diff --git "a/parse/train/N6JECD-PI5w/N6JECD-PI5w_middle.json" "b/parse/train/N6JECD-PI5w/N6JECD-PI5w_middle.json" new file mode 100644--- /dev/null +++ "b/parse/train/N6JECD-PI5w/N6JECD-PI5w_middle.json" @@ -0,0 +1,36326 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 443, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 445, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 445, + 97 + ], + "score": 1.0, + "content": "FAIRFIL: CONTRASTIVE NEURAL DEBIASING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 441, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 441, + 117 + ], + "score": 1.0, + "content": "METHOD FOR PRETRAINED TEXT ENCODERS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 493, + 169 + ], + "lines": [ + { + "bbox": [ + 111, + 135, + 428, + 148 + ], + "spans": [ + { + "bbox": [ + 111, + 135, + 428, + 148 + ], + "score": 1.0, + "content": "Pengyu Cheng∗ , Weituo Hao∗, Siyang Yuan , Shijing Si , Lawrence Carin", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 145, + 394, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 394, + 159 + ], + "score": 1.0, + "content": "Department of Electrical and Computer Engineering, Duke University", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 156, + 495, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 495, + 170 + ], + "score": 1.0, + "content": "{pengyu.cheng,weituo.hao,siyang.yuan,shijing.si,lcarin}@duke.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 277, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 197, + 335, + 210 + ], + "spans": [ + { + "bbox": [ + 276, + 197, + 335, + 210 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 218, + 468, + 384 + ], + "lines": [ + { + "bbox": [ + 141, + 218, + 469, + 231 + ], + "spans": [ + { + "bbox": [ + 141, + 218, + 469, + 231 + ], + "score": 1.0, + "content": "Pretrained text encoders, such as BERT, have been applied increasingly in various", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 230, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 141, + 230, + 469, + 243 + ], + "score": 1.0, + "content": "natural language processing (NLP) tasks, and have recently demonstrated signifi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 241, + 470, + 253 + ], + "spans": [ + { + "bbox": [ + 141, + 241, + 470, + 253 + ], + "score": 1.0, + "content": "cant performance gains. However, recent studies have demonstrated the existence", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 252, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 142, + 252, + 469, + 264 + ], + "score": 1.0, + "content": "of social bias in these pretrained NLP models. Although prior works have made", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "progress on word-level debiasing, improved sentence-level fairness of pretrained", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 273, + 470, + 286 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 470, + 286 + ], + "score": 1.0, + "content": "encoders still lacks exploration. In this paper, we proposed the first neural debi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 285, + 470, + 296 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 470, + 296 + ], + "score": 1.0, + "content": "asing method for a pretrained sentence encoder, which transforms the pretrained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 295, + 470, + 308 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 470, + 308 + ], + "score": 1.0, + "content": "encoder outputs into debiased representations via a fair filter (FairFil) network. To", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 307, + 470, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 470, + 319 + ], + "score": 1.0, + "content": "learn the FairFil, we introduce a contrastive learning framework that not only min-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 317, + 469, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 317, + 469, + 330 + ], + "score": 1.0, + "content": "imizes the correlation between filtered embeddings and bias words but also pre-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 329, + 469, + 340 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 469, + 340 + ], + "score": 1.0, + "content": "serves rich semantic information of the original sentences. On real-world datasets,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "score": 1.0, + "content": "our FairFil effectively reduces the bias degree of pretrained text encoders, while", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 351, + 469, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 351, + 469, + 362 + ], + "score": 1.0, + "content": "continuously showing desirable performance on downstream tasks. Moreover, our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 362, + 469, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 362, + 469, + 374 + ], + "score": 1.0, + "content": "post hoc method does not require any retraining of the text encoders, further en-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 373, + 283, + 385 + ], + "spans": [ + { + "bbox": [ + 142, + 373, + 283, + 385 + ], + "score": 1.0, + "content": "larging FairFil’s application space.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 406, + 206, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 208, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 208, + 421 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "score": 1.0, + "content": "Text encoders, which map raw-text data into low-dimensional embeddings, have become one of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "the fundamental tools for extensive tasks in natural language processing (Kiros et al., 2015; Lin", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "et al., 2017; Shen et al., 2019; Cheng et al., 2020b). With the development of deep learning, large-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 504, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 504, + 472 + ], + "score": 1.0, + "content": "scale neural sentence encoders pretrained on massive text corpora, such as Infersent (Conneau et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "score": 1.0, + "content": "2017), ELMo (Peters et al., 2018), BERT (Devlin et al., 2019), and GPT (Radford et al., 2018),", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "score": 1.0, + "content": "have become the mainstream to extract the sentence-level text representations, and have shown", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "desirable performance on many NLP downstream tasks (MacAvaney et al., 2019; Sun et al., 2019;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "score": 1.0, + "content": "Zhang et al., 2019). Although these pretrained models have been studied comprehensively from", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "many perspectives, such as performance (Joshi et al., 2020), efficiency (Sanh et al., 2019), and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "robustness (Liu et al., 2019), the fairness of pretrained text encoders has not received significant", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 537, + 182, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 182, + 548 + ], + "score": 1.0, + "content": "research attention.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "The fairness issue is also broadly recognized as social bias, which denotes the unbalanced model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "behaviors with respect to some socially sensitive topics, such as gender, race, and religion (Liang", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "et al., 2020). For data-driven NLP models, social bias is an intrinsic problem mainly caused by the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "score": 1.0, + "content": "unbalanced data of text corpora (Bolukbasi et al., 2016). To quantitatively measure the bias degree", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "of models, prior work proposed several statistical tests (Caliskan et al., 2017; Chaloner & Maldon-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "ado, 2019; Brunet et al., 2019), mostly focusing on word-level embedding models. To evaluate the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "sentence-level bias in the embedding space, May et al. (2019) extended the Word Embedding As-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "score": 1.0, + "content": "sociation Test (WEAT) (Caliskan et al., 2017) into a Sentence Encoder Association Test (SEAT).", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Based on the SEAT test, May et al. (2019) claimed the existence of social bias in the pretrained", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 653, + 184, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 184, + 663 + ], + "score": 1.0, + "content": "sentence encoders.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "Although related works have discussed the measurement of social bias in sentence embeddings,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "score": 1.0, + "content": "debiasing pretrained sentence encoders remains a challenge. Previous word embedding debiasing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 692, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 702 + ], + "score": 1.0, + "content": "methods (Bolukbasi et al., 2016; Kaneko & Bollegala, 2019; Manzini et al., 2019) have limited as-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 702, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 713 + ], + "score": 1.0, + "content": "sistance to sentence-level debiasing, because even if the social bias is eliminated at the word level,", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 722, + 192, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 194, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 194, + 733 + ], + "score": 1.0, + "content": "∗Equal contribution.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 443, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 445, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 445, + 97 + ], + "score": 1.0, + "content": "FAIRFIL: CONTRASTIVE NEURAL DEBIASING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 441, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 441, + 117 + ], + "score": 1.0, + "content": "METHOD FOR PRETRAINED TEXT ENCODERS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 493, + 169 + ], + "lines": [ + { + "bbox": [ + 111, + 135, + 428, + 148 + ], + "spans": [ + { + "bbox": [ + 111, + 135, + 428, + 148 + ], + "score": 1.0, + "content": "Pengyu Cheng∗ , Weituo Hao∗, Siyang Yuan , Shijing Si , Lawrence Carin", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 145, + 394, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 394, + 159 + ], + "score": 1.0, + "content": "Department of Electrical and Computer Engineering, Duke University", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 156, + 495, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 495, + 170 + ], + "score": 1.0, + "content": "{pengyu.cheng,weituo.hao,siyang.yuan,shijing.si,lcarin}@duke.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 111, + 135, + 495, + 170 + ] + }, + { + "type": "title", + "bbox": [ + 277, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 197, + 335, + 210 + ], + "spans": [ + { + "bbox": [ + 276, + 197, + 335, + 210 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 218, + 468, + 384 + ], + "lines": [ + { + "bbox": [ + 141, + 218, + 469, + 231 + ], + "spans": [ + { + "bbox": [ + 141, + 218, + 469, + 231 + ], + "score": 1.0, + "content": "Pretrained text encoders, such as BERT, have been applied increasingly in various", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 230, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 141, + 230, + 469, + 243 + ], + "score": 1.0, + "content": "natural language processing (NLP) tasks, and have recently demonstrated signifi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 241, + 470, + 253 + ], + "spans": [ + { + "bbox": [ + 141, + 241, + 470, + 253 + ], + "score": 1.0, + "content": "cant performance gains. However, recent studies have demonstrated the existence", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 252, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 142, + 252, + 469, + 264 + ], + "score": 1.0, + "content": "of social bias in these pretrained NLP models. Although prior works have made", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "progress on word-level debiasing, improved sentence-level fairness of pretrained", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 273, + 470, + 286 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 470, + 286 + ], + "score": 1.0, + "content": "encoders still lacks exploration. In this paper, we proposed the first neural debi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 285, + 470, + 296 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 470, + 296 + ], + "score": 1.0, + "content": "asing method for a pretrained sentence encoder, which transforms the pretrained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 295, + 470, + 308 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 470, + 308 + ], + "score": 1.0, + "content": "encoder outputs into debiased representations via a fair filter (FairFil) network. To", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 307, + 470, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 470, + 319 + ], + "score": 1.0, + "content": "learn the FairFil, we introduce a contrastive learning framework that not only min-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 317, + 469, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 317, + 469, + 330 + ], + "score": 1.0, + "content": "imizes the correlation between filtered embeddings and bias words but also pre-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 329, + 469, + 340 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 469, + 340 + ], + "score": 1.0, + "content": "serves rich semantic information of the original sentences. On real-world datasets,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 339, + 470, + 352 + ], + "score": 1.0, + "content": "our FairFil effectively reduces the bias degree of pretrained text encoders, while", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 351, + 469, + 362 + ], + "spans": [ + { + "bbox": [ + 141, + 351, + 469, + 362 + ], + "score": 1.0, + "content": "continuously showing desirable performance on downstream tasks. Moreover, our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 362, + 469, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 362, + 469, + 374 + ], + "score": 1.0, + "content": "post hoc method does not require any retraining of the text encoders, further en-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 373, + 283, + 385 + ], + "spans": [ + { + "bbox": [ + 142, + 373, + 283, + 385 + ], + "score": 1.0, + "content": "larging FairFil’s application space.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13, + "bbox_fs": [ + 141, + 218, + 470, + 385 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 406, + 206, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 208, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 208, + 421 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "score": 1.0, + "content": "Text encoders, which map raw-text data into low-dimensional embeddings, have become one of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "the fundamental tools for extensive tasks in natural language processing (Kiros et al., 2015; Lin", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "et al., 2017; Shen et al., 2019; Cheng et al., 2020b). With the development of deep learning, large-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 504, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 504, + 472 + ], + "score": 1.0, + "content": "scale neural sentence encoders pretrained on massive text corpora, such as Infersent (Conneau et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "score": 1.0, + "content": "2017), ELMo (Peters et al., 2018), BERT (Devlin et al., 2019), and GPT (Radford et al., 2018),", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 494 + ], + "score": 1.0, + "content": "have become the mainstream to extract the sentence-level text representations, and have shown", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "desirable performance on many NLP downstream tasks (MacAvaney et al., 2019; Sun et al., 2019;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 504, + 516 + ], + "score": 1.0, + "content": "Zhang et al., 2019). Although these pretrained models have been studied comprehensively from", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "many perspectives, such as performance (Joshi et al., 2020), efficiency (Sanh et al., 2019), and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "robustness (Liu et al., 2019), the fairness of pretrained text encoders has not received significant", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 537, + 182, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 182, + 548 + ], + "score": 1.0, + "content": "research attention.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 428, + 506, + 548 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "The fairness issue is also broadly recognized as social bias, which denotes the unbalanced model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "behaviors with respect to some socially sensitive topics, such as gender, race, and religion (Liang", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "et al., 2020). For data-driven NLP models, social bias is an intrinsic problem mainly caused by the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "score": 1.0, + "content": "unbalanced data of text corpora (Bolukbasi et al., 2016). To quantitatively measure the bias degree", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "of models, prior work proposed several statistical tests (Caliskan et al., 2017; Chaloner & Maldon-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "ado, 2019; Brunet et al., 2019), mostly focusing on word-level embedding models. To evaluate the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "sentence-level bias in the embedding space, May et al. (2019) extended the Word Embedding As-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "score": 1.0, + "content": "sociation Test (WEAT) (Caliskan et al., 2017) into a Sentence Encoder Association Test (SEAT).", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Based on the SEAT test, May et al. (2019) claimed the existence of social bias in the pretrained", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 653, + 184, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 184, + 663 + ], + "score": 1.0, + "content": "sentence encoders.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 553, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "Although related works have discussed the measurement of social bias in sentence embeddings,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "score": 1.0, + "content": "debiasing pretrained sentence encoders remains a challenge. Previous word embedding debiasing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 692, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 702 + ], + "score": 1.0, + "content": "methods (Bolukbasi et al., 2016; Kaneko & Bollegala, 2019; Manzini et al., 2019) have limited as-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 702, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 713 + ], + "score": 1.0, + "content": "sistance to sentence-level debiasing, because even if the social bias is eliminated at the word level,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "the sentence-level bias can still be caused by the unbalanced combination of words in the training", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "text. Besides, retraining a state-of-the-art sentence encoder for debiasing requires a massive amount", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "of computational resources, especially for large-scale deep models like BERT (Devlin et al., 2019)", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "and GPT (Radford et al., 2018). To the best of our knowledge, Liang et al. (2020) proposed the only", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "sentence-level debiasing method (Sent-Debias) for pretrained text encoders, in which the embed-", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "dings are revised by subtracting the latent biased direction vectors learned by Principal Component", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "Analysis (PCA) (Wold et al., 1987). However, Sent-Debias makes a strong assumption on the linear-", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 504, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 504, + 172 + ], + "score": 1.0, + "content": "ity of the bias in the sentence embedding space. Further, the calculation of bias directions depends", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "score": 1.0, + "content": "highly on the embeddings extracted from the training data and the number of principal components,", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 320, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 320, + 194 + ], + "score": 1.0, + "content": "preventing the method from adequate generalization.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 668, + 506, + 713 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 193 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "the sentence-level bias can still be caused by the unbalanced combination of words in the training", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "text. Besides, retraining a state-of-the-art sentence encoder for debiasing requires a massive amount", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "of computational resources, especially for large-scale deep models like BERT (Devlin et al., 2019)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "and GPT (Radford et al., 2018). To the best of our knowledge, Liang et al. (2020) proposed the only", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "sentence-level debiasing method (Sent-Debias) for pretrained text encoders, in which the embed-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "dings are revised by subtracting the latent biased direction vectors learned by Principal Component", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "Analysis (PCA) (Wold et al., 1987). However, Sent-Debias makes a strong assumption on the linear-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 504, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 504, + 172 + ], + "score": 1.0, + "content": "ity of the bias in the sentence embedding space. Further, the calculation of bias directions depends", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 505, + 184 + ], + "score": 1.0, + "content": "highly on the embeddings extracted from the training data and the number of principal components,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 320, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 320, + 194 + ], + "score": 1.0, + "content": "preventing the method from adequate generalization.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "In this paper, we proposed the first neural debiasing method for pretrained sentence encoders. For", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "a given pretrained encoder, our method learns a fair filter (FairFil) network, whose inputs are the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "original embeddings of the encoder, and outputs are the debiased embeddings. Inspired by the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "multi-view contrastive learning (Chen et al., 2020), for each training sentence, we first generate an", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "augmentation that has the same semantic meaning but in a different potential bias direction. We con-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "trastively train our FairFil by maximizing the mutual information between the debiased embeddings", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "score": 1.0, + "content": "of the original sentences and corresponding augmentations. To further eliminate bias from sensitive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 504, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 504, + 287 + ], + "score": 1.0, + "content": "words in sentences, we introduce a debiasing regularizer, which minimizes the mutual information", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "between debiased embeddings and the sensitive words’ embeddings. In the experiments, our Fair-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Fil outperforms Sent-Debias (Liang et al., 2020) in terms of the fairness and the representativeness", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "of debiased embeddings, indicating our FairFil not only effectively reduces the social bias in the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 482, + 332 + ], + "score": 1.0, + "content": "sentence embeddings, but also successfully preserves the rich semantic meaning of input text.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 208, + 354 + ], + "lines": [ + { + "bbox": [ + 104, + 340, + 209, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 209, + 356 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 503, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Mutual Information (MI) is a measure of the “amount of information” between two variables (Kull-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 372, + 308, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 308, + 384 + ], + "score": 1.0, + "content": "back, 1997). The mathematical definition of MI is", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 385, + 379, + 412 + ], + "lines": [ + { + "bbox": [ + 230, + 385, + 379, + 412 + ], + "spans": [ + { + "bbox": [ + 230, + 385, + 379, + 412 + ], + "score": 0.95, + "content": "\\mathcal { T } ( \\pmb { x } ; \\pmb { y } ) : = \\mathbb { E } _ { p ( \\pmb { x } , \\pmb { y } ) } \\Big [ \\log \\frac { p ( \\pmb { x } , \\pmb { y } ) } { p ( \\pmb { x } ) p ( \\pmb { y } ) } \\Big ] ,", + "type": "interline_equation", + "image_path": "0cfd62ec28d0d26275b94cfcdfb84f2e15e838caa664f0f44b3619ef87293c3b.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 385, + 379, + 398.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 230, + 398.5, + 379, + 412.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 413, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 133, + 427 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 414, + 165, + 426 + ], + "score": 0.92, + "content": "p ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 412, + 330, + 427 + ], + "score": 1.0, + "content": "is the joint distribution of two variables", + "type": "text" + }, + { + "bbox": [ + 331, + 414, + 356, + 426 + ], + "score": 0.92, + "content": "( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 412, + 378, + 427 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 379, + 414, + 423, + 426 + ], + "score": 0.93, + "content": "p ( { \\pmb x } ) , p ( { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 412, + 505, + 427 + ], + "score": 1.0, + "content": "are respectively the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 211, + 437 + ], + "score": 1.0, + "content": "marginal distributions of", + "type": "text" + }, + { + "bbox": [ + 211, + 426, + 230, + 436 + ], + "score": 0.87, + "content": "\\mathbf { \\nabla } _ { \\mathbf { x } , \\mathbf { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 424, + 506, + 437 + ], + "score": 1.0, + "content": ". Recently, mutual information has achieved considerable success", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "score": 1.0, + "content": "when applied as a learning criterion in diverse deep learning tasks, such as conditional genera-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "tion (Chen et al., 2016), domain adaptation (Gholami et al., 2020), representation learning (Chen", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "et al., 2020), and fairness (Song et al., 2019). However, the calculation of exact MI in (1) is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 358, + 482 + ], + "score": 1.0, + "content": "well-recognized as a challenge, because the expectation w.r.t", + "type": "text" + }, + { + "bbox": [ + 359, + 469, + 389, + 481 + ], + "score": 0.91, + "content": "p ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "is always intractable, espe-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 234, + 492 + ], + "score": 1.0, + "content": "cially when only samples from", + "type": "text" + }, + { + "bbox": [ + 234, + 480, + 265, + 492 + ], + "score": 0.93, + "content": "p ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "are provided. To this end, several upper and lower bounds", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "have been introduced to estimate the MI with samples. For MI maximization tasks (Hjelm et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "2018; Chen et al., 2020), Oord et al. (2018) derived a powerful MI estimator, InfoNCE, based on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 294, + 524 + ], + "score": 1.0, + "content": "noise contrastive estimation (NCE) (Gutmann", + "type": "text" + }, + { + "bbox": [ + 294, + 513, + 303, + 523 + ], + "score": 0.31, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "Hyvarinen, 2010). Given a batch of sample pairs ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 521, + 463, + 538 + ], + "spans": [ + { + "bbox": [ + 107, + 523, + 162, + 536 + ], + "score": 0.93, + "content": "\\{ ( \\pmb { x } _ { i } , \\pmb { y } _ { i } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 521, + 426, + 538 + ], + "score": 1.0, + "content": ", the InfoNCE estimator is defined with a learnable score function", + "type": "text" + }, + { + "bbox": [ + 426, + 523, + 457, + 536 + ], + "score": 0.95, + "content": "f ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 521, + 463, + 538 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32 + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 537, + 398, + 572 + ], + "lines": [ + { + "bbox": [ + 212, + 537, + 398, + 572 + ], + "spans": [ + { + "bbox": [ + 212, + 537, + 398, + 572 + ], + "score": 0.94, + "content": "\\mathcal { T } _ { \\mathrm { N C E } } : = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log \\frac { \\exp ( f ( { \\boldsymbol { { x } } } _ { i } , { \\boldsymbol { { y } } } _ { i } ) ) } { \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\exp ( f ( { \\boldsymbol { { x } } } _ { i } , { \\boldsymbol { { y } } } _ { j } ) ) } .", + "type": "interline_equation", + "image_path": "a1ae7f9599b66fe24c9b744111a316f5b9aaa17637b4554fad5e55747db8294b.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 537, + 398, + 554.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 212, + 554.5, + 398, + 572.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "For MI minimization tasks (Alemi et al., 2017; Song et al., 2019), Cheng et al. (2020a) introduced a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 431, + 596 + ], + "score": 1.0, + "content": "contrastive log-ratio upper bound (CLUB) based on a variational approximation", + "type": "text" + }, + { + "bbox": [ + 431, + 584, + 464, + 596 + ], + "score": 0.93, + "content": "q _ { \\boldsymbol { \\theta } } ( \\mathbf { \\boldsymbol { y } } | \\mathbf { \\boldsymbol { x } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "of condi-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 595, + 214, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 180, + 608 + ], + "score": 1.0, + "content": "tional distribution", + "type": "text" + }, + { + "bbox": [ + 180, + 595, + 209, + 607 + ], + "score": 0.93, + "content": "p ( \\pmb { y } | \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 595, + 214, + 608 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 608, + 423, + 644 + ], + "lines": [ + { + "bbox": [ + 187, + 608, + 423, + 644 + ], + "spans": [ + { + "bbox": [ + 187, + 608, + 423, + 644 + ], + "score": 0.95, + "content": "\\mathcal { T } _ { \\mathrm { C L U B } } : = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\bigg [ \\log q _ { \\theta } ( \\pmb { y } _ { i } | \\pmb { x } _ { i } ) - \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\log q _ { \\theta } ( \\pmb { y } _ { j } | \\pmb { x } _ { i } ) \\bigg ] .", + "type": "interline_equation", + "image_path": "7161a834b538a32bc3e6524fe9c0b125d82d294f47547adabb0eec1bc5b7eb53.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 608, + 423, + 626.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 187, + 626.0, + 423, + 644.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 645, + 504, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 658 + ], + "score": 1.0, + "content": "In the following, we use the above two MI estimators to induce the sentence encoder, eliminating", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 655, + 469, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 469, + 669 + ], + "score": 1.0, + "content": "the biased information and preserving the semantic information from the original raw text.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + }, + { + "type": "title", + "bbox": [ + 107, + 678, + 173, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 174, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 174, + 694 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 142, + 712 + ], + "score": 1.0, + "content": "Suppose", + "type": "text" + }, + { + "bbox": [ + 143, + 699, + 162, + 711 + ], + "score": 0.9, + "content": "E ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 698, + 409, + 712 + ], + "score": 1.0, + "content": "is a pretrained sentence encoder, which can encode a sentence", + "type": "text" + }, + { + "bbox": [ + 410, + 701, + 417, + 709 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "into low-dimensional", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 153, + 723 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 154, + 710, + 197, + 721 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\dot { \\boldsymbol { z } } = \\boldsymbol { E } ( \\boldsymbol { x } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 708, + 262, + 723 + ], + "score": 1.0, + "content": ". Each sentence", + "type": "text" + }, + { + "bbox": [ + 262, + 709, + 355, + 721 + ], + "score": 0.91, + "content": "\\pmb { x } = ( w ^ { 1 } , w ^ { 2 } , \\dots , w ^ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "is a sequence of words. The embed-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 161, + 733 + ], + "score": 1.0, + "content": "ding space of", + "type": "text" + }, + { + "bbox": [ + 161, + 723, + 168, + 730 + ], + "score": 0.78, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "has been recognized to have social bias in a series of studies (May et al., 2019; Kurita", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 193 + ], + "lines": [], + "index": 4.5, + "bbox_fs": [ + 105, + 81, + 506, + 194 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "In this paper, we proposed the first neural debiasing method for pretrained sentence encoders. For", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "a given pretrained encoder, our method learns a fair filter (FairFil) network, whose inputs are the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "original embeddings of the encoder, and outputs are the debiased embeddings. Inspired by the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "multi-view contrastive learning (Chen et al., 2020), for each training sentence, we first generate an", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "augmentation that has the same semantic meaning but in a different potential bias direction. We con-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "trastively train our FairFil by maximizing the mutual information between the debiased embeddings", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "score": 1.0, + "content": "of the original sentences and corresponding augmentations. To further eliminate bias from sensitive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 504, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 504, + 287 + ], + "score": 1.0, + "content": "words in sentences, we introduce a debiasing regularizer, which minimizes the mutual information", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "between debiased embeddings and the sensitive words’ embeddings. In the experiments, our Fair-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Fil outperforms Sent-Debias (Liang et al., 2020) in terms of the fairness and the representativeness", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "of debiased embeddings, indicating our FairFil not only effectively reduces the social bias in the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 482, + 332 + ], + "score": 1.0, + "content": "sentence embeddings, but also successfully preserves the rich semantic meaning of input text.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 198, + 506, + 332 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 208, + 354 + ], + "lines": [ + { + "bbox": [ + 104, + 340, + 209, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 209, + 356 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 503, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Mutual Information (MI) is a measure of the “amount of information” between two variables (Kull-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 372, + 308, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 308, + 384 + ], + "score": 1.0, + "content": "back, 1997). The mathematical definition of MI is", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 361, + 505, + 384 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 385, + 379, + 412 + ], + "lines": [ + { + "bbox": [ + 230, + 385, + 379, + 412 + ], + "spans": [ + { + "bbox": [ + 230, + 385, + 379, + 412 + ], + "score": 0.95, + "content": "\\mathcal { T } ( \\pmb { x } ; \\pmb { y } ) : = \\mathbb { E } _ { p ( \\pmb { x } , \\pmb { y } ) } \\Big [ \\log \\frac { p ( \\pmb { x } , \\pmb { y } ) } { p ( \\pmb { x } ) p ( \\pmb { y } ) } \\Big ] ,", + "type": "interline_equation", + "image_path": "0cfd62ec28d0d26275b94cfcdfb84f2e15e838caa664f0f44b3619ef87293c3b.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 385, + 379, + 398.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 230, + 398.5, + 379, + 412.