@163.com"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.261,
+ 0.253,
+ 0.341,
+ 0.267
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.142,
+ 0.276,
+ 0.461,
+ 0.546
+ ],
+ "angle": 0,
+ "content": "Existing accent transfer works rely on parallel data or speech recognition models. This paper focuses on the practical application of accent transfer and aims to implement accent transfer using non-parallel datasets. The study has encountered the challenge of speech representation disentanglement and modeling accents. In our accent modeling transfer framework, we manage to solve these problems by two proposed methods. First, we learn the suprasegmental information associated with tone to finely model the accents in terms of tone and rhythm. Second, we propose to use mutual information learning to disentangle the accent features and control the accent of the generated speech during the inference time. Experiments show that the proposed framework attains superior performance to the baseline models in terms of accentedness and audio quality."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.554,
+ 0.26,
+ 0.568
+ ],
+ "angle": 0,
+ "content": "1 Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.578,
+ 0.49,
+ 0.754
+ ],
+ "angle": 0,
+ "content": "The accent transfer task refers to the synthesis of speech with an accent, such as British English for North Americans. Accent pronunciation is a distinctive form of expression influenced by the native language, the speaker's social group or speaking in a particular region (Loots and Niesler, 2011). In general, people find it easier to talk with others in their own accent group. The use of speech is now widely adopted, for example in the field of chatbots and film dubbing requires research on accent transfer of speech."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.756,
+ 0.49,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "At present, the accent transfer task for parallel data has achieved sound research results and increasing performance. Divided according to training data, the methods are specifically divided into: (1) Parallel corpus of different accents of the same speaker using source and target speech content and time alignment (Finkelstein et al., 2022; Liu et al., 2022; Hida et al., 2022; Toda et al., 2007; Oyamada et al., 2017). (2) Non-parallel corpus of"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.254,
+ 0.885,
+ 0.606
+ ],
+ "angle": 0,
+ "content": "multiple speakers with multiple accents using inconsistent source and target speech content (Wang et al., 2021; Zhao et al., 2018, 2019; Kaneko and Kameoka, 2017; Kaneko et al., 2019, 2020a, 2021; Finkelstein et al., 2022) used a multi-stage trained tts model to achieve transfer of North American accents, Australian accents, and British accents, and used a CHiVE-BERT pre-training model to enhance the audio effect of accent generation. Liu et al. (2022) added an accent variance adaptor to model the rhythmicity of accent variance, and also enhanced the accent generation audio by using a consistency constraint module. The use of phonetic posteriorgrams (PPG) is an essential idea in the application of non-parallel data (Wang et al., 2021; Zhao et al., 2018, 2019). Wang et al. (2021) extracted PPG from a Chinese pre-trained speech recognition model and then used them in an end-to-end speech conversion model based on adversarial learning disentangling. This approach achieved accent transfer from Chinese Mandarin to Tianjin and obtained decent results."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.611,
+ 0.884,
+ 0.852
+ ],
+ "angle": 0,
+ "content": "However, existing accent transfer works highly rely on a large amount of labelled parallel data or advanced speech recognition models. Working with an enormous amount of labeling data is always hectic, labor-intensive, and time-consuming, which is more severe for low-resource languages. This limitation hinders the wider application of accent transfer in low-resource scenarios. Hence, it is a timely question: Is it feasible to do non-parallel accent transfer task under a unified framework? It is challenging because the speech representation containing various components, including speaker timbre, accent characteristics, and linguistic content, which are difficult to disentangle, especially for non-parallel accent transfer task."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.855,
+ 0.884,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "The dataset for the task in this paper can be represented as \\(\\{S_a(A_a), S_b(A_b)\\}\\), where the \\(a\\) speaker \\(S_a\\) can only speak the accent \\(A_a\\), and the \\(b\\) speaker \\(S_b\\) can only speak the accent \\(A_b\\). The objec"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.142,
+ 0.905,
+ 0.295,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "* Corresponding author"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.48,
+ 0.928,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "9288"
+ },
+ {
+ "type": "footer",
+ "bbox": [
+ 0.218,
+ 0.946,
+ 0.78,
+ 0.973
+ ],
+ "angle": 0,
+ "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9288-9298 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.085,
+ 0.49,
+ 0.229
+ ],
+ "angle": 0,
+ "content": "tives of this paper are to achieve respectively: (1) \\( S_{a}(A_{a}) \\to S_{b}(A_{a}) \\), where the \\( A_{a} \\) accent transfer to the \\( S_{b} \\) speaker without changing the linguistic content of the speech itself. (2) \\( S_{b}(A_{b}) \\to S_{a}(A_{b}) \\), where the \\( A_{b} \\) accent transfer to the \\( S_{a} \\) speaker without changing the linguistic content of the speech itself. (3) \\( S_{a}(A_{a}) \\Longleftrightarrow S_{b}(A_{b}) \\), two-way speaker timbre and accent transfer between the \\( S_{a} \\) speaker and the \\( S_{b} \\) speaker."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.233,
+ 0.49,
+ 0.538
+ ],
+ "angle": 0,
+ "content": "Following the success of mutual information learning to disentangle speaker information in the One-shot voice conversion (VC) task (Yang et al., 2022), this paper applies the non-parallel databased voice conversion model MaskCycleGANVC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. The most challenging task is to achieve effective disentangling of accent features, linguistic features, speaker timbre features and fine-grained embodied modeling of accents in a unified model architecture, and to achieve controlled and effective speaker timbre and accent transfer in the prediction phase of the model. The accurate modeling of phonetic pronunciation tones in the task of accent transfer is crucial. The contributions of this paper are as follows."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.543,
+ 0.49,
+ 0.671
+ ],
+ "angle": 0,
+ "content": "(1) For accents being difficult to model fine-grained concretely, this paper fine-grained concretely models accents in terms of phonetic intonation, rhythmic pauses and other pronunciation features. Then, an accent feature encoder is proposed, which can effectively extract accent features in the inference stage and realise accent controllability modeling."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.675,
+ 0.49,
+ 0.803
+ ],
+ "angle": 0,
+ "content": "(2) To address the problem of difficult speaker information disentangling in non-parallel data sets with different speakers with different accents, this paper proposes mutual information learning to maximize the mutual information upper bound of accent features and speech features. It can effectively disentangle the speaker features, accent features and phonetic features of speech."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.807,
+ 0.491,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "(3) Experimental results show that method converts the speech up to a MOS score of 4.12. Achieving optimal results compared to baselines on the English accent transfer task for the public VCTK dataset and on the Lao accent transfer for the self-constructed Lao dataset. It significantly improves the accentedness and audio quality."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.113,
+ 0.543,
+ 0.491,
+ 0.919
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.51,
+ 0.084,
+ 0.614,
+ 0.099
+ ],
+ "angle": 0,
+ "content": "2 Method"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.111,
+ 0.885,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "In this section, we first describe our model architecture. Then we introduce the fine-grained accent modeling and adversarial mutual information learning and show how accent transfer between speaker \\( S_{a} \\) and speaker \\( S_{b} \\) on non-parallel datasets with different speakers and accents is achieved."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.22,
+ 0.834,
+ 0.235
+ ],
+ "angle": 0,
+ "content": "2.1 Architecture of the proposed model"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.241,
+ 0.885,
+ 0.499
+ ],
+ "angle": 0,
+ "content": "The generator structure of the model is shown in Figure. 1. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The generator part is composed of five parts: an accent encoder \\( E_{ac} \\), a speaker encoder \\( E_{s} \\), a speech content encoder \\( E_{c} \\), a speech generator \\( G \\), the feature disentanglements \\( C1 \\) and \\( C2 \\). The model training strategy adopts a non-parallel voice conversion approach. Given a non-parallel corpus \\( D(x,y) \\), the training mechanism involves mapping source speech \\( x \\) to converted speech \\( y \\) and then back to \\( x' \\), with the primary training objective being the minimization of the mean square error between \\( x \\) and \\( x' \\). The baseline models training details and our model parameters please refer to Appendix 5."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.509,
+ 0.511,
+ 0.792,
+ 0.526
+ ],
+ "angle": 0,
+ "content": "2.2 Encoder and accent modeling"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.532,
+ 0.885,
+ 0.757
+ ],
+ "angle": 0,
+ "content": "Encoder: The accent encoder \\( E_{ac} \\) takes the mel-spectrogram \\( S \\) and normalized pitch contour \\( P \\) of the speech as inputs. The accent encoder provides global speech accent features to control the speaker's accent. We use a vector quantized variational autoencoder (VQ-VAE) (Van Den Oord et al., 2017; Polyak et al., 2021) model to learn suprasegmental representations related to tone. The speaker encoder is the same structure in (Chen et al., 2021), using Conv1D as the main structure to extract speaker features. The speech content encoder is the downsampling module in MaskCycleGAN (Kaneko et al., 2021), the rhythm encoder \\( E_r \\) extract speech rhythm features, we have:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.554,
+ 0.77,
+ 0.882,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "\\[\n\\begin{array}{l} Z _ {a c} = E _ {a c} \\left(E _ {r} (S), V Q - V A E (P)\\right), \\\\ Z _ {c} = E _ {c} (S), \\tag {1} \\\\ Z _ {t} = E _ {s} (S), \\\\ \\end{array}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.839,
+ 0.884,
+ 0.886
+ ],
+ "angle": 0,
+ "content": "where \\( Z_{ac} \\) represents the accent features, \\( Z_{c} \\) represents the speech content features, \\( Z_{t} \\) represents the speaker features."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.509,
+ 0.888,
+ 0.884,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "Accent modeling: Tones alteration stands out as one of the primary distinguishing features of"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "9289"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.128,
+ 0.086,
+ 0.885,
+ 0.452
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.112,
+ 0.46,
+ 0.884,
+ 0.518
+ ],
+ "angle": 0,
+ "content": "Figure 1: Framework of proposed model. The \\( Z_{ac} \\) is the output hidden states of accent encoder \\( E_{ac} \\), \\( Z_{t} \\) is the output hidden states of speaker encoder \\( E_{s} \\), \\( Z_{c} \\) is the output hidden states of content encoder \\( E_{c} \\), \\( Z_{r} \\) is the output hidden states of rhythm encoder \\( E_{r} \\). \\( C1 \\) is speaker identity classify with linear, \\( C2 \\) is accent classify with gradient reverse linear."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.543,
+ 0.49,
+ 0.706
+ ],
+ "angle": 0,
+ "content": "accents, and these different tones are reflected in the spectrum by different fundamental frequencies, so tone features can be represented by the discrete speech feature F0. To better model the tone at fine-grained degree, we adopt the VQ-VAE framework to train a pre-training model for speech tone feature F0. The yet another algorithm for pitch tracking (YAAPT) (Kasi and Zahorian, 2002) algorithm is used to extract the F0 from the input signal, \\( x \\), generating \\( P = (P_{1},\\dots,P_{T^{\\prime}}) \\), we have:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.228,
