{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 78, 432, 116 ], "lines": [ { "bbox": [ 106, 78, 433, 97 ], "spans": [ { "bbox": [ 106, 78, 433, 97 ], "score": 1.0, "content": "TEXTLESS PHRASE STRUCTURE INDUCTION", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 99, 385, 117 ], "spans": [ { "bbox": [ 106, 99, 385, 117 ], "score": 1.0, "content": "FROM VISUALLY-GROUNDED SPEECH", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 112, 136, 244, 157 ], "lines": [ { "bbox": [ 113, 136, 201, 147 ], "spans": [ { "bbox": [ 113, 136, 201, 147 ], "score": 1.0, "content": "Anonymous authors", "type": "text" } ], "index": 2 }, { "bbox": [ 112, 146, 245, 159 ], "spans": [ { "bbox": [ 112, 146, 245, 159 ], "score": 1.0, "content": "Paper under double-blind review", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "title", "bbox": [ 278, 187, 333, 199 ], "lines": [ { "bbox": [ 276, 185, 336, 201 ], "spans": [ { "bbox": [ 276, 185, 336, 201 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 143, 203, 468, 335 ], "lines": [ { "bbox": [ 142, 203, 469, 215 ], "spans": [ { "bbox": [ 142, 203, 469, 215 ], "score": 1.0, "content": "We study phrase structure induction from visually-grounded speech without in-", "type": "text" } ], "index": 5 }, { "bbox": [ 141, 213, 470, 226 ], "spans": [ { "bbox": [ 141, 213, 470, 226 ], "score": 1.0, "content": "termediate text or text pre-trained models. The core idea is to first segment the", "type": "text" } ], "index": 6 }, { "bbox": [ 141, 225, 470, 237 ], "spans": [ { "bbox": [ 141, 225, 470, 237 ], "score": 1.0, "content": "speech waveform into sequences of word segments, then induce phrase structure", "type": "text" } ], "index": 7 }, { "bbox": [ 141, 235, 470, 249 ], "spans": [ { "bbox": [ 141, 235, 470, 249 ], "score": 1.0, "content": "based on the inferred segment-level continuous representations. To this end, we", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 247, 470, 259 ], "spans": [ { "bbox": [ 141, 247, 470, 259 ], "score": 1.0, "content": "present the Audio-Visual Neural Syntax Learner (AV-NSL) that learns non-trivial", "type": "text" } ], "index": 9 }, { "bbox": [ 142, 258, 469, 270 ], "spans": [ { "bbox": [ 142, 258, 469, 270 ], "score": 1.0, "content": "phrase structure by listening to audio and looking at images, without ever read-", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 269, 469, 281 ], "spans": [ { "bbox": [ 141, 269, 469, 281 ], "score": 1.0, "content": "ing text. Experiments on SpokenCOCO, the spoken version of MSCOCO with", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 280, 469, 291 ], "spans": [ { "bbox": [ 141, 280, 469, 291 ], "score": 1.0, "content": "paired images and spoken captions, show that AV-NSL infers meaningful phrase", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 291, 469, 303 ], "spans": [ { "bbox": [ 141, 291, 469, 303 ], "score": 1.0, "content": "structures similar to those learned from naturally-supervised text parsing, quanti-", "type": "text" } ], "index": 13 }, { "bbox": [ 142, 302, 469, 314 ], "spans": [ { "bbox": [ 142, 302, 469, 314 ], "score": 1.0, "content": "tatively and qualitatively. The findings in this paper extend prior work in unsu-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 313, 469, 324 ], "spans": [ { "bbox": [ 141, 313, 469, 324 ], "score": 1.0, "content": "pervised language acquisition from speech and grounded grammar induction, and", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 324, 413, 336 ], "spans": [ { "bbox": [ 141, 324, 413, 336 ], "score": 1.0, "content": "manifest one possibility of bridging the gap between the two fields.", "type": "text" } ], "index": 16 } ], "index": 10.5 }, { "type": "title", "bbox": [ 108, 343, 206, 355 ], "lines": [ { "bbox": [ 105, 342, 208, 358 ], "spans": [ { "bbox": [ 105, 342, 208, 358 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 360, 505, 426 ], "lines": [ { "bbox": [ 105, 360, 505, 373 ], "spans": [ { "bbox": [ 105, 360, 505, 373 ], "score": 1.0, "content": "Toddlers learn their first language through listening, talking, and interacting with the world through", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 371, 505, 385 ], "spans": [ { "bbox": [ 105, 371, 505, 385 ], "score": 1.0, "content": "multi-sensory inputs. Different levels of early language acquisition happen without supervisory", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 381, 506, 395 ], "spans": [ { "bbox": [ 105, 381, 506, 395 ], "score": 1.0, "content": "feedback (Dupoux, 2018): phonetics, phonology, morphology, syntax, semantics, pragmatics. It is", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 392, 506, 406 ], "spans": [ { "bbox": [ 105, 392, 506, 406 ], "score": 