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 413, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 133, + 427 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 414, + 165, + 426 + ], + "score": 0.92, + "content": "p ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 412, + 330, + 427 + ], + "score": 1.0, + "content": "is the joint distribution of two variables", + "type": "text" + }, + { + "bbox": [ + 331, + 414, + 356, + 426 + ], + "score": 0.92, + "content": "( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 412, + 378, + 427 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 379, + 414, + 423, + 426 + ], + "score": 0.93, + "content": "p ( { \\pmb x } ) , p ( { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 412, + 505, + 427 + ], + "score": 1.0, + "content": "are respectively the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 211, + 437 + ], + "score": 1.0, + "content": "marginal distributions of", + "type": "text" + }, + { + "bbox": [ + 211, + 426, + 230, + 436 + ], + "score": 0.87, + "content": "\\mathbf { \\nabla } _ { \\mathbf { x } , \\mathbf { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 424, + 506, + 437 + ], + "score": 1.0, + "content": ". Recently, mutual information has achieved considerable success", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "score": 1.0, + "content": "when applied as a learning criterion in diverse deep learning tasks, such as conditional genera-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "tion (Chen et al., 2016), domain adaptation (Gholami et al., 2020), representation learning (Chen", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "et al., 2020), and fairness (Song et al., 2019). However, the calculation of exact MI in (1) is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 358, + 482 + ], + "score": 1.0, + "content": "well-recognized as a challenge, because the expectation w.r.t", + "type": "text" + }, + { + "bbox": [ + 359, + 469, + 389, + 481 + ], + "score": 0.91, + "content": "p ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "is always intractable, espe-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 234, + 492 + ], + "score": 1.0, + "content": "cially when only samples from", + "type": "text" + }, + { + "bbox": [ + 234, + 480, + 265, + 492 + ], + "score": 0.93, + "content": "p ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "are provided. To this end, several upper and lower bounds", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "have been introduced to estimate the MI with samples. For MI maximization tasks (Hjelm et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "2018; Chen et al., 2020), Oord et al. (2018) derived a powerful MI estimator, InfoNCE, based on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 294, + 524 + ], + "score": 1.0, + "content": "noise contrastive estimation (NCE) (Gutmann", + "type": "text" + }, + { + "bbox": [ + 294, + 513, + 303, + 523 + ], + "score": 0.31, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "Hyvarinen, 2010). Given a batch of sample pairs ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 521, + 463, + 538 + ], + "spans": [ + { + "bbox": [ + 107, + 523, + 162, + 536 + ], + "score": 0.93, + "content": "\\{ ( \\pmb { x } _ { i } , \\pmb { y } _ { i } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 521, + 426, + 538 + ], + "score": 1.0, + "content": ", the InfoNCE estimator is defined with a learnable score function", + "type": "text" + }, + { + "bbox": [ + 426, + 523, + 457, + 536 + ], + "score": 0.95, + "content": "f ( { \\pmb x } , { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 521, + 463, + 538 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 412, + 506, + 538 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 537, + 398, + 572 + ], + "lines": [ + { + "bbox": [ + 212, + 537, + 398, + 572 + ], + "spans": [ + { + "bbox": [ + 212, + 537, + 398, + 572 + ], + "score": 0.94, + "content": "\\mathcal { T } _ { \\mathrm { N C E } } : = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log \\frac { \\exp ( f ( { \\boldsymbol { { x } } } _ { i } , { \\boldsymbol { { y } } } _ { i } ) ) } { \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\exp ( f ( { \\boldsymbol { { x } } } _ { i } , { \\boldsymbol { { y } } } _ { j } ) ) } .", + "type": "interline_equation", + "image_path": "a1ae7f9599b66fe24c9b744111a316f5b9aaa17637b4554fad5e55747db8294b.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 537, + 398, + 554.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 212, + 554.5, + 398, + 572.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "For MI minimization tasks (Alemi et al., 2017; Song et al., 2019), Cheng et al. (2020a) introduced a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 431, + 596 + ], + "score": 1.0, + "content": "contrastive log-ratio upper bound (CLUB) based on a variational approximation", + "type": "text" + }, + { + "bbox": [ + 431, + 584, + 464, + 596 + ], + "score": 0.93, + "content": "q _ { \\boldsymbol { \\theta } } ( \\mathbf { \\boldsymbol { y } } | \\mathbf { \\boldsymbol { x } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "of condi-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 595, + 214, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 180, + 608 + ], + "score": 1.0, + "content": "tional distribution", + "type": "text" + }, + { + "bbox": [ + 180, + 595, + 209, + 607 + ], + "score": 0.93, + "content": "p ( \\pmb { y } | \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 595, + 214, + 608 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 572, + 506, + 608 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 608, + 423, + 644 + ], + "lines": [ + { + "bbox": [ + 187, + 608, + 423, + 644 + ], + "spans": [ + { + "bbox": [ + 187, + 608, + 423, + 644 + ], + "score": 0.95, + "content": "\\mathcal { T } _ { \\mathrm { C L U B } } : = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\bigg [ \\log q _ { \\theta } ( \\pmb { y } _ { i } | \\pmb { x } _ { i } ) - \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\log q _ { \\theta } ( \\pmb { y } _ { j } | \\pmb { x } _ { i } ) \\bigg ] .", + "type": "interline_equation", + "image_path": "7161a834b538a32bc3e6524fe9c0b125d82d294f47547adabb0eec1bc5b7eb53.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 608, + 423, + 626.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 187, + 626.0, + 423, + 644.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 645, + 504, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 658 + ], + "score": 1.0, + "content": "In the following, we use the above two MI estimators to induce the sentence encoder, eliminating", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 655, + 469, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 469, + 669 + ], + "score": 1.0, + "content": "the biased information and preserving the semantic information from the original raw text.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 643, + 505, + 669 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 678, + 173, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 174, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 174, + 694 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 142, + 712 + ], + "score": 1.0, + "content": "Suppose", + "type": "text" + }, + { + "bbox": [ + 143, + 699, + 162, + 711 + ], + "score": 0.9, + "content": "E ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 698, + 409, + 712 + ], + "score": 1.0, + "content": "is a pretrained sentence encoder, which can encode a sentence", + "type": "text" + }, + { + "bbox": [ + 410, + 701, + 417, + 709 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "into low-dimensional", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 153, + 723 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 154, + 710, + 197, + 721 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\dot { \\boldsymbol { z } } = \\boldsymbol { E } ( \\boldsymbol { x } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 708, + 262, + 723 + ], + "score": 1.0, + "content": ". Each sentence", + "type": "text" + }, + { + "bbox": [ + 262, + 709, + 355, + 721 + ], + "score": 0.91, + "content": "\\pmb { x } = ( w ^ { 1 } , w ^ { 2 } , \\dots , w ^ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "is a sequence of words. The embed-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 161, + 733 + ], + "score": 1.0, + "content": "ding space of", + "type": "text" + }, + { + "bbox": [ + 161, + 723, + 168, + 730 + ], + "score": 0.78, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "has been recognized to have social bias in a series of studies (May et al., 2019; Kurita", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "et al., 2019; Liang et al., 2020). To eliminate the social bias in the embedding space, we aim to learn", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 161, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 189, + 173 + ], + "score": 1.0, + "content": "a fair filter network", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 190, + 161, + 207, + 173 + ], + "score": 0.89, + "content": "f ( \\cdot )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 207, + 161, + 338, + 173 + ], + "score": 1.0, + "content": "on top of the sentence encoder", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 338, + 161, + 357, + 173 + ], + "score": 0.9, + "content": "E ( \\cdot )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 358, + 161, + 505, + 173 + ], + "score": 1.0, + "content": ", such that the output embedding of", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 160, + 185 + ], + "score": 1.0, + "content": "our fair filter", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 161, + 172, + 201, + 184 + ], + "score": 0.93, + "content": "\\begin{array} { r } { d = f ( z ) } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 202, + 171, + 506, + 185 + ], + "score": 1.0, + "content": "can be debiased. To train the fair filter, we design a multi-view contrastive", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 430, + 196 + ], + "score": 1.0, + "content": "learning framework, which consists of three steps. First, for each input sentence", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 430, + 185, + 438, + 193 + ], + "score": 0.66, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 438, + 182, + 505, + 196 + ], + "score": 1.0, + "content": ", we generate an", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 189, + 207 + ], + "score": 1.0, + "content": "augmented sentence", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 190, + 194, + 200, + 204 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 200, + 194, + 357, + 207 + ], + "score": 1.0, + "content": "that has the same semantic meaning as", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 358, + 196, + 366, + 204 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 366, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "but in a different potential bias di-", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 446, + 217 + ], + "score": 1.0, + "content": "rection. Then, we maximize the mutual information between the original embedding", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 446, + 205, + 487, + 217 + ], + "score": 0.93, + "content": "z = f ( { \\pmb x } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 487, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 215, + 228 + ], + "score": 1.0, + "content": "the augmented embedding", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 215, + 216, + 263, + 228 + ], + "score": 0.93, + "content": "z ^ { \\prime } = f ( x ^ { \\prime } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 263, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "with the InfoNCE (Oord et al., 2018) contrastive loss. Fur-", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 442, + 239 + ], + "score": 1.0, + "content": "ther, we design a debiasing regularizer to minimize the mutual information between", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 443, + 227, + 450, + 237 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 451, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "and sensitive", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 411, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 179, + 250 + ], + "score": 1.0, + "content": "attribute words in", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 179, + 240, + 187, + 248 + ], + "score": 0.69, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 187, + 237, + 411, + 250 + ], + "score": 1.0, + "content": ". In the following, we discuss these three steps in detail.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 49, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 97, + 504, + 135 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 81, + 494, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 78, + 496, + 94 + ], + "spans": [ + { + "bbox": [ + 114, + 78, + 496, + 94 + ], + "score": 1.0, + "content": "Table 1: Examples of generating an augmentation sentence under the sensitive topic “gender”.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 109, + 97, + 504, + 135 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 97, + 504, + 135 + ], + "spans": [ + { + "bbox": [ + 109, + 97, + 504, + 135 + ], + "score": 0.923, + "html": "
Bias directionSensitive Attribute wordsText content
Originalmalehe,his{He} is good at playing {his} basketball.
Augmentationfemaleshe,her{She{She} is good at playing {her} basketball.
", + "type": "table", + "image_path": "e51de30e34c044121ed00266c3dc4486583b6a350b45ba45fbe871497c37818f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 109, + 97, + 504, + 109.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 109.66666666666667, + 504, + 122.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 122.33333333333334, + 504, + 135.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 505, + 249 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "et al., 2019; Liang et al., 2020). To eliminate the social bias in the embedding space, we aim to learn", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 161, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 189, + 173 + ], + "score": 1.0, + "content": "a fair filter network", + "type": "text" + }, + { + "bbox": [ + 190, + 161, + 207, + 173 + ], + "score": 0.89, + "content": "f ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 161, + 338, + 173 + ], + "score": 1.0, + "content": "on top of the sentence encoder", + "type": "text" + }, + { + "bbox": [ + 338, + 161, + 357, + 173 + ], + "score": 0.9, + "content": "E ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 161, + 505, + 173 + ], + "score": 1.0, + "content": ", such that the output embedding of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 160, + 185 + ], + "score": 1.0, + "content": "our fair filter", + "type": "text" + }, + { + "bbox": [ + 161, + 172, + 201, + 184 + ], + "score": 0.93, + "content": "\\begin{array} { r } { d = f ( z ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 171, + 506, + 185 + ], + "score": 1.0, + "content": "can be debiased. To train the fair filter, we design a multi-view contrastive", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 430, + 196 + ], + "score": 1.0, + "content": "learning framework, which consists of three steps. First, for each input sentence", + "type": "text" + }, + { + "bbox": [ + 430, + 185, + 438, + 193 + ], + "score": 0.66, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 182, + 505, + 196 + ], + "score": 1.0, + "content": ", we generate an", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 189, + 207 + ], + "score": 1.0, + "content": "augmented sentence", + "type": "text" + }, + { + "bbox": [ + 190, + 194, + 200, + 204 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 194, + 357, + 207 + ], + "score": 1.0, + "content": "that has the same semantic meaning as", + "type": "text" + }, + { + "bbox": [ + 358, + 196, + 366, + 204 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "but in a different potential bias di-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 446, + 217 + ], + "score": 1.0, + "content": "rection. Then, we maximize the mutual information between the original embedding", + "type": "text" + }, + { + "bbox": [ + 446, + 205, + 487, + 217 + ], + "score": 0.93, + "content": "z = f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 215, + 228 + ], + "score": 1.0, + "content": "the augmented embedding", + "type": "text" + }, + { + "bbox": [ + 215, + 216, + 263, + 228 + ], + "score": 0.93, + "content": "z ^ { \\prime } = f ( x ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "with the InfoNCE (Oord et al., 2018) contrastive loss. Fur-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 442, + 239 + ], + "score": 1.0, + "content": "ther, we design a debiasing regularizer to minimize the mutual information between", + "type": "text" + }, + { + "bbox": [ + 443, + 227, + 450, + 237 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "and sensitive", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 411, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 179, + 250 + ], + "score": 1.0, + "content": "attribute words in", + "type": "text" + }, + { + "bbox": [ + 179, + 240, + 187, + 248 + ], + "score": 0.69, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 237, + 411, + 250 + ], + "score": 1.0, + "content": ". In the following, we discuss these three steps in detail.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 106, + 260, + 363, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 364, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 364, + 273 + ], + "score": 1.0, + "content": "3.1 DATA AUGMENTATIONS WITH SENSITIVE ATTRIBUTES", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 505, + 344 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "We first describe the sentence data augmentation process for our FairFil contrastive learning. De-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 234, + 301 + ], + "score": 1.0, + "content": "note a social sensitive topic as", + "type": "text" + }, + { + "bbox": [ + 234, + 288, + 335, + 300 + ], + "score": 0.92, + "content": "\\mathcal { T } = \\{ \\mathcal { D } _ { 1 } , \\mathcal { D } _ { 2 } , \\ldots , \\mathcal { D } _ { K } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 288, + 367, + 301 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 367, + 288, + 381, + 299 + ], + "score": 0.79, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 288, + 449, + 299 + ], + "score": 0.69, + "content": "( k = 1 , \\ldots , K )", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "is one of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 346, + 313 + ], + "score": 1.0, + "content": "potential bias directions under the topic. For example, if", + "type": "text" + }, + { + "bbox": [ + 346, + 300, + 356, + 309 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 298, + 506, + 313 + ], + "score": 1.0, + "content": "represents the sensitive topic “gen-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 151, + 323 + ], + "score": 1.0, + "content": "der”, then", + "type": "text" + }, + { + "bbox": [ + 152, + 310, + 162, + 320 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 309, + 317, + 323 + ], + "score": 1.0, + "content": "consists two potential bias directions", + "type": "text" + }, + { + "bbox": [ + 318, + 310, + 456, + 322 + ], + "score": 0.81, + "content": "\\{ \\mathcal { D } _ { 1 } , \\mathcal { D } _ { 2 } \\} \\stackrel { - } { = } \\{ { } ^ { } m a l e ^ { , \\prime } , \\ { } ^ { } f e m a l e ^ { , \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 309, + 505, + 323 + ], + "score": 1.0, + "content": ". Similarly,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 504, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 117, + 334 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 117, + 321, + 127, + 331 + ], + "score": 0.82, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 320, + 344, + 334 + ], + "score": 1.0, + "content": "is set as the major “religions” of the world, then", + "type": "text" + }, + { + "bbox": [ + 344, + 322, + 354, + 331 + ], + "score": 0.81, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 320, + 416, + 334 + ], + "score": 1.0, + "content": "could contain", + "type": "text" + }, + { + "bbox": [ + 417, + 321, + 504, + 333 + ], + "score": 0.91, + "content": "\\left\\{ \\mathcal { D } _ { 1 } , \\mathcal { D } _ { 2 } , \\mathcal { D } _ { 3 } , \\mathcal { D } _ { 4 } \\right\\} \\ =", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 331, + 404, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 404, + 344 + ], + "score": 1.0, + "content": "{“Christianity”, “Islam”, “Judaism”, “Buddhism”} as four components.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 241, + 362 + ], + "score": 1.0, + "content": "For a given social sensitive topic", + "type": "text" + }, + { + "bbox": [ + 241, + 348, + 320, + 361 + ], + "score": 0.92, + "content": "\\mathcal { T } = \\{ { D } _ { 1 } , . . . { D } _ { K } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 348, + 363, + 362 + ], + "score": 1.0, + "content": ", if a word", + "type": "text" + }, + { + "bbox": [ + 364, + 351, + 373, + 359 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 348, + 506, + 362 + ], + "score": 1.0, + "content": "is related to one of the potential", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 163, + 372 + ], + "score": 1.0, + "content": "bias direction", + "type": "text" + }, + { + "bbox": [ + 164, + 360, + 177, + 371 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 360, + 223, + 372 + ], + "score": 1.0, + "content": "(denote as", + "type": "text" + }, + { + "bbox": [ + 223, + 361, + 258, + 371 + ], + "score": 0.89, + "content": "w \\in \\mathcal { D } _ { k } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 360, + 296, + 372 + ], + "score": 1.0, + "content": "), we call", + "type": "text" + }, + { + "bbox": [ + 297, + 362, + 305, + 370 + ], + "score": 0.72, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 360, + 422, + 372 + ], + "score": 1.0, + "content": "a sensitive attribute word of", + "type": "text" + }, + { + "bbox": [ + 423, + 360, + 436, + 371 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "(also called bias", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 374, + 384 + ], + "score": 1.0, + "content": "attribute word in Liang et al. (2020)). For a sensitive attribute word", + "type": "text" + }, + { + "bbox": [ + 374, + 371, + 407, + 382 + ], + "score": 0.92, + "content": "w \\in \\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 370, + 505, + 384 + ], + "score": 1.0, + "content": ", suppose we can always", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 251, + 395 + ], + "score": 1.0, + "content": "find another sensitive attribute word", + "type": "text" + }, + { + "bbox": [ + 251, + 382, + 282, + 394 + ], + "score": 0.93, + "content": "u \\in { \\mathcal { D } } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 380, + 323, + 395 + ], + "score": 1.0, + "content": ", such that", + "type": "text" + }, + { + "bbox": [ + 323, + 383, + 332, + 392 + ], + "score": 0.79, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 380, + 349, + 395 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 349, + 384, + 357, + 392 + ], + "score": 0.74, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "has the equivalent semantic meaning", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 297, + 405 + ], + "score": 1.0, + "content": "but in a different bias direction. Then we call", + "type": "text" + }, + { + "bbox": [ + 297, + 395, + 304, + 403 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 392, + 410, + 405 + ], + "score": 1.0, + "content": "as a replaceable word of", + "type": "text" + }, + { + "bbox": [ + 410, + 394, + 419, + 403 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 392, + 470, + 405 + ], + "score": 1.0, + "content": "in direction", + "type": "text" + }, + { + "bbox": [ + 470, + 393, + 483, + 405 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 392, + 505, + 405 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 147, + 416 + ], + "score": 1.0, + "content": "denote as", + "type": "text" + }, + { + "bbox": [ + 147, + 404, + 193, + 416 + ], + "score": 0.93, + "content": "u = r _ { j } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 403, + 286, + 416 + ], + "score": 1.0, + "content": ". For the topic “gende", + "type": "text" + }, + { + "bbox": [ + 287, + 404, + 392, + 416 + ], + "score": 0.58, + "content": "\\therefore \\prime = \\{ \\stackrel { } { \\cdot } m a l e ^ { \\prime \\prime } , \\stackrel { } { \\cdot } f e m a l e ^ { \\prime \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 403, + 435, + 416 + ], + "score": 1.0, + "content": ", the word", + "type": "text" + }, + { + "bbox": [ + 435, + 405, + 457, + 414 + ], + "score": 0.79, + "content": "w =", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "“boy” is in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 415, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 218, + 427 + ], + "score": 1.0, + "content": "the potential bias direction", + "type": "text" + }, + { + "bbox": [ + 219, + 415, + 279, + 426 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { 1 } = \\ ^ { \\ast } m a l e ^ { \\prime \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 415, + 506, + 427 + ], + "score": 1.0, + "content": "; a replaceable word of “boy” in “female” direction is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 425, + 196, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 192, + 438 + ], + "score": 0.91, + "content": "r _ { 2 } ( \\bar { w } ) = \\cdots \\mathrm { g i r l } ^ { * } \\in \\mathcal { D } _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 426, + 196, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 287, + 455 + ], + "score": 1.0, + "content": "With the above definitions, for each sentence", + "type": "text" + }, + { + "bbox": [ + 288, + 445, + 295, + 452 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 442, + 443, + 455 + ], + "score": 1.0, + "content": ", we generate an augmented sentence", + "type": "text" + }, + { + "bbox": [ + 443, + 443, + 454, + 452 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 442, + 493, + 455 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 493, + 442, + 504, + 452 + ], + "score": 0.85, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 244, + 466 + ], + "score": 1.0, + "content": "has the same semantic meaning as", + "type": "text" + }, + { + "bbox": [ + 244, + 455, + 252, + 464 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "but in a different potential bias direction. More specifically, for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 150, + 477 + ], + "score": 1.0, + "content": "a sentence", + "type": "text" + }, + { + "bbox": [ + 151, + 464, + 243, + 476 + ], + "score": 0.91, + "content": "\\pmb { x } = ( w ^ { 1 } , w ^ { 2 } , \\dots , \\breve { w } ^ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 462, + 471, + 477 + ], + "score": 1.0, + "content": ", we first find the sensitive word positions as an index set", + "type": "text" + }, + { + "bbox": [ + 471, + 465, + 480, + 475 + ], + "score": 0.81, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 462, + 506, + 477 + ], + "score": 1.0, + "content": ", such", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 145, + 488 + ], + "score": 1.0, + "content": "that each", + "type": "text" + }, + { + "bbox": [ + 146, + 476, + 159, + 486 + ], + "score": 0.76, + "content": "w ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 477, + 193, + 487 + ], + "score": 0.77, + "content": "{ \\bf \\Phi } _ { p } \\in { \\mathcal { P } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 475, + 362, + 488 + ], + "score": 1.0, + "content": ") is a sensitive attribute words in direction", + "type": "text" + }, + { + "bbox": [ + 363, + 476, + 376, + 487 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 475, + 506, + 488 + ], + "score": 1.0, + "content": ". We further make a reasonable", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 484, + 504, + 502 + ], + "spans": [ + { + "bbox": [ + 104, + 484, + 299, + 502 + ], + "score": 1.0, + "content": "assumption that the embedding bias of direction", + "type": "text" + }, + { + "bbox": [ + 300, + 487, + 314, + 497 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 484, + 464, + 502 + ], + "score": 1.0, + "content": "is only caused by the sensitive words", + "type": "text" + }, + { + "bbox": [ + 465, + 486, + 504, + 498 + ], + "score": 0.92, + "content": "\\{ w ^ { p } \\} _ { p \\in { \\mathcal { P } } }", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 118, + 510 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 118, + 499, + 126, + 507 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 497, + 263, + 510 + ], + "score": 1.0, + "content": ". To sample an augmentation to", + "type": "text" + }, + { + "bbox": [ + 263, + 500, + 271, + 507 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 497, + 469, + 510 + ], + "score": 1.0, + "content": ", we first select another potential bias direction", + "type": "text" + }, + { + "bbox": [ + 470, + 498, + 482, + 510 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 497, + 505, + 510 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 506, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 451, + 522 + ], + "score": 1.0, + "content": "then replace all sensitive attribute words by their replaceable words in the direction", + "type": "text" + }, + { + "bbox": [ + 451, + 509, + 464, + 521 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 506, + 506, + 522 + ], + "score": 1.0, + "content": ". That is,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 518, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 198, + 533 + ], + "score": 0.93, + "content": "\\pmb { x } ^ { \\prime } = \\{ v ^ { 1 } , v ^ { 2 } , \\bot \\bot , v ^ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 518, + 229, + 534 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 230, + 519, + 263, + 531 + ], + "score": 0.92, + "content": "v ^ { l } = w ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 518, + 274, + 534 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 274, + 520, + 299, + 532 + ], + "score": 0.91, + "content": "l \\notin \\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 518, + 321, + 534 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 321, + 519, + 371, + 533 + ], + "score": 0.94, + "content": "v ^ { l } = r _ { j } ( w ^ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 518, + 381, + 534 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 381, + 520, + 406, + 531 + ], + "score": 0.89, + "content": "l \\in \\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 518, + 506, + 534 + ], + "score": 1.0, + "content": ". In Table 1, we provide", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 531, + 369, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 369, + 545 + ], + "score": 1.0, + "content": "an example for sentence augmentation under the “gender” topic.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 107, + 554, + 302, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 303, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 303, + 567 + ], + "score": 1.0, + "content": "3.2 CONTRASTIVE LEARNING FRAMEWORK", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 570, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 241, + 584 + ], + "score": 1.0, + "content": "After obtaining the sentence pair", + "type": "text" + }, + { + "bbox": [ + 241, + 571, + 270, + 583 + ], + "score": 0.92, + "content": "( { \\pmb x } , { \\pmb x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 570, + 505, + 584 + ], + "score": 1.0, + "content": "with the augmentation strategy from Section 3.1, we con-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 407, + 594 + ], + "score": 1.0, + "content": "struct a contrastive learning framework to learn our debiasing fair filter", + "type": "text" + }, + { + "bbox": [ + 408, + 582, + 425, + 594 + ], + "score": 0.91, + "content": "f ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 582, + 505, + 594 + ], + "score": 1.0, + "content": ". As shown in the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 592, + 361, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 361, + 606 + ], + "score": 1.0, + "content": "Figure 1(a), our framework consists of the following two steps:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 210, + 623 + ], + "score": 1.0, + "content": "(1) We encode sentences", + "type": "text" + }, + { + "bbox": [ + 210, + 610, + 239, + 622 + ], + "score": 0.92, + "content": "( { \\pmb x } , { \\pmb x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 608, + 311, + 623 + ], + "score": 1.0, + "content": "into embeddings", + "type": "text" + }, + { + "bbox": [ + 311, + 610, + 339, + 622 + ], + "score": 0.92, + "content": "( z , z ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 608, + 454, + 623 + ], + "score": 1.0, + "content": "with the pretrained encoder", + "type": "text" + }, + { + "bbox": [ + 455, + 610, + 474, + 621 + ], + "score": 0.89, + "content": "E ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 608, + 505, + 623 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 623, + 114, + 631 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 620, + 132, + 633 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 133, + 621, + 143, + 631 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 620, + 458, + 633 + ], + "score": 1.0, + "content": "have the same meaning but different potential bias directions, the embeddings", + "type": "text" + }, + { + "bbox": [ + 458, + 621, + 486, + 633 + ], + "score": 0.92, + "content": "( z , z ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 620, + 506, + 633 + ], + "score": 1.0, + "content": "will", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 632, + 480, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 439, + 644 + ], + "score": 1.0, + "content": "have different bias directions, which are caused by the sensitive attributed words in", + "type": "text" + }, + { + "bbox": [ + 439, + 634, + 447, + 642 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 632, + 465, + 644 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 466, + 633, + 476, + 642 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 632, + 480, + 644 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 279, + 660 + ], + "score": 1.0, + "content": "(2) We then feed the sentence embeddings", + "type": "text" + }, + { + "bbox": [ + 279, + 649, + 307, + 660 + ], + "score": 0.92, + "content": "( z , z ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 648, + 396, + 660 + ], + "score": 1.0, + "content": "through our fair filter", + "type": "text" + }, + { + "bbox": [ + 396, + 649, + 413, + 661 + ], + "score": 0.91, + "content": "f ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "to obtain the debiased", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 187, + 672 + ], + "score": 1.0, + "content": "embedding outputs", + "type": "text" + }, + { + "bbox": [ + 187, + 660, + 215, + 672 + ], + "score": 0.93, + "content": "( d , d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 660, + 255, + 672 + ], + "score": 1.0, + "content": ". Ideally,", + "type": "text" + }, + { + "bbox": [ + 256, + 660, + 264, + 670 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 660, + 283, + 672 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 283, + 660, + 293, + 670 + ], + "score": 0.83, + "content": "\\mathbf { \\Delta } d ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "should represent the same semantic meaning with-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "out social bias. Inspired by SimCLR (Chen et al., 2020), we encourage the overlapped semantic", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 681, + 430, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 191, + 694 + ], + "score": 1.0, + "content": "information between", + "type": "text" + }, + { + "bbox": [ + 192, + 682, + 200, + 691 + ], + "score": 0.78, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 681, + 217, + 694 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 682, + 227, + 691 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } d ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 681, + 391, + 694 + ], + "score": 1.0, + "content": "by maximizing their mutual information", + "type": "text" + }, + { + "bbox": [ + 392, + 681, + 426, + 694 + ], + "score": 0.93, + "content": "\\bar { \\mathcal { T } } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 681, + 430, + 694 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 221, + 712 + ], + "score": 1.0, + "content": "However, the calculation of", + "type": "text" + }, + { + "bbox": [ + 221, + 699, + 255, + 711 + ], + "score": 0.93, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 698, + 496, + 712 + ], + "score": 1.0, + "content": "is practically difficult because only embedding samples of", + "type": "text" + }, + { + "bbox": [ + 497, + 699, + 505, + 709 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 124, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 710, + 135, + 720 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "are available. Therefore, we use the InfoNCE mutual information estimator (Oord et al.,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 261, + 734 + ], + "score": 1.0, + "content": "2018) to minimize the lower bound of", + "type": "text" + }, + { + "bbox": [ + 262, + 721, + 296, + 732 + ], + "score": 0.93, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 720, + 477, + 734 + ], + "score": 1.0, + "content": "instead. Based on a learnable score function", + "type": "text" + }, + { + "bbox": [ + 477, + 721, + 501, + 732 + ], + "score": 0.91, + "content": "g ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 720, + 506, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 97, + 504, + 135 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 81, + 494, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 78, + 496, + 94 + ], + "spans": [ + { + "bbox": [ + 114, + 78, + 496, + 94 + ], + "score": 1.0, + "content": "Table 1: Examples of generating an augmentation sentence under the sensitive topic “gender”.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 109, + 97, + 504, + 135 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 97, + 504, + 135 + ], + "spans": [ + { + "bbox": [ + 109, + 97, + 504, + 135 + ], + "score": 0.923, + "html": "
Bias directionSensitive Attribute wordsText content
Originalmalehe,his{He} is good at playing {his} basketball.