+ 0.715,
+ 0.443,
+ 0.735
+ ],
+ "angle": 0,
+ "content": "\\[\nz _ {1: L ^ {\\prime}} ^ {F} = \\mathrm {E n c o d e r} _ {\\Phi_ {1} F} (P _ {1: T ^ {\\prime}}),\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.14,
+ 0.737,
+ 0.488,
+ 0.756
+ ],
+ "angle": 0,
+ "content": "\\[\ne _ {1: L ^ {\\prime}} ^ {F} = \\operatorname {V e c t o r Q u a n t i z a t i o n} _ {\\Phi_ {2} F} \\left(z _ {1: L ^ {\\prime}}\\right), \\quad (2)\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.24,
+ 0.759,
+ 0.443,
+ 0.779
+ ],
+ "angle": 0,
+ "content": "\\[\n\\hat {P} _ {1: L ^ {\\prime}} = \\mathrm {D e c o d e r} _ {\\Phi_ {3}} (e _ {1: L ^ {\\prime}} ^ {F}),\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.79,
+ 0.491,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "Each element in \\( z_{1:L'}^F \\) is an integer \\( z_s \\in \\{0,1,\\dots,K'\\} \\), where \\( K' \\) is the encoder dictionary size. The VectorQuantization \\( \\Phi_{2F} \\), a bottleneck with a learned codebook \\( C = (e_1,\\dots,e_{K'}) \\), where each item in \\( C \\) is a 128-dimensional vector. The encoder extracts a sequence of latent vectors \\( \\mathrm{Encoder}_{\\Phi_1F}(P) = (h_1,\\dots,h_{L'}) \\) from the raw audio, where \\( h_i \\in \\mathbb{R}^{128} \\), for all \\( 1 \\leq i \\leq L' \\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.543,
+ 0.885,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "Then, the bottleneck maps each latent vector to its nearest vector in the codebook \\( C \\). The embedded latent vectors are then being fed into the decoder \\( \\mathrm{Decoder}_{\\Phi_3}(e_{1:L'}^F) = \\hat{P} \\) which reconstructs the original F0 signal. Similar to (Dhariwal et al., 2020), we use Exponential Moving Average updates to learn the codebook and employ random restarts for unused embeddings, we use the indices of the mapped latent vectors to generate \\( Z_p \\). The rhythm of the speech is extracted using the same structure as \\( E_r \\) in SpeechSplit (Qian et al., 2020) to obtain rhythm features output \\( Z_r \\). At last, the \\( Z_p \\oplus Z_r \\) as a representation of accent \\( Z_{ac} \\), it is incorporated into the speech generation model. We chose \\( E_p \\) and \\( E_r \\) as components of the accent encoder \\( E_{ac} \\) because pitch and rhythm represent the most significant aspects of accent variation. Additionally, recent works (Qian et al., 2020; Dhariwal et al., 2020) have demonstrated their effectiveness in extracting pitch and rhythm features. We hope that the discrete representations learned from F0 capture pitch patterns and/or other suprasegmental information. The proposed extension is straightforward"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.522,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "9290"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.085,
+ 0.489,
+ 0.133
+ ],
+ "angle": 0,
+ "content": "ward, but we observe that it results in impressive improvements for accent modeling, especially F0 of reconstructed speech waveforms in Lao."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.145,
+ 0.4,
+ 0.158
+ ],
+ "angle": 0,
+ "content": "2.3 Generator and accent transfer"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.165,
+ 0.49,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "Generator: The speech generator \\(G\\) is based on the architecture of MaskCycleGAN-VC (Kaneko et al., 2021) and uses the Filling in Frames (FIF) strategy during training, where a random part of the spectrogram is masked. The mel-spectrogram is downsampled and mapped from high dimension to low dimension \\(Z_{c}\\). Then, the upsampled \\(Z_{t}\\) (speaker features) and \\(Z_{ac}\\) (accent features) are combined to map these features into the mel-spectrogram of the target speaker. The discriminator has the same structure as MaskCycleGAN-VC and takes as input the mel-spectrogram generated by the generator for the target speaker. We have:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.221,
+ 0.388,
+ 0.488,
+ 0.407
+ ],
+ "angle": 0,
+ "content": "\\[\n\\hat {S} = G \\left(Z _ {a c}, Z _ {c}, Z _ {t}\\right), \\tag {3}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.419,
+ 0.423,
+ 0.436
+ ],
+ "angle": 0,
+ "content": "where \\(\\hat{S}\\) represents the converted speech."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.437,
+ 0.487,
+ 0.485
+ ],
+ "angle": 0,
+ "content": "Controllable accent transfer: During the model inference stage, different target accents can be achieved by controlling the accent feature \\( Z_{ac} \\)."
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.138,
+ 0.51,
+ 0.488,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "\\[\n\\begin{array}{l} \\hat {S} _ {S _ {b} \\left(A _ {a}\\right)} = G \\left(Z _ {a c} \\left(A _ {a}\\right), Z _ {c} \\left(S _ {a}\\right), Z _ {t} \\left(S _ {b}\\right)\\right), \\tag {4} \\\\ \\hat {S} _ {S _ {a} (A _ {b})} = G \\left(Z _ {a c} (A _ {b}), Z _ {c} (S _ {b}), Z _ {t} (S _ {a})\\right), \\\\ \\end{array}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.565,
+ 0.489,
+ 0.614
+ ],
+ "angle": 0,
+ "content": "where \\(\\hat{S}_{S_b(A_a)}\\) represents the speaker \\(S_{b}\\) with accent \\(A_{a}\\), which is not present in training data (only present the speaker \\(S_{b}\\) with accent \\(A_{b}\\))."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.626,
+ 0.362,
+ 0.64
+ ],
+ "angle": 0,
+ "content": "2.4 Disentanglement and loss"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.647,
+ 0.49,
+ 0.854
+ ],
+ "angle": 0,
+ "content": "Speaker information disentanglement: As shown in Fig. 1, we use common classifier \\(C_1\\) and adversarial speaker classifier \\(C_2\\) with gradient reverse linear (GRL) (Ganin et al., 2016) to recognize the identity of speaker, The Fig. 2 illustrates the model architecture of the two classifiers. Both classifiers, \\(C_1\\) and \\(C_2\\), utilize speaker IDs as supervisory labels during the training process. The objective of \\(C_1\\) is to accurately classify the speaker ID associated with \\(Z_{t}\\) as the training progresses. On the contrary, as the training progresses, \\(C_2\\) is designed to gradually struggle in correctly classifying the speaker ID for \\(Z_{ac}\\) and \\(Z_{c}\\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.856,
+ 0.489,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "The variational contrastive log-ratio upper bound (vCLUB) (Cheng et al., 2020) is used to compute the upper bound of mutual information (MI) for irrelevant information of the speaker, decreasing"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.527,
+ 0.083,
+ 0.883,
+ 0.228
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.508,
+ 0.237,
+ 0.883,
+ 0.266
+ ],
+ "angle": 0,
+ "content": "Figure 2: Framework of common classifier \\( C1 \\) and adversarial speaker classifier \\( C2 \\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.292,
+ 0.882,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "the correlation among different speaker-irrelevant speech representations:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.58,
+ 0.337,
+ 0.882,
+ 0.424
+ ],
+ "angle": 0,
+ "content": "\\[\n\\begin{array}{l} \\hat {\\mathcal {I}} (Z _ {a c}, Z _ {c}) _ {m i n} \\\\ = \\frac {1}{N ^ {2}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left[ \\log q _ {\\theta} \\left(Z _ {a c _ {i}} \\mid Z _ {c _ {i}}\\right) \\right. \\tag {5} \\\\ - \\log q _ {\\theta} (Z _ {a c _ {j}} | Z _ {c _ {i}}) ], \\\\ \\end{array}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.509,
+ 0.435,
+ 0.882,
+ 0.515
+ ],
+ "angle": 0,
+ "content": "where \\(q_{\\theta}(Z_{ac}|Z_c)\\) is a variational distribution with parameter \\(\\theta\\) to approximate \\(p(Z_{ac}|Z_c)\\). \\(\\hat{\\mathcal{I}}\\) is the unbiased estimator for vCLUB with samples \\(\\{Z_{ac_i}, Z_{c_i}\\}\\). The indexes \\(i\\) and \\(j\\) are the samples of \\(Z_{ac}\\) and \\(Z_c\\). The MI loss is:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.621,
+ 0.53,
+ 0.882,
+ 0.548
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} _ {M I} = \\hat {\\mathcal {I}} \\left(Z _ {a c}, Z _ {c}\\right). \\tag {6}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.563,
+ 0.884,
+ 0.644
+ ],
+ "angle": 0,
+ "content": "Loss: The final training objective of the proposed model is to train the generator \\( G_{(S_a(A_a) \\longleftrightarrow (S_b(A_b))} \\), achieving bidirectional conversion of speech and accent between speakers \\( S_a, S_b \\) and accents \\( A_a, A_b \\). A full objective \\( \\mathcal{L}_{full} \\) is written as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.66,
+ 0.657,
+ 0.84,
+ 0.678
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} _ {f u l l} = \\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})} ^ {m a s k - a d v}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.71,
+ 0.681,
+ 0.84,
+ 0.702
+ ],
+ "angle": 0,
+ "content": "\\[\n+ \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {\\text {m a s k - a d v}}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.612,
+ 0.704,
+ 0.84,
+ 0.725
+ ],
+ "angle": 0,
+ "content": "\\[\n+ \\lambda_ {c y c} \\left(\\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {c y c}\\right)\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.646,
+ 0.728,
+ 0.882,
+ 0.753
+ ],
+ "angle": 0,
+ "content": "\\[\n\\left. + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})}\\right) \\tag {7}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.533,
+ 0.751,
+ 0.841,
+ 0.836
+ ],
+ "angle": 0,
+ "content": "\\[\n\\begin{array}{l} + \\lambda_ {i d} \\left(\\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})} ^ {i d} + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {i d}\\right) \\\\ + \\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {a d v 2} \\\\ \\left. + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})}\\right) \\\\ + \\alpha \\mathcal {L} _ {\\text {c o m} - C _ {1}} + \\beta \\mathcal {L} _ {\\text {a d v} - C _ {2}} + \\gamma \\mathcal {L} _ {M I}, \\\\ \\end{array}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.853,
+ 0.884,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathcal{L}^{mask - adv}\\), \\(\\mathcal{L}^{cyc}\\), \\(\\mathcal{L}^{id}\\) and \\(\\mathcal{L}^{adv2}\\) are loss function defined in MaskCycleGAN-VC (Kaneko et al., 2021). \\(\\mathcal{L}_{com - C_1}\\) and \\(\\mathcal{L}_{adv - C_2}\\) are the cross-entropy loss of the classifiers \\(C_1\\) and \\(C_2\\). The"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.519,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "9291"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.085,
+ 0.49,
+ 0.135
+ ],
+ "angle": 0,
+ "content": "\\(\\mathcal{L}_{MI}\\) loss corresponds to minimizing mutual information of \\(Z_{ac}\\) and \\(Z_{c}\\). \\(\\lambda_{cyc}, \\lambda_{id}, \\alpha, \\beta\\) and \\(\\gamma\\) are the hyperparameters."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.145,
+ 0.262,
+ 0.162
+ ],
+ "angle": 0,
+ "content": "3 Experiments"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.171,
+ 0.2,
+ 0.185
+ ],
+ "angle": 0,
+ "content": "3.1 Data"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.192,
+ 0.491,
+ 0.385
+ ],
+ "angle": 0,
+ "content": "The experiments are conducted on the VCTK (Veaux et al., 2016) corpus. For English accent transfer, we conducted speaker selection involving individuals with diverse accents to validate our approach. In addition, we conducted comprehensive experiments on accent transfer in the low-resource language of Laotian. We utilize a total of 1000 Lao Vientiane accent and 1000 HuaPhan accent utterances, with 100 samples in the validation and testing set, totaling about 1 hour. All audio data used in the experiment have a sampling rate of \\(22.05\\mathrm{kHz}\\)."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.114,
+ 0.396,
+ 0.361,
+ 0.412
+ ],
+ "angle": 0,
+ "content": "3.2 Model and training setup"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.416,
+ 0.489,
+ 0.497
+ ],
+ "angle": 0,
+ "content": "In the experiments, \\( a \\) accent \\( S_{a}(A_{a}) \\) is used as the source speech and \\( b \\) accent \\( S_{b}(A_{b}) \\) is used as the target speech for training. We compare our proposed method with the current best non-parallel speech conversion models:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.502,