1.0, "content": "therefore crucial to think about learning language, from identifying lower-level phones or words to", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 403, 506, 417 ], "spans": [ { "bbox": [ 105, 403, 506, 417 ], "score": 1.0, "content": "inducing high-level linguistic structure like grammar, in natural settings.1 To this end, there have", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 415, 254, 427 ], "spans": [ { "bbox": [ 105, 415, 254, 427 ], "score": 1.0, "content": "been two ongoing efforts in parallel:", "type": "text" } ], "index": 23 } ], "index": 20.5 }, { "type": "text", "bbox": [ 106, 430, 505, 500 ], "lines": [ { "bbox": [ 105, 429, 504, 443 ], "spans": [ { "bbox": [ 105, 429, 504, 443 ], "score": 1.0, "content": "• Zero-resource speech processing, where speech models are constructed without any textual in-", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 442, 504, 453 ], "spans": [ { "bbox": [ 115, 442, 504, 453 ], "score": 1.0, "content": "termediates, with the goal of mimicking how children learn to speak before learning to read or", "type": "text" } ], "index": 25 }, { "bbox": [ 114, 452, 505, 465 ], "spans": [ { "bbox": [ 114, 452, 505, 465 ], "score": 1.0, "content": "write. The modeling tasks are constrained to unsupervised learning of subphones, phones, and", "type": "text" } ], "index": 26 }, { "bbox": [ 114, 462, 227, 476 ], "spans": [ { "bbox": [ 114, 462, 227, 476 ], "score": 1.0, "content": "words (Jansen et al., 2013).", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 477, 505, 491 ], "spans": [ { "bbox": [ 106, 477, 505, 491 ], "score": 1.0, "content": "• Grammar induction, which aims to learn latent syntactic structures, including constituency trees", "type": "text" } ], "index": 28 }, { "bbox": [ 114, 489, 434, 502 ], "spans": [ { "bbox": [ 114, 489, 434, 502 ], "score": 1.0, "content": "and dependency trees, with no annotation of syntactic structures as supervision.", "type": "text" } ], "index": 29 } ], "index": 26.5 }, { "type": "text", "bbox": [ 107, 504, 505, 581 ], "lines": [ { "bbox": [ 105, 503, 505, 516 ], "spans": [ { "bbox": [ 105, 503, 505, 516 ], "score": 1.0, "content": "Notably in recent years, multi-modal induction has emerged as a promising and effective objective", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 515, 505, 527 ], "spans": [ { "bbox": [ 105, 515, 505, 527 ], "score": 1.0, "content": "for both efforts. In speech, Harwath (2018) proposed to leverage parallel image-speech data to", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 525, 506, 539 ], "spans": [ { "bbox": [ 105, 525, 506, 539 ], "score": 1.0, "content": "acquire associated words (Harwath & Glass, 2017) and phones (Harwath et al., 2020) from raw", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 536, 505, 549 ], "spans": [ { "bbox": [ 105, 536, 505, 549 ], "score": 1.0, "content": "waveforms. In syntax induction, Shi et al. (2019) proposed to induce phrase-structure grammar", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 547, 505, 560 ], "spans": [ { "bbox": [ 105, 547, 505, 560 ], "score": 1.0, "content": "from parallel image-text data. The above observations motivated us to build a computational model", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 559, 505, 571 ], "spans": [ { "bbox": [ 105, 559, 505, 571 ], "score": 1.0, "content": "that leverages the visual modality to acquire low-level words up to high-level phrase-structure from", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 570, 467, 582 ], "spans": [ { "bbox": [ 105, 570, 467, 582 ], "score": 1.0, "content": "raw speech waveforms, without any intermediate textual forms or any direct supervision.2", "type": "text" } ], "index": 36 } ], "index": 33 }, { "type": "text", "bbox": [ 107, 586, 505, 653 ], "lines": [ { "bbox": [ 105, 586, 506, 599 ], "spans": [ { "bbox": [ 105, 586, 506, 599 ], "score": 1.0, "content": "In this paper, we present the Audio-Visual Neural Syntax Learner (AV-NSL), an approach toward", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 597, 505, 609 ], "spans": [ { "bbox": [ 105, 597, 505, 609 ], "score": 1.0, "content": "learning phrase structure from raw speech waveforms without relying on any kind of intermediate", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 609, 505, 621 ], "spans": [ { "bbox": [ 106, 609, 505, 621 ], "score": 1.0, "content": "textual form or text pre-trained models (Figure 1). In a nutshell, AV-NSL trains a visually-grounded", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 619, 506, 632 ], "spans": [ { "bbox": [ 105, 619, 506, 632 ], "score": 1.0, "content": "syntax learner directly on a sequence of continuous speech representations given by an