Augmentationfemaleshe,her{She{She} is good at playing {her} basketball.
", + "type": "table", + "image_path": "e51de30e34c044121ed00266c3dc4486583b6a350b45ba45fbe871497c37818f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 109, + 97, + 504, + 109.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 109.66666666666667, + 504, + 122.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 122.33333333333334, + 504, + 135.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 505, + 249 + ], + "lines": [], + "index": 8, + "bbox_fs": [ + 105, + 150, + 506, + 250 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 106, + 260, + 363, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 364, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 364, + 273 + ], + "score": 1.0, + "content": "3.1 DATA AUGMENTATIONS WITH SENSITIVE ATTRIBUTES", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 505, + 344 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "We first describe the sentence data augmentation process for our FairFil contrastive learning. De-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 234, + 301 + ], + "score": 1.0, + "content": "note a social sensitive topic as", + "type": "text" + }, + { + "bbox": [ + 234, + 288, + 335, + 300 + ], + "score": 0.92, + "content": "\\mathcal { T } = \\{ \\mathcal { D } _ { 1 } , \\mathcal { D } _ { 2 } , \\ldots , \\mathcal { D } _ { K } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 288, + 367, + 301 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 367, + 288, + 381, + 299 + ], + "score": 0.79, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 288, + 449, + 299 + ], + "score": 0.69, + "content": "( k = 1 , \\ldots , K )", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "is one of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 346, + 313 + ], + "score": 1.0, + "content": "potential bias directions under the topic. For example, if", + "type": "text" + }, + { + "bbox": [ + 346, + 300, + 356, + 309 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 298, + 506, + 313 + ], + "score": 1.0, + "content": "represents the sensitive topic “gen-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 151, + 323 + ], + "score": 1.0, + "content": "der”, then", + "type": "text" + }, + { + "bbox": [ + 152, + 310, + 162, + 320 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 309, + 317, + 323 + ], + "score": 1.0, + "content": "consists two potential bias directions", + "type": "text" + }, + { + "bbox": [ + 318, + 310, + 456, + 322 + ], + "score": 0.81, + "content": "\\{ \\mathcal { D } _ { 1 } , \\mathcal { D } _ { 2 } \\} \\stackrel { - } { = } \\{ { } ^ { } m a l e ^ { , \\prime } , \\ { } ^ { } f e m a l e ^ { , \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 309, + 505, + 323 + ], + "score": 1.0, + "content": ". Similarly,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 504, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 117, + 334 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 117, + 321, + 127, + 331 + ], + "score": 0.82, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 320, + 344, + 334 + ], + "score": 1.0, + "content": "is set as the major “religions” of the world, then", + "type": "text" + }, + { + "bbox": [ + 344, + 322, + 354, + 331 + ], + "score": 0.81, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 320, + 416, + 334 + ], + "score": 1.0, + "content": "could contain", + "type": "text" + }, + { + "bbox": [ + 417, + 321, + 504, + 333 + ], + "score": 0.91, + "content": "\\left\\{ \\mathcal { D } _ { 1 } , \\mathcal { D } _ { 2 } , \\mathcal { D } _ { 3 } , \\mathcal { D } _ { 4 } \\right\\} \\ =", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 331, + 404, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 404, + 344 + ], + "score": 1.0, + "content": "{“Christianity”, “Islam”, “Judaism”, “Buddhism”} as four components.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 276, + 506, + 344 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 241, + 362 + ], + "score": 1.0, + "content": "For a given social sensitive topic", + "type": "text" + }, + { + "bbox": [ + 241, + 348, + 320, + 361 + ], + "score": 0.92, + "content": "\\mathcal { T } = \\{ { D } _ { 1 } , . . . { D } _ { K } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 348, + 363, + 362 + ], + "score": 1.0, + "content": ", if a word", + "type": "text" + }, + { + "bbox": [ + 364, + 351, + 373, + 359 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 348, + 506, + 362 + ], + "score": 1.0, + "content": "is related to one of the potential", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 163, + 372 + ], + "score": 1.0, + "content": "bias direction", + "type": "text" + }, + { + "bbox": [ + 164, + 360, + 177, + 371 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 360, + 223, + 372 + ], + "score": 1.0, + "content": "(denote as", + "type": "text" + }, + { + "bbox": [ + 223, + 361, + 258, + 371 + ], + "score": 0.89, + "content": "w \\in \\mathcal { D } _ { k } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 360, + 296, + 372 + ], + "score": 1.0, + "content": "), we call", + "type": "text" + }, + { + "bbox": [ + 297, + 362, + 305, + 370 + ], + "score": 0.72, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 360, + 422, + 372 + ], + "score": 1.0, + "content": "a sensitive attribute word of", + "type": "text" + }, + { + "bbox": [ + 423, + 360, + 436, + 371 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "(also called bias", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 374, + 384 + ], + "score": 1.0, + "content": "attribute word in Liang et al. (2020)). For a sensitive attribute word", + "type": "text" + }, + { + "bbox": [ + 374, + 371, + 407, + 382 + ], + "score": 0.92, + "content": "w \\in \\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 370, + 505, + 384 + ], + "score": 1.0, + "content": ", suppose we can always", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 251, + 395 + ], + "score": 1.0, + "content": "find another sensitive attribute word", + "type": "text" + }, + { + "bbox": [ + 251, + 382, + 282, + 394 + ], + "score": 0.93, + "content": "u \\in { \\mathcal { D } } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 380, + 323, + 395 + ], + "score": 1.0, + "content": ", such that", + "type": "text" + }, + { + "bbox": [ + 323, + 383, + 332, + 392 + ], + "score": 0.79, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 380, + 349, + 395 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 349, + 384, + 357, + 392 + ], + "score": 0.74, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "has the equivalent semantic meaning", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 297, + 405 + ], + "score": 1.0, + "content": "but in a different bias direction. Then we call", + "type": "text" + }, + { + "bbox": [ + 297, + 395, + 304, + 403 + ], + "score": 0.75, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 392, + 410, + 405 + ], + "score": 1.0, + "content": "as a replaceable word of", + "type": "text" + }, + { + "bbox": [ + 410, + 394, + 419, + 403 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 392, + 470, + 405 + ], + "score": 1.0, + "content": "in direction", + "type": "text" + }, + { + "bbox": [ + 470, + 393, + 483, + 405 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 392, + 505, + 405 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 147, + 416 + ], + "score": 1.0, + "content": "denote as", + "type": "text" + }, + { + "bbox": [ + 147, + 404, + 193, + 416 + ], + "score": 0.93, + "content": "u = r _ { j } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 403, + 286, + 416 + ], + "score": 1.0, + "content": ". For the topic “gende", + "type": "text" + }, + { + "bbox": [ + 287, + 404, + 392, + 416 + ], + "score": 0.58, + "content": "\\therefore \\prime = \\{ \\stackrel { } { \\cdot } m a l e ^ { \\prime \\prime } , \\stackrel { } { \\cdot } f e m a l e ^ { \\prime \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 403, + 435, + 416 + ], + "score": 1.0, + "content": ", the word", + "type": "text" + }, + { + "bbox": [ + 435, + 405, + 457, + 414 + ], + "score": 0.79, + "content": "w =", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "“boy” is in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 415, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 218, + 427 + ], + "score": 1.0, + "content": "the potential bias direction", + "type": "text" + }, + { + "bbox": [ + 219, + 415, + 279, + 426 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { 1 } = \\ ^ { \\ast } m a l e ^ { \\prime \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 415, + 506, + 427 + ], + "score": 1.0, + "content": "; a replaceable word of “boy” in “female” direction is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 425, + 196, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 192, + 438 + ], + "score": 0.91, + "content": "r _ { 2 } ( \\bar { w } ) = \\cdots \\mathrm { g i r l } ^ { * } \\in \\mathcal { D } _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 426, + 196, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 348, + 506, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 287, + 455 + ], + "score": 1.0, + "content": "With the above definitions, for each sentence", + "type": "text" + }, + { + "bbox": [ + 288, + 445, + 295, + 452 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 442, + 443, + 455 + ], + "score": 1.0, + "content": ", we generate an augmented sentence", + "type": "text" + }, + { + "bbox": [ + 443, + 443, + 454, + 452 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 442, + 493, + 455 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 493, + 442, + 504, + 452 + ], + "score": 0.85, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 244, + 466 + ], + "score": 1.0, + "content": "has the same semantic meaning as", + "type": "text" + }, + { + "bbox": [ + 244, + 455, + 252, + 464 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "but in a different potential bias direction. More specifically, for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 150, + 477 + ], + "score": 1.0, + "content": "a sentence", + "type": "text" + }, + { + "bbox": [ + 151, + 464, + 243, + 476 + ], + "score": 0.91, + "content": "\\pmb { x } = ( w ^ { 1 } , w ^ { 2 } , \\dots , \\breve { w } ^ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 462, + 471, + 477 + ], + "score": 1.0, + "content": ", we first find the sensitive word positions as an index set", + "type": "text" + }, + { + "bbox": [ + 471, + 465, + 480, + 475 + ], + "score": 0.81, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 462, + 506, + 477 + ], + "score": 1.0, + "content": ", such", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 145, + 488 + ], + "score": 1.0, + "content": "that each", + "type": "text" + }, + { + "bbox": [ + 146, + 476, + 159, + 486 + ], + "score": 0.76, + "content": "w ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 477, + 193, + 487 + ], + "score": 0.77, + "content": "{ \\bf \\Phi } _ { p } \\in { \\mathcal { P } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 475, + 362, + 488 + ], + "score": 1.0, + "content": ") is a sensitive attribute words in direction", + "type": "text" + }, + { + "bbox": [ + 363, + 476, + 376, + 487 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 475, + 506, + 488 + ], + "score": 1.0, + "content": ". We further make a reasonable", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 484, + 504, + 502 + ], + "spans": [ + { + "bbox": [ + 104, + 484, + 299, + 502 + ], + "score": 1.0, + "content": "assumption that the embedding bias of direction", + "type": "text" + }, + { + "bbox": [ + 300, + 487, + 314, + 497 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 484, + 464, + 502 + ], + "score": 1.0, + "content": "is only caused by the sensitive words", + "type": "text" + }, + { + "bbox": [ + 465, + 486, + 504, + 498 + ], + "score": 0.92, + "content": "\\{ w ^ { p } \\} _ { p \\in { \\mathcal { P } } }", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 118, + 510 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 118, + 499, + 126, + 507 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 497, + 263, + 510 + ], + "score": 1.0, + "content": ". To sample an augmentation to", + "type": "text" + }, + { + "bbox": [ + 263, + 500, + 271, + 507 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 497, + 469, + 510 + ], + "score": 1.0, + "content": ", we first select another potential bias direction", + "type": "text" + }, + { + "bbox": [ + 470, + 498, + 482, + 510 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 497, + 505, + 510 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 506, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 451, + 522 + ], + "score": 1.0, + "content": "then replace all sensitive attribute words by their replaceable words in the direction", + "type": "text" + }, + { + "bbox": [ + 451, + 509, + 464, + 521 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 506, + 506, + 522 + ], + "score": 1.0, + "content": ". That is,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 518, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 198, + 533 + ], + "score": 0.93, + "content": "\\pmb { x } ^ { \\prime } = \\{ v ^ { 1 } , v ^ { 2 } , \\bot \\bot , v ^ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 518, + 229, + 534 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 230, + 519, + 263, + 531 + ], + "score": 0.92, + "content": "v ^ { l } = w ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 518, + 274, + 534 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 274, + 520, + 299, + 532 + ], + "score": 0.91, + "content": "l \\notin \\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 518, + 321, + 534 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 321, + 519, + 371, + 533 + ], + "score": 0.94, + "content": "v ^ { l } = r _ { j } ( w ^ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 518, + 381, + 534 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 381, + 520, + 406, + 531 + ], + "score": 0.89, + "content": "l \\in \\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 518, + 506, + 534 + ], + "score": 1.0, + "content": ". In Table 1, we provide", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 531, + 369, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 369, + 545 + ], + "score": 1.0, + "content": "an example for sentence augmentation under the “gender” topic.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32, + "bbox_fs": [ + 104, + 442, + 506, + 545 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 554, + 302, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 303, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 303, + 567 + ], + "score": 1.0, + "content": "3.2 CONTRASTIVE LEARNING FRAMEWORK", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 570, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 241, + 584 + ], + "score": 1.0, + "content": "After obtaining the sentence pair", + "type": "text" + }, + { + "bbox": [ + 241, + 571, + 270, + 583 + ], + "score": 0.92, + "content": "( { \\pmb x } , { \\pmb x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 570, + 505, + 584 + ], + "score": 1.0, + "content": "with the augmentation strategy from Section 3.1, we con-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 407, + 594 + ], + "score": 1.0, + "content": "struct a contrastive learning framework to learn our debiasing fair filter", + "type": "text" + }, + { + "bbox": [ + 408, + 582, + 425, + 594 + ], + "score": 0.91, + "content": "f ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 582, + 505, + 594 + ], + "score": 1.0, + "content": ". As shown in the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 592, + 361, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 361, + 606 + ], + "score": 1.0, + "content": "Figure 1(a), our framework consists of the following two steps:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 570, + 505, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 210, + 623 + ], + "score": 1.0, + "content": "(1) We encode sentences", + "type": "text" + }, + { + "bbox": [ + 210, + 610, + 239, + 622 + ], + "score": 0.92, + "content": "( { \\pmb x } , { \\pmb x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 608, + 311, + 623 + ], + "score": 1.0, + "content": "into embeddings", + "type": "text" + }, + { + "bbox": [ + 311, + 610, + 339, + 622 + ], + "score": 0.92, + "content": "( z , z ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 608, + 454, + 623 + ], + "score": 1.0, + "content": "with the pretrained encoder", + "type": "text" + }, + { + "bbox": [ + 455, + 610, + 474, + 621 + ], + "score": 0.89, + "content": "E ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 608, + 505, + 623 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 623, + 114, + 631 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 620, + 132, + 633 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 133, + 621, + 143, + 631 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 620, + 458, + 633 + ], + "score": 1.0, + "content": "have the same meaning but different potential bias directions, the embeddings", + "type": "text" + }, + { + "bbox": [ + 458, + 621, + 486, + 633 + ], + "score": 0.92, + "content": "( z , z ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 620, + 506, + 633 + ], + "score": 1.0, + "content": "will", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 632, + 480, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 439, + 644 + ], + "score": 1.0, + "content": "have different bias directions, which are caused by the sensitive attributed words in", + "type": "text" + }, + { + "bbox": [ + 439, + 634, + 447, + 642 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 632, + 465, + 644 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 466, + 633, + 476, + 642 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 632, + 480, + 644 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 608, + 506, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 279, + 660 + ], + "score": 1.0, + "content": "(2) We then feed the sentence embeddings", + "type": "text" + }, + { + "bbox": [ + 279, + 649, + 307, + 660 + ], + "score": 0.92, + "content": "( z , z ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 648, + 396, + 660 + ], + "score": 1.0, + "content": "through our fair filter", + "type": "text" + }, + { + "bbox": [ + 396, + 649, + 413, + 661 + ], + "score": 0.91, + "content": "f ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "to obtain the debiased", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 187, + 672 + ], + "score": 1.0, + "content": "embedding outputs", + "type": "text" + }, + { + "bbox": [ + 187, + 660, + 215, + 672 + ], + "score": 0.93, + "content": "( d , d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 660, + 255, + 672 + ], + "score": 1.0, + "content": ". Ideally,", + "type": "text" + }, + { + "bbox": [ + 256, + 660, + 264, + 670 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 660, + 283, + 672 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 283, + 660, + 293, + 670 + ], + "score": 0.83, + "content": "\\mathbf { \\Delta } d ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "should represent the same semantic meaning with-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "out social bias. Inspired by SimCLR (Chen et al., 2020), we encourage the overlapped semantic", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 681, + 430, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 191, + 694 + ], + "score": 1.0, + "content": "information between", + "type": "text" + }, + { + "bbox": [ + 192, + 682, + 200, + 691 + ], + "score": 0.78, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 681, + 217, + 694 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 682, + 227, + 691 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } d ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 681, + 391, + 694 + ], + "score": 1.0, + "content": "by maximizing their mutual information", + "type": "text" + }, + { + "bbox": [ + 392, + 681, + 426, + 694 + ], + "score": 0.93, + "content": "\\bar { \\mathcal { T } } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 681, + 430, + 694 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 648, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 221, + 712 + ], + "score": 1.0, + "content": "However, the calculation of", + "type": "text" + }, + { + "bbox": [ + 221, + 699, + 255, + 711 + ], + "score": 0.93, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 698, + 496, + 712 + ], + "score": 1.0, + "content": "is practically difficult because only embedding samples of", + "type": "text" + }, + { + "bbox": [ + 497, + 699, + 505, + 709 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 124, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 710, + 135, + 720 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "are available. Therefore, we use the InfoNCE mutual information estimator (Oord et al.,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 261, + 734 + ], + "score": 1.0, + "content": "2018) to minimize the lower bound of", + "type": "text" + }, + { + "bbox": [ + 262, + 721, + 296, + 732 + ], + "score": 0.93, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 720, + 477, + 734 + ], + "score": 1.0, + "content": "instead. Based on a learnable score function", + "type": "text" + }, + { + "bbox": [ + 477, + 721, + 501, + 732 + ], + "score": 0.91, + "content": "g ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 720, + 506, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 297, + 457, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 402, + 315 + ], + "score": 1.0, + "content": "the contrastive InfoNCE estimator is calculated within a batch of samples", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 402, + 299, + 457, + 312 + ], + "score": 0.92, + "content": "\\{ ( d _ { i } , d _ { i } ^ { \\prime } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 49, + "bbox_fs": [ + 105, + 698, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 80, + 485, + 203 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 80, + 485, + 203 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 80, + 485, + 203 + ], + "spans": [ + { + "bbox": [ + 126, + 80, + 485, + 203 + ], + "score": 0.973, + "type": "image", + "image_path": "8c1f7a0b55452733b6dca530e9279647bdf58ae19e81215c4ef29b41a8378f5b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 126, + 80, + 485, + 121.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 126, + 121.0, + 485, + 162.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 126, + 162.0, + 485, + 203.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 210, + 506, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 379, + 222 + ], + "score": 1.0, + "content": "Figure 1: (a) Contrastive learning framework of FairFil: Sentence", + "type": "text" + }, + { + "bbox": [ + 380, + 213, + 388, + 220 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 209, + 477, + 222 + ], + "score": 1.0, + "content": "and its augmentation", + "type": "text" + }, + { + "bbox": [ + 478, + 210, + 488, + 221 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 211, + 234 + ], + "score": 1.0, + "content": "encoded into embeddings", + "type": "text" + }, + { + "bbox": [ + 212, + 222, + 219, + 231 + ], + "score": 0.78, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 221, + 237, + 234 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 221, + 247, + 231 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 221, + 304, + 234 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + }, + { + "bbox": [ + 304, + 222, + 319, + 231 + ], + "score": 0.86, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "is the embedding of a sensitive attribute word", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 164, + 244 + ], + "score": 1.0, + "content": "selected from", + "type": "text" + }, + { + "bbox": [ + 164, + 234, + 172, + 243 + ], + "score": 0.56, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 233, + 176, + 244 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 177, + 232, + 198, + 243 + ], + "score": 0.85, + "content": "\\mathcal { T } _ { \\mathrm { N C E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 233, + 378, + 244 + ], + "score": 1.0, + "content": "maximizes the mutual information between", + "type": "text" + }, + { + "bbox": [ + 378, + 233, + 386, + 242 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 233, + 404, + 244 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 232, + 414, + 242 + ], + "score": 0.78, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 233, + 418, + 244 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 419, + 232, + 445, + 243 + ], + "score": 0.82, + "content": "\\scriptstyle { \\mathcal { T } } _ { \\mathrm { C L U B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 233, + 505, + 244 + ], + "score": 1.0, + "content": "eliminates the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 185, + 255 + ], + "score": 1.0, + "content": "bias information of", + "type": "text" + }, + { + "bbox": [ + 186, + 244, + 200, + 253 + ], + "score": 0.88, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 243, + 223, + 255 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 223, + 244, + 231, + 253 + ], + "score": 0.73, + "content": "\\mathbf { \\delta } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 243, + 366, + 255 + ], + "score": 1.0, + "content": ". (b) Illustration of information in", + "type": "text" + }, + { + "bbox": [ + 366, + 244, + 374, + 253 + ], + "score": 0.79, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 243, + 392, + 255 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 392, + 243, + 402, + 253 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 243, + 505, + 255 + ], + "score": 1.0, + "content": ": The blue and red circles", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 226, + 266 + ], + "score": 1.0, + "content": "represent the information in", + "type": "text" + }, + { + "bbox": [ + 226, + 255, + 234, + 264 + ], + "score": 0.77, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 254, + 254, + 266 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 254, + 254, + 264, + 264 + ], + "score": 0.84, + "content": "\\mathbf { { \\mathbf { { \\mathit { d } } } } ^ { \\prime } } \\mathbf { \\Sigma }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 254, + 505, + 266 + ], + "score": 1.0, + "content": ", respectively. The intersection is the mutual information", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 264, + 461, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 142, + 278 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 142, + 266, + 150, + 275 + ], + "score": 0.79, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 264, + 168, + 278 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 168, + 265, + 177, + 275 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 264, + 461, + 278 + ], + "score": 1.0, + "content": ". The shadow area represents the bias information of both embeddings.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 104, + 299, + 461, + 312 + ], + "lines": [ + { + "bbox": [ + 104, + 297, + 457, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 402, + 315 + ], + "score": 1.0, + "content": "the contrastive InfoNCE estimator is calculated within a batch of samples", + "type": "text" + }, + { + "bbox": [ + 402, + 299, + 457, + 312 + ], + "score": 0.92, + "content": "\\{ ( d _ { i } , d _ { i } ^ { \\prime } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 318, + 397, + 353 + ], + "lines": [ + { + "bbox": [ + 213, + 318, + 397, + 353 + ], + "spans": [ + { + "bbox": [ + 213, + 318, + 397, + 353 + ], + "score": 0.94, + "content": "\\mathcal { T } _ { \\mathrm { N C E } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log \\frac { \\exp ( g ( d _ { i } , d _ { i } ^ { \\prime } ) ) } { \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\exp ( g ( d _ { i } , d _ { j } ^ { \\prime } ) ) } .", + "type": "interline_equation", + "image_path": "fe4473099a6a0c44da9d37613a02a0710a1f912c63b5fdb0cf97a893bc1aa48d.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 213, + 318, + 397, + 335.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 213, + 335.5, + 397, + 353.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 359, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 163, + 371 + ], + "score": 1.0, + "content": "By maximize", + "type": "text" + }, + { + "bbox": [ + 163, + 360, + 185, + 370 + ], + "score": 0.89, + "content": "\\mathcal { T } _ { \\mathrm { N C E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 359, + 434, + 371 + ], + "score": 1.0, + "content": ", we encourage the difference between the positive pair score", + "type": "text" + }, + { + "bbox": [ + 434, + 359, + 471, + 371 + ], + "score": 0.92, + "content": "g ( d _ { i } , d _ { i } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 370, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 186, + 384 + ], + "score": 1.0, + "content": "negative pair score", + "type": "text" + }, + { + "bbox": [ + 186, + 370, + 224, + 384 + ], + "score": 0.93, + "content": "g ( d _ { i } , d _ { j } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 370, + 260, + 384 + ], + "score": 1.0, + "content": ", so that", + "type": "text" + }, + { + "bbox": [ + 260, + 371, + 271, + 381 + ], + "score": 0.87, + "content": "\\mathbf { \\ b { d } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 370, + 448, + 384 + ], + "score": 1.0, + "content": "can share more semantic information with", + "type": "text" + }, + { + "bbox": [ + 449, + 371, + 459, + 383 + ], + "score": 0.88, + "content": "\\mathbf { \\Delta } d _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 370, + 506, + 384 + ], + "score": 1.0, + "content": "than other", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 381, + 182, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 182, + 397 + ], + "score": 1.0, + "content": "embeddings d0j6=i.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 409, + 245, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 246, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 246, + 422 + ], + "score": 1.0, + "content": "3.3 DEBIASING REGULARIZER", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "Practically, the contrastive learning framework in Section 3.2 can already show encouraging debi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 442, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 423, + 454 + ], + "score": 1.0, + "content": "asing performance (as shown in the Experiments). However, the embedding", + "type": "text" + }, + { + "bbox": [ + 423, + 442, + 431, + 452 + ], + "score": 0.67, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 442, + 506, + 454 + ], + "score": 1.0, + "content": "can contain extra", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 206, + 465 + ], + "score": 1.0, + "content": "biased information from", + "type": "text" + }, + { + "bbox": [ + 206, + 454, + 213, + 462 + ], + "score": 0.72, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 452, + 305, + 465 + ], + "score": 1.0, + "content": ", that only maximizing", + "type": "text" + }, + { + "bbox": [ + 305, + 452, + 339, + 464 + ], + "score": 0.94, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "fails to eliminate. To encourage no extra", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 137, + 477 + ], + "score": 1.0, + "content": "bias in", + "type": "text" + }, + { + "bbox": [ + 137, + 464, + 144, + 473 + ], + "score": 0.74, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 463, + 506, + 477 + ], + "score": 1.0, + "content": ", we introduce a debiasing regularizer which minimizes the mutual information between", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 474, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 153, + 486 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 154, + 475, + 162, + 484 + ], + "score": 0.74, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 474, + 320, + 486 + ], + "score": 1.0, + "content": "and the potential bias from embedding", + "type": "text" + }, + { + "bbox": [ + 321, + 476, + 328, + 484 + ], + "score": 0.66, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 474, + 504, + 486 + ], + "score": 1.0, + "content": ". As discussed in Section 3.1, in our frame-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 211, + 497 + ], + "score": 1.0, + "content": "work the potential bias of", + "type": "text" + }, + { + "bbox": [ + 212, + 487, + 219, + 495 + ], + "score": 0.77, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 486, + 448, + 497 + ], + "score": 1.0, + "content": "is assumed to come from the sensitive attribute words in", + "type": "text" + }, + { + "bbox": [ + 448, + 487, + 456, + 495 + ], + "score": 0.68, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 486, + 505, + 497 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 497, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 429, + 508 + ], + "score": 1.0, + "content": "we should reduce the bias word information from the debiased representation", + "type": "text" + }, + { + "bbox": [ + 429, + 497, + 437, + 506 + ], + "score": 0.75, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 497, + 460, + 508 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 461, + 497, + 475, + 506 + ], + "score": 0.87, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 497, + 505, + 508 + ], + "score": 1.0, + "content": "be the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 267, + 519 + ], + "score": 1.0, + "content": "embedding of a sensitive attribute word", + "type": "text" + }, + { + "bbox": [ + 268, + 507, + 281, + 517 + ], + "score": 0.87, + "content": "w ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 506, + 329, + 519 + ], + "score": 1.0, + "content": "in sentence", + "type": "text" + }, + { + "bbox": [ + 330, + 509, + 337, + 517 + ], + "score": 0.71, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 506, + 430, + 519 + ], + "score": 1.0, + "content": ". The word embedding", + "type": "text" + }, + { + "bbox": [ + 431, + 507, + 445, + 517 + ], + "score": 0.86, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "can always be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 308, + 530 + ], + "score": 1.0, + "content": "obtained from the pretrained text encoders (Bordia", + "type": "text" + }, + { + "bbox": [ + 308, + 519, + 317, + 528 + ], + "score": 0.3, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "Bowman, 2019). We then minimize the mutual", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 156, + 541 + ], + "score": 1.0, + "content": "information", + "type": "text" + }, + { + "bbox": [ + 157, + 529, + 195, + 541 + ], + "score": 0.91, + "content": "\\mathcal { T } ( w ^ { p } ; d )", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 528, + 505, + 541 + ], + "score": 1.0, + "content": ", using the CLUB mutual information upper bound (Cheng et al., 2020a) to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 102, + 535, + 508, + 557 + ], + "spans": [ + { + "bbox": [ + 102, + 535, + 142, + 557 + ], + "score": 1.0, + "content": "estimate", + "type": "text" + }, + { + "bbox": [ + 142, + 540, + 181, + 552 + ], + "score": 0.92, + "content": "\\mathcal { T } ( w ^ { p } ; d )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 535, + 428, + 557 + ], + "score": 1.0, + "content": "with embedding samples. Given a batch of embedding pairs", + "type": "text" + }, + { + "bbox": [ + 428, + 540, + 487, + 552 + ], + "score": 0.93, + "content": "\\bar { \\{ ( d _ { i } , \\boldsymbol { w } ^ { p } ) \\} } _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 535, + 508, + 557 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 551, + 275, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 275, + 563 + ], + "score": 1.0, + "content": "can calculate the debiasing regularizer as:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 569, + 424, + 604 + ], + "lines": [ + { + "bbox": [ + 186, + 569, + 424, + 604 + ], + "spans": [ + { + "bbox": [ + 186, + 569, + 424, + 604 + ], + "score": 0.95, + "content": "\\mathcal { T } _ { \\mathrm { C L U B } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\Big [ \\log q _ { \\theta } ( { \\pmb w } _ { i } ^ { p } | \\pmb d _ { i } ) - \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\log q _ { \\theta } ( { \\pmb w } _ { j } ^ { p } | \\pmb d _ { i } ) \\Big ] ,", + "type": "interline_equation", + "image_path": "a63c40840b1faa97447066e2f554e7dfffa8be0097875d90080e8ba1df426f47.