+ 0.49,
+ 0.55
+ ],
+ "angle": 0,
+ "content": "CycleGAN-VC2 (Kaneko et al., 2019). A GAN-based speech conversion model that uses mel cepstrum as input and output."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.555,
+ 0.49,
+ 0.635
+ ],
+ "angle": 0,
+ "content": "CycleGAN-VC3 (Kaneko et al., 2020a). This model uses mel spectrogram as input and output instead of mel cepstrum, and incorporates a time-frequency adaptive normalization (TFAN) module on the basis of CycleGAN-VC2."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.641,
+ 0.489,
+ 0.689
+ ],
+ "angle": 0,
+ "content": "MaskCycleGAN-VC (Kaneko et al., 2021). This model adds a mask mechanism on the basis of CycleGAN-VC2."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.695,
+ 0.49,
+ 0.743
+ ],
+ "angle": 0,
+ "content": "SRD (Yang et al., 2022). A method that disentangles speaker information based on mutual information learning."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.743,
+ 0.49,
+ 0.919
+ ],
+ "angle": 0,
+ "content": "In the experimental application, the feature of the speech is an 80-dimensional Mel-spectrogram, and the tone feature is represented F0. The F0 is extracted from the raw audio using a window size of \\(20ms\\) and a \\(5ms\\) hop. The VQ-VAE quantization described at Sec. 2.2, is applied using an F0 codebook of \\(K0 = 20\\) tokens and an encoder that downsamples the signal by \\(\\times 16\\). The configurations of MaskCycleGAN-VC followed its original paper (Kaneko et al., 2021). We chose the model checkpoint with the lowest loss on the"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.083,
+ 0.885,
+ 0.268
+ ],
+ "angle": 0,
+ "content": "Table 1: Evaluation results of different models. Average MCD, RMSE, SSIM, MOS with \\(95\\%\\) confidence between the converted speech and the ground truth reference. “*” denotes the proposed model. “\\(w/o\\)” is short for “without” in ablation study. The “\\(A_{p243}\\)”, “\\(A_{p329}\\)”, “\\(A_{p248}\\)” and “\\(A_{p244}\\)” represent the London accent, American accent, Indian accent and Manchester accent of English, respectively. \\(\\Delta\\): the difference value between our model and the best baseline model, \\(\\uparrow\\): improved performance compared to the best baseline model, Bold: the best performance under each category, underline: the second best performance, “-”: results are not available."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.514,
+ 0.277,
+ 0.88,
+ 0.926
+ ],
+ "angle": 0,
+ "content": "| Symbol | Definition |
| x | Original data of speech a |
| y | Original data of speech b |
| x' | Generate new sample of speech a |
| y' | Generate new sample of speech b |
| Gθ1X→Y | Forward conversion from speech a to speech b with parameters θ1 |
| Gθ2Y→X | Inverse conversion from speech b to speech a with parameters θ2 |
| S | The mel-spectrogram |
| P | The normalized pitch contour |
| Sa | The speaker identity of speech a |
| Sb | The speaker identity of speech b |
| Aa | The accent identity of speech a |
| Ab | The accent identity of speech b |
| Eac | An accent encoder |
| Es | A speaker encoder |
| Ec | A speech content encoder |
| C1 | A speaker identity classifier with linear |
| C2 | A speaker identity classifier with gradient reverse linear |
| Zac | The accent feature from Eac |
| Zt | The speaker feature from Es |
| Zc | The speaker feature from Ec |
| Zac | The accent feature from Ac |
| I | The unbiased estimator for vCLUB (Cheng et al., 2020) |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.509,
+ 0.49,
+ 0.572
+ ],
+ "angle": 0,
+ "content": "the input mel-spectrogram and encourage the converter to fill in missing frames based on surrounding frames. These adjustments add some structure to the text and make it even more reader-friendly."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.573,
+ 0.49,
+ 0.751
+ ],
+ "angle": 0,
+ "content": "This paper applies the non-parallel database-based voice conversion model MaskCycleGAN-VC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The comprehensive list of the primary symbols used throughout this paper is presented in Table 3."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.928,
+ 0.521,
+ 0.941
+ ],
+ "angle": 0,
+ "content": "9298"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_origin.pdf b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..012d6469dfca0b0c19b1b078e4d9867814afd4a6
--- /dev/null
+++ b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e6f1d89d9062a30f4567fbdc60112e99ae6c8fbb517473a785fbe7b8309e7eb
+size 1799240
diff --git a/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/full.md b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..dd54803b95925d530f9d10a026624c4ca5b4cd3d
--- /dev/null
+++ b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/full.md
@@ -0,0 +1,280 @@
+# Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modeling
+
+Linqin Wang $^{1,2}$ , Zhengtao Yu $^{1,2}$ , *Yuanzhang Yang $^{1,2}$ , Shengxiang Gao $^{1,2}$ , Cunli Mao $^{1,2}$ , Yuxin Huang $^{1,2}$
+
+$^{1}$ Faculty of Information Engineering and Automation,
+
+Kunming University of Science and Technology, Kunming, China
+
+2 Yunnan Key Laboratory of Artificial Intelligence, Kunming, China {linqinwang7767}@163.com, {ztyu,gaoshengxiang.yn}@ hotmail.com, {maocunli, yyz845935161, huangyuxin2004}@163.com
+
+# Abstract
+
+Existing accent transfer works rely on parallel data or speech recognition models. This paper focuses on the practical application of accent transfer and aims to implement accent transfer using non-parallel datasets. The study has encountered the challenge of speech representation disentanglement and modeling accents. In our accent modeling transfer framework, we manage to solve these problems by two proposed methods. First, we learn the suprasegmental information associated with tone to finely model the accents in terms of tone and rhythm. Second, we propose to use mutual information learning to disentangle the accent features and control the accent of the generated speech during the inference time. Experiments show that the proposed framework attains superior performance to the baseline models in terms of accentedness and audio quality.
+
+# 1 Introduction
+
+The accent transfer task refers to the synthesis of speech with an accent, such as British English for North Americans. Accent pronunciation is a distinctive form of expression influenced by the native language, the speaker's social group or speaking in a particular region (Loots and Niesler, 2011). In general, people find it easier to talk with others in their own accent group. The use of speech is now widely adopted, for example in the field of chatbots and film dubbing requires research on accent transfer of speech.
+
+At present, the accent transfer task for parallel data has achieved sound research results and increasing performance. Divided according to training data, the methods are specifically divided into: (1) Parallel corpus of different accents of the same speaker using source and target speech content and time alignment (Finkelstein et al., 2022; Liu et al., 2022; Hida et al., 2022; Toda et al., 2007; Oyamada et al., 2017). (2) Non-parallel corpus of
+
+multiple speakers with multiple accents using inconsistent source and target speech content (Wang et al., 2021; Zhao et al., 2018, 2019; Kaneko and Kameoka, 2017; Kaneko et al., 2019, 2020a, 2021; Finkelstein et al., 2022) used a multi-stage trained tts model to achieve transfer of North American accents, Australian accents, and British accents, and used a CHiVE-BERT pre-training model to enhance the audio effect of accent generation. Liu et al. (2022) added an accent variance adaptor to model the rhythmicity of accent variance, and also enhanced the accent generation audio by using a consistency constraint module. The use of phonetic posteriorgrams (PPG) is an essential idea in the application of non-parallel data (Wang et al., 2021; Zhao et al., 2018, 2019). Wang et al. (2021) extracted PPG from a Chinese pre-trained speech recognition model and then used them in an end-to-end speech conversion model based on adversarial learning disentangling. This approach achieved accent transfer from Chinese Mandarin to Tianjin and obtained decent results.
+
+However, existing accent transfer works highly rely on a large amount of labelled parallel data or advanced speech recognition models. Working with an enormous amount of labeling data is always hectic, labor-intensive, and time-consuming, which is more severe for low-resource languages. This limitation hinders the wider application of accent transfer in low-resource scenarios. Hence, it is a timely question: Is it feasible to do non-parallel accent transfer task under a unified framework? It is challenging because the speech representation containing various components, including speaker timbre, accent characteristics, and linguistic content, which are difficult to disentangle, especially for non-parallel accent transfer task.
+
+The dataset for the task in this paper can be represented as $\{S_a(A_a), S_b(A_b)\}$ , where the $a$ speaker $S_a$ can only speak the accent $A_a$ , and the $b$ speaker $S_b$ can only speak the accent $A_b$ . The objec
+
+tives of this paper are to achieve respectively: (1) $S_{a}(A_{a}) \to S_{b}(A_{a})$ , where the $A_{a}$ accent transfer to the $S_{b}$ speaker without changing the linguistic content of the speech itself. (2) $S_{b}(A_{b}) \to S_{a}(A_{b})$ , where the $A_{b}$ accent transfer to the $S_{a}$ speaker without changing the linguistic content of the speech itself. (3) $S_{a}(A_{a}) \Longleftrightarrow S_{b}(A_{b})$ , two-way speaker timbre and accent transfer between the $S_{a}$ speaker and the $S_{b}$ speaker.
+
+Following the success of mutual information learning to disentangle speaker information in the One-shot voice conversion (VC) task (Yang et al., 2022), this paper applies the non-parallel databased voice conversion model MaskCycleGANVC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. The most challenging task is to achieve effective disentangling of accent features, linguistic features, speaker timbre features and fine-grained embodied modeling of accents in a unified model architecture, and to achieve controlled and effective speaker timbre and accent transfer in the prediction phase of the model. The accurate modeling of phonetic pronunciation tones in the task of accent transfer is crucial. The contributions of this paper are as follows.
+
+(1) For accents being difficult to model fine-grained concretely, this paper fine-grained concretely models accents in terms of phonetic intonation, rhythmic pauses and other pronunciation features. Then, an accent feature encoder is proposed, which can effectively extract accent features in the inference stage and realise accent controllability modeling.
+(2) To address the problem of difficult speaker information disentangling in non-parallel data sets with different speakers with different accents, this paper proposes mutual information learning to maximize the mutual information upper bound of accent features and speech features. It can effectively disentangle the speaker features, accent features and phonetic features of speech.
+(3) Experimental results show that method converts the speech up to a MOS score of 4.12. Achieving optimal results compared to baselines on the English accent transfer task for the public VCTK dataset and on the Lao accent transfer for the self-constructed Lao dataset. It significantly improves the accentedness and audio quality.
+
+# 2 Method
+
+In this section, we first describe our model architecture. Then we introduce the fine-grained accent modeling and adversarial mutual information learning and show how accent transfer between speaker $S_{a}$ and speaker $S_{b}$ on non-parallel datasets with different speakers and accents is achieved.
+
+# 2.1 Architecture of the proposed model
+
+The generator structure of the model is shown in Figure. 1. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The generator part is composed of five parts: an accent encoder $E_{ac}$ , a speaker encoder $E_{s}$ , a speech content encoder $E_{c}$ , a speech generator $G$ , the feature disentanglements $C1$ and $C2$ . The model training strategy adopts a non-parallel voice conversion approach. Given a non-parallel corpus $D(x,y)$ , the training mechanism involves mapping source speech $x$ to converted speech $y$ and then back to $x'$ , with the primary training objective being the minimization of the mean square error between $x$ and $x'$ . The baseline models training details and our model parameters please refer to Appendix 5.