audio-visual", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 629, 506, 644 ], "spans": [ { "bbox": [ 105, 629, 506, 644 ], "score": 1.0, "content": "word segmentation model. We also introduce a self-training process and an unsupervised decoding", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 641, 505, 654 ], "spans": [ { "bbox": [ 105, 641, 505, 654 ], "score": 1.0, "content": "method to improve the final output of in AV-NSL. To measure the effectiveness of AV-NSL, we", "type": "text" } ], "index": 42 } ], "index": 39.5 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 106, 661, 505, 732 ], "lines": [ { "bbox": [ 119, 659, 506, 673 ], "spans": [ { "bbox": [ 119, 659, 506, 673 ], "score": 1.0, "content": "1Natural settings here means situations that are similar to human language learning; that is, we are able to", "type": "text" } ] }, { "bbox": [ 105, 670, 402, 681 ], "spans": [ { "bbox": [ 105, 670, 402, 681 ], "score": 1.0, "content": "access parallel data from different modalities, while the amount of data is limited.", "type": "text" } ] }, { "bbox": [ 118, 679, 505, 694 ], "spans": [ { "bbox": [ 118, 679, 505, 694 ], "score": 1.0, "content": "2It is worth noting that human language acquisition in similar settings has also attracted the attention from", "type": "text" } ] }, { "bbox": [ 106, 691, 506, 704 ], "spans": [ { "bbox": [ 106, 691, 506, 704 ], "score": 1.0, "content": "the developmental psychology community (Mason, 1980; Naigles, 1990; Dupoux, 2018, inter alia). For in-", "type": "text" } ] }, { "bbox": [ 105, 700, 506, 714 ], "spans": [ { "bbox": [ 105, 700, 506, 714 ], "score": 1.0, "content": "stance, from Dupoux (2018): “Yet . . . there is still a large gap between models that learn from speech, which", "type": "text" } ] }, { "bbox": [ 106, 711, 505, 723 ], "spans": [ { "bbox": [ 106, 711, 505, 723 ], "score": 1.0, "content": "are limited to the discovery of phonemes and word forms, and models that learn syntax and semantics, which", "type": "text" } ] }, { "bbox": [ 105, 720, 217, 733 ], "spans": [ { "bbox": [ 105, 720, 217, 733 ], "score": 1.0, "content": "only work from textual input.”", "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, 432, 116 ], "lines": [ { "bbox": [ 106, 78, 433, 97 ], "spans": [ { "bbox": [ 106, 78, 433, 97 ], "score": 1.0, "content": "TEXTLESS PHRASE STRUCTURE INDUCTION", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 99, 385, 117 ], "spans": [ { "bbox": [ 106, 99, 385, 117 ], "score": 1.0, "content": "FROM VISUALLY-GROUNDED SPEECH", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 112, 136, 244, 157 ], "lines": [ { "bbox": [ 113, 136, 201, 147 ], "spans": [ { "bbox": [ 113, 136, 201, 147 ], "score": 1.0, "content": "Anonymous authors", "type": "text" } ], "index": 2 }, { "bbox": [ 112, 146, 245, 159 ], "spans": [ { "bbox": [ 112, 146, 245, 159 ], "score": 1.0, "content": "Paper under double-blind review", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 112, 136, 245, 159 ] }, { "type": "title", "bbox": [ 278, 187, 333, 199 ], "lines": [ { "bbox": [ 276, 185, 336, 201 ], "spans": [ { "bbox": [ 276, 185, 336, 201 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 143, 203, 468, 335 ], "lines": [ { "bbox": [ 142, 203, 469, 215 ], "spans": [ { "bbox": [ 142, 203, 469, 215 ], "score": 1.0, "content": "We study phrase structure induction from visually-grounded speech without in-", "type": "text" } ], "index": 5 }, { "bbox": [ 141, 213, 470, 226 ], "spans": [ { "bbox": [ 141, 213, 470, 226 ], "score": 1.0, "content": "termediate text or text pre-trained models. The core idea is to first segment the", "type": "text" } ], "index": 6 }, { "bbox": [ 141, 225, 470, 237 ], "spans": [ { "bbox": [ 141, 225, 470, 237 ], "score": 1.0, "content": "speech waveform into sequences of word segments, then induce phrase structure", "type": "text" } ], "index": 7 }, { "bbox": [ 141, 235, 470, 249 ], "spans": [ { "bbox": [ 141, 235, 470, 249 ], "score": 1.0, "content": "based on the inferred segment-level continuous representations. To this end, we", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 247, 470, 259 ], "spans": [ { "bbox": [ 141, 247, 470, 259 ], "score": 1.0, "content": "present the Audio-Visual Neural Syntax Learner (AV-NSL) that learns non-trivial", "type": "text" } ], "index": 9 }, { "bbox": [ 142, 258, 469, 270 ], "spans": [ { "bbox": [ 142, 258, 469, 270 ], "score": 1.0, "content": "phrase structure by listening to audio and looking at images, without ever read-", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 269, 469, 281 ], "spans": [ { "bbox": [ 141, 269, 469, 281 ], "score": 1.0, "content": "ing text. Experiments on SpokenCOCO, the spoken version of MSCOCO with", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 280, 469, 291 ], "spans": [ { "bbox": [ 141, 280, 469, 291 ], "score": 1.0, "content": "paired images and spoken captions, show that AV-NSL infers meaningful phrase", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 291, 469, 303 ], "spans": [ { "bbox": [ 141, 291, 469, 303 ], "score": 1.0, "content": "structures similar to those learned from naturally-supervised text parsing, quanti-", "type": "text" } ], "index": 13 }, { "bbox": [ 142, 302, 469, 314 ], "spans": [ { "bbox": [ 142, 302, 469, 314 ], "score": 1.0, "content": "tatively and qualitatively. The findings in this paper extend prior work in unsu-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 313, 469, 324 ], "spans": [ { "bbox": [ 141, 313, 469, 324 ], "score": 1.0, "content": "pervised language acquisition from speech and grounded grammar induction, and", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 324, 413, 336 ], "spans": [ { "bbox": [ 141, 324, 413, 336 ], "score": 1.0, "content": "manifest one possibility of bridging the gap between the two fields.", "type": "text" } ], "index": 16 } ], "index": 10.5, "bbox_fs": [ 141, 203, 470, 336 ] }, { "type": "title", "bbox": [ 108, 343, 206, 355 ], "lines": [ { "bbox": [ 105, 342, 208, 358 ], "spans": [ { "bbox": [ 105, 342, 208, 358 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 360, 505, 426 ], "lines": [ { "bbox": [ 105, 360, 505, 373 ], "spans": [ { "bbox": [ 105, 360, 505, 373 ], "score": 1.0, "content": "Toddlers learn their first language through listening, talking, and interacting with the world through", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 371, 505, 385 ], "spans": [ { "bbox": [ 105, 371, 505, 385 ], "score": 1.0, "content": "multi-sensory inputs. Different levels of early language acquisition happen without supervisory", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 381, 506, 395 ], "spans": [ { "bbox": [ 105, 381, 506, 395 ], "score": 1.0, "content": "feedback (Dupoux, 2018): phonetics, phonology, morphology, syntax, semantics, pragmatics. It is", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 392, 506, 406 ], "spans": [ { "bbox": [ 105, 392, 506, 406 ], "score": 1.0, "content": "therefore crucial to think about learning language, from identifying lower-level phones or words to", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 403, 506, 417 ], "spans": [ { "bbox": [ 105, 403, 506, 417 ], "score": 1.0, "content": "inducing high-level linguistic structure like grammar, in natural settings.1 To this end, there have", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 415, 254, 427 ], "spans": [ { "bbox": [ 105, 415, 254, 427 ], "score": 1.0, "content": "been two ongoing efforts in parallel:", "type": "text" } ], "index": 23 } ], "index": 20.5, "bbox_fs": [ 105, 360, 506, 427 ] }, { "type": "text", "bbox": [ 106, 430, 505, 500 ], "lines": [ { "bbox": [ 105, 429, 504, 443 ], "spans": [ { "bbox": [ 105, 429, 504, 443 ], "score": 1.0, "content": "• Zero-resource speech processing, where speech models are constructed without any textual in-", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 442, 504, 453 ], "spans": [ { "bbox": [ 115, 442, 504, 453 ], "score": 1.0, "content": "termediates, with the goal of mimicking how children learn to speak before learning to read or", "type": "text" } ], "index": 25 }, { "bbox": [ 114, 452, 505, 465 ], "spans": [ { "bbox": [ 114, 452, 505, 465 ], "score": 1.0, "content": "write. The modeling tasks are constrained to unsupervised learning of subphones, phones, and", "type": "text" } ], "index": 26 }, { "bbox": [ 114, 462, 227, 476 ], "spans": [ { "bbox": [ 114, 462, 227, 476 ], "score": 1.0, "content": "words (Jansen et al., 2013).", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 477, 505, 491 ], "spans": [ { "bbox": [ 106, 477, 505, 491 ], "score": 1.0, "content": "• Grammar induction, which aims to learn latent syntactic structures, including constituency trees", "type": "text" } ], "index": 28 }, { "bbox": [ 114, 489, 434, 502 ], "spans": [ { "bbox": [ 114, 489, 434, 502 ], "score": 1.0, "content": "and dependency trees, with no annotation of syntactic structures as supervision.", "type": "text" } ], "index": 29 } ], "index": 26.5, "bbox_fs": [ 105, 429, 505, 502 ] }, { "type": "text", "bbox": [ 107, 504, 505, 581 ], "lines": [ { "bbox": [ 105, 503, 505, 516 ], "spans": [ { "bbox": [ 105, 503, 505, 516 ], "score": 1.0, "content": "Notably in recent years, multi-modal induction has emerged as a promising and effective objective", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 515, 505, 527 ], "spans": [ { "bbox": [ 105, 515, 505, 527 ], "score": 1.0, "content": "for both efforts. In speech, Harwath (2018) proposed to leverage parallel image-speech data to", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 525, 506, 539 ], "spans": [ { "bbox": [ 105, 525, 506, 539 ], "score": 1.0, "content": "acquire associated words (Harwath & Glass, 2017) and phones (Harwath et al., 2020) from raw", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 536, 505, 549 ], "spans": [ { "bbox": [ 105, 536, 505, 549 ], "score": 1.0, "content": "waveforms. In syntax induction, Shi et al. (2019) proposed to induce phrase-structure grammar", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 547, 505, 560 ], "spans": [ { "bbox": [ 105, 547, 505, 560 ], "score": 1.0, "content": "from parallel image-text data. The above observations motivated us to build a computational model", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 559, 505, 571 ], "spans": [ { "bbox": [ 105, 559, 505, 571 ], "score": 1.0, "content": "that leverages the visual modality to acquire low-level words up to high-level phrase-structure from", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 570, 467, 582 ], "spans": [ { "bbox": [ 105, 570, 467, 582 ], "score": 1.0, "content": "raw speech waveforms, without any intermediate textual forms or any direct supervision.2", "type": "text" } ], "index": 36 } ], "index": 33, "bbox_fs": [ 105, 503, 506, 582 ] }, { "type": "text", "bbox": [ 107, 586, 505, 653 ], "lines": [ { "bbox": [ 105, 586, 506, 599 ], "spans": [ { "bbox": [ 105, 586, 506, 599 ], "score": 1.0, "content": "In this paper, we present the Audio-Visual Neural Syntax Learner (AV-NSL), an approach toward", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 597, 505, 609 ], "spans": [ { "bbox": [ 105, 597, 505, 609 ], "score": 1.0, "content": "learning phrase structure from raw speech waveforms without relying on any kind of intermediate", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 609, 505, 621 ], "spans": [ { "bbox": [ 106, 609, 505, 621 ], "score": 1.0, "content": "textual form or text pre-trained models (Figure 1). In a nutshell, AV-NSL trains a visually-grounded", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 619, 506, 632 ], "spans": [ { "bbox": [ 105, 619, 506, 632 ], "score": 1.0, "content": "syntax learner directly on a sequence of continuous speech representations given by an audio-visual", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 629, 506, 644 ], "spans": [ { "bbox": [ 105, 629, 506, 644 ], "score": 1.0, "content": "word segmentation model. We also introduce a self-training process and an unsupervised decoding", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 641, 505, 654 ], "spans": [ { "bbox": [ 105, 641, 505, 654 ], "score": 1.0, "content": "method to improve the final output of in AV-NSL. To measure the effectiveness of AV-NSL, we", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 223, 505, 236 ], "spans": [ { "bbox": [ 105, 223, 505, 236 ], "score": 1.0, "content": "compare it to text-based syntax learner VG-NSL (Shi et al., 2019) and further introduce a novel", "type": "text", "cross_page": true } ], "index": 14 }, { "bbox": [ 106, 235, 505, 246 ], "spans": [ { "bbox": [ 106, 235, 505, 246 ], "score": 1.0, "content": "evaluation metric, SAIOU, that accounts for structure differences when the number of tree nodes", "type": "text", "cross_page": true } ], "index": 15 }, { "bbox": [ 105, 245, 505, 258 ], "spans": [ { "bbox": [ 105, 245, 505, 258 ], "score": 1.0, "content": "are mismatched. To validate our design choice of AV-NSL, we construct several baselines and", "type": "text", "cross_page": true } ], "index": 16 }, { "bbox": [ 105, 257, 505, 270 ], "spans": [ { "bbox": [ 105, 257, 505, 270 ], "score": 1.0, "content": "introduce alternative modeling choices, including acoustic compound-PCFG (Kim et al., 2019a).", "type": "text", "cross_page": true } ], "index": 17 }, { "bbox": [ 106, 268, 505, 280 ], "spans": [ { "bbox": [ 106, 268, 505, 280 ], "score": 1.0, "content": "Qualitatively, we provide constituency recall analyses and the visualizations of the inferred word", "type": "text", "cross_page": true } ], "index": 18 }, { "bbox": [ 105, 279, 240, 291 ], "spans": [ { "bbox": [ 105, 279, 240, 291 ], "score": 1.0, "content": "segmentation and tree structures.", "type": "text", "cross_page": true } ], "index": 19 } ], "index": 39.5, "bbox_fs": [ 105, 586, 506, 654 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 109, 81, 372, 201 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 81, 372, 201 ], "group_id": 0, "lines": [ { "bbox": [ 109, 81, 372, 201 ], "spans": [ { "bbox": [ 109, 81, 372, 201 ], "score": 0.973, "type": "image", "image_path": "f0d70eacb042ff33533595a02ebee55672bac6319f24de2f807d7ec9aa263d95.