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 569, + 424, + 586.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 186, + 586.5, + 424, + 604.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 505, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 132, + 624 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 613, + 142, + 623 + ], + "score": 0.84, + "content": "q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 611, + 419, + 624 + ], + "score": 1.0, + "content": "is a variational approximation to ground-truth conditional distribution", + "type": "text" + }, + { + "bbox": [ + 420, + 611, + 450, + 623 + ], + "score": 0.9, + "content": "p ( \\pmb { w } | \\pmb { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 611, + 505, + 624 + ], + "score": 1.0, + "content": ". We parame-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 131, + 635 + ], + "score": 1.0, + "content": "terize", + "type": "text" + }, + { + "bbox": [ + 131, + 624, + 141, + 633 + ], + "score": 0.84, + "content": "q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 622, + 435, + 635 + ], + "score": 1.0, + "content": "with another neural network. As proved in Cheng et al. (2020a), the better", + "type": "text" + }, + { + "bbox": [ + 436, + 623, + 470, + 634 + ], + "score": 0.94, + "content": "q _ { \\theta } ( { \\pmb w } | { \\pmb d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "approx-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 135, + 646 + ], + "score": 1.0, + "content": "imates", + "type": "text" + }, + { + "bbox": [ + 135, + 633, + 166, + 645 + ], + "score": 0.93, + "content": "p ( \\pmb { w } | \\pmb { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 632, + 244, + 646 + ], + "score": 1.0, + "content": ", the more accurate", + "type": "text" + }, + { + "bbox": [ + 244, + 633, + 270, + 645 + ], + "score": 0.88, + "content": "\\scriptstyle { \\mathcal { T } } _ { \\mathrm { C L U B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "serves as the mutual information upper bound. Therefore,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 103, + 641, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 103, + 641, + 352, + 660 + ], + "score": 1.0, + "content": "besides the loss in (5), we also maximize the log-likelihood of", + "type": "text" + }, + { + "bbox": [ + 352, + 644, + 387, + 656 + ], + "score": 0.92, + "content": "q _ { \\theta } ( { \\pmb w } | { \\pmb d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 641, + 442, + 660 + ], + "score": 1.0, + "content": "with samples", + "type": "text" + }, + { + "bbox": [ + 443, + 644, + 501, + 657 + ], + "score": 0.92, + "content": "\\{ ( d _ { i } , { \\boldsymbol { w } } _ { i } ^ { p } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 641, + 506, + 660 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 660, + 505, + 739 + ], + "lines": [ + { + "bbox": [ + 105, + 661, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 504, + 673 + ], + "score": 1.0, + "content": "Based on the above sections, the overall learning scheme of our fair filter (FairFil) is described in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 673, + 504, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 504, + 684 + ], + "score": 1.0, + "content": "Algorithm 1. Also, we provide an intuitive explanation to the two loss terms in our framework.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 308, + 695 + ], + "score": 1.0, + "content": "In Figure 1(b), the blue and red circles represent", + "type": "text" + }, + { + "bbox": [ + 309, + 684, + 317, + 693 + ], + "score": 0.8, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 682, + 336, + 695 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 336, + 683, + 345, + 693 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 682, + 505, + 695 + ], + "score": 1.0, + "content": ", respectively, in the embedding space.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 693, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 174, + 706 + ], + "score": 1.0, + "content": "The intersection", + "type": "text" + }, + { + "bbox": [ + 174, + 694, + 209, + 706 + ], + "score": 0.93, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 693, + 464, + 706 + ], + "score": 1.0, + "content": "is the common semantic information extracted from sentences", + "type": "text" + }, + { + "bbox": [ + 464, + 696, + 472, + 704 + ], + "score": 0.78, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 693, + 491, + 706 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 491, + 694, + 501, + 704 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 693, + 505, + 706 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 705, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 505, + 717 + ], + "score": 1.0, + "content": "while the two shadow parts are the extra bias. Note that the perfect debiased embeddings lead", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 716, + 505, + 729 + ], + "spans": [ + { + "bbox": [ + 105, + 716, + 261, + 729 + ], + "score": 1.0, + "content": "to coincident circles. By maximizing", + "type": "text" + }, + { + "bbox": [ + 262, + 716, + 283, + 727 + ], + "score": 0.83, + "content": "\\mathcal { T } _ { \\mathrm { N C E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 716, + 450, + 729 + ], + "score": 1.0, + "content": "term, we enlarge the overlapped area of", + "type": "text" + }, + { + "bbox": [ + 451, + 717, + 458, + 726 + ], + "score": 0.83, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 716, + 477, + 729 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 477, + 716, + 487, + 726 + ], + "score": 0.86, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 716, + 505, + 729 + ], + "score": 1.0, + "content": "; by", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 727, + 325, + 740 + ], + "spans": [ + { + "bbox": [ + 105, + 727, + 154, + 740 + ], + "score": 1.0, + "content": "minimizing", + "type": "text" + }, + { + "bbox": [ + 155, + 727, + 180, + 738 + ], + "score": 0.86, + "content": "\\scriptstyle { \\mathcal { T } } _ { \\mathrm { C L U B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 727, + 325, + 740 + ], + "score": 1.0, + "content": ", we shrink the biased shadow parts.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 80, + 485, + 203 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 80, + 485, + 203 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 80, + 485, + 203 + ], + "spans": [ + { + "bbox": [ + 126, + 80, + 485, + 203 + ], + "score": 0.973, + "type": "image", + "image_path": "8c1f7a0b55452733b6dca530e9279647bdf58ae19e81215c4ef29b41a8378f5b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 126, + 80, + 485, + 121.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 126, + 121.0, + 485, + 162.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 126, + 162.0, + 485, + 203.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 210, + 506, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 379, + 222 + ], + "score": 1.0, + "content": "Figure 1: (a) Contrastive learning framework of FairFil: Sentence", + "type": "text" + }, + { + "bbox": [ + 380, + 213, + 388, + 220 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 209, + 477, + 222 + ], + "score": 1.0, + "content": "and its augmentation", + "type": "text" + }, + { + "bbox": [ + 478, + 210, + 488, + 221 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 211, + 234 + ], + "score": 1.0, + "content": "encoded into embeddings", + "type": "text" + }, + { + "bbox": [ + 212, + 222, + 219, + 231 + ], + "score": 0.78, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 221, + 237, + 234 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 221, + 247, + 231 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 221, + 304, + 234 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + }, + { + "bbox": [ + 304, + 222, + 319, + 231 + ], + "score": 0.86, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "is the embedding of a sensitive attribute word", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 164, + 244 + ], + "score": 1.0, + "content": "selected from", + "type": "text" + }, + { + "bbox": [ + 164, + 234, + 172, + 243 + ], + "score": 0.56, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 233, + 176, + 244 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 177, + 232, + 198, + 243 + ], + "score": 0.85, + "content": "\\mathcal { T } _ { \\mathrm { N C E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 233, + 378, + 244 + ], + "score": 1.0, + "content": "maximizes the mutual information between", + "type": "text" + }, + { + "bbox": [ + 378, + 233, + 386, + 242 + ], + "score": 0.76, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 233, + 404, + 244 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 232, + 414, + 242 + ], + "score": 0.78, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 233, + 418, + 244 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 419, + 232, + 445, + 243 + ], + "score": 0.82, + "content": "\\scriptstyle { \\mathcal { T } } _ { \\mathrm { C L U B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 233, + 505, + 244 + ], + "score": 1.0, + "content": "eliminates the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 185, + 255 + ], + "score": 1.0, + "content": "bias information of", + "type": "text" + }, + { + "bbox": [ + 186, + 244, + 200, + 253 + ], + "score": 0.88, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 243, + 223, + 255 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 223, + 244, + 231, + 253 + ], + "score": 0.73, + "content": "\\mathbf { \\delta } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 243, + 366, + 255 + ], + "score": 1.0, + "content": ". (b) Illustration of information in", + "type": "text" + }, + { + "bbox": [ + 366, + 244, + 374, + 253 + ], + "score": 0.79, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 243, + 392, + 255 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 392, + 243, + 402, + 253 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 243, + 505, + 255 + ], + "score": 1.0, + "content": ": The blue and red circles", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 226, + 266 + ], + "score": 1.0, + "content": "represent the information in", + "type": "text" + }, + { + "bbox": [ + 226, + 255, + 234, + 264 + ], + "score": 0.77, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 254, + 254, + 266 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 254, + 254, + 264, + 264 + ], + "score": 0.84, + "content": "\\mathbf { { \\mathbf { { \\mathit { d } } } } ^ { \\prime } } \\mathbf { \\Sigma }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 254, + 505, + 266 + ], + "score": 1.0, + "content": ", respectively. The intersection is the mutual information", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 264, + 461, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 142, + 278 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 142, + 266, + 150, + 275 + ], + "score": 0.79, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 264, + 168, + 278 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 168, + 265, + 177, + 275 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 264, + 461, + 278 + ], + "score": 1.0, + "content": ". The shadow area represents the bias information of both embeddings.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 104, + 299, + 461, + 312 + ], + "lines": [], + "index": 9, + "bbox_fs": [ + 104, + 297, + 457, + 315 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 318, + 397, + 353 + ], + "lines": [ + { + "bbox": [ + 213, + 318, + 397, + 353 + ], + "spans": [ + { + "bbox": [ + 213, + 318, + 397, + 353 + ], + "score": 0.94, + "content": "\\mathcal { T } _ { \\mathrm { N C E } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log \\frac { \\exp ( g ( d _ { i } , d _ { i } ^ { \\prime } ) ) } { \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\exp ( g ( d _ { i } , d _ { j } ^ { \\prime } ) ) } .", + "type": "interline_equation", + "image_path": "fe4473099a6a0c44da9d37613a02a0710a1f912c63b5fdb0cf97a893bc1aa48d.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 213, + 318, + 397, + 335.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 213, + 335.5, + 397, + 353.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 359, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 163, + 371 + ], + "score": 1.0, + "content": "By maximize", + "type": "text" + }, + { + "bbox": [ + 163, + 360, + 185, + 370 + ], + "score": 0.89, + "content": "\\mathcal { T } _ { \\mathrm { N C E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 359, + 434, + 371 + ], + "score": 1.0, + "content": ", we encourage the difference between the positive pair score", + "type": "text" + }, + { + "bbox": [ + 434, + 359, + 471, + 371 + ], + "score": 0.92, + "content": "g ( d _ { i } , d _ { i } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 370, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 186, + 384 + ], + "score": 1.0, + "content": "negative pair score", + "type": "text" + }, + { + "bbox": [ + 186, + 370, + 224, + 384 + ], + "score": 0.93, + "content": "g ( d _ { i } , d _ { j } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 370, + 260, + 384 + ], + "score": 1.0, + "content": ", so that", + "type": "text" + }, + { + "bbox": [ + 260, + 371, + 271, + 381 + ], + "score": 0.87, + "content": "\\mathbf { \\ b { d } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 370, + 448, + 384 + ], + "score": 1.0, + "content": "can share more semantic information with", + "type": "text" + }, + { + "bbox": [ + 449, + 371, + 459, + 383 + ], + "score": 0.88, + "content": "\\mathbf { \\Delta } d _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 370, + 506, + 384 + ], + "score": 1.0, + "content": "than other", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 381, + 182, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 182, + 397 + ], + "score": 1.0, + "content": "embeddings d0j6=i.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 359, + 506, + 397 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 409, + 245, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 246, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 246, + 422 + ], + "score": 1.0, + "content": "3.3 DEBIASING REGULARIZER", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "Practically, the contrastive learning framework in Section 3.2 can already show encouraging debi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 442, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 423, + 454 + ], + "score": 1.0, + "content": "asing performance (as shown in the Experiments). However, the embedding", + "type": "text" + }, + { + "bbox": [ + 423, + 442, + 431, + 452 + ], + "score": 0.67, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 442, + 506, + 454 + ], + "score": 1.0, + "content": "can contain extra", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 206, + 465 + ], + "score": 1.0, + "content": "biased information from", + "type": "text" + }, + { + "bbox": [ + 206, + 454, + 213, + 462 + ], + "score": 0.72, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 452, + 305, + 465 + ], + "score": 1.0, + "content": ", that only maximizing", + "type": "text" + }, + { + "bbox": [ + 305, + 452, + 339, + 464 + ], + "score": 0.94, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "fails to eliminate. To encourage no extra", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 137, + 477 + ], + "score": 1.0, + "content": "bias in", + "type": "text" + }, + { + "bbox": [ + 137, + 464, + 144, + 473 + ], + "score": 0.74, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 463, + 506, + 477 + ], + "score": 1.0, + "content": ", we introduce a debiasing regularizer which minimizes the mutual information between", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 474, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 153, + 486 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 154, + 475, + 162, + 484 + ], + "score": 0.74, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 474, + 320, + 486 + ], + "score": 1.0, + "content": "and the potential bias from embedding", + "type": "text" + }, + { + "bbox": [ + 321, + 476, + 328, + 484 + ], + "score": 0.66, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 474, + 504, + 486 + ], + "score": 1.0, + "content": ". As discussed in Section 3.1, in our frame-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 211, + 497 + ], + "score": 1.0, + "content": "work the potential bias of", + "type": "text" + }, + { + "bbox": [ + 212, + 487, + 219, + 495 + ], + "score": 0.77, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 486, + 448, + 497 + ], + "score": 1.0, + "content": "is assumed to come from the sensitive attribute words in", + "type": "text" + }, + { + "bbox": [ + 448, + 487, + 456, + 495 + ], + "score": 0.68, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 486, + 505, + 497 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 497, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 429, + 508 + ], + "score": 1.0, + "content": "we should reduce the bias word information from the debiased representation", + "type": "text" + }, + { + "bbox": [ + 429, + 497, + 437, + 506 + ], + "score": 0.75, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 497, + 460, + 508 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 461, + 497, + 475, + 506 + ], + "score": 0.87, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 497, + 505, + 508 + ], + "score": 1.0, + "content": "be the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 267, + 519 + ], + "score": 1.0, + "content": "embedding of a sensitive attribute word", + "type": "text" + }, + { + "bbox": [ + 268, + 507, + 281, + 517 + ], + "score": 0.87, + "content": "w ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 506, + 329, + 519 + ], + "score": 1.0, + "content": "in sentence", + "type": "text" + }, + { + "bbox": [ + 330, + 509, + 337, + 517 + ], + "score": 0.71, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 506, + 430, + 519 + ], + "score": 1.0, + "content": ". The word embedding", + "type": "text" + }, + { + "bbox": [ + 431, + 507, + 445, + 517 + ], + "score": 0.86, + "content": "\\pmb { w } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "can always be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 308, + 530 + ], + "score": 1.0, + "content": "obtained from the pretrained text encoders (Bordia", + "type": "text" + }, + { + "bbox": [ + 308, + 519, + 317, + 528 + ], + "score": 0.3, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "Bowman, 2019). We then minimize the mutual", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 156, + 541 + ], + "score": 1.0, + "content": "information", + "type": "text" + }, + { + "bbox": [ + 157, + 529, + 195, + 541 + ], + "score": 0.91, + "content": "\\mathcal { T } ( w ^ { p } ; d )", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 528, + 505, + 541 + ], + "score": 1.0, + "content": ", using the CLUB mutual information upper bound (Cheng et al., 2020a) to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 102, + 535, + 508, + 557 + ], + "spans": [ + { + "bbox": [ + 102, + 535, + 142, + 557 + ], + "score": 1.0, + "content": "estimate", + "type": "text" + }, + { + "bbox": [ + 142, + 540, + 181, + 552 + ], + "score": 0.92, + "content": "\\mathcal { T } ( w ^ { p } ; d )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 535, + 428, + 557 + ], + "score": 1.0, + "content": "with embedding samples. Given a batch of embedding pairs", + "type": "text" + }, + { + "bbox": [ + 428, + 540, + 487, + 552 + ], + "score": 0.93, + "content": "\\bar { \\{ ( d _ { i } , \\boldsymbol { w } ^ { p } ) \\} } _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 535, + 508, + 557 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 551, + 275, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 275, + 563 + ], + "score": 1.0, + "content": "can calculate the debiasing regularizer as:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 21.5, + "bbox_fs": [ + 102, + 430, + 508, + 563 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 569, + 424, + 604 + ], + "lines": [ + { + "bbox": [ + 186, + 569, + 424, + 604 + ], + "spans": [ + { + "bbox": [ + 186, + 569, + 424, + 604 + ], + "score": 0.95, + "content": "\\mathcal { T } _ { \\mathrm { C L U B } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\Big [ \\log q _ { \\theta } ( { \\pmb w } _ { i } ^ { p } | \\pmb d _ { i } ) - \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\log q _ { \\theta } ( { \\pmb w } _ { j } ^ { p } | \\pmb d _ { i } ) \\Big ] ,", + "type": "interline_equation", + "image_path": "a63c40840b1faa97447066e2f554e7dfffa8be0097875d90080e8ba1df426f47.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 569, + 424, + 586.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 186, + 586.5, + 424, + 604.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 505, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 132, + 624 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 613, + 142, + 623 + ], + "score": 0.84, + "content": "q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 611, + 419, + 624 + ], + "score": 1.0, + "content": "is a variational approximation to ground-truth conditional distribution", + "type": "text" + }, + { + "bbox": [ + 420, + 611, + 450, + 623 + ], + "score": 0.9, + "content": "p ( \\pmb { w } | \\pmb { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 611, + 505, + 624 + ], + "score": 1.0, + "content": ". We parame-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 131, + 635 + ], + "score": 1.0, + "content": "terize", + "type": "text" + }, + { + "bbox": [ + 131, + 624, + 141, + 633 + ], + "score": 0.84, + "content": "q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 622, + 435, + 635 + ], + "score": 1.0, + "content": "with another neural network. As proved in Cheng et al. (2020a), the better", + "type": "text" + }, + { + "bbox": [ + 436, + 623, + 470, + 634 + ], + "score": 0.94, + "content": "q _ { \\theta } ( { \\pmb w } | { \\pmb d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "approx-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 135, + 646 + ], + "score": 1.0, + "content": "imates", + "type": "text" + }, + { + "bbox": [ + 135, + 633, + 166, + 645 + ], + "score": 0.93, + "content": "p ( \\pmb { w } | \\pmb { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 632, + 244, + 646 + ], + "score": 1.0, + "content": ", the more accurate", + "type": "text" + }, + { + "bbox": [ + 244, + 633, + 270, + 645 + ], + "score": 0.88, + "content": "\\scriptstyle { \\mathcal { T } } _ { \\mathrm { C L U B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "serves as the mutual information upper bound. Therefore,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 103, + 641, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 103, + 641, + 352, + 660 + ], + "score": 1.0, + "content": "besides the loss in (5), we also maximize the log-likelihood of", + "type": "text" + }, + { + "bbox": [ + 352, + 644, + 387, + 656 + ], + "score": 0.92, + "content": "q _ { \\theta } ( { \\pmb w } | { \\pmb d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 641, + 442, + 660 + ], + "score": 1.0, + "content": "with samples", + "type": "text" + }, + { + "bbox": [ + 443, + 644, + 501, + 657 + ], + "score": 0.92, + "content": "\\{ ( d _ { i } , { \\boldsymbol { w } } _ { i } ^ { p } ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 641, + 506, + 660 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 103, + 611, + 506, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 660, + 505, + 739 + ], + "lines": [ + { + "bbox": [ + 105, + 661, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 504, + 673 + ], + "score": 1.0, + "content": "Based on the above sections, the overall learning scheme of our fair filter (FairFil) is described in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 673, + 504, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 504, + 684 + ], + "score": 1.0, + "content": "Algorithm 1. Also, we provide an intuitive explanation to the two loss terms in our framework.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 308, + 695 + ], + "score": 1.0, + "content": "In Figure 1(b), the blue and red circles represent", + "type": "text" + }, + { + "bbox": [ + 309, + 684, + 317, + 693 + ], + "score": 0.8, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 682, + 336, + 695 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 336, + 683, + 345, + 693 + ], + "score": 0.85, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 682, + 505, + 695 + ], + "score": 1.0, + "content": ", respectively, in the embedding space.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 693, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 174, + 706 + ], + "score": 1.0, + "content": "The intersection", + "type": "text" + }, + { + "bbox": [ + 174, + 694, + 209, + 706 + ], + "score": 0.93, + "content": "\\mathcal { T } ( d ; d ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 693, + 464, + 706 + ], + "score": 1.0, + "content": "is the common semantic information extracted from sentences", + "type": "text" + }, + { + "bbox": [ + 464, + 696, + 472, + 704 + ], + "score": 0.78, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 693, + 491, + 706 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 491, + 694, + 501, + 704 + ], + "score": 0.86, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 693, + 505, + 706 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 705, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 505, + 717 + ], + "score": 1.0, + "content": "while the two shadow parts are the extra bias. Note that the perfect debiased embeddings lead", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 716, + 505, + 729 + ], + "spans": [ + { + "bbox": [ + 105, + 716, + 261, + 729 + ], + "score": 1.0, + "content": "to coincident circles. By maximizing", + "type": "text" + }, + { + "bbox": [ + 262, + 716, + 283, + 727 + ], + "score": 0.83, + "content": "\\mathcal { T } _ { \\mathrm { N C E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 716, + 450, + 729 + ], + "score": 1.0, + "content": "term, we enlarge the overlapped area of", + "type": "text" + }, + { + "bbox": [ + 451, + 717, + 458, + 726 + ], + "score": 0.83, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 716, + 477, + 729 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 477, + 716, + 487, + 726 + ], + "score": 0.86, + "content": "\\pmb { d } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 716, + 505, + 729 + ], + "score": 1.0, + "content": "; by", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 727, + 325, + 740 + ], + "spans": [ + { + "bbox": [ + 105, + 727, + 154, + 740 + ], + "score": 1.0, + "content": "minimizing", + "type": "text" + }, + { + "bbox": [ + 155, + 727, + 180, + 738 + ], + "score": 0.86, + "content": "\\scriptstyle { \\mathcal { T } } _ { \\mathrm { C L U B } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 727, + 325, + 740 + ], + "score": 1.0, + "content": ", we shrink the biased shadow parts.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 661, + 505, + 740 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 92, + 503, + 245 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 82, + 326, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 81, + 327, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 327, + 96 + ], + "score": 1.0, + "content": "Algorithm 1 Updating the FairFil with a sample batch", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 110, + 92, + 503, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 92, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 110, + 92, + 503, + 245 + ], + "score": 0.899, + "html": "
Begin with the pretrained text encoder E()and a batch of sentences {x1. Find the sensitive attribute words {wP} and corresponding embeddings {wP}.
Generate augmentation x' from xi,by replacing {wP} with {rj(wp)}.
Encode (xi,x) into embeddings di=f(E(xi),d𝑖= f(E(x')).
Calculate INcE with {(di,di)}=1 and score function g. if adding debiasing regularizer then
Update the variational approximation qe(w|d) by maximizing log-likelihood with {(di,w )}
Calculate IcLuB with qe(wld) and {(di,w)}1
Learning loss L= -INcE + βIcLUB· else
Learning loss L = -INcE·
end if
Update FairFil f and score function g by gradient descent with respect to L.
", + "type": "table", + "image_path": "03a1adeb3cc52d169b4f8ecb0c39e58cde5e3c402f96562d90cae020d0fa9fb9.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 92, + 503, + 143.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 143.0, + 503, + 194.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 194.0, + 503, + 245.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "title", + "bbox": [ + 108, + 265, + 211, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 213, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 213, + 280 + ], + "score": 1.0, + "content": "4 RELATED WORK", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 287, + 317, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 287, + 318, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 318, + 299 + ], + "score": 1.0, + "content": "4.1 BIAS IN NATURAL LANGUAGE PROCESSING", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "Social bias has recently been recognized as an important issue in natural language processing (NLP)", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "systems. The studies on bias in NLP are mainly delineated into two categories: bias in the embed-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "ding spaces, and bias in downstream tasks (Blodgett et al., 2020). For bias in downstream tasks, the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "analyses cover comprehensive topics, including machine translation (Stanovsky et al., 2019), lan-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "guage modeling (Bordia & Bowman, 2019), sentiment analysis (Kiritchenko & Mohammad, 2018)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "and toxicity detection (Dixon et al., 2018). The social bias in embedding spaces has been stud-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "ied from two important perspectives: bias measurements and and debiasing methods. To measure", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "the bias in an embedding space, Caliskan et al. (2017) proposed a Word Embedding Association", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "Test (WEAT), which compares the similarity between two sets of target words and two sets of at-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "score": 1.0, + "content": "tribute words. May et al. (2019) further extended the WEAT to a Sentence Encoder Association Test", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "(SEAT), which replaces the word embeddings by sentence embeddings encoded from pre-defined", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 424, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 505, + 439 + ], + "score": 1.0, + "content": "biased sentence templates. For debiasing methods, most of the prior works focus on word-level rep-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 435, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 505, + 451 + ], + "score": 1.0, + "content": "resentations (Bolukbasi et al., 2016; Bordia & Bowman, 2019). The only sentence-level debiasing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "method is proposed by Liang et al. (2020), which learns bias directions by PCA and subtracts them", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 459, + 206, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 206, + 471 + ], + "score": 1.0, + "content": "in the embedding space.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 107, + 479, + 243, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 243, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 243, + 492 + ], + "score": 1.0, + "content": "4.2 CONTRASTIVE LEARNING", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "Contrastive learning is a broad class of training strategies that learns meaningful representations by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "score": 1.0, + "content": "making positive and negative embedding pairs more distinguishable. Usually, contrastive learning", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "requires a pairwise embedding critic as a similarity/distance of data pairs. Then the learning ob-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "jective is constructed by maximizing the margin between the critic values of positive data pairs and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "negative data pairs. Previously contrastive learning has shown encouraging performance in many", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "tasks, including metric learning (Weinberger et al., 2006; Davis et al., 2007), word representation", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "learning (Mikolov et al., 2013), graph learning (Tang et al., 2015; Grover & Leskovec, 2016), etc.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 570, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 505, + 587 + ], + "score": 1.0, + "content": "Recently, contrastive learning has been applied to the unsupervised visual representation learning", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "task, and significantly reduced the performance gap between supervised and unsupervised learn-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "score": 1.0, + "content": "ing (He et al., 2020; Chen et al., 2020; Qian et al., 2020). Among these unsupervised methods, Chen", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "et al. (2020) proposed a simple multi-view contrastive learning framework (SimCLR). For each im-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "age data, SimCLR generates two augmented images, and then the mutual information of the two", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 394, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 394, + 640 + ], + "score": 1.0, + "content": "augmentation embeddings is maximized within a batch of training data.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 200, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 647, + 201, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 201, + 663 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "We first describe the experimental setup in detail, including the pretrained encoders, the training of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "FairFil, and the downstream tasks. The results of our FairFil are reported and analyzed, along with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "the previous Sent-Debias method. In general, we evaluate our neural debiasing method from two", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "perspectives: (1) fairness: we compare the bias degree of the original and debiased sentence embed-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "score": 1.0, + "content": "dings for debiasing performance; and (2) representativeness: we apply the debiased embeddings", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 721, + 429, + 736 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 429, + 736 + ], + "score": 1.0, + "content": "into downstream tasks, and compare the performance with original embeddings.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 92, + 503, + 245 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 82, + 326, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 81, + 327, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 327, + 96 + ], + "score": 1.0, + "content": "Algorithm 1 Updating the FairFil with a sample batch", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 110, + 92, + 503, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 92, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 110, + 92, + 503, + 245 + ], + "score": 0.899, + "html": "