+
+# 2.2 Encoder and accent modeling
+
+Encoder: The accent encoder $E_{ac}$ takes the mel-spectrogram $S$ and normalized pitch contour $P$ of the speech as inputs. The accent encoder provides global speech accent features to control the speaker's accent. We use a vector quantized variational autoencoder (VQ-VAE) (Van Den Oord et al., 2017; Polyak et al., 2021) model to learn suprasegmental representations related to tone. The speaker encoder is the same structure in (Chen et al., 2021), using Conv1D as the main structure to extract speaker features. The speech content encoder is the downsampling module in MaskCycleGAN (Kaneko et al., 2021), the rhythm encoder $E_r$ extract speech rhythm features, we have:
+
+$$
+\begin{array}{l} Z _ {a c} = E _ {a c} \left(E _ {r} (S), V Q - V A E (P)\right), \\ Z _ {c} = E _ {c} (S), \tag {1} \\ Z _ {t} = E _ {s} (S), \\ \end{array}
+$$
+
+where $Z_{ac}$ represents the accent features, $Z_{c}$ represents the speech content features, $Z_{t}$ represents the speaker features.
+
+Accent modeling: Tones alteration stands out as one of the primary distinguishing features of
+
+
+Figure 1: Framework of proposed model. The $Z_{ac}$ is the output hidden states of accent encoder $E_{ac}$ , $Z_{t}$ is the output hidden states of speaker encoder $E_{s}$ , $Z_{c}$ is the output hidden states of content encoder $E_{c}$ , $Z_{r}$ is the output hidden states of rhythm encoder $E_{r}$ . $C1$ is speaker identity classify with linear, $C2$ is accent classify with gradient reverse linear.
+
+accents, and these different tones are reflected in the spectrum by different fundamental frequencies, so tone features can be represented by the discrete speech feature F0. To better model the tone at fine-grained degree, we adopt the VQ-VAE framework to train a pre-training model for speech tone feature F0. The yet another algorithm for pitch tracking (YAAPT) (Kasi and Zahorian, 2002) algorithm is used to extract the F0 from the input signal, $x$ , generating $P = (P_{1},\dots,P_{T^{\prime}})$ , we have:
+
+$$
+z _ {1: L ^ {\prime}} ^ {F} = \mathrm {E n c o d e r} _ {\Phi_ {1} F} (P _ {1: T ^ {\prime}}),
+$$
+
+$$
+e _ {1: L ^ {\prime}} ^ {F} = \operatorname {V e c t o r Q u a n t i z a t i o n} _ {\Phi_ {2} F} \left(z _ {1: L ^ {\prime}}\right), \quad (2)
+$$
+
+$$
+\hat {P} _ {1: L ^ {\prime}} = \mathrm {D e c o d e r} _ {\Phi_ {3}} (e _ {1: L ^ {\prime}} ^ {F}),
+$$
+
+Each element in $z_{1:L'}^F$ is an integer $z_s \in \{0,1,\dots,K'\}$ , where $K'$ is the encoder dictionary size. The VectorQuantization $\Phi_{2F}$ , a bottleneck with a learned codebook $C = (e_1,\dots,e_{K'})$ , where each item in $C$ is a 128-dimensional vector. The encoder extracts a sequence of latent vectors $\mathrm{Encoder}_{\Phi_1F}(P) = (h_1,\dots,h_{L'})$ from the raw audio, where $h_i \in \mathbb{R}^{128}$ , for all $1 \leq i \leq L'$ .
+
+Then, the bottleneck maps each latent vector to its nearest vector in the codebook $C$ . The embedded latent vectors are then being fed into the decoder $\mathrm{Decoder}_{\Phi_3}(e_{1:L'}^F) = \hat{P}$ which reconstructs the original F0 signal. Similar to (Dhariwal et al., 2020), we use Exponential Moving Average updates to learn the codebook and employ random restarts for unused embeddings, we use the indices of the mapped latent vectors to generate $Z_p$ . The rhythm of the speech is extracted using the same structure as $E_r$ in SpeechSplit (Qian et al., 2020) to obtain rhythm features output $Z_r$ . At last, the $Z_p \oplus Z_r$ as a representation of accent $Z_{ac}$ , it is incorporated into the speech generation model. We chose $E_p$ and $E_r$ as components of the accent encoder $E_{ac}$ because pitch and rhythm represent the most significant aspects of accent variation. Additionally, recent works (Qian et al., 2020; Dhariwal et al., 2020) have demonstrated their effectiveness in extracting pitch and rhythm features. We hope that the discrete representations learned from F0 capture pitch patterns and/or other suprasegmental information. The proposed extension is straightforward
+
+ward, but we observe that it results in impressive improvements for accent modeling, especially F0 of reconstructed speech waveforms in Lao.
+
+# 2.3 Generator and accent transfer
+
+Generator: The speech generator $G$ is based on the architecture of MaskCycleGAN-VC (Kaneko et al., 2021) and uses the Filling in Frames (FIF) strategy during training, where a random part of the spectrogram is masked. The mel-spectrogram is downsampled and mapped from high dimension to low dimension $Z_{c}$ . Then, the upsampled $Z_{t}$ (speaker features) and $Z_{ac}$ (accent features) are combined to map these features into the mel-spectrogram of the target speaker. The discriminator has the same structure as MaskCycleGAN-VC and takes as input the mel-spectrogram generated by the generator for the target speaker. We have:
+
+$$
+\hat {S} = G \left(Z _ {a c}, Z _ {c}, Z _ {t}\right), \tag {3}
+$$
+
+where $\hat{S}$ represents the converted speech.
+
+Controllable accent transfer: During the model inference stage, different target accents can be achieved by controlling the accent feature $Z_{ac}$ .
+
+$$
+\begin{array}{l} \hat {S} _ {S _ {b} \left(A _ {a}\right)} = G \left(Z _ {a c} \left(A _ {a}\right), Z _ {c} \left(S _ {a}\right), Z _ {t} \left(S _ {b}\right)\right), \tag {4} \\ \hat {S} _ {S _ {a} (A _ {b})} = G \left(Z _ {a c} (A _ {b}), Z _ {c} (S _ {b}), Z _ {t} (S _ {a})\right), \\ \end{array}
+$$
+
+where $\hat{S}_{S_b(A_a)}$ represents the speaker $S_{b}$ with accent $A_{a}$ , which is not present in training data (only present the speaker $S_{b}$ with accent $A_{b}$ ).
+
+# 2.4 Disentanglement and loss
+
+Speaker information disentanglement: As shown in Fig. 1, we use common classifier $C_1$ and adversarial speaker classifier $C_2$ with gradient reverse linear (GRL) (Ganin et al., 2016) to recognize the identity of speaker, The Fig. 2 illustrates the model architecture of the two classifiers. Both classifiers, $C_1$ and $C_2$ , utilize speaker IDs as supervisory labels during the training process. The objective of $C_1$ is to accurately classify the speaker ID associated with $Z_{t}$ as the training progresses. On the contrary, as the training progresses, $C_2$ is designed to gradually struggle in correctly classifying the speaker ID for $Z_{ac}$ and $Z_{c}$ .
+
+The variational contrastive log-ratio upper bound (vCLUB) (Cheng et al., 2020) is used to compute the upper bound of mutual information (MI) for irrelevant information of the speaker, decreasing
+
+
+Figure 2: Framework of common classifier $C1$ and adversarial speaker classifier $C2$ .
+
+the correlation among different speaker-irrelevant speech representations:
+
+$$
+\begin{array}{l} \hat {\mathcal {I}} (Z _ {a c}, Z _ {c}) _ {m i n} \\ = \frac {1}{N ^ {2}} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \left[ \log q _ {\theta} \left(Z _ {a c _ {i}} \mid Z _ {c _ {i}}\right) \right. \tag {5} \\ - \log q _ {\theta} (Z _ {a c _ {j}} | Z _ {c _ {i}}) ], \\ \end{array}
+$$
+
+where $q_{\theta}(Z_{ac}|Z_c)$ is a variational distribution with parameter $\theta$ to approximate $p(Z_{ac}|Z_c)$ . $\hat{\mathcal{I}}$ is the unbiased estimator for vCLUB with samples $\{Z_{ac_i}, Z_{c_i}\}$ . The indexes $i$ and $j$ are the samples of $Z_{ac}$ and $Z_c$ . The MI loss is:
+
+$$
+\mathcal {L} _ {M I} = \hat {\mathcal {I}} \left(Z _ {a c}, Z _ {c}\right). \tag {6}
+$$
+
+Loss: The final training objective of the proposed model is to train the generator $G_{(S_a(A_a) \longleftrightarrow (S_b(A_b))}$ , achieving bidirectional conversion of speech and accent between speakers $S_a, S_b$ and accents $A_a, A_b$ . A full objective $\mathcal{L}_{full}$ is written as follows:
+
+$$
+\mathcal {L} _ {f u l l} = \mathcal {L} _ {S _ {a} (A _ {a}) \rightarrow S _ {b} (A _ {b})} ^ {m a s k - a d v}
+$$
+
+$$
++ \mathcal {L} _ {S _ {b} (A _ {b}) \rightarrow S _ {a} (A _ {a})} ^ {\text {m a s k - a d v}}
+$$
+
+$$
++ \lambda_ {c y c} \left(\mathcal {L} _ {S _ {a} (A _ {a}) \rightarrow S _ {b} (A _ {b}) \rightarrow S _ {a} (A _ {a})} ^ {c y c}\right)
+$$
+
+$$
+\left. + \mathcal {L} _ {S _ {b} (A _ {b}) \rightarrow S _ {a} (A _ {a}) \rightarrow S _ {b} (A _ {b})}\right) \tag {7}
+$$
+
+$$
+\begin{array}{l} + \lambda_ {i d} \left(\mathcal {L} _ {S _ {a} (A _ {a}) \rightarrow S _ {b} (A _ {b})} ^ {i d} + \mathcal {L} _ {S _ {b} (A _ {b}) \rightarrow S _ {a} (A _ {a})} ^ {i d}\right) \\ + \mathcal {L} _ {S _ {a} (A _ {a}) \rightarrow S _ {b} (A _ {b}) \rightarrow S _ {a} (A _ {a})} ^ {a d v 2} \\ \left. + \mathcal {L} _ {S _ {b} (A _ {b}) \rightarrow S _ {a} (A _ {a}) \rightarrow S _ {b} (A _ {b})}\right) \\ + \alpha \mathcal {L} _ {\text {c o m} - C _ {1}} + \beta \mathcal {L} _ {\text {a d v} - C _ {2}} + \gamma \mathcal {L} _ {M I}, \\ \end{array}
+$$
+
+where $\mathcal{L}^{mask - adv}$ , $\mathcal{L}^{cyc}$ , $\mathcal{L}^{id}$ and $\mathcal{L}^{adv2}$ are loss function defined in MaskCycleGAN-VC (Kaneko et al., 2021). $\mathcal{L}_{com - C_1}$ and $\mathcal{L}_{adv - C_2}$ are the cross-entropy loss of the classifiers $C_1$ and $C_2$ . The
+
+$\mathcal{L}_{MI}$ loss corresponds to minimizing mutual information of $Z_{ac}$ and $Z_{c}$ . $\lambda_{cyc}, \lambda_{id}, \alpha, \beta$ and $\gamma$ are the hyperparameters.