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 81, 372, 121.0 ], "spans": [], "index": 0 }, { "bbox": [ 109, 121.0, 372, 161.0 ], "spans": [], "index": 1 }, { "bbox": [ 109, 161.0, 372, 201.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 382, 82, 504, 203 ], "group_id": 0, "lines": [ { "bbox": [ 382, 81, 505, 95 ], "spans": [ { "bbox": [ 382, 81, 505, 95 ], "score": 1.0, "content": "Figure 1: We study the pro-", "type": "text" } ], "index": 3 }, { "bbox": [ 381, 92, 505, 105 ], "spans": [ { "bbox": [ 381, 92, 505, 105 ], "score": 1.0, "content": "cess of inducing phrase struc-", "type": "text" } ], "index": 4 }, { "bbox": [ 381, 103, 505, 116 ], "spans": [ { "bbox": [ 381, 103, 505, 116 ], "score": 1.0, "content": "ture, in the form of con-", "type": "text" } ], "index": 5 }, { "bbox": [ 383, 116, 505, 127 ], "spans": [ { "bbox": [ 383, 116, 505, 127 ], "score": 1.0, "content": "stituency parse tree, on un-", "type": "text" } ], "index": 6 }, { "bbox": [ 381, 124, 505, 138 ], "spans": [ { "bbox": [ 381, 124, 505, 138 ], "score": 1.0, "content": "supervised inferred word seg-", "type": "text" } ], "index": 7 }, { "bbox": [ 382, 136, 505, 149 ], "spans": [ { "bbox": [ 382, 136, 505, 149 ], "score": 1.0, "content": "ments from raw speech wave-", "type": "text" } ], "index": 8 }, { "bbox": [ 382, 147, 505, 159 ], "spans": [ { "bbox": [ 382, 147, 505, 159 ], "score": 1.0, "content": "form. No intermediate text to-", "type": "text" } ], "index": 9 }, { "bbox": [ 381, 157, 505, 170 ], "spans": [ { "bbox": [ 381, 157, 505, 170 ], "score": 1.0, "content": "kens or ASR is needed. For", "type": "text" } ], "index": 10 }, { "bbox": [ 382, 169, 505, 182 ], "spans": [ { "bbox": [ 382, 169, 505, 182 ], "score": 1.0, "content": "illustration purpose, here we", "type": "text" } ], "index": 11 }, { "bbox": [ 381, 180, 505, 193 ], "spans": [ { "bbox": [ 381, 180, 505, 193 ], "score": 1.0, "content": "show the gold parse tree from", "type": "text" } ], "index": 12 }, { "bbox": [ 382, 192, 473, 204 ], "spans": [ { "bbox": [ 382, 192, 473, 204 ], "score": 1.0, "content": "the given text caption.", "type": "text" } ], "index": 13 } ], "index": 8 } ], "index": 4.5 }, { "type": "text", "bbox": [ 107, 223, 505, 289 ], "lines": [ { "bbox": [ 105, 223, 505, 236 ], "spans": [ { "bbox": [ 105, 223, 505, 236 ], "score": 1.0, "content": "compare it to text-based syntax learner VG-NSL (Shi et al., 2019) and further introduce a novel", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 235, 505, 246 ], "spans": [ { "bbox": [ 106, 235, 505, 246 ], "score": 1.0, "content": "evaluation metric, SAIOU, that accounts for structure differences when the number of tree nodes", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 245, 505, 258 ], "spans": [ { "bbox": [ 105, 245, 505, 258 ], "score": 1.0, "content": "are mismatched. To validate our design choice of AV-NSL, we construct several baselines and", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 257, 505, 270 ], "spans": [ { "bbox": [ 105, 257, 505, 270 ], "score": 1.0, "content": "introduce alternative modeling choices, including acoustic compound-PCFG (Kim et al., 2019a).", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 268, 505, 280 ], "spans": [ { "bbox": [ 106, 268, 505, 280 ], "score": 1.0, "content": "Qualitatively, we provide constituency recall analyses and the visualizations of the inferred word", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 279, 240, 291 ], "spans": [ { "bbox": [ 105, 279, 240, 291 ], "score": 1.0, "content": "segmentation and tree structures.", "type": "text" } ], "index": 19 } ], "index": 16.5 }, { "type": "text", "bbox": [ 108, 295, 504, 340 ], "lines": [ { "bbox": [ 106, 295, 505, 307 ], "spans": [ { "bbox": [ 106, 295, 505, 307 ], "score": 1.0, "content": "In summary, we present the first study on inducing phrase structure from visually-grounded speech", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 306, 506, 319 ], "spans": [ { "bbox": [ 106, 306, 506, 319 ], "score": 1.0, "content": "without relying on text, introducing the AV-NSL model (§3) with comprehensive experiments (§4)", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 317, 506, 331 ], "spans": [ { "bbox": [ 105, 317, 506, 331 ], "score": 1.0, "content": "and analysis (§5). As a by product, we improve over the previous state of the art in unsupervised", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 329, 214, 340 ], "spans": [ { "bbox": [ 106, 329, 214, 340 ], "score": 1.0, "content": "word segmentation (§4.4).", "type": "text" } ], "index": 23 } ], "index": 21.5 }, { "type": "title", "bbox": [ 108, 350, 211, 362 ], "lines": [ { "bbox": [ 105, 349, 213, 365 ], "spans": [ { "bbox": [ 105, 349, 213, 365 