Begin with the pretrained text encoder E()and a batch of sentences {x1. Find the sensitive attribute words {wP} and corresponding embeddings {wP}.
Generate augmentation x' from xi,by replacing {wP} with {rj(wp)}.
Encode (xi,x) into embeddings di=f(E(xi),d𝑖= f(E(x')).
Calculate INcE with {(di,di)}=1 and score function g. if adding debiasing regularizer then
Update the variational approximation qe(w|d) by maximizing log-likelihood with {(di,w )}
Calculate IcLuB with qe(wld) and {(di,w)}1
Learning loss L= -INcE + βIcLUB· else
Learning loss L = -INcE·
end if
Update FairFil f and score function g by gradient descent with respect to L.
", + "type": "table", + "image_path": "03a1adeb3cc52d169b4f8ecb0c39e58cde5e3c402f96562d90cae020d0fa9fb9.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 92, + 503, + 143.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 143.0, + 503, + 194.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 194.0, + 503, + 245.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "title", + "bbox": [ + 108, + 265, + 211, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 213, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 213, + 280 + ], + "score": 1.0, + "content": "4 RELATED WORK", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 287, + 317, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 287, + 318, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 318, + 299 + ], + "score": 1.0, + "content": "4.1 BIAS IN NATURAL LANGUAGE PROCESSING", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "Social bias has recently been recognized as an important issue in natural language processing (NLP)", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "systems. The studies on bias in NLP are mainly delineated into two categories: bias in the embed-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "ding spaces, and bias in downstream tasks (Blodgett et al., 2020). For bias in downstream tasks, the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "analyses cover comprehensive topics, including machine translation (Stanovsky et al., 2019), lan-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "guage modeling (Bordia & Bowman, 2019), sentiment analysis (Kiritchenko & Mohammad, 2018)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "and toxicity detection (Dixon et al., 2018). The social bias in embedding spaces has been stud-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "ied from two important perspectives: bias measurements and and debiasing methods. To measure", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "the bias in an embedding space, Caliskan et al. (2017) proposed a Word Embedding Association", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "Test (WEAT), which compares the similarity between two sets of target words and two sets of at-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "score": 1.0, + "content": "tribute words. May et al. (2019) further extended the WEAT to a Sentence Encoder Association Test", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "(SEAT), which replaces the word embeddings by sentence embeddings encoded from pre-defined", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 424, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 505, + 439 + ], + "score": 1.0, + "content": "biased sentence templates. For debiasing methods, most of the prior works focus on word-level rep-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 435, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 505, + 451 + ], + "score": 1.0, + "content": "resentations (Bolukbasi et al., 2016; Bordia & Bowman, 2019). The only sentence-level debiasing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "method is proposed by Liang et al. (2020), which learns bias directions by PCA and subtracts them", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 459, + 206, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 206, + 471 + ], + "score": 1.0, + "content": "in the embedding space.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13, + "bbox_fs": [ + 104, + 304, + 505, + 471 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 479, + 243, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 243, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 243, + 492 + ], + "score": 1.0, + "content": "4.2 CONTRASTIVE LEARNING", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "Contrastive learning is a broad class of training strategies that learns meaningful representations by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "score": 1.0, + "content": "making positive and negative embedding pairs more distinguishable. Usually, contrastive learning", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "requires a pairwise embedding critic as a similarity/distance of data pairs. Then the learning ob-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "jective is constructed by maximizing the margin between the critic values of positive data pairs and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "negative data pairs. Previously contrastive learning has shown encouraging performance in many", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "tasks, including metric learning (Weinberger et al., 2006; Davis et al., 2007), word representation", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "learning (Mikolov et al., 2013), graph learning (Tang et al., 2015; Grover & Leskovec, 2016), etc.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 570, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 505, + 587 + ], + "score": 1.0, + "content": "Recently, contrastive learning has been applied to the unsupervised visual representation learning", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "task, and significantly reduced the performance gap between supervised and unsupervised learn-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "score": 1.0, + "content": "ing (He et al., 2020; Chen et al., 2020; Qian et al., 2020). Among these unsupervised methods, Chen", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "et al. (2020) proposed a simple multi-view contrastive learning framework (SimCLR). For each im-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "age data, SimCLR generates two augmented images, and then the mutual information of the two", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 394, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 394, + 640 + ], + "score": 1.0, + "content": "augmentation embeddings is maximized within a batch of training data.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 495, + 506, + 640 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 200, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 647, + 201, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 201, + 663 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "We first describe the experimental setup in detail, including the pretrained encoders, the training of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "FairFil, and the downstream tasks. The results of our FairFil are reported and analyzed, along with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "the previous Sent-Debias method. In general, we evaluate our neural debiasing method from two", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "perspectives: (1) fairness: we compare the bias degree of the original and debiased sentence embed-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 725 + ], + "score": 1.0, + "content": "dings for debiasing performance; and (2) representativeness: we apply the debiased embeddings", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 721, + 429, + 736 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 429, + 736 + ], + "score": 1.0, + "content": "into downstream tasks, and compare the performance with original embeddings.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 667, + 506, + 736 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 249, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 250, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 250, + 95 + ], + "score": 1.0, + "content": "5.1 BIAS EVALUATION METRIC", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 98, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 99, + 506, + 110 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 506, + 110 + ], + "score": 1.0, + "content": "To evaluate the bias in sentence embeddings, we use the Sentence Encoder Association Test", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 110, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 122 + ], + "score": 1.0, + "content": "(SEAT) (May et al., 2019), which is an extension of the Word Embedding Association Test", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "(WEAT) (Caliskan et al., 2017). The WEAT test measures the bias in word embeddings by com-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "paring the distances of two sets of target words to two sets of attribute words. More specifically,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 136, + 156 + ], + "score": 1.0, + "content": "denote", + "type": "text" + }, + { + "bbox": [ + 136, + 143, + 146, + 153 + ], + "score": 0.83, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 142, + 166, + 156 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 143, + 176, + 153 + ], + "score": 0.82, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 142, + 362, + 156 + ], + "score": 1.0, + "content": "as two sets of target word embeddings (e.g.,", + "type": "text" + }, + { + "bbox": [ + 362, + 143, + 372, + 153 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "includes “male” words such as", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 180, + 166 + ], + "score": 1.0, + "content": "“boy” and “man”;", + "type": "text" + }, + { + "bbox": [ + 180, + 154, + 190, + 165 + ], + "score": 0.81, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 152, + 469, + 166 + ], + "score": 1.0, + "content": "contains “female” words like “girl” and “woman”). The attribute sets", + "type": "text" + }, + { + "bbox": [ + 469, + 154, + 478, + 164 + ], + "score": 0.81, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 152, + 495, + 166 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 496, + 154, + 504, + 164 + ], + "score": 0.8, + "content": "\\boldsymbol { B }", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 362, + 177 + ], + "score": 1.0, + "content": "are selected from some social concepts that should be “equal” to", + "type": "text" + }, + { + "bbox": [ + 362, + 165, + 372, + 174 + ], + "score": 0.85, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 165, + 389, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 390, + 165, + 399, + 176 + ], + "score": 0.83, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 165, + 504, + 177 + ], + "score": 1.0, + "content": "(e.g., career or personality", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 475, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 284, + 188 + ], + "score": 1.0, + "content": "words). Then the bias degree w.r.t attributes", + "type": "text" + }, + { + "bbox": [ + 284, + 176, + 311, + 187 + ], + "score": 0.93, + "content": "( A , B )", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 176, + 414, + 188 + ], + "score": 1.0, + "content": "of each word embedding", + "type": "text" + }, + { + "bbox": [ + 414, + 177, + 419, + 185 + ], + "score": 0.74, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 176, + 475, + 188 + ], + "score": 1.0, + "content": "is defined as:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 192, + 415, + 206 + ], + "lines": [ + { + "bbox": [ + 195, + 192, + 415, + 206 + ], + "spans": [ + { + "bbox": [ + 195, + 192, + 415, + 206 + ], + "score": 0.88, + "content": "s ( t , \\mathcal { A } , \\mathcal { B } ) = \\mathrm { m e a n } _ { a \\in \\mathcal { A } } \\cos ( t , a ) - \\mathrm { m e a n } _ { b \\in \\mathcal { B } } \\cos ( t , b ) ,", + "type": "interline_equation", + "image_path": "e34930275b800071a949db344a09873492ef9a369806dcd04f9678b87e84d1b2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 195, + 192, + 415, + 206 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 465, + 223 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 465, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 133, + 224 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 211, + 165, + 223 + ], + "score": 0.91, + "content": "\\cos ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 210, + 465, + 224 + ], + "score": 1.0, + "content": "is the cosine similarity. Based on (6), the normalized WEAT effect size is:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 228, + 416, + 256 + ], + "lines": [ + { + "bbox": [ + 195, + 228, + 416, + 256 + ], + "spans": [ + { + "bbox": [ + 195, + 228, + 416, + 256 + ], + "score": 0.92, + "content": "d _ { \\mathrm { W E A T } } = \\frac { \\mathrm { m e a n } _ { x \\in \\mathcal { X } } s ( x , \\mathcal { A } , \\mathcal { B } ) - \\mathrm { m e a n } _ { y \\in \\mathcal { Y } } s ( y , \\mathcal { A } , \\mathcal { B } ) } { \\mathrm { s t d } _ { t \\in \\mathcal { X } \\cup \\mathcal { Y } } s ( t , \\mathcal { A } , \\mathcal { B } ) } .", + "type": "interline_equation", + "image_path": "681a0f99acde358feb5acbd3dcc422754a639d34602fcd02bf4ec03d0d0ff340.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 195, + 228, + 416, + 242.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 195, + 242.0, + 416, + 256.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 260, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 272 + ], + "score": 1.0, + "content": "The SEAT test extends WEAT by replacing the word embeddings with sentence embeddings. Both", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "target words and attribute words are converted into sentences with several semantically bleached", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 243, + 294 + ], + "score": 1.0, + "content": "sentence templates (e.g., “This is", + "type": "text" + }, + { + "bbox": [ + 243, + 283, + 286, + 293 + ], + "score": 0.27, + "content": "{ < } \\mathrm { w o r d } { > } ^ { \\mathrm { w } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "). Then the SEAT statistic is similarly calculated with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "score": 1.0, + "content": "(7) based on the embeddings of converted sentences. The closer the effect size is to zero, the more", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 466, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 466, + 316 + ], + "score": 1.0, + "content": "fair the embeddings are. Therefore, we report the absolute effect size as the bias measure.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 107, + 325, + 237, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 239, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 239, + 338 + ], + "score": 1.0, + "content": "5.2 PRETRAINED ENCODERS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 341, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "We test our neural debiasing method on BERT (Devlin et al., 2019). Since the pretrained BERT", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "requires the additional fine-tuning process for downstream tasks, we report the performance of our", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "FairFil under two scenarios: (1) pretrained BERT: we directly learn our FairFil network based", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "on pretrained BERT without any additional fine-tuning; and (2) BERT post tasks: we fix the pa-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 470, + 397 + ], + "score": 1.0, + "content": "rameters of the FairFil network learned on pretrained BERT, and then fine-tune the BERT", + "type": "text" + }, + { + "bbox": [ + 470, + 386, + 477, + 395 + ], + "score": 0.44, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "FairFil", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "together on task-specific data. Note that when fine-tuning, our FairFil will no longer update, which", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 407, + 355, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 355, + 420 + ], + "score": 1.0, + "content": "satisfies a fair comparison to Sent-Debias (Liang et al., 2020).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 504, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 504, + 436 + ], + "score": 1.0, + "content": "For the downstream tasks of BERT, we follow the setup from Sent-Debias (Liang et al., 2020) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "conduct experiments on the following three downstream tasks: (1) SST-2: A sentiment classifi-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "cation task on the Stanford Sentiment Treebank (SST-2) dataset (Socher et al., 2013), on which", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "sentence embeddings are used to predict the corresponding sentiment labels; (2) CoLA: Another", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "sentiment classification task on the Corpus of Linguistic Acceptability (CoLA) grammatical accept-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "ability judgment (Warstadt et al., 2019); and (3) QNLI: A binary question answering task on the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 489, + 403, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 403, + 502 + ], + "score": 1.0, + "content": "Question Natural Language Inference (QNLI) dataset (Wang et al., 2018).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 108, + 510, + 226, + 521 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 228, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 228, + 523 + ], + "score": 1.0, + "content": "5.3 TRAINING OF FAIRFIL", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "score": 1.0, + "content": "We parameterize the fair filter network with one-layer fully-connected neural networks with the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 301, + 550 + ], + "score": 1.0, + "content": "ReLU activation function. The score function", + "type": "text" + }, + { + "bbox": [ + 301, + 540, + 308, + 550 + ], + "score": 0.78, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "in the InfoNCE estimator is set to a two-layer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 453, + 561 + ], + "score": 1.0, + "content": "fully-connected network with one-dimensional output. The variational approximation", + "type": "text" + }, + { + "bbox": [ + 453, + 550, + 464, + 560 + ], + "score": 0.79, + "content": "q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "in CLUB", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 381, + 573 + ], + "score": 1.0, + "content": "estimator is parameterized by a multi-variate Gaussian distribution", + "type": "text" + }, + { + "bbox": [ + 382, + 560, + 501, + 572 + ], + "score": 0.92, + "content": "q _ { \\theta } ( w | d ) = N ( \\mu ( d ) , \\sigma ^ { 2 } ( d ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 559, + 505, + 573 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 133, + 584 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 570, + 153, + 583 + ], + "score": 0.91, + "content": "\\mu ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 570, + 172, + 584 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 173, + 570, + 191, + 582 + ], + "score": 0.9, + "content": "\\sigma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 570, + 506, + 584 + ], + "score": 1.0, + "content": "are also two-layer fully-connected neural nets. The batch size is set to 128.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 580, + 378, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 186, + 595 + ], + "score": 1.0, + "content": "The learning rate is", + "type": "text" + }, + { + "bbox": [ + 186, + 581, + 225, + 592 + ], + "score": 0.91, + "content": "1 \\times 1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 580, + 378, + 595 + ], + "score": 1.0, + "content": ". We train the fair filter for 10 epochs.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "For an appropriate comparison, we follow the setup of Sent-Debias (Liang et al., 2020) and select the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "same training data for the training of FairFil. The training corpora consist 183,060 sentences from", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 620, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 634 + ], + "score": 1.0, + "content": "the following five datasets: WikiText-2 (Merity et al., 201y), Stanford Sentiment Treebank (Socher", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "et al., 2013), Reddit (V”olske et al., 2017), MELD (Poria et al., 2019) and POM (Park et al., 2014).", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 422, + 655 + ], + "score": 1.0, + "content": "Following Liang et al. (2020), we mainly select “gender” as the sensitive topic", + "type": "text" + }, + { + "bbox": [ + 422, + 643, + 432, + 653 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 642, + 506, + 655 + ], + "score": 1.0, + "content": ", and use the same", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "pre-defined word sets of sensitive attribute words and their replaceable words as Sent-Debias did.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "The word embeddings for training the debiasing regularizer is selected from the token embedding", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 675, + 204, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 204, + 687 + ], + "score": 1.0, + "content": "of the pretrained BERT.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + }, + { + "type": "title", + "bbox": [ + 107, + 693, + 222, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 223, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 223, + 706 + ], + "score": 1.0, + "content": "5.4 DEBIASING RESULTS", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "In Tables 2 and 3 we report the evaluation results of debiased embeddings on both the absolute", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "SEAT effect size and the downstream classification accuracy. For the SEAT test, we follow the", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 249, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 250, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 250, + 95 + ], + "score": 1.0, + "content": "5.1 BIAS EVALUATION METRIC", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 98, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 99, + 506, + 110 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 506, + 110 + ], + "score": 1.0, + "content": "To evaluate the bias in sentence embeddings, we use the Sentence Encoder Association Test", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 110, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 122 + ], + "score": 1.0, + "content": "(SEAT) (May et al., 2019), which is an extension of the Word Embedding Association Test", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "(WEAT) (Caliskan et al., 2017). The WEAT test measures the bias in word embeddings by com-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "paring the distances of two sets of target words to two sets of attribute words. More specifically,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 136, + 156 + ], + "score": 1.0, + "content": "denote", + "type": "text" + }, + { + "bbox": [ + 136, + 143, + 146, + 153 + ], + "score": 0.83, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 142, + 166, + 156 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 143, + 176, + 153 + ], + "score": 0.82, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 142, + 362, + 156 + ], + "score": 1.0, + "content": "as two sets of target word embeddings (e.g.,", + "type": "text" + }, + { + "bbox": [ + 362, + 143, + 372, + 153 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "includes “male” words such as", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 180, + 166 + ], + "score": 1.0, + "content": "“boy” and “man”;", + "type": "text" + }, + { + "bbox": [ + 180, + 154, + 190, + 165 + ], + "score": 0.81, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 152, + 469, + 166 + ], + "score": 1.0, + "content": "contains “female” words like “girl” and “woman”). The attribute sets", + "type": "text" + }, + { + "bbox": [ + 469, + 154, + 478, + 164 + ], + "score": 0.81, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 152, + 495, + 166 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 496, + 154, + 504, + 164 + ], + "score": 0.8, + "content": "\\boldsymbol { B }", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 362, + 177 + ], + "score": 1.0, + "content": "are selected from some social concepts that should be “equal” to", + "type": "text" + }, + { + "bbox": [ + 362, + 165, + 372, + 174 + ], + "score": 0.85, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 165, + 389, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 390, + 165, + 399, + 176 + ], + "score": 0.83, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 165, + 504, + 177 + ], + "score": 1.0, + "content": "(e.g., career or personality", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 475, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 284, + 188 + ], + "score": 1.0, + "content": "words). Then the bias degree w.r.t attributes", + "type": "text" + }, + { + "bbox": [ + 284, + 176, + 311, + 187 + ], + "score": 0.93, + "content": "( A , B )", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 176, + 414, + 188 + ], + "score": 1.0, + "content": "of each word embedding", + "type": "text" + }, + { + "bbox": [ + 414, + 177, + 419, + 185 + ], + "score": 0.74, + "content": "\\pmb { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 176, + 475, + 188 + ], + "score": 1.0, + "content": "is defined as:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 99, + 506, + 188 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 192, + 415, + 206 + ], + "lines": [ + { + "bbox": [ + 195, + 192, + 415, + 206 + ], + "spans": [ + { + "bbox": [ + 195, + 192, + 415, + 206 + ], + "score": 0.88, + "content": "s ( t , \\mathcal { A } , \\mathcal { B } ) = \\mathrm { m e a n } _ { a \\in \\mathcal { A } } \\cos ( t , a ) - \\mathrm { m e a n } _ { b \\in \\mathcal { B } } \\cos ( t , b ) ,", + "type": "interline_equation", + "image_path": "e34930275b800071a949db344a09873492ef9a369806dcd04f9678b87e84d1b2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 195, + 192, + 415, + 206 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 465, + 223 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 465, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 133, + 224 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 211, + 165, + 223 + ], + "score": 0.91, + "content": "\\cos ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 210, + 465, + 224 + ], + "score": 1.0, + "content": "is the cosine similarity. Based on (6), the normalized WEAT effect size is:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 210, + 465, + 224 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 228, + 416, + 256 + ], + "lines": [ + { + "bbox": [ + 195, + 228, + 416, + 256 + ], + "spans": [ + { + "bbox": [ + 195, + 228, + 416, + 256 + ], + "score": 0.92, + "content": "d _ { \\mathrm { W E A T } } = \\frac { \\mathrm { m e a n } _ { x \\in \\mathcal { X } } s ( x , \\mathcal { A } , \\mathcal { B } ) - \\mathrm { m e a n } _ { y \\in \\mathcal { Y } } s ( y , \\mathcal { A } , \\mathcal { B } ) } { \\mathrm { s t d } _ { t \\in \\mathcal { X } \\cup \\mathcal { Y } } s ( t , \\mathcal { A } , \\mathcal { B } ) } .", + "type": "interline_equation", + "image_path": "681a0f99acde358feb5acbd3dcc422754a639d34602fcd02bf4ec03d0d0ff340.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 195, + 228, + 416, + 242.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 195, + 242.0, + 416, + 256.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 260, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 272 + ], + "score": 1.0, + "content": "The SEAT test extends WEAT by replacing the word embeddings with sentence embeddings. Both", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "target words and attribute words are converted into sentences with several semantically bleached", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 243, + 294 + ], + "score": 1.0, + "content": "sentence templates (e.g., “This is", + "type": "text" + }, + { + "bbox": [ + 243, + 283, + 286, + 293 + ], + "score": 0.27, + "content": "{ < } \\mathrm { w o r d } { > } ^ { \\mathrm { w } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "). Then the SEAT statistic is similarly calculated with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "score": 1.0, + "content": "(7) based on the embeddings of converted sentences. The closer the effect size is to zero, the more", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 466, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 466, + 316 + ], + "score": 1.0, + "content": "fair the embeddings are. Therefore, we report the absolute effect size as the bias measure.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 261, + 506, + 316 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 325, + 237, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 239, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 239, + 338 + ], + "score": 1.0, + "content": "5.2 PRETRAINED ENCODERS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 341, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "We test our neural debiasing method on BERT (Devlin et al., 2019). Since the pretrained BERT", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "requires the additional fine-tuning process for downstream tasks, we report the performance of our", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "FairFil under two scenarios: (1) pretrained BERT: we directly learn our FairFil network based", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "on pretrained BERT without any additional fine-tuning; and (2) BERT post tasks: we fix the pa-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 470, + 397 + ], + "score": 1.0, + "content": "rameters of the FairFil network learned on pretrained BERT, and then fine-tune the BERT", + "type": "text" + }, + { + "bbox": [ + 470, + 386, + 477, + 395 + ], + "score": 0.44, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "FairFil", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "together on task-specific data. Note that when fine-tuning, our FairFil will no longer update, which", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 407, + 355, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 355, + 420 + ], + "score": 1.0, + "content": "satisfies a fair comparison to Sent-Debias (Liang et al., 2020).