+
+# 3 Experiments
+
+# 3.1 Data
+
+The experiments are conducted on the VCTK (Veaux et al., 2016) corpus. For English accent transfer, we conducted speaker selection involving individuals with diverse accents to validate our approach. In addition, we conducted comprehensive experiments on accent transfer in the low-resource language of Laotian. We utilize a total of 1000 Lao Vientiane accent and 1000 HuaPhan accent utterances, with 100 samples in the validation and testing set, totaling about 1 hour. All audio data used in the experiment have a sampling rate of $22.05\mathrm{kHz}$ .
+
+# 3.2 Model and training setup
+
+In the experiments, $a$ accent $S_{a}(A_{a})$ is used as the source speech and $b$ accent $S_{b}(A_{b})$ is used as the target speech for training. We compare our proposed method with the current best non-parallel speech conversion models:
+
+CycleGAN-VC2 (Kaneko et al., 2019). A GAN-based speech conversion model that uses mel cepstrum as input and output.
+
+CycleGAN-VC3 (Kaneko et al., 2020a). This model uses mel spectrogram as input and output instead of mel cepstrum, and incorporates a time-frequency adaptive normalization (TFAN) module on the basis of CycleGAN-VC2.
+
+MaskCycleGAN-VC (Kaneko et al., 2021). This model adds a mask mechanism on the basis of CycleGAN-VC2.
+
+SRD (Yang et al., 2022). A method that disentangles speaker information based on mutual information learning.
+
+In the experimental application, the feature of the speech is an 80-dimensional Mel-spectrogram, and the tone feature is represented F0. The F0 is extracted from the raw audio using a window size of $20ms$ and a $5ms$ hop. The VQ-VAE quantization described at Sec. 2.2, is applied using an F0 codebook of $K0 = 20$ tokens and an encoder that downsamples the signal by $\times 16$ . The configurations of MaskCycleGAN-VC followed its original paper (Kaneko et al., 2021). We chose the model checkpoint with the lowest loss on the
+
+Table 1: Evaluation results of different models. Average MCD, RMSE, SSIM, MOS with $95\%$ confidence between the converted speech and the ground truth reference. “*” denotes the proposed model. “ $w/o$ ” is short for “without” in ablation study. The “ $A_{p243}$ ”, “ $A_{p329}$ ”, “ $A_{p248}$ ” and “ $A_{p244}$ ” represent the London accent, American accent, Indian accent and Manchester accent of English, respectively. $\Delta$ : the difference value between our model and the best baseline model, $\uparrow$ : improved performance compared to the best baseline model, Bold: the best performance under each category, underline: the second best performance, “-”: results are not available.
+
+| Symbol | Definition |
| x | Original data of speech a |
| y | Original data of speech b |
| x' | Generate new sample of speech a |
| y' | Generate new sample of speech b |
| Gθ1X→Y | Forward conversion from speech a to speech b with parameters θ1 |
| Gθ2Y→X | Inverse conversion from speech b to speech a with parameters θ2 |
| S | The mel-spectrogram |
| P | The normalized pitch contour |
| Sa | The speaker identity of speech a |
| Sb | The speaker identity of speech b |
| Aa | The accent identity of speech a |
| Ab | The accent identity of speech b |
| Eac | An accent encoder |
| Es | A speaker encoder |
| Ec | A speech content encoder |
| C1 | A speaker identity classifier with linear |
| C2 | A speaker identity classifier with gradient reverse linear |
| Zac | The accent feature from Eac |
| Zt | The speaker feature from Es |
| Zc | The speaker feature from Ec |
| Zac | The accent feature from Ac |
| I | The unbiased estimator for vCLUB (Cheng et al., 2020) |
+
+the input mel-spectrogram and encourage the converter to fill in missing frames based on surrounding frames. These adjustments add some structure to the text and make it even more reader-friendly.
+
+This paper applies the non-parallel database-based voice conversion model MaskCycleGAN-VC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The comprehensive list of the primary symbols used throughout this paper is presented in Table 3.
\ No newline at end of file
diff --git a/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/images.zip b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/images.zip
new file mode 100644
index 0000000000000000000000000000000000000000..7e5eeb87ff9cc33c08ce23d9039957855533d526
--- /dev/null
+++ b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/images.zip
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e912a906b8b346be7e4ae4eb1a2c5543b0f1d2dedc5c32e2af628c7bfddf904e
+size 707469
diff --git a/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/layout.json b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..87d5c4264a8ac27aa2cbe66233fbc55574b1a387
--- /dev/null
+++ b/2023/Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modelling/layout.json
@@ -0,0 +1,9456 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 95,
+ 66,
+ 498,
+ 100
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 95,
+ 66,
+ 498,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 95,
+ 66,
+ 498,
+ 100
+ ],
+ "type": "text",
+ "content": "Non-parallel Accent Transfer based on Fine-grained Controllable Accent Modeling"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "text",
+ "content": "Linqin Wang"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "text",
+ "content": ", Zhengtao Yu"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "text",
+ "content": ", *Yuanzhang Yang"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "text",
+ "content": ", Shengxiang Gao"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "text",
+ "content": ", Cunli Mao"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "text",
+ "content": ", Yuxin Huang"
+ },
+ {
+ "bbox": [
+ 161,
+ 105,
+ 434,
+ 133
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 167,
+ 134,
+ 428,
+ 148
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 134,
+ 428,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 134,
+ 428,
+ 148
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 167,
+ 134,
+ 428,
+ 148
+ ],
+ "type": "text",
+ "content": " Faculty of Information Engineering and Automation,"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 139,
+ 148,
+ 456,
+ 162
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 148,
+ 456,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 148,
+ 456,
+ 162
+ ],
+ "type": "text",
+ "content": "Kunming University of Science and Technology, Kunming, China"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 117,
+ 162,
+ 480,
+ 203
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 162,
+ 480,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 162,
+ 480,
+ 203
+ ],
+ "type": "text",
+ "content": "2 Yunnan Key Laboratory of Artificial Intelligence, Kunming, China {linqinwang7767}@163.com, {ztyu,gaoshengxiang.yn}@ hotmail.com, {maocunli, yyz845935161, huangyuxin2004}@163.com"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 155,
+ 212,
+ 202,
+ 224
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 155,
+ 212,
+ 202,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 155,
+ 212,
+ 202,
+ 224
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 84,
+ 232,
+ 274,
+ 459
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 84,
+ 232,
+ 274,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 84,
+ 232,
+ 274,
+ 459
+ ],
+ "type": "text",
+ "content": "Existing accent transfer works rely on parallel data or speech recognition models. This paper focuses on the practical application of accent transfer and aims to implement accent transfer using non-parallel datasets. The study has encountered the challenge of speech representation disentanglement and modeling accents. In our accent modeling transfer framework, we manage to solve these problems by two proposed methods. First, we learn the suprasegmental information associated with tone to finely model the accents in terms of tone and rhythm. Second, we propose to use mutual information learning to disentangle the accent features and control the accent of the generated speech during the inference time. Experiments show that the proposed framework attains superior performance to the baseline models in terms of accentedness and audio quality."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 68,
+ 465,
+ 154,
+ 477
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 465,
+ 154,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 465,
+ 154,
+ 477
+ ],
+ "type": "text",
+ "content": "1 Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 67,
+ 486,
+ 291,
+ 634
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 486,
+ 291,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 486,
+ 291,
+ 634
+ ],
+ "type": "text",
+ "content": "The accent transfer task refers to the synthesis of speech with an accent, such as British English for North Americans. Accent pronunciation is a distinctive form of expression influenced by the native language, the speaker's social group or speaking in a particular region (Loots and Niesler, 2011). In general, people find it easier to talk with others in their own accent group. The use of speech is now widely adopted, for example in the field of chatbots and film dubbing requires research on accent transfer of speech."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 67,
+ 635,
+ 291,
+ 757
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 635,
+ 291,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 635,
+ 291,
+ 757
+ ],
+ "type": "text",
+ "content": "At present, the accent transfer task for parallel data has achieved sound research results and increasing performance. Divided according to training data, the methods are specifically divided into: (1) Parallel corpus of different accents of the same speaker using source and target speech content and time alignment (Finkelstein et al., 2022; Liu et al., 2022; Hida et al., 2022; Toda et al., 2007; Oyamada et al., 2017). (2) Non-parallel corpus of"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 213,
+ 526,
+ 509
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 213,
+ 526,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 213,
+ 526,
+ 509
+ ],
+ "type": "text",
+ "content": "multiple speakers with multiple accents using inconsistent source and target speech content (Wang et al., 2021; Zhao et al., 2018, 2019; Kaneko and Kameoka, 2017; Kaneko et al., 2019, 2020a, 2021; Finkelstein et al., 2022) used a multi-stage trained tts model to achieve transfer of North American accents, Australian accents, and British accents, and used a CHiVE-BERT pre-training model to enhance the audio effect of accent generation. Liu et al. (2022) added an accent variance adaptor to model the rhythmicity of accent variance, and also enhanced the accent generation audio by using a consistency constraint module. The use of phonetic posteriorgrams (PPG) is an essential idea in the application of non-parallel data (Wang et al., 2021; Zhao et al., 2018, 2019). Wang et al. (2021) extracted PPG from a Chinese pre-trained speech recognition model and then used them in an end-to-end speech conversion model based on adversarial learning disentangling. This approach achieved accent transfer from Chinese Mandarin to Tianjin and obtained decent results."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 513,
+ 525,
+ 716
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 513,
+ 525,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 513,
+ 525,
+ 716
+ ],
+ "type": "text",
+ "content": "However, existing accent transfer works highly rely on a large amount of labelled parallel data or advanced speech recognition models. Working with an enormous amount of labeling data is always hectic, labor-intensive, and time-consuming, which is more severe for low-resource languages. This limitation hinders the wider application of accent transfer in low-resource scenarios. Hence, it is a timely question: Is it feasible to do non-parallel accent transfer task under a unified framework? It is challenging because the speech representation containing various components, including speaker timbre, accent characteristics, and linguistic content, which are difficult to disentangle, especially for non-parallel accent transfer task."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": "The dataset for the task in this paper can be represented as "
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\{S_a(A_a), S_b(A_b)\\}"
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": ", where the "
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "a"
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " speaker "
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "S_a"
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " can only speak the accent "
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "A_a"
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": ", and the "
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "b"
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " speaker "
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "S_b"
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " can only speak the accent "
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "A_b"
+ },
+ {
+ "bbox": [
+ 302,
+ 719,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": ". The objec"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 84,
+ 761,
+ 175,
+ 772
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 84,
+ 761,
+ 175,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 84,
+ 761,
+ 175,
+ 772
+ ],
+ "type": "text",
+ "content": "* Corresponding author"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 285,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 285,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "9288"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 818
+ ],
+ "type": "footer",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 818
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 795,
+ 464,
+ 818
+ ],
+ "type": "text",
+ "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9288-9298 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": "tives of this paper are to achieve respectively: (1) "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "S_{a}(A_{a}) \\to S_{b}(A_{a})"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": ", where the "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "A_{a}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " accent transfer to the "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "S_{b}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " speaker without changing the linguistic content of the speech itself. (2) "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "S_{b}(A_{b}) \\to S_{a}(A_{b})"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": ", where the "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "A_{b}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " accent transfer to the "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "S_{a}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " speaker without changing the linguistic content of the speech itself. (3) "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "S_{a}(A_{a}) \\Longleftrightarrow S_{b}(A_{b})"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": ", two-way speaker timbre and accent transfer between the "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "S_{a}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " speaker and the "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "inline_equation",