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "title", "bbox": [ 107, 374, 428, 386 ], "lines": [ { "bbox": [ 105, 373, 430, 388 ], "spans": [ { "bbox": [ 105, 373, 430, 388 ], "score": 1.0, "content": "2.1 UNSUPERVISED AND DISTANTLY SUPERVISED GRAMMAR INDUCTION", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 395, 505, 527 ], "lines": [ { "bbox": [ 105, 394, 505, 408 ], "spans": [ { "bbox": [ 105, 394, 505, 408 ], "score": 1.0, "content": "Much work has been proposed to induce grammar from different sources of distant supervision,", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 406, 505, 419 ], "spans": [ { "bbox": [ 105, 406, 505, 419 ], "score": 1.0, "content": "including language modeling (Shen et al., 2018; 2019; Kim et al., 2019a;b), masked language mod-", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 417, 505, 430 ], "spans": [ { "bbox": [ 106, 417, 505, 430 ], "score": 1.0, "content": "eling (Drozdov et al., 2019), natural language inference (Li et al., 2019), and, more recently, visual", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 428, 506, 441 ], "spans": [ { "bbox": [ 105, 428, 506, 441 ], "score": 1.0, "content": "grounding via image-caption matching (Shi et al., 2019; Zhao & Titov, 2020; Hong et al., 2021;", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 439, 505, 452 ], "spans": [ { "bbox": [ 106, 439, 505, 452 ], "score": 1.0, "content": "Wan et al., 2022, inter alia). There has also been extensive study directly targeting unsupervised", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 450, 505, 462 ], "spans": [ { "bbox": [ 105, 450, 505, 462 ], "score": 1.0, "content": "constituency parsing (Klein & Manning, 2002; 2004; Bod, 2006; Spitkovsky et al., 2013, inter alia).", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 461, 506, 474 ], "spans": [ { "bbox": [ 105, 461, 506, 474 ], "score": 1.0, "content": "To the best of our knowledge, existing work on grammar induction from distant supervision has", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 470, 505, 486 ], "spans": [ { "bbox": [ 105, 470, 505, 486 ], "score": 1.0, "content": "been based almost exclusively on text input. The most relevant work to ours is MMC-PCFG (Zhang", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 482, 505, 496 ], "spans": [ { "bbox": [ 105, 482, 505, 496 ], "score": 1.0, "content": "et al., 2021), where speech features are treated as an auxiliary input for video-text grammar in-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 494, 505, 506 ], "spans": [ { "bbox": [ 105, 494, 505, 506 ], "score": 1.0, "content": "duction. However, text data and an off-the-shelf automatic speech recognition (ASR) model are", "type": "text" } ], "index": 35 }, { "bbox": [ 104, 504, 505, 519 ], "spans": [ { "bbox": [ 104, 504, 505, 519 ], "score": 1.0, "content": "required. In contrast to them, AV-NSL induces constituency parse trees from raw speech bypassing", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 516, 352, 528 ], "spans": [ { "bbox": [ 105, 516, 352, 528 ], "score": 1.0, "content": "text, with distant supervision from parallel audio-visual data.", "type": "text" } ], "index": 37 } ], "index": 31.5 }, { "type": "title", "bbox": [ 107, 541, 369, 552 ], "lines": [ { "bbox": [ 106, 540, 370, 553 ], "spans": [ { "bbox": [ 106, 540, 370, 553 ], "score": 1.0, "content": "2.2 UNSUPERVISED LANGUAGE ACQUISITION FROM SPEECH", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 561, 505, 693 ], "lines": [ { "bbox": [ 105, 560, 505, 574 ], "spans": [ { "bbox": [ 105, 560, 505, 574 ], "score": 1.0, "content": "The earliest work (de Sa, 1994; De Marcken, 1996; Roy & Pentland, 2002) on language acquisition", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 572, 505, 585 ], "spans": [ { "bbox": [ 105, 572, 505, 585 ], "score": 1.0, "content": "from speech required phonetic lexicon/labels in the process. The idea of spoken term discovery, i.e.,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 583, 505, 595 ], "spans": [ { "bbox": [ 106, 583, 505, 595 ], "score": 1.0, "content": "discovering repetitive patterns or keywords from unannotated speech, was first addressed by Park", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 594, 505, 606 ], "spans": [ { "bbox": [ 105, 594, 505, 606 ], "score": 1.0, "content": "& Glass (2007). Thereafter, subsequent work improved upon the original (Zhang & Glass, 2009;", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 604, 505, 617 ], "spans": [ { "bbox": [ 105, 604, 505, 617 ], "score": 1.0, "content": "Jansen & Van Durme, 2011; McInnes & Goldwater, 2011; Zhang, 2013, inter alia). Other related", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 616, 506, 630 ], "spans": [ { "bbox": [ 105, 