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 340, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 504, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 504, + 436 + ], + "score": 1.0, + "content": "For the downstream tasks of BERT, we follow the setup from Sent-Debias (Liang et al., 2020) and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "conduct experiments on the following three downstream tasks: (1) SST-2: A sentiment classifi-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "cation task on the Stanford Sentiment Treebank (SST-2) dataset (Socher et al., 2013), on which", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "sentence embeddings are used to predict the corresponding sentiment labels; (2) CoLA: Another", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "sentiment classification task on the Corpus of Linguistic Acceptability (CoLA) grammatical accept-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "ability judgment (Warstadt et al., 2019); and (3) QNLI: A binary question answering task on the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 489, + 403, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 403, + 502 + ], + "score": 1.0, + "content": "Question Natural Language Inference (QNLI) dataset (Wang et al., 2018).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 425, + 506, + 502 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 510, + 226, + 521 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 228, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 228, + 523 + ], + "score": 1.0, + "content": "5.3 TRAINING OF FAIRFIL", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 539 + ], + "score": 1.0, + "content": "We parameterize the fair filter network with one-layer fully-connected neural networks with the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 301, + 550 + ], + "score": 1.0, + "content": "ReLU activation function. The score function", + "type": "text" + }, + { + "bbox": [ + 301, + 540, + 308, + 550 + ], + "score": 0.78, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "in the InfoNCE estimator is set to a two-layer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 453, + 561 + ], + "score": 1.0, + "content": "fully-connected network with one-dimensional output. The variational approximation", + "type": "text" + }, + { + "bbox": [ + 453, + 550, + 464, + 560 + ], + "score": 0.79, + "content": "q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "in CLUB", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 381, + 573 + ], + "score": 1.0, + "content": "estimator is parameterized by a multi-variate Gaussian distribution", + "type": "text" + }, + { + "bbox": [ + 382, + 560, + 501, + 572 + ], + "score": 0.92, + "content": "q _ { \\theta } ( w | d ) = N ( \\mu ( d ) , \\sigma ^ { 2 } ( d ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 559, + 505, + 573 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 570, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 133, + 584 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 570, + 153, + 583 + ], + "score": 0.91, + "content": "\\mu ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 570, + 172, + 584 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 173, + 570, + 191, + 582 + ], + "score": 0.9, + "content": "\\sigma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 570, + 506, + 584 + ], + "score": 1.0, + "content": "are also two-layer fully-connected neural nets. The batch size is set to 128.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 580, + 378, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 186, + 595 + ], + "score": 1.0, + "content": "The learning rate is", + "type": "text" + }, + { + "bbox": [ + 186, + 581, + 225, + 592 + ], + "score": 0.91, + "content": "1 \\times 1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 580, + 378, + 595 + ], + "score": 1.0, + "content": ". We train the fair filter for 10 epochs.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 526, + 506, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "For an appropriate comparison, we follow the setup of Sent-Debias (Liang et al., 2020) and select the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "same training data for the training of FairFil. The training corpora consist 183,060 sentences from", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 620, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 634 + ], + "score": 1.0, + "content": "the following five datasets: WikiText-2 (Merity et al., 201y), Stanford Sentiment Treebank (Socher", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "et al., 2013), Reddit (V”olske et al., 2017), MELD (Poria et al., 2019) and POM (Park et al., 2014).", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 422, + 655 + ], + "score": 1.0, + "content": "Following Liang et al. (2020), we mainly select “gender” as the sensitive topic", + "type": "text" + }, + { + "bbox": [ + 422, + 643, + 432, + 653 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 642, + 506, + 655 + ], + "score": 1.0, + "content": ", and use the same", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "pre-defined word sets of sensitive attribute words and their replaceable words as Sent-Debias did.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "The word embeddings for training the debiasing regularizer is selected from the token embedding", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 675, + 204, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 204, + 687 + ], + "score": 1.0, + "content": "of the pretrained BERT.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 599, + 506, + 687 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 693, + 222, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 223, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 223, + 706 + ], + "score": 1.0, + "content": "5.4 DEBIASING RESULTS", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "In Tables 2 and 3 we report the evaluation results of debiased embeddings on both the absolute", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "SEAT effect size and the downstream classification accuracy. For the SEAT test, we follow the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "setup in Liang et al. (2020), and test the sentence templates of Terms/Names under different domains", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 385, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 504, + 396 + ], + "score": 1.0, + "content": "designed by Caliskan et al. (2017). The column name Origin refers to the original BERT results,", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 433, + 408 + ], + "score": 1.0, + "content": "and Sent-D is short for Sent-Debias (Liang et al., 2020). FairFil− and FairFil (as", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 433, + 396, + 462, + 406 + ], + "score": 0.27, + "content": "\\mathrm { F a i r F ^ { - } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 462, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "and FairF", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "in the tables) are our method without/with the debiasing regularizer in Section 3.3. The best results", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 416, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 431 + ], + "score": 1.0, + "content": "of effect size (the lower the better) and classification accuracy (the higher the better) are bold among", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Sent-D, FairFil−, and FairFil. Since the pretrained BERT does not correspond to any downstream", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 439, + 319, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 319, + 452 + ], + "score": 1.0, + "content": "task, the classification accuracy is not reported for it.", + "type": "text", + "cross_page": true + } + ], + "index": 14 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 97, + 505, + 211 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 123, + 80, + 485, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 79, + 487, + 93 + ], + "spans": [ + { + "bbox": [ + 122, + 79, + 487, + 93 + ], + "score": 1.0, + "content": "Table 2: Performance of debiased embeddings on Pretrained BERT and BERT post SST-2.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 109, + 97, + 505, + 211 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 97, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 109, + 97, + 505, + 211 + ], + "score": 0.984, + "html": "
Pretrained BERTBERT post SST-2
OriginSent-DFairF-FairFOriginSent-DFairFFairF
Names,Career/Family0.4770.0960.2180.1820.0360.1090.2370.218
Terms,Career/Family0.1080.4370.0860.0760.0100.0570.3760.377
Terms,Math/Arts0.2530.1940.1330.1240.2190.2210.3010.263
Names,Math/Arts0.2540.1940.1010.0821.1530.7550.0840.099
Terms, Science/Arts0.3990.0750.2180.2040.1030.0810.1330.127
Names, Science/Arts0.6360.5400.3200.2350.2220.0470.0170.005
Avg. Abs. Effect Size0.3540.2560.1790.1500.2910.2120.1910.182
Classification Acc.1-1192.789.191.791.6
", + "type": "table", + "image_path": "34e9e64ff39ac335347ce8e54b7dbdc9e8b6fdf1638f6ecd2440b494110833cc.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 109, + 97, + 505, + 135.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 135.0, + 505, + 173.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 173.0, + 505, + 211.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 109, + 243, + 505, + 356 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 121, + 225, + 487, + 237 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 121, + 223, + 489, + 239 + ], + "spans": [ + { + "bbox": [ + 121, + 223, + 489, + 239 + ], + "score": 1.0, + "content": "Table 3: Performance of debiased embeddings on BERT post CoLA and BERT post QNLI.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 109, + 243, + 505, + 356 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 243, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 109, + 243, + 505, + 356 + ], + "score": 0.983, + "html": "
BERT post CoLABERT post QNLI
OriginSent-DFairF-FairFOriginSent-DFairF-FairF
Names, Career/Family0.0090.1490.2730.0340.2610.0540.1960.103
Terms, Career/Family0.1990.1860.1560.1190.1550.0040.0500.206
Terms,Math/Arts0.2680.3110.0080.0920.5840.0830.3060.323
Names,Math/Arts0.1500.3080.0600.1010.5810.6290.1680.288
Terms, Science/Arts0.4250.1630.2450.2490.0870.7160.5000.245
Names,Science/Arts0.0320.1920.1020.1270.5210.4430.3780.167
Avg. Abs.Effect Size0.1810.2170.1410.1200.3650.3210.2660.222
Classification Acc.57.655.456.556.591.390.691.090.8
", + "type": "table", + "image_path": "f5d8fc9a49cc5afb531787f31ea81ae3b037ec4af7301043842334f2f54e715a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 243, + 505, + 280.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 280.6666666666667, + 505, + 318.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 318.33333333333337, + 505, + 356.00000000000006 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 106, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "setup in Liang et al. (2020), and test the sentence templates of Terms/Names under different domains", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 385, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 504, + 396 + ], + "score": 1.0, + "content": "designed by Caliskan et al. (2017). The column name Origin refers to the original BERT results,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 433, + 408 + ], + "score": 1.0, + "content": "and Sent-D is short for Sent-Debias (Liang et al., 2020). FairFil− and FairFil (as", + "type": "text" + }, + { + "bbox": [ + 433, + 396, + 462, + 406 + ], + "score": 0.27, + "content": "\\mathrm { F a i r F ^ { - } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "and FairF", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "in the tables) are our method without/with the debiasing regularizer in Section 3.3. The best results", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 416, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 431 + ], + "score": 1.0, + "content": "of effect size (the lower the better) and classification accuracy (the higher the better) are bold among", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Sent-D, FairFil−, and FairFil. Since the pretrained BERT does not correspond to any downstream", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 439, + 319, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 319, + 452 + ], + "score": 1.0, + "content": "task, the classification accuracy is not reported for it.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 301, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 456, + 302, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 302, + 468 + ], + "score": 1.0, + "content": "From the SEAT test results, our contrastive learn-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 468, + 302, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 302, + 479 + ], + "score": 1.0, + "content": "ing framework effectively reduces the gender", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 479, + 302, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 302, + 490 + ], + "score": 1.0, + "content": "bias for both pretrained BERT and fine-tuned", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 488, + 302, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 302, + 501 + ], + "score": 1.0, + "content": "BERT under most test scenarios. Comparing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 500, + 303, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 303, + 512 + ], + "score": 1.0, + "content": "with Sent-Debias, our FairFil reaches a lower", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 511, + 302, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 302, + 523 + ], + "score": 1.0, + "content": "bias degree on the majority of the individual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 522, + 302, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 302, + 534 + ], + "score": 1.0, + "content": "SEAT tests. Considering the average of abso-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 533, + 302, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 302, + 545 + ], + "score": 1.0, + "content": "lute effect size, our FairFil is distinguished by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 545, + 302, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 302, + 556 + ], + "score": 1.0, + "content": "a significant margin to Sent-Debias. Moreover,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 555, + 302, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 302, + 567 + ], + "score": 1.0, + "content": "our FairFil achieves higher downstream classi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 566, + 302, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 302, + 577 + ], + "score": 1.0, + "content": "fication accuracy than Sent-Debias, which indi-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 578, + 302, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 302, + 589 + ], + "score": 1.0, + "content": "cates learning neural filter networks can preserve", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 588, + 302, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 302, + 600 + ], + "score": 1.0, + "content": "more semantic meaning than subtracting bias di-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 599, + 218, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 218, + 610 + ], + "score": 1.0, + "content": "rections learned from PCA.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5 + }, + { + "type": "table", + "bbox": [ + 313, + 488, + 503, + 589 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 309, + 457, + 504, + 479 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 456, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 309, + 456, + 504, + 470 + ], + "score": 1.0, + "content": "Table 4: Comparison of average debiasing per-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 309, + 469, + 432, + 479 + ], + "spans": [ + { + "bbox": [ + 309, + 469, + 432, + 479 + ], + "score": 1.0, + "content": "formance on pretrained BERT", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "table_body", + "bbox": [ + 313, + 488, + 503, + 589 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 313, + 488, + 503, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 503, + 589 + ], + "score": 0.98, + "html": "
MethodBias Degree
BERT origin (Devlin et al.,2019) FastText (Bojanowski etal., 2017)0.354
0.565
BERT word (Bolukbasi et al., 2016)0.861
BERT simple (May et al., 2019)0.298
Sent-Debias (Liang et al.,2020)0.256
FairFil- (Ours)0.179
FairFil (Ours)0.150
", + "type": "table", + "image_path": "d5cec6c5fa8e102545c3d64716e74071e9781ac802d7f69b45a869964a36d06d.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 313, + 488, + 503, + 500.625 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 313, + 500.625, + 503, + 513.25 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 313, + 513.25, + 503, + 525.875 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 313, + 525.875, + 503, + 538.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 313, + 538.5, + 503, + 551.125 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 313, + 551.125, + 503, + 563.75 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 313, + 563.75, + 503, + 576.375 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 313, + 576.375, + 503, + 589.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 32.0 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "For the ablation study, we also report the results of FairFil without the debiasing regularizer, as in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 329, + 639 + ], + "score": 1.0, + "content": "FairF−. Only with the contrastive learning framework,", + "type": "text" + }, + { + "bbox": [ + 330, + 627, + 360, + 637 + ], + "score": 0.49, + "content": "\\mathrm { F a i r F ^ { - } }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "already reduces the bias effectively", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 636, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 652 + ], + "score": 1.0, + "content": "and even achieves better effect size than the FairF on some of the SEAT tests. With the debiasing", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "regularizer, FairF has better average SEAT effect sizes but slightly loses in terms of the downstream", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "performance. However, the overall performance of FairF and FairF− shows a trade-off between", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 316, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 316, + 682 + ], + "score": 1.0, + "content": "fairness and representativeness of the filter network.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "We also compare the debiasing performance on a broader class of baselines, including word-level", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "debiasing methods, and report the average absolute SEAT effect size on the pretrained BERT en-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "coder. Both FairF− and FairF achieve a lower bias degree than other baselines. The word-level", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "debiasing methods (FastText (Bojanowski et al., 2017) and BERT word (Bolukbasi et al., 2016))", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 97, + 505, + 211 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 123, + 80, + 485, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 79, + 487, + 93 + ], + "spans": [ + { + "bbox": [ + 122, + 79, + 487, + 93 + ], + "score": 1.0, + "content": "Table 2: Performance of debiased embeddings on Pretrained BERT and BERT post SST-2.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 109, + 97, + 505, + 211 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 97, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 109, + 97, + 505, + 211 + ], + "score": 0.984, + "html": "
Pretrained BERTBERT post SST-2
OriginSent-DFairF-FairFOriginSent-DFairFFairF
Names,Career/Family0.4770.0960.2180.1820.0360.1090.2370.218
Terms,Career/Family0.1080.4370.0860.0760.0100.0570.3760.377
Terms,Math/Arts0.2530.1940.1330.1240.2190.2210.3010.263
Names,Math/Arts0.2540.1940.1010.0821.1530.7550.0840.099
Terms, Science/Arts0.3990.0750.2180.2040.1030.0810.1330.127
Names, Science/Arts0.6360.5400.3200.2350.2220.0470.0170.005
Avg. Abs. Effect Size0.3540.2560.1790.1500.2910.2120.1910.182
Classification Acc.1-1192.789.191.791.6
", + "type": "table", + "image_path": "34e9e64ff39ac335347ce8e54b7dbdc9e8b6fdf1638f6ecd2440b494110833cc.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 109, + 97, + 505, + 135.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 135.0, + 505, + 173.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 173.0, + 505, + 211.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 109, + 243, + 505, + 356 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 121, + 225, + 487, + 237 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 121, + 223, + 489, + 239 + ], + "spans": [ + { + "bbox": [ + 121, + 223, + 489, + 239 + ], + "score": 1.0, + "content": "Table 3: Performance of debiased embeddings on BERT post CoLA and BERT post QNLI.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 109, + 243, + 505, + 356 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 243, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 109, + 243, + 505, + 356 + ], + "score": 0.983, + "html": "
BERT post CoLABERT post QNLI
OriginSent-DFairF-FairFOriginSent-DFairF-FairF
Names, Career/Family0.0090.1490.2730.0340.2610.0540.1960.103
Terms, Career/Family0.1990.1860.1560.1190.1550.0040.0500.206
Terms,Math/Arts0.2680.3110.0080.0920.5840.0830.3060.323
Names,Math/Arts0.1500.3080.0600.1010.5810.6290.1680.288
Terms, Science/Arts0.4250.1630.2450.2490.0870.7160.5000.245
Names,Science/Arts0.0320.1920.1020.1270.5210.4430.3780.167
Avg. Abs.Effect Size0.1810.2170.1410.1200.3650.3210.2660.222
Classification Acc.57.655.456.556.591.390.691.090.8
", + "type": "table", + "image_path": "f5d8fc9a49cc5afb531787f31ea81ae3b037ec4af7301043842334f2f54e715a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 243, + 505, + 280.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 280.6666666666667, + 505, + 318.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 318.33333333333337, + 505, + 356.00000000000006 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 451 + ], + "lines": [], + "index": 11, + "bbox_fs": [ + 105, + 374, + 505, + 452 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 301, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 456, + 302, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 302, + 468 + ], + "score": 1.0, + "content": "From the SEAT test results, our contrastive learn-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 468, + 302, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 302, + 479 + ], + "score": 1.0, + "content": "ing framework effectively reduces the gender", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 479, + 302, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 302, + 490 + ], + "score": 1.0, + "content": "bias for both pretrained BERT and fine-tuned", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 488, + 302, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 302, + 501 + ], + "score": 1.0, + "content": "BERT under most test scenarios. Comparing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 500, + 303, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 303, + 512 + ], + "score": 1.0, + "content": "with Sent-Debias, our FairFil reaches a lower", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 511, + 302, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 302, + 523 + ], + "score": 1.0, + "content": "bias degree on the majority of the individual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 522, + 302, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 302, + 534 + ], + "score": 1.0, + "content": "SEAT tests. Considering the average of abso-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 533, + 302, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 302, + 545 + ], + "score": 1.0, + "content": "lute effect size, our FairFil is distinguished by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 545, + 302, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 302, + 556 + ], + "score": 1.0, + "content": "a significant margin to Sent-Debias. Moreover,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 555, + 302, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 302, + 567 + ], + "score": 1.0, + "content": "our FairFil achieves higher downstream classi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 566, + 302, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 302, + 577 + ], + "score": 1.0, + "content": "fication accuracy than Sent-Debias, which indi-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 578, + 302, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 302, + 589 + ], + "score": 1.0, + "content": "cates learning neural filter networks can preserve", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 588, + 302, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 302, + 600 + ], + "score": 1.0, + "content": "more semantic meaning than subtracting bias di-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 599, + 218, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 218, + 610 + ], + "score": 1.0, + "content": "rections learned from PCA.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 456, + 303, + 610 + ] + }, + { + "type": "table", + "bbox": [ + 313, + 488, + 503, + 589 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 309, + 457, + 504, + 479 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 456, + 504, + 470 + ], + "spans": [ + { + "bbox": [ + 309, + 456, + 504, + 470 + ], + "score": 1.0, + "content": "Table 4: Comparison of average debiasing per-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 309, + 469, + 432, + 479 + ], + "spans": [ + { + "bbox": [ + 309, + 469, + 432, + 479 + ], + "score": 1.0, + "content": "formance on pretrained BERT", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "table_body", + "bbox": [ + 313, + 488, + 503, + 589 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 313, + 488, + 503, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 503, + 589 + ], + "score": 0.98, + "html": "
MethodBias Degree
BERT origin (Devlin et al.,2019) FastText (Bojanowski etal., 2017)0.354
0.565
BERT word (Bolukbasi et al., 2016)0.861
BERT simple (May et al., 2019)0.298
Sent-Debias (Liang et al.,2020)0.256
FairFil- (Ours)0.179
FairFil (Ours)0.150
", + "type": "table", + "image_path": "d5cec6c5fa8e102545c3d64716e74071e9781ac802d7f69b45a869964a36d06d.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 313, + 488, + 503, + 500.625 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 313, + 500.625, + 503, + 513.25 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 313, + 513.25, + 503, + 525.875 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 313, + 525.875, + 503, + 538.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 313, + 538.5, + 503, + 551.125 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 313, + 551.125, + 503, + 563.75 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 313, + 563.75, + 503, + 576.375 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 313, + 576.375, + 503, + 589.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 32.0 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "For the ablation study, we also report the results of FairFil without the debiasing regularizer, as in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 329, + 639 + ], + "score": 1.0, + "content": "FairF−. Only with the contrastive learning framework,", + "type": "text" + }, + { + "bbox": [ + 330, + 627, + 360, + 637 + ], + "score": 0.49, + "content": "\\mathrm { F a i r F ^ { - } }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "already reduces the bias effectively", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 636, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 652 + ], + "score": 1.0, + "content": "and even achieves better effect size than the FairF on some of the SEAT tests. With the debiasing", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "regularizer, FairF has better average SEAT effect sizes but slightly loses in terms of the downstream", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "performance. However, the overall performance of FairF and FairF− shows a trade-off between", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 316, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 316, + 682 + ], + "score": 1.0, + "content": "fairness and representativeness of the filter network.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 615, + 506, + 682 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "We also compare the debiasing performance on a broader class of baselines, including word-level", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "debiasing methods, and report the average absolute SEAT effect size on the pretrained BERT en-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "coder. Both FairF− and FairF achieve a lower bias degree than other baselines. The word-level", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "debiasing methods (FastText (Bojanowski et al., 2017) and BERT word (Bolukbasi et al., 2016))", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 81, + 492, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 81, + 492, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 81, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 117, + 81, + 492, + 186 + ], + "score": 0.97, + "type": "image", + "image_path": "b1af18f81a0530ed7cdc54ea708ebd8933b60de57ffa2ef5a690d953331c31cd.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 81, + 492, + 116.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 116.0, + 492, + 151.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 151.0, + 492, + 186.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 152, + 189, + 459, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 189, + 460, + 203 + ], + "spans": [ + { + "bbox": [ + 151, + 189, + 460, + 203 + ], + "score": 1.0, + "content": "Figure 2: Influence of the training data proportion to debias degree of BERT.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 167, + 213, + 440, + 303 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 167, + 213, + 440, + 303 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 167, + 213, + 440, + 303 + ], + "spans": [ + { + "bbox": [ + 167, + 213, + 440, + 303 + ], + "score": 0.965, + "type": "image", + "image_path": "73e47c5a70cdb751cd0be57c568fba5cb6b2d7062571501831b7227b38e7b83d.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 167, + 213, + 440, + 243.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 167, + 243.0, + 440, + 273.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 167, + 273.0, + 440, + 303.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 309, + 503, + 332 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "Figure 3: T-SNE plots of sentence embedding mean of each words contextualized in templates. The", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 320, + 471, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 471, + 332 + ], + "score": 1.0, + "content": "left-hand side is from the original pretrained BERT; the right-hand side is from our FairFil.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 501, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 503, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 503, + 355 + ], + "score": 1.0, + "content": "have the worst debiasing performance, which validates our observation that the word-level debias-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 354, + 385, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 385, + 366 + ], + "score": 1.0, + "content": "ing methods cannot reduce sentence-level social bias in NLP models.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 375, + 176, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 177, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 177, + 387 + ], + "score": 1.0, + "content": "5.5 ANALYSIS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "To test the influence of data proportion on the model’s debiasing performance, we select WikiText-2", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "with 13,750 sentences as the training corpora following the setup in Liang et al. (2020). Then we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "randomly divide the training data into 5 equal-sized partitions. We evaluate the bias degree of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "score": 1.0, + "content": "sentence debiasing methods on different combinations of the partitions, specifically with training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 177, + 447 + ], + "score": 1.0, + "content": "data proportions", + "type": "text" + }, + { + "bbox": [ + 177, + 434, + 197, + 444 + ], + "score": 0.81, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 433, + 200, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 200, + 434, + 220, + 444 + ], + "score": 0.8, + "content": "40 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 433, + 223, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 223, + 434, + 243, + 444 + ], + "score": 0.82, + "content": "60 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 433, + 246, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 246, + 434, + 266, + 444 + ], + "score": 0.82, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 433, + 270, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 271, + 434, + 295, + 444 + ], + "score": 0.83, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "). Under each data proportion, we repeat the training", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "5 times to obtain the mean and variance of the absolute SEAT effect size. In Figure 2, we plot the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "bias degree of BERT post tasks with different training data proportions. In general, both Sent-Debias", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "score": 1.0, + "content": "and FairFil achieve better performance and smaller variance when the proportion of training data is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 170, + 490 + ], + "score": 1.0, + "content": "larger. Under a", + "type": "text" + }, + { + "bbox": [ + 171, + 478, + 190, + 488 + ], + "score": 0.87, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "training proportion, our FairFil can better remove bias in text encoder, which", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 487, + 429, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 429, + 501 + ], + "score": 1.0, + "content": "shows FairFil has better data efficiency with the contrastive learning framework.