+ "content": "S_{b}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 192
+ ],
+ "type": "text",
+ "content": " speaker."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 69,
+ 195,
+ 291,
+ 452
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 195,
+ 291,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 195,
+ 291,
+ 452
+ ],
+ "type": "text",
+ "content": "Following the success of mutual information learning to disentangle speaker information in the One-shot voice conversion (VC) task (Yang et al., 2022), this paper applies the non-parallel databased voice conversion model MaskCycleGANVC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. The most challenging task is to achieve effective disentangling of accent features, linguistic features, speaker timbre features and fine-grained embodied modeling of accents in a unified model architecture, and to achieve controlled and effective speaker timbre and accent transfer in the prediction phase of the model. The accurate modeling of phonetic pronunciation tones in the task of accent transfer is crucial. The contributions of this paper are as follows."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 67,
+ 456,
+ 292,
+ 772
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 5,
+ "blocks": [
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 564
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 564
+ ],
+ "type": "text",
+ "content": "(1) For accents being difficult to model fine-grained concretely, this paper fine-grained concretely models accents in terms of phonetic intonation, rhythmic pauses and other pronunciation features. Then, an accent feature encoder is proposed, which can effectively extract accent features in the inference stage and realise accent controllability modeling."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 567,
+ 291,
+ 675
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 567,
+ 291,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 567,
+ 291,
+ 675
+ ],
+ "type": "text",
+ "content": "(2) To address the problem of difficult speaker information disentangling in non-parallel data sets with different speakers with different accents, this paper proposes mutual information learning to maximize the mutual information upper bound of accent features and speech features. It can effectively disentangle the speaker features, accent features and phonetic features of speech."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 678,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 678,
+ 292,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 678,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": "(3) Experimental results show that method converts the speech up to a MOS score of 4.12. Achieving optimal results compared to baselines on the English accent transfer task for the public VCTK dataset and on the Lao accent transfer for the self-constructed Lao dataset. It significantly improves the accentedness and audio quality."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 70,
+ 365,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 303,
+ 70,
+ 365,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 303,
+ 70,
+ 365,
+ 83
+ ],
+ "type": "text",
+ "content": "2 Method"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 302,
+ 93,
+ 526,
+ 174
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 93,
+ 526,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 93,
+ 526,
+ 174
+ ],
+ "type": "text",
+ "content": "In this section, we first describe our model architecture. Then we introduce the fine-grained accent modeling and adversarial mutual information learning and show how accent transfer between speaker "
+ },
+ {
+ "bbox": [
+ 302,
+ 93,
+ 526,
+ 174
+ ],
+ "type": "inline_equation",
+ "content": "S_{a}"
+ },
+ {
+ "bbox": [
+ 302,
+ 93,
+ 526,
+ 174
+ ],
+ "type": "text",
+ "content": " and speaker "
+ },
+ {
+ "bbox": [
+ 302,
+ 93,
+ 526,
+ 174
+ ],
+ "type": "inline_equation",
+ "content": "S_{b}"
+ },
+ {
+ "bbox": [
+ 302,
+ 93,
+ 526,
+ 174
+ ],
+ "type": "text",
+ "content": " on non-parallel datasets with different speakers and accents is achieved."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 302,
+ 185,
+ 496,
+ 197
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 185,
+ 496,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 185,
+ 496,
+ 197
+ ],
+ "type": "text",
+ "content": "2.1 Architecture of the proposed model"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": "The generator structure of the model is shown in Figure. 1. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The generator part is composed of five parts: an accent encoder "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "E_{ac}"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ", a speaker encoder "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "E_{s}"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ", a speech content encoder "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "E_{c}"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ", a speech generator "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "G"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ", the feature disentanglements "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "C1"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "C2"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ". The model training strategy adopts a non-parallel voice conversion approach. Given a non-parallel corpus "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "D(x,y)"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ", the training mechanism involves mapping source speech "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "x"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": " to converted speech "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "y"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": " and then back to "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "x'"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ", with the primary training objective being the minimization of the mean square error between "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "x"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "inline_equation",
+ "content": "x'"
+ },
+ {
+ "bbox": [
+ 302,
+ 202,
+ 526,
+ 419
+ ],
+ "type": "text",
+ "content": ". The baseline models training details and our model parameters please refer to Appendix 5."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 302,
+ 429,
+ 471,
+ 442
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 429,
+ 471,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 429,
+ 471,
+ 442
+ ],
+ "type": "text",
+ "content": "2.2 Encoder and accent modeling"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "content": "Encoder: The accent encoder "
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "inline_equation",
+ "content": "E_{ac}"
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "content": " takes the mel-spectrogram "
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "inline_equation",
+ "content": "S"
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "content": " and normalized pitch contour "
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "inline_equation",
+ "content": "P"
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "content": " of the speech as inputs. The accent encoder provides global speech accent features to control the speaker's accent. We use a vector quantized variational autoencoder (VQ-VAE) (Van Den Oord et al., 2017; Polyak et al., 2021) model to learn suprasegmental representations related to tone. The speaker encoder is the same structure in (Chen et al., 2021), using Conv1D as the main structure to extract speaker features. The speech content encoder is the downsampling module in MaskCycleGAN (Kaneko et al., 2021), the rhythm encoder "
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "inline_equation",
+ "content": "E_r"
+ },
+ {
+ "bbox": [
+ 302,
+ 447,
+ 526,
+ 636
+ ],
+ "type": "text",
+ "content": " extract speech rhythm features, we have:"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 329,
+ 647,
+ 524,
+ 695
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 329,
+ 647,
+ 524,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 329,
+ 647,
+ 524,
+ 695
+ ],
+ "type": "interline_equation",
+ "content": "\\begin{array}{l} Z _ {a c} = E _ {a c} \\left(E _ {r} (S), V Q - V A E (P)\\right), \\\\ Z _ {c} = E _ {c} (S), \\tag {1} \\\\ Z _ {t} = E _ {s} (S), \\\\ \\end{array}",
+ "image_path": "c83c30d04d675aa93169211d4be1df1debe83f9cf0b5a8f06cf7652efe2f3c2c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "text",
+ "content": " represents the accent features, "
+ },
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "inline_equation",
+ "content": "Z_{c}"
+ },
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "text",
+ "content": " represents the speech content features, "
+ },
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "inline_equation",
+ "content": "Z_{t}"
+ },
+ {
+ "bbox": [
+ 302,
+ 705,
+ 525,
+ 745
+ ],
+ "type": "text",
+ "content": " represents the speaker features."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 746,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": "Accent modeling: Tones alteration stands out as one of the primary distinguishing features of"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "9289"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 76,
+ 72,
+ 526,
+ 380
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 76,
+ 72,
+ 526,
+ 380
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 76,
+ 72,
+ 526,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 76,
+ 72,
+ 526,
+ 380
+ ],
+ "type": "image",
+ "image_path": "f2920a1775a0213513da8d4a3440e392413387912e4c4eb3378654955912ce50.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": "Figure 1: Framework of proposed model. The "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": " is the output hidden states of accent encoder "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "E_{ac}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "Z_{t}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": " is the output hidden states of speaker encoder "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "E_{s}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "Z_{c}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": " is the output hidden states of content encoder "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "E_{c}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "Z_{r}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": " is the output hidden states of rhythm encoder "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "E_{r}"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": ". "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "C1"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": " is speaker identity classify with linear, "
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "inline_equation",
+ "content": "C2"
+ },
+ {
+ "bbox": [
+ 66,
+ 386,
+ 525,
+ 435
+ ],
+ "type": "text",
+ "content": " is accent classify with gradient reverse linear."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 593
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 593
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 593
+ ],
+ "type": "text",
+ "content": "accents, and these different tones are reflected in the spectrum by different fundamental frequencies, so tone features can be represented by the discrete speech feature F0. To better model the tone at fine-grained degree, we adopt the VQ-VAE framework to train a pre-training model for speech tone feature F0. The yet another algorithm for pitch tracking (YAAPT) (Kasi and Zahorian, 2002) algorithm is used to extract the F0 from the input signal, "
+ },
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 593
+ ],
+ "type": "inline_equation",
+ "content": "x"
+ },
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 593
+ ],
+ "type": "text",
+ "content": ", generating "
+ },
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 593
+ ],
+ "type": "inline_equation",
+ "content": "P = (P_{1},\\dots,P_{T^{\\prime}})"
+ },
+ {
+ "bbox": [
+ 67,
+ 456,
+ 291,
+ 593
+ ],
+ "type": "text",
+ "content": ", we have:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 135,
+ 601,
+ 263,
+ 618
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 601,
+ 263,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 601,
+ 263,
+ 618
+ ],
+ "type": "interline_equation",
+ "content": "z _ {1: L ^ {\\prime}} ^ {F} = \\mathrm {E n c o d e r} _ {\\Phi_ {1} F} (P _ {1: T ^ {\\prime}}),",
+ "image_path": "c61316fbfb76e6720b7fd075a58972a62a2eb88e03e52cce41f8e03d6d04544e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 83,
+ 619,
+ 290,
+ 635
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 83,
+ 619,
+ 290,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 83,
+ 619,
+ 290,
+ 635
+ ],
+ "type": "interline_equation",
+ "content": "e _ {1: L ^ {\\prime}} ^ {F} = \\operatorname {V e c t o r Q u a n t i z a t i o n} _ {\\Phi_ {2} F} \\left(z _ {1: L ^ {\\prime}}\\right), \\quad (2)",
+ "image_path": "94981b87841570bd5bfb64a670d556037bfcd826889b5b3e9e0560fc29f11901.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 142,
+ 638,
+ 263,
+ 655
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 638,
+ 263,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 638,
+ 263,
+ 655
+ ],
+ "type": "interline_equation",
+ "content": "\\hat {P} _ {1: L ^ {\\prime}} = \\mathrm {D e c o d e r} _ {\\Phi_ {3}} (e _ {1: L ^ {\\prime}} ^ {F}),",
+ "image_path": "dc602f3fab58ceff6802a4afba8bc2663b439de36002e36db0c64041b6a42a58.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": "Each element in "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "z_{1:L'}^F"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": " is an integer "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "z_s \\in \\{0,1,\\dots,K'\\}"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "K'"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": " is the encoder dictionary size. The VectorQuantization "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\Phi_{2F}"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": ", a bottleneck with a learned codebook "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "C = (e_1,\\dots,e_{K'})"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": ", where each item in "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "C"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": " is a 128-dimensional vector. The encoder extracts a sequence of latent vectors "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Encoder}_{\\Phi_1F}(P) = (h_1,\\dots,h_{L'})"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": " from the raw audio, where "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "h_i \\in \\mathbb{R}^{128}"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": ", for all "
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "1 \\leq i \\leq L'"
+ },
+ {
+ "bbox": [
+ 67,
+ 664,
+ 292,
+ 772