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "To further study output debiased sentence embedding, we visualize the relative distances of at-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "tributes and targets of SEAT before/after our debiasing process. We choose the target words as “he”", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "and “she.” Attributes are selected from different social domains. We first contextualize the selected", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "words into sentence templates as described in Section 5.1. We then average the original/debiased em-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "score": 1.0, + "content": "beddings of these sentence template and plot the t-SNE (Maaten & Hinton, 2008) in Figure 3. From", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "the t-SNE, the debiased encoder provides more balanced distances from gender targets “he/she” to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 570, + 197, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 197, + 584 + ], + "score": 1.0, + "content": "the attribute concepts.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 108, + 592, + 202, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 203, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 203, + 608 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "This paper has developed a novel debiasing method for large-scale pretrained text encoder neural", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "networks. We proposed a fair filter (FairFil) network, which takes the original sentence embeddings", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "as input and outputs the debiased sentence embeddings. To train the fair filter, we constructed a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "multi-view contrast learning framework, which maximizes the mutual information between each", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "sentence and its augmentation. The augmented sentence is generated by replacing sensitive words", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "in the original sentence with words in a similar semantic but different bias directions. Further,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "we designed a debiasing regularizer that minimizes the mutual information between the debiased", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "embeddings and the corresponding sensitive words in sentences. Experimental results demonstrate", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the proposed FairFil not only reduces the bias in sentence embedding space, but also maintains the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "semantic meaning of the embeddings. This post hoc method does not require access to the training", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 721, + 497, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 497, + 733 + ], + "score": 1.0, + "content": "corpora, or any retraining process of the pretrained text encoder, which enhances its applicability.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 81, + 492, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 81, + 492, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 81, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 117, + 81, + 492, + 186 + ], + "score": 0.97, + "type": "image", + "image_path": "b1af18f81a0530ed7cdc54ea708ebd8933b60de57ffa2ef5a690d953331c31cd.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 81, + 492, + 116.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 116.0, + 492, + 151.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 151.0, + 492, + 186.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 152, + 189, + 459, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 189, + 460, + 203 + ], + "spans": [ + { + "bbox": [ + 151, + 189, + 460, + 203 + ], + "score": 1.0, + "content": "Figure 2: Influence of the training data proportion to debias degree of BERT.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 167, + 213, + 440, + 303 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 167, + 213, + 440, + 303 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 167, + 213, + 440, + 303 + ], + "spans": [ + { + "bbox": [ + 167, + 213, + 440, + 303 + ], + "score": 0.965, + "type": "image", + "image_path": "73e47c5a70cdb751cd0be57c568fba5cb6b2d7062571501831b7227b38e7b83d.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 167, + 213, + 440, + 243.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 167, + 243.0, + 440, + 273.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 167, + 273.0, + 440, + 303.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 309, + 503, + 332 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "Figure 3: T-SNE plots of sentence embedding mean of each words contextualized in templates. The", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 320, + 471, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 471, + 332 + ], + "score": 1.0, + "content": "left-hand side is from the original pretrained BERT; the right-hand side is from our FairFil.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 501, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 503, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 503, + 355 + ], + "score": 1.0, + "content": "have the worst debiasing performance, which validates our observation that the word-level debias-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 354, + 385, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 385, + 366 + ], + "score": 1.0, + "content": "ing methods cannot reduce sentence-level social bias in NLP models.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 343, + 503, + 366 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 375, + 176, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 177, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 177, + 387 + ], + "score": 1.0, + "content": "5.5 ANALYSIS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "To test the influence of data proportion on the model’s debiasing performance, we select WikiText-2", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "with 13,750 sentences as the training corpora following the setup in Liang et al. (2020). Then we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "randomly divide the training data into 5 equal-sized partitions. We evaluate the bias degree of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "score": 1.0, + "content": "sentence debiasing methods on different combinations of the partitions, specifically with training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 177, + 447 + ], + "score": 1.0, + "content": "data proportions", + "type": "text" + }, + { + "bbox": [ + 177, + 434, + 197, + 444 + ], + "score": 0.81, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 433, + 200, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 200, + 434, + 220, + 444 + ], + "score": 0.8, + "content": "40 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 433, + 223, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 223, + 434, + 243, + 444 + ], + "score": 0.82, + "content": "60 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 433, + 246, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 246, + 434, + 266, + 444 + ], + "score": 0.82, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 433, + 270, + 447 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 271, + 434, + 295, + 444 + ], + "score": 0.83, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "). Under each data proportion, we repeat the training", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "5 times to obtain the mean and variance of the absolute SEAT effect size. In Figure 2, we plot the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "bias degree of BERT post tasks with different training data proportions. In general, both Sent-Debias", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "score": 1.0, + "content": "and FairFil achieve better performance and smaller variance when the proportion of training data is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 170, + 490 + ], + "score": 1.0, + "content": "larger. Under a", + "type": "text" + }, + { + "bbox": [ + 171, + 478, + 190, + 488 + ], + "score": 0.87, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "training proportion, our FairFil can better remove bias in text encoder, which", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 487, + 429, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 429, + 501 + ], + "score": 1.0, + "content": "shows FairFil has better data efficiency with the contrastive learning framework.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 389, + 506, + 501 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "To further study output debiased sentence embedding, we visualize the relative distances of at-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "tributes and targets of SEAT before/after our debiasing process. We choose the target words as “he”", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "and “she.” Attributes are selected from different social domains. We first contextualize the selected", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "words into sentence templates as described in Section 5.1. We then average the original/debiased em-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "score": 1.0, + "content": "beddings of these sentence template and plot the t-SNE (Maaten & Hinton, 2008) in Figure 3. From", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "the t-SNE, the debiased encoder provides more balanced distances from gender targets “he/she” to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 570, + 197, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 197, + 584 + ], + "score": 1.0, + "content": "the attribute concepts.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 505, + 505, + 584 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 592, + 202, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 203, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 203, + 608 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "This paper has developed a novel debiasing method for large-scale pretrained text encoder neural", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "networks. We proposed a fair filter (FairFil) network, which takes the original sentence embeddings", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "as input and outputs the debiased sentence embeddings. To train the fair filter, we constructed a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "multi-view contrast learning framework, which maximizes the mutual information between each", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "sentence and its augmentation. The augmented sentence is generated by replacing sensitive words", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "in the original sentence with words in a similar semantic but different bias directions. Further,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "we designed a debiasing regularizer that minimizes the mutual information between the debiased", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "embeddings and the corresponding sensitive words in sentences. Experimental results demonstrate", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the proposed FairFil not only reduces the bias in sentence embedding space, but also maintains the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "semantic meaning of the embeddings. This post hoc method does not require access to the training", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 721, + 497, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 497, + 733 + ], + "score": 1.0, + "content": "corpora, or any retraining process of the pretrained text encoder, which enhances its applicability.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 610, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 225, + 94 + ], + "lines": [ + { + "bbox": [ + 107, + 81, + 226, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 226, + 96 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 365, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 367, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 367, + 120 + ], + "score": 1.0, + "content": "This research was supported in part by the DOE, NSF and ONR.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 136, + 176, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 176, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 176, + 149 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 504, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 165, + 424, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 165, + 424, + 178 + ], + "score": 1.0, + "content": "bottleneck. In International Conference on Learning Representations, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 186, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "Su Lin Blodgett, Solon Barocas, Hal Daume III, and Hanna Wallach. Language (technology) is ´", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 197, + 445, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 445, + 210 + ], + "score": 1.0, + "content": "power: A critical survey of” bias” in nlp. arXiv preprint arXiv:2005.14050, 2020.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 218, + 506, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 228, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 115, + 228, + 506, + 243 + ], + "score": 1.0, + "content": "subword information. Transactions of the Association for Computational Linguistics, 5:135–146,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 240, + 142, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 142, + 252 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 108, + 260, + 504, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 506, + 274 + ], + "score": 1.0, + "content": "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 283, + 375, + 295 + ], + "spans": [ + { + "bbox": [ + 116, + 283, + 375, + 295 + ], + "score": 1.0, + "content": "in neural information processing systems, pp. 4349–4357, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Shikha Bordia and Samuel R. Bowman. Identifying and reducing gender bias in word-level language", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 116, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "models. In Proceedings of the 2019 Conference of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 114, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "for Computational Linguistics: Student Research Workshop, pp. 7–15, Minneapolis, Minnesota,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 336, + 143, + 349 + ], + "spans": [ + { + "bbox": [ + 114, + 336, + 143, + 349 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "Marc-Etienne Brunet, Colleen Alkalay-Houlihan, Ashton Anderson, and Richard Zemel. Under-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 368, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 504, + 380 + ], + "score": 1.0, + "content": "standing the origins of bias in word embeddings. In International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "spans": [ + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "score": 1.0, + "content": "ing, pp. 803–811, 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 105, + 399, + 504, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 399, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 413 + ], + "score": 1.0, + "content": "Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. Semantics derived automatically from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 410, + 443, + 423 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 443, + 423 + ], + "score": 1.0, + "content": "language corpora contain human-like biases. Science, 356(6334):183–186, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 504, + 444 + ], + "score": 1.0, + "content": "Kaytlin Chaloner and Alfredo Maldonado. Measuring gender bias in word embeddings across do-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "mains and discovering new gender bias word categories. In Proceedings of the First Workshop on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 453, + 374, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 453, + 374, + 466 + ], + "score": 1.0, + "content": "Gender Bias in Natural Language Processing, pp. 25–32, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 485, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 485, + 504, + 497 + ], + "score": 1.0, + "content": "contrastive learning of visual representations. In Proceedings of the 37th international conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 496, + 229, + 508 + ], + "spans": [ + { + "bbox": [ + 115, + 496, + 229, + 508 + ], + "score": 1.0, + "content": "on Machine learning, 2020.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 116, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "Interpretable representation learning by information maximizing generative adversarial nets. In", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 538, + 416, + 552 + ], + "spans": [ + { + "bbox": [ + 115, + 538, + 416, + 552 + ], + "score": 1.0, + "content": "Advances in neural information processing systems, pp. 2172–2180, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 506, + 573 + ], + "score": 1.0, + "content": "Pengyu Cheng, Weituo Hao, Shuyang Dai, Jiachang Liu, Zhe Gan, and Lawrence Carin. Club: A", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "contrastive log-ratio upper bound of mutual information. In Proceedings of the 37th international", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 581, + 280, + 594 + ], + "spans": [ + { + "bbox": [ + 116, + 581, + 280, + 594 + ], + "score": 1.0, + "content": "conference on Machine learning, 2020a.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "Pengyu Cheng, Martin Renqiang Min, Dinghan Shen, Christopher Malon, Yizhe Zhang, Yitong", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 614, + 504, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 614, + 504, + 626 + ], + "score": 1.0, + "content": "Li, and Lawrence Carin. Improving disentangled text representation learning with information-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 115, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "theoretic guidance. In Proceedings of the 58th Annual Meeting of the Association for Computa-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 635, + 285, + 648 + ], + "spans": [ + { + "bbox": [ + 115, + 635, + 285, + 648 + ], + "score": 1.0, + "content": "tional Linguistics, pp. 7530–7541, 2020b.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 503, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "Alexis Conneau, Douwe Kiela, Holger Schwenk, Lo¨ıc Barrault, and Antoine Bordes. Supervised", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "learning of universal sentence representations from natural language inference data. In EMNLP,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 677, + 144, + 690 + ], + "spans": [ + { + "bbox": [ + 115, + 677, + 144, + 690 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Jason V Davis, Brian Kulis, Prateek Jain, Suvrit Sra, and Inderjit S Dhillon. Information-theoretic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "metric learning. In Proceedings of the 24th international conference on Machine learning, pp.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 720, + 182, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 182, + 731 + ], + "score": 1.0, + "content": "209–216, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 225, + 94 + ], + "lines": [ + { + "bbox": [ + 107, + 81, + 226, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 226, + 96 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 365, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 367, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 367, + 120 + ], + "score": 1.0, + "content": "This research was supported in part by the DOE, NSF and ONR.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 106, + 367, + 120 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 136, + 176, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 176, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 176, + 149 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 504, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 165, + 424, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 165, + 424, + 178 + ], + "score": 1.0, + "content": "bottleneck. In International Conference on Learning Representations, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 154, + 505, + 178 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 186, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "Su Lin Blodgett, Solon Barocas, Hal Daume III, and Hanna Wallach. Language (technology) is ´", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 197, + 445, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 445, + 210 + ], + "score": 1.0, + "content": "power: A critical survey of” bias” in nlp. arXiv preprint arXiv:2005.14050, 2020.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 185, + 505, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 218, + 506, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 228, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 115, + 228, + 506, + 243 + ], + "score": 1.0, + "content": "subword information. Transactions of the Association for Computational Linguistics, 5:135–146,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 240, + 142, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 142, + 252 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 218, + 506, + 252 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 260, + 504, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 506, + 274 + ], + "score": 1.0, + "content": "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 283, + 375, + 295 + ], + "spans": [ + { + "bbox": [ + 116, + 283, + 375, + 295 + ], + "score": 1.0, + "content": "in neural information processing systems, pp. 4349–4357, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 260, + 506, + 295 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Shikha Bordia and Samuel R. Bowman. Identifying and reducing gender bias in word-level language", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 116, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "models. In Proceedings of the 2019 Conference of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 114, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "for Computational Linguistics: Student Research Workshop, pp. 7–15, Minneapolis, Minnesota,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 336, + 143, + 349 + ], + "spans": [ + { + "bbox": [ + 114, + 336, + 143, + 349 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 303, + 506, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "Marc-Etienne Brunet, Colleen Alkalay-Houlihan, Ashton Anderson, and Richard Zemel. Under-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 368, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 504, + 380 + ], + "score": 1.0, + "content": "standing the origins of bias in word embeddings. In International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "spans": [ + { + "bbox": [ + 115, + 379, + 214, + 390 + ], + "score": 1.0, + "content": "ing, pp. 803–811, 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 357, + 505, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 399, + 504, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 399, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 413 + ], + "score": 1.0, + "content": "Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. Semantics derived automatically from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 410, + 443, + 423 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 443, + 423 + ], + "score": 1.0, + "content": "language corpora contain human-like biases. Science, 356(6334):183–186, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 399, + 505, + 423 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 504, + 444 + ], + "score": 1.0, + "content": "Kaytlin Chaloner and Alfredo Maldonado. Measuring gender bias in word embeddings across do-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "mains and discovering new gender bias word categories. In Proceedings of the First Workshop on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 453, + 374, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 453, + 374, + 466 + ], + "score": 1.0, + "content": "Gender Bias in Natural Language Processing, pp. 25–32, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 431, + 505, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 485, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 485, + 504, + 497 + ], + "score": 1.0, + "content": "contrastive learning of visual representations. In Proceedings of the 37th international conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 496, + 229, + 508 + ], + "spans": [ + { + "bbox": [ + 115, + 496, + 229, + 508 + ], + "score": 1.0, + "content": "on Machine learning, 2020.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 474, + 505, + 508 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 116, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "Interpretable representation learning by information maximizing generative adversarial nets. In", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 538, + 416, + 552 + ], + "spans": [ + { + "bbox": [ + 115, + 538, + 416, + 552 + ], + "score": 1.0, + "content": "Advances in neural information processing systems, pp. 2172–2180, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 517, + 505, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 506, + 573 + ], + "score": 1.0, + "content": "Pengyu Cheng, Weituo Hao, Shuyang Dai, Jiachang Liu, Zhe Gan, and Lawrence Carin. Club: A", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "contrastive log-ratio upper bound of mutual information. In Proceedings of the 37th international", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 581, + 280, + 594 + ], + "spans": [ + { + "bbox": [ + 116, + 581, + 280, + 594 + ], + "score": 1.0, + "content": "conference on Machine learning, 2020a.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 559, + 506, + 594 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "Pengyu Cheng, Martin Renqiang Min, Dinghan Shen, Christopher Malon, Yizhe Zhang, Yitong", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 614, + 504, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 614, + 504, + 626 + ], + "score": 1.0, + "content": "Li, and Lawrence Carin. Improving disentangled text representation learning with information-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 115, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "theoretic guidance. In Proceedings of the 58th Annual Meeting of the Association for Computa-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 635, + 285, + 648 + ], + "spans": [ + { + "bbox": [ + 115, + 635, + 285, + 648 + ], + "score": 1.0, + "content": "tional Linguistics, pp. 7530–7541, 2020b.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 601, + 505, + 648 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 503, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "Alexis Conneau, Douwe Kiela, Holger Schwenk, Lo¨ıc Barrault, and Antoine Bordes. Supervised", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "learning of universal sentence representations from natural language inference data. In EMNLP,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 677, + 144, + 690 + ], + "spans": [ + { + "bbox": [ + 115, + 677, + 144, + 690 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 655, + 505, + 690 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Jason V Davis, Brian Kulis, Prateek Jain, Suvrit Sra, and Inderjit S Dhillon. Information-theoretic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "metric learning. In Proceedings of the 24th international conference on Machine learning, pp.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 720, + 182, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 182, + 731 + ], + "score": 1.0, + "content": "209–216, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 699, + 506, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "the North American Chapter of the Association for Computational Linguistics, pp. 4171–4186,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 142, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 142, + 127 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 133, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Measuring and miti-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 159 + ], + "score": 1.0, + "content": "gating unintended bias in text classification. In Proceedings of the 2018 AAAI/ACM Conference", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 294, + 168 + ], + "score": 1.0, + "content": "on AI, Ethics, and Society, pp. 67–73, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 188 + ], + "score": 1.0, + "content": "Behnam Gholami, Pritish Sahu, Ognjen Rudovic, Konstantinos Bousmalis, and Vladimir Pavlovic.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 117, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "Unsupervised multi-target domain adaptation: An information theoretic approach. IEEE Trans-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 196, + 323, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 196, + 323, + 209 + ], + "score": 1.0, + "content": "actions on Image Processing, 29:3993–4002, 2020.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 506, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 213, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 228 + ], + "score": 1.0, + "content": "Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 225, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 506, + 240 + ], + "score": 1.0, + "content": "of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 237, + 196, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 196, + 248 + ], + "score": 1.0, + "content": "pp. 855–864, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 266, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 505, + 280 + ], + "score": 1.0, + "content": "for unnormalized statistical models. In Proceedings of the Thirteenth International Conference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 277, + 354, + 290 + ], + "spans": [ + { + "bbox": [ + 116, + 277, + 354, + 290 + ], + "score": 1.0, + "content": "on Artificial Intelligence and Statistics, pp. 297–304, 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 506, + 309 + ], + "score": 1.0, + "content": "Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 317, + 379, + 332 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 379, + 332 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition, pp. 9729–9738, 2020.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 109, + 336, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 336, + 504, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 504, + 349 + ], + "score": 1.0, + "content": "R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 116, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 358, + 454, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 358, + 454, + 371 + ], + "score": 1.0, + "content": "and maximization. In International Conference on Learning Representations, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 504, + 390 + ], + "score": 1.0, + "content": "Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Span-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 115, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "bert: Improving pre-training by representing and predicting spans. Transactions of the Association", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 399, + 304, + 411 + ], + "spans": [ + { + "bbox": [ + 114, + 399, + 304, + 411 + ], + "score": 1.0, + "content": "for Computational Linguistics, 8:64–77, 2020.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 104, + 417, + 504, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 432 + ], + "score": 1.0, + "content": "Masahiro Kaneko and Danushka Bollegala. Gender-preserving debiasing for pre-trained word em-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 428, + 321, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 321, + 441 + ], + "score": 1.0, + "content": "beddings. arXiv preprint arXiv:1906.00742, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 108, + 447, + 504, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 504, + 460 + ], + "score": 1.0, + "content": "Svetlana Kiritchenko and Saif Mohammad. Examining gender and race bias in two hundred senti-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "ment analysis systems. In Proceedings of the Seventh Joint Conference on Lexical and Computa-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 469, + 257, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 257, + 482 + ], + "score": 1.0, + "content": "tional Semantics, pp. 43–53, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Ryan Kiros, Yukun Zhu, Russ R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 498, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 498, + 506, + 513 + ], + "score": 1.0, + "content": "and Sanja Fidler. Skip-thought vectors. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 510, + 208, + 522 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 208, + 522 + ], + "score": 1.0, + "content": "pp. 3294–3302, 2015.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 105, + 528, + 435, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 435, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 435, + 542 + ], + "score": 1.0, + "content": "Solomon Kullback. Information theory and statistics. Courier Corporation, 1997.