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": "Then, the bottleneck maps each latent vector to its nearest vector in the codebook "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "C"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": ". The embedded latent vectors are then being fed into the decoder "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Decoder}_{\\Phi_3}(e_{1:L'}^F) = \\hat{P}"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": " which reconstructs the original F0 signal. Similar to (Dhariwal et al., 2020), we use Exponential Moving Average updates to learn the codebook and employ random restarts for unused embeddings, we use the indices of the mapped latent vectors to generate "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "Z_p"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": ". The rhythm of the speech is extracted using the same structure as "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "E_r"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": " in SpeechSplit (Qian et al., 2020) to obtain rhythm features output "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "Z_r"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": ". At last, the "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "Z_p \\oplus Z_r"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": " as a representation of accent "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": ", it is incorporated into the speech generation model. We chose "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "E_p"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "E_r"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": " as components of the accent encoder "
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "inline_equation",
+ "content": "E_{ac}"
+ },
+ {
+ "bbox": [
+ 302,
+ 456,
+ 526,
+ 767
+ ],
+ "type": "text",
+ "content": " because pitch and rhythm represent the most significant aspects of accent variation. Additionally, recent works (Qian et al., 2020; Dhariwal et al., 2020) have demonstrated their effectiveness in extracting pitch and rhythm features. We hope that the discrete representations learned from F0 capture pitch patterns and/or other suprasegmental information. The proposed extension is straightforward"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 310,
+ 791
+ ],
+ "type": "text",
+ "content": "9290"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 290,
+ 111
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 290,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 290,
+ 111
+ ],
+ "type": "text",
+ "content": "ward, but we observe that it results in impressive improvements for accent modeling, especially F0 of reconstructed speech waveforms in Lao."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 67,
+ 121,
+ 238,
+ 132
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 121,
+ 238,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 121,
+ 238,
+ 132
+ ],
+ "type": "text",
+ "content": "2.3 Generator and accent transfer"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "text",
+ "content": "Generator: The speech generator "
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "inline_equation",
+ "content": "G"
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "text",
+ "content": " is based on the architecture of MaskCycleGAN-VC (Kaneko et al., 2021) and uses the Filling in Frames (FIF) strategy during training, where a random part of the spectrogram is masked. The mel-spectrogram is downsampled and mapped from high dimension to low dimension "
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "inline_equation",
+ "content": "Z_{c}"
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "text",
+ "content": ". Then, the upsampled "
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "inline_equation",
+ "content": "Z_{t}"
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "text",
+ "content": " (speaker features) and "
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 67,
+ 138,
+ 291,
+ 314
+ ],
+ "type": "text",
+ "content": " (accent features) are combined to map these features into the mel-spectrogram of the target speaker. The discriminator has the same structure as MaskCycleGAN-VC and takes as input the mel-spectrogram generated by the generator for the target speaker. We have:"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 131,
+ 326,
+ 290,
+ 342
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 326,
+ 290,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 326,
+ 290,
+ 342
+ ],
+ "type": "interline_equation",
+ "content": "\\hat {S} = G \\left(Z _ {a c}, Z _ {c}, Z _ {t}\\right), \\tag {3}",
+ "image_path": "b959275b1aefc43721bc3ad9d607ac2939acbb2906f591d721f10a0b39643416.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 352,
+ 251,
+ 366
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 352,
+ 251,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 352,
+ 251,
+ 366
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 67,
+ 352,
+ 251,
+ 366
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{S}"
+ },
+ {
+ "bbox": [
+ 67,
+ 352,
+ 251,
+ 366
+ ],
+ "type": "text",
+ "content": " represents the converted speech."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 367,
+ 289,
+ 407
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 367,
+ 289,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 367,
+ 289,
+ 407
+ ],
+ "type": "text",
+ "content": "Controllable accent transfer: During the model inference stage, different target accents can be achieved by controlling the accent feature "
+ },
+ {
+ "bbox": [
+ 67,
+ 367,
+ 289,
+ 407
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 67,
+ 367,
+ 289,
+ 407
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 82,
+ 428,
+ 290,
+ 463
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 82,
+ 428,
+ 290,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 82,
+ 428,
+ 290,
+ 463
+ ],
+ "type": "interline_equation",
+ "content": "\\begin{array}{l} \\hat {S} _ {S _ {b} \\left(A _ {a}\\right)} = G \\left(Z _ {a c} \\left(A _ {a}\\right), Z _ {c} \\left(S _ {a}\\right), Z _ {t} \\left(S _ {b}\\right)\\right), \\tag {4} \\\\ \\hat {S} _ {S _ {a} (A _ {b})} = G \\left(Z _ {a c} (A _ {b}), Z _ {c} (S _ {b}), Z _ {t} (S _ {a})\\right), \\\\ \\end{array}",
+ "image_path": "650205149ec4388b3d0b4441f892b09e9fba305512bda3aa946037503a634539.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{S}_{S_b(A_a)}"
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "text",
+ "content": " represents the speaker "
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "S_{b}"
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "text",
+ "content": " with accent "
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "A_{a}"
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "text",
+ "content": ", which is not present in training data (only present the speaker "
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "S_{b}"
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "text",
+ "content": " with accent "
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "A_{b}"
+ },
+ {
+ "bbox": [
+ 67,
+ 475,
+ 290,
+ 516
+ ],
+ "type": "text",
+ "content": ")."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 67,
+ 526,
+ 215,
+ 538
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 526,
+ 215,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 526,
+ 215,
+ 538
+ ],
+ "type": "text",
+ "content": "2.4 Disentanglement and loss"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": "Speaker information disentanglement: As shown in Fig. 1, we use common classifier "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "C_1"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": " and adversarial speaker classifier "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "C_2"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": " with gradient reverse linear (GRL) (Ganin et al., 2016) to recognize the identity of speaker, The Fig. 2 illustrates the model architecture of the two classifiers. Both classifiers, "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "C_1"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "C_2"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": ", utilize speaker IDs as supervisory labels during the training process. The objective of "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "C_1"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": " is to accurately classify the speaker ID associated with "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "Z_{t}"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": " as the training progresses. On the contrary, as the training progresses, "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "C_2"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": " is designed to gradually struggle in correctly classifying the speaker ID for "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "inline_equation",
+ "content": "Z_{c}"
+ },
+ {
+ "bbox": [
+ 67,
+ 544,
+ 291,
+ 718
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 67,
+ 719,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 719,
+ 290,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 719,
+ 290,
+ 772
+ ],
+ "type": "text",
+ "content": "The variational contrastive log-ratio upper bound (vCLUB) (Cheng et al., 2020) is used to compute the upper bound of mutual information (MI) for irrelevant information of the speaker, decreasing"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 313,
+ 69,
+ 525,
+ 191
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 313,
+ 69,
+ 525,
+ 191
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 69,
+ 525,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 69,
+ 525,
+ 191
+ ],
+ "type": "image",
+ "image_path": "aaf86af8ca179b4ece27e6cb1958a38f80f64ef4712e600391877753dbd2a458.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 302,
+ 199,
+ 525,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 199,
+ 525,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 199,
+ 525,
+ 223
+ ],
+ "type": "text",
+ "content": "Figure 2: Framework of common classifier "
+ },
+ {
+ "bbox": [
+ 302,
+ 199,
+ 525,
+ 223
+ ],
+ "type": "inline_equation",
+ "content": "C1"
+ },
+ {
+ "bbox": [
+ 302,
+ 199,
+ 525,
+ 223
+ ],
+ "type": "text",
+ "content": " and adversarial speaker classifier "
+ },
+ {
+ "bbox": [
+ 302,
+ 199,
+ 525,
+ 223
+ ],
+ "type": "inline_equation",
+ "content": "C2"
+ },
+ {
+ "bbox": [
+ 302,
+ 199,
+ 525,
+ 223
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 302,
+ 245,
+ 524,
+ 272
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 245,
+ 524,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 245,
+ 524,
+ 272
+ ],
+ "type": "text",
+ "content": "the correlation among different speaker-irrelevant speech representations:"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 345,
+ 283,
+ 524,
+ 356
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 345,
+ 283,
+ 524,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 283,
+ 524,
+ 356
+ ],
+ "type": "interline_equation",
+ "content": "\\begin{array}{l} \\hat {\\mathcal {I}} (Z _ {a c}, Z _ {c}) _ {m i n} \\\\ = \\frac {1}{N ^ {2}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left[ \\log q _ {\\theta} \\left(Z _ {a c _ {i}} \\mid Z _ {c _ {i}}\\right) \\right. \\tag {5} \\\\ - \\log q _ {\\theta} (Z _ {a c _ {j}} | Z _ {c _ {i}}) ], \\\\ \\end{array}",
+ "image_path": "971caff4fd3f225d537f4474af13af202dc0e5af5fd6536a9b3ebb61e3428a35.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "q_{\\theta}(Z_{ac}|Z_c)"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": " is a variational distribution with parameter "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "\\theta"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": " to approximate "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "p(Z_{ac}|Z_c)"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": ". "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{\\mathcal{I}}"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": " is the unbiased estimator for vCLUB with samples "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "\\{Z_{ac_i}, Z_{c_i}\\}"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": ". The indexes "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "i"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "j"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": " are the samples of "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "inline_equation",
+ "content": "Z_c"
+ },
+ {
+ "bbox": [
+ 302,
+ 365,
+ 524,
+ 433
+ ],
+ "type": "text",
+ "content": ". The MI loss is:"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 369,
+ 445,
+ 524,
+ 460
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 369,
+ 445,
+ 524,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 369,
+ 445,
+ 524,
+ 460
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} _ {M I} = \\hat {\\mathcal {I}} \\left(Z _ {a c}, Z _ {c}\\right). \\tag {6}",
+ "image_path": "92e43fc4568758c0f131a3a7bb80ec5268e9a0f2f93d275b63037044de5bdee6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "text",
+ "content": "Loss: The final training objective of the proposed model is to train the generator "
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "inline_equation",
+ "content": "G_{(S_a(A_a) \\longleftrightarrow (S_b(A_b))}"
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "text",
+ "content": ", achieving bidirectional conversion of speech and accent between speakers "
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "inline_equation",
+ "content": "S_a, S_b"
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "text",
+ "content": " and accents "
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "inline_equation",
+ "content": "A_a, A_b"
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "text",
+ "content": ". A full objective "
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}_{full}"
+ },
+ {
+ "bbox": [
+ 302,
+ 473,
+ 525,
+ 541
+ ],
+ "type": "text",
+ "content": " is written as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 392,
+ 552,
+ 499,
+ 570
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 392,
+ 552,
+ 499,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 392,
+ 552,
+ 499,
+ 570
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} _ {f u l l} = \\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})} ^ {m a s k - a d v}",
+ "image_path": "0b2be0f8334606a3f29481ead832163d1cd293ecbd12a8bf0ca37464ffd9a295.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 422,
+ 572,
+ 499,
+ 590
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 422,
+ 572,
+ 499,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 422,
+ 572,
+ 499,
+ 590
+ ],
+ "type": "interline_equation",
+ "content": "+ \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {\\text {m a s k - a d v}}",
+ "image_path": "9390895df361ca628af3dbcaa68021c5ac810fb3f86d2b41079d69148eb3645d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 364,
+ 592,
+ 499,
+ 609
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 364,
+ 592,
+ 499,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 364,
+ 592,
+ 499,
+ 609
+ ],
+ "type": "interline_equation",
+ "content": "+ \\lambda_ {c y c} \\left(\\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {c y c}\\right)",
+ "image_path": "95b2d53233ab4114f9f9e6a3cff6bcb9fd0d92265bc1d1281b64345b44b78a67.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 384,
+ 612,
+ 524,
+ 633
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 384,
+ 612,
+ 524,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 384,
+ 612,
+ 524,
+ 633
+ ],
+ "type": "interline_equation",