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "score": 1.0, + "content": "Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. Measuring bias in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 116, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "contextualized word representations. In Proceedings of the First Workshop on Gender Bias in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 569, + 320, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 569, + 320, + 582 + ], + "score": 1.0, + "content": "Natural Language Processing, pp. 166–172, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "Paul Pu Liang, Irene Mengze Li, Emily Zheng, Yao Chong Lim, Ruslan Salakhutdinov, and Louis-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "Philippe Morency. Towards debiasing sentence representations. In Proceedings of the 58th Annual", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 609, + 440, + 623 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 440, + 623 + ], + "score": 1.0, + "content": "Meeting of the Association for Computational Linguistics, pp. 5502–5515, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 504, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 650, + 220, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 220, + 661 + ], + "score": 1.0, + "content": "arXiv:1703.03130, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 692, + 322, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 322, + 703 + ], + "score": 1.0, + "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 298, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 298, + 732 + ], + "score": 1.0, + "content": "learning research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "the North American Chapter of the Association for Computational Linguistics, pp. 4171–4186,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 142, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 142, + 127 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 127 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 133, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Measuring and miti-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 159 + ], + "score": 1.0, + "content": "gating unintended bias in text classification. In Proceedings of the 2018 AAAI/ACM Conference", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 294, + 168 + ], + "score": 1.0, + "content": "on AI, Ethics, and Society, pp. 67–73, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 133, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 188 + ], + "score": 1.0, + "content": "Behnam Gholami, Pritish Sahu, Ognjen Rudovic, Konstantinos Bousmalis, and Vladimir Pavlovic.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 117, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "Unsupervised multi-target domain adaptation: An information theoretic approach. IEEE Trans-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 196, + 323, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 196, + 323, + 209 + ], + "score": 1.0, + "content": "actions on Image Processing, 29:3993–4002, 2020.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 173, + 505, + 209 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 506, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 213, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 228 + ], + "score": 1.0, + "content": "Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 225, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 506, + 240 + ], + "score": 1.0, + "content": "of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 237, + 196, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 196, + 248 + ], + "score": 1.0, + "content": "pp. 855–864, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 213, + 506, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 266, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 505, + 280 + ], + "score": 1.0, + "content": "for unnormalized statistical models. In Proceedings of the Thirteenth International Conference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 277, + 354, + 290 + ], + "spans": [ + { + "bbox": [ + 116, + 277, + 354, + 290 + ], + "score": 1.0, + "content": "on Artificial Intelligence and Statistics, pp. 297–304, 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 255, + 505, + 290 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 506, + 309 + ], + "score": 1.0, + "content": "Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 317, + 379, + 332 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 379, + 332 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition, pp. 9729–9738, 2020.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 295, + 506, + 332 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 336, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 336, + 504, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 504, + 349 + ], + "score": 1.0, + "content": "R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 116, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 358, + 454, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 358, + 454, + 371 + ], + "score": 1.0, + "content": "and maximization. In International Conference on Learning Representations, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 336, + 505, + 371 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 504, + 390 + ], + "score": 1.0, + "content": "Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Span-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 115, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "bert: Improving pre-training by representing and predicting spans. Transactions of the Association", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 399, + 304, + 411 + ], + "spans": [ + { + "bbox": [ + 114, + 399, + 304, + 411 + ], + "score": 1.0, + "content": "for Computational Linguistics, 8:64–77, 2020.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 376, + 506, + 411 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 417, + 504, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 432 + ], + "score": 1.0, + "content": "Masahiro Kaneko and Danushka Bollegala. Gender-preserving debiasing for pre-trained word em-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 428, + 321, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 321, + 441 + ], + "score": 1.0, + "content": "beddings. arXiv preprint arXiv:1906.00742, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 416, + 505, + 441 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 447, + 504, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 504, + 460 + ], + "score": 1.0, + "content": "Svetlana Kiritchenko and Saif Mohammad. Examining gender and race bias in two hundred senti-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "ment analysis systems. In Proceedings of the Seventh Joint Conference on Lexical and Computa-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 469, + 257, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 257, + 482 + ], + "score": 1.0, + "content": "tional Semantics, pp. 43–53, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 447, + 505, + 482 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Ryan Kiros, Yukun Zhu, Russ R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 498, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 498, + 506, + 513 + ], + "score": 1.0, + "content": "and Sanja Fidler. Skip-thought vectors. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 510, + 208, + 522 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 208, + 522 + ], + "score": 1.0, + "content": "pp. 3294–3302, 2015.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 487, + 506, + 522 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 528, + 435, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 435, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 435, + 542 + ], + "score": 1.0, + "content": "Solomon Kullback. Information theory and statistics. Courier Corporation, 1997.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 527, + 435, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "score": 1.0, + "content": "Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. Measuring bias in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 116, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "contextualized word representations. In Proceedings of the First Workshop on Gender Bias in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 569, + 320, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 569, + 320, + 582 + ], + "score": 1.0, + "content": "Natural Language Processing, pp. 166–172, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 545, + 505, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "Paul Pu Liang, Irene Mengze Li, Emily Zheng, Yao Chong Lim, Ruslan Salakhutdinov, and Louis-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "Philippe Morency. Towards debiasing sentence representations. In Proceedings of the 58th Annual", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 609, + 440, + 623 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 440, + 623 + ], + "score": 1.0, + "content": "Meeting of the Association for Computational Linguistics, pp. 5502–5515, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 587, + 505, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 504, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 650, + 220, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 220, + 661 + ], + "score": 1.0, + "content": "arXiv:1703.03130, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 627, + 505, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 692, + 322, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 322, + 703 + ], + "score": 1.0, + "content": "approach. arXiv preprint arXiv:1907.11692, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 668, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 298, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 298, + 732 + ], + "score": 1.0, + "content": "learning research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 104, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Sean MacAvaney, Andrew Yates, Arman Cohan, and Nazli Goharian. Cedr: Contextualized embed-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "dings for document ranking. In Proceedings of the 42nd International ACM SIGIR Conference", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 432, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 432, + 117 + ], + "score": 1.0, + "content": "on Research and Development in Information Retrieval, pp. 1101–1104, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 157 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 504, + 136 + ], + "score": 1.0, + "content": "Thomas Manzini, Yao Chong Lim, Yulia Tsvetkov, and Alan W Black. Black is to criminal as cau-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "score": 1.0, + "content": "casian is to police: Detecting and removing multiclass bias in word embeddings. arXiv preprint", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 146, + 219, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 146, + 219, + 156 + ], + "score": 1.0, + "content": "arXiv:1904.04047, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. On mea-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 115, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "suring social biases in sentence encoders. In Proceedings of the 2019 Conference of the North", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "American Chapter of the Association for Computational Linguistics: Human Language Technolo-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 198, + 360, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 198, + 360, + 210 + ], + "score": 1.0, + "content": "gies, Volume 1 (Long and Short Papers), pp. 622–628, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 230 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 213, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 213, + 241 + ], + "score": 1.0, + "content": "models. In ICLR, 201y.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 504, + 260 + ], + "score": 1.0, + "content": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word represen-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 258, + 368, + 270 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 368, + 270 + ], + "score": 1.0, + "content": "tations in vector space. arXiv preprint arXiv:1301.3781, 2013.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 503, + 300 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 504, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 504, + 289 + ], + "score": 1.0, + "content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 288, + 329, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 288, + 329, + 300 + ], + "score": 1.0, + "content": "tive coding. arXiv preprint arXiv:1807.03748, 2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "Sunghyun Park, Han Suk Shim, Moitreya Chatterjee, Kenji Sagae, and Louis-Philippe Morency.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 116, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "Computational analysis of persuasiveness in social multimedia: A novel dataset and multimodal", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "prediction approach. In Proceedings of the 16th International Conference on Multimodal Inter-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 340, + 217, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 340, + 217, + 352 + ], + "score": 1.0, + "content": "action, pp. 50–57, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of NAACL-HLT,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 381, + 207, + 393 + ], + "spans": [ + { + "bbox": [ + 115, + 381, + 207, + 393 + ], + "score": 1.0, + "content": "pp. 2227–2237, 2018.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 115, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "Mihalcea. MELD: A multimodal multi-party dataset for emotion recognition in conversations.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 421, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 505, + 436 + ], + "score": 1.0, + "content": "In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 433, + 181, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 433, + 181, + 444 + ], + "score": 1.0, + "content": "527–536, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "score": 1.0, + "content": "Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "and Yin Cui. Spatiotemporal contrastive video representation learning. arXiv preprint", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 474, + 219, + 485 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 219, + 485 + ], + "score": 1.0, + "content": "arXiv:2008.03800, 2020.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 504, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 504, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 504, + 506 + ], + "score": 1.0, + "content": "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language under-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 504, + 285, + 516 + ], + "spans": [ + { + "bbox": [ + 116, + 504, + 285, + 516 + ], + "score": 1.0, + "content": "standing by generative pre-training, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 503, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 505, + 535 + ], + "score": 1.0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 534, + 443, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 534, + 443, + 546 + ], + "score": 1.0, + "content": "bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 553, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 568 + ], + "score": 1.0, + "content": "Dinghan Shen, Pengyu Cheng, Dhanasekar Sundararaman, Xinyuan Zhang, Qian Yang, Meng Tang,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 564, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 504, + 577 + ], + "score": 1.0, + "content": "Asli Celikyilmaz, and Lawrence Carin. Learning compressed sentence representations for on-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "device text processing. In Proceedings of the 57th Annual Meeting of the Association for Compu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 587, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 279, + 599 + ], + "score": 1.0, + "content": "tational Linguistics, pp. 107–116, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 477, + 618 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew", + "type": "text" + }, + { + "bbox": [ + 478, + 605, + 502, + 617 + ], + "score": 0.34, + "content": "\\mathrm { ~ Y ~ N ~ g ~ } _ { }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 605, + 505, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 115, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 conference on empirical methods in natural language pro-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 639, + 240, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 240, + 649 + ], + "score": 1.0, + "content": "cessing, pp. 1631–1642, 2013.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 105, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 671 + ], + "score": 1.0, + "content": "Jiaming Song, Pratyusha Kalluri, Aditya Grover, Shengjia Zhao, and Stefano Ermon. Learning", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "controllable fair representations. In The 22nd International Conference on Artificial Intelligence", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 679, + 264, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 679, + 264, + 691 + ], + "score": 1.0, + "content": "and Statistics, pp. 2164–2173, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Gabriel Stanovsky, Noah A Smith, and Luke Zettlemoyer. Evaluating gender bias in machine trans-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "lation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguis-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 225, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 225, + 732 + ], + "score": 1.0, + "content": "tics, pp. 1679–1684, 2019.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Sean MacAvaney, Andrew Yates, Arman Cohan, and Nazli Goharian. Cedr: Contextualized embed-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "dings for document ranking. In Proceedings of the 42nd International ACM SIGIR Conference", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 432, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 432, + 117 + ], + "score": 1.0, + "content": "on Research and Development in Information Retrieval, pp. 1101–1104, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 503, + 157 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 504, + 136 + ], + "score": 1.0, + "content": "Thomas Manzini, Yao Chong Lim, Yulia Tsvetkov, and Alan W Black. Black is to criminal as cau-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 505, + 147 + ], + "score": 1.0, + "content": "casian is to police: Detecting and removing multiclass bias in word embeddings. arXiv preprint", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 146, + 219, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 146, + 219, + 156 + ], + "score": 1.0, + "content": "arXiv:1904.04047, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 122, + 505, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. On mea-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 115, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "suring social biases in sentence encoders. In Proceedings of the 2019 Conference of the North", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "American Chapter of the Association for Computational Linguistics: Human Language Technolo-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 198, + 360, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 198, + 360, + 210 + ], + "score": 1.0, + "content": "gies, Volume 1 (Long and Short Papers), pp. 622–628, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 164, + 505, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 230 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 213, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 213, + 241 + ], + "score": 1.0, + "content": "models. In ICLR, 201y.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 215, + 505, + 241 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 504, + 260 + ], + "score": 1.0, + "content": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word represen-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 258, + 368, + 270 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 368, + 270 + ], + "score": 1.0, + "content": "tations in vector space. arXiv preprint arXiv:1301.3781, 2013.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 245, + 504, + 270 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 503, + 300 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 504, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 504, + 289 + ], + "score": 1.0, + "content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 288, + 329, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 288, + 329, + 300 + ], + "score": 1.0, + "content": "tive coding. arXiv preprint arXiv:1807.03748, 2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 276, + 504, + 300 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "Sunghyun Park, Han Suk Shim, Moitreya Chatterjee, Kenji Sagae, and Louis-Philippe Morency.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 116, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "Computational analysis of persuasiveness in social multimedia: A novel dataset and multimodal", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "prediction approach. In Proceedings of the 16th International Conference on Multimodal Inter-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 340, + 217, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 340, + 217, + 352 + ], + "score": 1.0, + "content": "action, pp. 50–57, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 306, + 506, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of NAACL-HLT,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 381, + 207, + 393 + ], + "spans": [ + { + "bbox": [ + 115, + 381, + 207, + 393 + ], + "score": 1.0, + "content": "pp. 2227–2237, 2018.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 358, + 505, + 393 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 115, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "Mihalcea. MELD: A multimodal multi-party dataset for emotion recognition in conversations.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 421, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 505, + 436 + ], + "score": 1.0, + "content": "In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 433, + 181, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 433, + 181, + 444 + ], + "score": 1.0, + "content": "527–536, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 400, + 505, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 466 + ], + "score": 1.0, + "content": "Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "and Yin Cui. Spatiotemporal contrastive video representation learning. arXiv preprint", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 474, + 219, + 485 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 219, + 485 + ], + "score": 1.0, + "content": "arXiv:2008.03800, 2020.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 451, + 505, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 504, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 504, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 504, + 506 + ], + "score": 1.0, + "content": "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language under-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 504, + 285, + 516 + ], + "spans": [ + { + "bbox": [ + 116, + 504, + 285, + 516 + ], + "score": 1.0, + "content": "standing by generative pre-training, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 492, + 504, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 503, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 505, + 535 + ], + "score": 1.0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 534, + 443, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 534, + 443, + 546 + ], + "score": 1.0, + "content": "bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 524, + 505, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 553, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 568 + ], + "score": 1.0, + "content": "Dinghan Shen, Pengyu Cheng, Dhanasekar Sundararaman, Xinyuan Zhang, Qian Yang, Meng Tang,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 564, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 504, + 577 + ], + "score": 1.0, + "content": "Asli Celikyilmaz, and Lawrence Carin. Learning compressed sentence representations for on-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "device text processing. In Proceedings of the 57th Annual Meeting of the Association for Compu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 587, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 279, + 599 + ], + "score": 1.0, + "content": "tational Linguistics, pp. 107–116, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 551, + 506, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 477, + 618 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew", + "type": "text" + }, + { + "bbox": [ + 478, + 605, + 502, + 617 + ], + "score": 0.34, + "content": "\\mathrm { ~ Y ~ N ~ g ~ } _ { }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 605, + 505, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 115, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 conference on empirical methods in natural language pro-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 639, + 240, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 240, + 649 + ], + "score": 1.0, + "content": "cessing, pp. 1631–1642, 2013.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 605, + 505, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 671 + ], + "score": 1.0, + "content": "Jiaming Song, Pratyusha Kalluri, Aditya Grover, Shengjia Zhao, and Stefano Ermon. Learning", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "controllable fair representations. In The 22nd International Conference on Artificial Intelligence", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 679, + 264, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 679, + 264, + 691 + ], + "score": 1.0, + "content": "and Statistics, pp. 2164–2173, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 656, + 505, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Gabriel Stanovsky, Noah A Smith, and Luke Zettlemoyer. Evaluating gender bias in machine trans-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "lation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguis-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 225, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 225, + 732 + ], + "score": 1.0, + "content": "tics, pp. 1679–1684, 2019.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequen-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "tial recommendation with bidirectional encoder representations from transformer. In Proceed-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "ings of the 28th ACM International Conference on Information and Knowledge Management, pp.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 115, + 191, + 126 + ], + "spans": [ + { + "bbox": [ + 117, + 115, + 191, + 126 + ], + "score": 1.0, + "content": "1441–1450, 2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 133, + 504, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. Line: Large-scale", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "information network embedding. In Proceedings of the 24th international conference on world", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 156, + 249, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 249, + 168 + ], + "score": 1.0, + "content": "wide web, pp. 1067–1077, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 105, + 174, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "Michael V”olske, Martin Potthast, Shahbaz Syed, and Benno Stein. TL;DR: Mining Reddit to learn", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "automatic summarization. In Proceedings of the Workshop on New Frontiers in Summarization,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 198, + 186, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 198, + 186, + 208 + ], + "score": 1.0, + "content": "pp. 59–63, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 506, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. Glue:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "A multi-task benchmark and analysis platform for natural language understanding. In Proceedings", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 115, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 249, + 219, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 249, + 219, + 261 + ], + "score": 1.0, + "content": "NLP, pp. 353–355, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 502, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 504, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 504, + 281 + ], + "score": 1.0, + "content": "Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. Neural network acceptability judgments.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 278, + 442, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 278, + 442, + 291 + ], + "score": 1.0, + "content": "Transactions of the Association for Computational Linguistics, 7:625–641, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 504, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "Kilian Q Weinberger, John Blitzer, and Lawrence K Saul. Distance metric learning for large margin", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "nearest neighbor classification. In Advances in neural information processing systems, pp. 1473–", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 318, + 167, + 332 + ], + "spans": [ + { + "bbox": [ + 116, + 318, + 167, + 332 + ], + "score": 1.0, + "content": "1480, 2006.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 104, + 338, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 350, + 322, + 361 + ], + "spans": [ + { + "bbox": [ + 116, + 350, + 322, + 361 + ], + "score": 1.0, + "content": "intelligent laboratory systems, 2(1-3):37–52, 1987.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 368, + 504, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "score": 1.0, + "content": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluat-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 380, + 398, + 392 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 398, + 392 + ], + "score": 1.0, + "content": "ing text generation with bert. arXiv preprint arXiv:1904.09675, 2019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequen-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "tial recommendation with bidirectional encoder representations from transformer. In Proceed-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "ings of the 28th ACM International Conference on Information and Knowledge Management, pp.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 115, + 191, + 126 + ], + "spans": [ + { + "bbox": [ + 117, + 115, + 191, + 126 + ], + "score": 1.0, + "content": "1441–1450, 2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 81, + 505, + 126 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 133, + 504, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. Line: Large-scale", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "information network embedding. In Proceedings of the 24th international conference on world", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 156, + 249, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 249, + 168 + ], + "score": 1.0, + "content": "wide web, pp. 1067–1077, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 133, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 174, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "Michael V”olske, Martin Potthast, Shahbaz Syed, and Benno Stein. TL;DR: Mining Reddit to learn", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "automatic summarization. In Proceedings of the Workshop on New Frontiers in Summarization,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 198, + 186, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 198, + 186, + 208 + ], + "score": 1.0, + "content": "pp. 59–63, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 174, + 505, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 506, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. Glue:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "A multi-task benchmark and analysis platform for natural language understanding. In Proceedings", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 115, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 249, + 219, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 249, + 219, + 261 + ], + "score": 1.0, + "content": "NLP, pp. 353–355, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 216, + 506, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 502, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 504, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 504, + 281 + ], + "score": 1.0, + "content": "Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. Neural network acceptability judgments.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 278, + 442, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 278, + 442, + 291 + ], + "score": 1.0, + "content": "Transactions of the Association for Computational Linguistics, 7:625–641, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 267, + 504, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 504, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "Kilian Q Weinberger, John Blitzer, and Lawrence K Saul. Distance metric learning for large margin", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "nearest neighbor classification. In Advances in neural information processing systems, pp. 1473–", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 318, + 167, + 332 + ], + "spans": [ + { + "bbox": [ + 116, + 318, + 167, + 332 + ], + "score": 1.0, + "content": "1480, 2006.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 297, + 505, + 332 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 338, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 350, + 322, + 361 + ], + "spans": [ + { + "bbox": [ + 116, + 350, + 322, + 361 + ], + "score": 1.0, + "content": "intelligent laboratory systems, 2(1-3):37–52, 1987.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 337, + 506, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 368, + 504, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 504, + 381 + ], + "score": 1.0, + "content": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluat-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 380, + 398, + 392 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 398, + 392 + ], + "score": 1.0, + "content": "ing text generation with bert. arXiv preprint arXiv:1904.09675, 2019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 106, + 368, + 504, + 392 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file