+ "content": "\\left. + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})}\\right) \\tag {7}",
+ "image_path": "4913f94020e7ae385e43cd0f48268d0130c0ecd9f4d537840d11a657ead94254.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 317,
+ 631,
+ 500,
+ 703
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 631,
+ 500,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 631,
+ 500,
+ 703
+ ],
+ "type": "interline_equation",
+ "content": "\\begin{array}{l} + \\lambda_ {i d} \\left(\\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})} ^ {i d} + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {i d}\\right) \\\\ + \\mathcal {L} _ {S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a})} ^ {a d v 2} \\\\ \\left. + \\mathcal {L} _ {S _ {b} (A _ {b}) \\rightarrow S _ {a} (A _ {a}) \\rightarrow S _ {b} (A _ {b})}\\right) \\\\ + \\alpha \\mathcal {L} _ {\\text {c o m} - C _ {1}} + \\beta \\mathcal {L} _ {\\text {a d v} - C _ {2}} + \\gamma \\mathcal {L} _ {M I}, \\\\ \\end{array}",
+ "image_path": "4cf85ccf40e93d0be067088eff7954ab41ddc143202e7cb0f5b65e030a295278.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}^{mask - adv}"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}^{cyc}"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}^{id}"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}^{adv2}"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " are loss function defined in MaskCycleGAN-VC (Kaneko et al., 2021). "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}_{com - C_1}"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}_{adv - C_2}"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " are the cross-entropy loss of the classifiers "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "C_1"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "C_2"
+ },
+ {
+ "bbox": [
+ 302,
+ 717,
+ 525,
+ 772
+ ],
+ "type": "text",
+ "content": ". The"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 308,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 308,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 308,
+ 791
+ ],
+ "type": "text",
+ "content": "9291"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{L}_{MI}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "text",
+ "content": " loss corresponds to minimizing mutual information of "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "inline_equation",
+ "content": "Z_{ac}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "inline_equation",
+ "content": "Z_{c}"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "text",
+ "content": ". "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "inline_equation",
+ "content": "\\lambda_{cyc}, \\lambda_{id}, \\alpha, \\beta"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "inline_equation",
+ "content": "\\gamma"
+ },
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 113
+ ],
+ "type": "text",
+ "content": " are the hyperparameters."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 68,
+ 121,
+ 155,
+ 136
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 121,
+ 155,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 121,
+ 155,
+ 136
+ ],
+ "type": "text",
+ "content": "3 Experiments"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 68,
+ 143,
+ 119,
+ 155
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 143,
+ 119,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 143,
+ 119,
+ 155
+ ],
+ "type": "text",
+ "content": "3.1 Data"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 161,
+ 292,
+ 323
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 161,
+ 292,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 161,
+ 292,
+ 323
+ ],
+ "type": "text",
+ "content": "The experiments are conducted on the VCTK (Veaux et al., 2016) corpus. For English accent transfer, we conducted speaker selection involving individuals with diverse accents to validate our approach. In addition, we conducted comprehensive experiments on accent transfer in the low-resource language of Laotian. We utilize a total of 1000 Lao Vientiane accent and 1000 HuaPhan accent utterances, with 100 samples in the validation and testing set, totaling about 1 hour. All audio data used in the experiment have a sampling rate of "
+ },
+ {
+ "bbox": [
+ 67,
+ 161,
+ 292,
+ 323
+ ],
+ "type": "inline_equation",
+ "content": "22.05\\mathrm{kHz}"
+ },
+ {
+ "bbox": [
+ 67,
+ 161,
+ 292,
+ 323
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 67,
+ 333,
+ 214,
+ 346
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 333,
+ 214,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 333,
+ 214,
+ 346
+ ],
+ "type": "text",
+ "content": "3.2 Model and training setup"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "text",
+ "content": "In the experiments, "
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "a"
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "text",
+ "content": " accent "
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "S_{a}(A_{a})"
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "text",
+ "content": " is used as the source speech and "
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "b"
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "text",
+ "content": " accent "
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "S_{b}(A_{b})"
+ },
+ {
+ "bbox": [
+ 67,
+ 349,
+ 290,
+ 417
+ ],
+ "type": "text",
+ "content": " is used as the target speech for training. We compare our proposed method with the current best non-parallel speech conversion models:"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 67,
+ 422,
+ 291,
+ 462
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 422,
+ 291,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 422,
+ 291,
+ 462
+ ],
+ "type": "text",
+ "content": "CycleGAN-VC2 (Kaneko et al., 2019). A GAN-based speech conversion model that uses mel cepstrum as input and output."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 67,
+ 466,
+ 291,
+ 534
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 466,
+ 291,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 466,
+ 291,
+ 534
+ ],
+ "type": "text",
+ "content": "CycleGAN-VC3 (Kaneko et al., 2020a). This model uses mel spectrogram as input and output instead of mel cepstrum, and incorporates a time-frequency adaptive normalization (TFAN) module on the basis of CycleGAN-VC2."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 67,
+ 539,
+ 290,
+ 579
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 539,
+ 290,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 539,
+ 290,
+ 579
+ ],
+ "type": "text",
+ "content": "MaskCycleGAN-VC (Kaneko et al., 2021). This model adds a mask mechanism on the basis of CycleGAN-VC2."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 67,
+ 584,
+ 291,
+ 624
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 584,
+ 291,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 584,
+ 291,
+ 624
+ ],
+ "type": "text",
+ "content": "SRD (Yang et al., 2022). A method that disentangles speaker information based on mutual information learning."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": "In the experimental application, the feature of the speech is an 80-dimensional Mel-spectrogram, and the tone feature is represented F0. The F0 is extracted from the raw audio using a window size of "
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "20ms"
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": " and a "
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "5ms"
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": " hop. The VQ-VAE quantization described at Sec. 2.2, is applied using an F0 codebook of "
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "K0 = 20"
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": " tokens and an encoder that downsamples the signal by "
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "inline_equation",
+ "content": "\\times 16"
+ },
+ {
+ "bbox": [
+ 67,
+ 624,
+ 291,
+ 772
+ ],
+ "type": "text",
+ "content": ". The configurations of MaskCycleGAN-VC followed its original paper (Kaneko et al., 2021). We chose the model checkpoint with the lowest loss on the"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": "Table 1: Evaluation results of different models. Average MCD, RMSE, SSIM, MOS with "
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "95\\%"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": " confidence between the converted speech and the ground truth reference. “*” denotes the proposed model. “"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "w/o"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": "” is short for “without” in ablation study. The “"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "A_{p243}"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": "”, “"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "A_{p329}"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": "”, “"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "A_{p248}"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": "” and “"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "A_{p244}"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": "” represent the London accent, American accent, Indian accent and Manchester accent of English, respectively. "
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "\\Delta"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": ": the difference value between our model and the best baseline model, "
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "inline_equation",
+ "content": "\\uparrow"
+ },
+ {
+ "bbox": [
+ 302,
+ 69,
+ 526,
+ 225
+ ],
+ "type": "text",
+ "content": ": improved performance compared to the best baseline model, Bold: the best performance under each category, underline: the second best performance, “-”: results are not available."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 305,
+ 232,
+ 523,
+ 778
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 305,
+ 232,
+ 523,
+ 778
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 232,
+ 523,
+ 778
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 232,
+ 523,
+ 778
+ ],
+ "type": "table",
+ "html": "| Symbol | Definition |
| x | Original data of speech a |
| y | Original data of speech b |
| x' | Generate new sample of speech a |
| y' | Generate new sample of speech b |
| Gθ1X→Y | Forward conversion from speech a to speech b with parameters θ1 |
| Gθ2Y→X | Inverse conversion from speech b to speech a with parameters θ2 |
| S | The mel-spectrogram |
| P | The normalized pitch contour |
| Sa | The speaker identity of speech a |
| Sb | The speaker identity of speech b |
| Aa | The accent identity of speech a |
| Ab | The accent identity of speech b |
| Eac | An accent encoder |
| Es | A speaker encoder |
| Ec | A speech content encoder |
| C1 | A speaker identity classifier with linear |
| C2 | A speaker identity classifier with gradient reverse linear |
| Zac | The accent feature from Eac |
| Zt | The speaker feature from Es |
| Zc | The speaker feature from Ec |
| Zac | The accent feature from Ac |
| I | The unbiased estimator for vCLUB (Cheng et al., 2020) |
",
+ "image_path": "429987b15999e2013a5110aa58a4499fce765d905720eba1421359d4779700c9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 67,
+ 428,
+ 291,
+ 481
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 428,
+ 291,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 428,
+ 291,
+ 481
+ ],
+ "type": "text",
+ "content": "the input mel-spectrogram and encourage the converter to fill in missing frames based on surrounding frames. These adjustments add some structure to the text and make it even more reader-friendly."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 67,
+ 481,
+ 291,
+ 631
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 481,
+ 291,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 481,
+ 291,
+ 631
+ ],
+ "type": "text",
+ "content": "This paper applies the non-parallel database-based voice conversion model MaskCycleGAN-VC (Kaneko et al., 2021) to a more challenging task: voice and accent joint conversion. The source speaker's accent can be converted to the target speaker's accent without changing the linguistic content of the speech. We improve the generator part of the MaskCycleGAN-VC (Kaneko et al., 2021) for specific data and application scenarios. The comprehensive list of the primary symbols used throughout this paper is presented in Table 3."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 286,
+ 780,
+ 309,
+ 791
+ ],
+ "type": "text",
+ "content": "9298"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 10
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2023/Normal-Abnormal Decoupling Memory for Medical Report Generation/b90e4221-98d5-45c3-a7ce-3e01b0ad4ff9_content_list.json b/2023/Normal-Abnormal Decoupling Memory for Medical Report Generation/b90e4221-98d5-45c3-a7ce-3e01b0ad4ff9_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..789f21eb17b0dc40934c22b93eb226e15eace0c7
--- /dev/null
+++ b/2023/Normal-Abnormal Decoupling Memory for Medical Report Generation/b90e4221-98d5-45c3-a7ce-3e01b0ad4ff9_content_list.json
@@ -0,0 +1,2418 @@
+[
+ {
+ "type": "text",
+ "text": "Normal-Abnormal Decoupling Memory for Medical Report Generation",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 89,
+ 868,
+ 110
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Guosheng Zhao $^{1}$ , Yan Yan $^{2}$ , Zijian Zhao $^{1*}$",
+ "bbox": [
+ 319,
+ 135,
+ 685,
+ 153
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "$^{1}$ School of Control Science and Engineering, Shandong University",
+ "$^{2}$ Department of Computer Science, Illinois Institute of Technology."
+ ],
+ "bbox": [
+ 228,
+ 153,
+ 771,
+ 187
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "{zhgs,zhaozijian}@sdu.edu.cn, yyan34@iit.edu",
+ "bbox": [
+ 284,
+ 187,
+ 717,
+ 202
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 260,
+ 252,
+ 339,
+ 266
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "The automatic generation of medical reports plays a crucial role in clinical automation. In contrast to natural images, radiological images exhibit a high degree of similarity, while medical data are prone to data bias and complex noise, posing challenges for existing methods in capturing nuanced visual information. To address these challenges, we introduce a novel normal-abnormal semantic decoupling network that utilizes abnormal pattern memory. Different from directly optimizing the network using medical reports, we optimize visual extraction through the extraction of abnormal semantics from the reports. Moreover, we independently learn normal semantics based on abnormal semantics, ensuring that the optimization of the visual network remains unaffected by normal semantics learning. Then, we divided the words in the report into four parts: normal/abnormal sentences and normal/abnormal semantics, optimizing the network with distinct weights for each partition. The two semantic components, along with visual information, are seamlessly integrated to facilitate the generation of precise and coherent reports. This approach mitigates the impact of noisy normal semantics and reports. Moreover, we develop a novel encoder for abnormal pattern memory, which improves the network's ability to detect anomalies by capturing and embedding the abnormal patterns of images in the visual encoder. This approach demonstrates excellent performance on the benchmark MIMIC-CXR, surpassing the current state-of-the-art methods.