616, 506, 630 ], "score": 1.0, "content": "work has considered tasks like unsupervised word segmentation and unsupervised ASR, sometimes", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 626, 506, 640 ], "spans": [ { "bbox": [ 105, 626, 506, 640 ], "score": 1.0, "content": "jointly with spoken term discovery (Lee & Glass, 2012; Lee et al., 2015; Kamper et al., 2015; 2017;", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 636, 506, 651 ], "spans": [ { "bbox": [ 105, 636, 506, 651 ], "score": 1.0, "content": "Kamper & van Niekerk, 2021; Chorowski et al., 2021; Bhati et al., 2021; Kamper, 2022; Algayres", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "et al., 2022) The discovery of lexical units was applied to text-free language modeling (Nguyen", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 660, 506, 672 ], "spans": [ { "bbox": [ 105, 660, 506, 672 ], "score": 1.0, "content": "et al., 2020; Peng & Harwath, 2022a) and speech generation (Lakhotia et al., 2021; Polyak et al.,", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 670, 506, 684 ], "spans": [ { "bbox": [ 105, 670, 506, 684 ], "score": 1.0, "content": "2021; Kharitonov et al., 2022). The ZeroSpeech challenges (Versteegh et al., 2015; Dunbar et al.,", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 681, 443, 695 ], "spans": [ { "bbox": [ 105, 681, 443, 695 ], "score": 1.0, "content": "2017; 2019; 2020; Nguyen et al., 2020) have been a major driving force in the field.", "type": "text" } ], "index": 50 } ], "index": 44.5 }, { "type": "text", "bbox": [ 108, 699, 504, 731 ], "lines": [ { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Harwath (2018) opened up a new direction in visually grounded language acquisition, showing", "type": "text" } ], "index": 51 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "word-like (Harwath & Glass, 2017) and phone-like (Harwath et al., 2020) units are acquired from", "type": "text" } ], "index": 52 }, { "bbox": [ 106, 720, 506, 733 ], "spans": [ { "bbox": [ 106, 720, 506, 733 ], "score": 1.0, "content": "speech by analyzing audio-visual retrieval models. Numerous works have studied the character-", "type": "text" } ], "index": 53 } ], "index": 52 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "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": "image", "bbox": [ 109, 81, 372, 201 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 81, 372, 201 ], "group_id": 0, "lines": [ { "bbox": [ 109, 81, 372, 201 ], "spans": [ { "bbox": [ 109, 81, 372, 201 ], "score": 0.973, "type": "image", "image_path": "f0d70eacb042ff33533595a02ebee55672bac6319f24de2f807d7ec9aa263d95.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 81, 372, 121.0 ], "spans": [], "index": 0 }, { "bbox": [ 109, 121.0, 372, 161.0 ], "spans": [], "index": 1 }, { "bbox": [ 109, 161.0, 372, 201.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 382, 82, 504, 203 ], "group_id": 0, "lines": [ { "bbox": [ 382, 81, 505, 95 ], "spans": [ { "bbox": [ 382, 81, 505, 95 ], "score": 1.0, "content": "Figure 1: We study the pro-", "type": "text" } ], "index": 3 }, { "bbox": [ 381, 92, 505, 105 ], "spans": [ { "bbox": [ 381, 92, 505, 105 ], "score": 1.0, "content": "cess of inducing phrase struc-", "type": "text" } ], "index": 4 }, { "bbox": [ 381, 103, 505, 116 ], "spans": [ { "bbox": [ 381, 103, 505, 116 ], "score": 1.0, "content": "ture, in the form of con-", "type": "text" } ], "index": 5 }, { "bbox": [ 383, 116, 505, 127 ], "spans": [ { "bbox": [ 383, 116, 505, 127 ], "score": 1.0, "content": "stituency parse tree, on un-", "type": "text" } ], "index": 6 }, { "bbox": [ 381, 124, 505, 138 ], "spans": [ { "bbox": [ 381, 124, 505, 138 ], "score": 1.0, "content": "supervised inferred word seg-", "type": "text" } ], "index": 7 }, { "bbox": [ 382, 136, 505, 149 ], "spans": [ { "bbox": [ 382, 136, 505, 149 ], "score": 1.0, "content": "ments from raw speech wave-", "type": "text" } ], "index": 8 }, { "bbox": [ 382, 147, 505, 159 ], "spans": [ { "bbox": [ 382, 147, 505, 159 ], "score": 1.0, "content": "form. No intermediate text to-", "type": "text" } ], "index": 9 }, { "bbox": [ 381, 157, 505, 170 ], "spans": [ { "bbox": [ 381, 157, 505, 170 ], "score": 1.0, "content": "kens or ASR is needed. For", "type": "text" } ], "index": 10 }, { "bbox": [ 382, 169, 505, 182 ], "spans": [ { "bbox": [ 382, 169, 505, 182 ], "score": 1.0, "content": "illustration purpose, here we", "type": "text" } ], "index": 11 }, { "bbox": [ 381, 180, 505, 193 ], "spans": [ { "bbox": [ 381, 180, 505, 193 ], "score": 1.0, "content": "show the gold parse tree from", "type": "text" } ], "index": 12 }, { "bbox": [ 382, 192, 473, 204 ], "spans": [ { "bbox": [ 382, 192, 473, 204 ], "score": 1.0, "content": "the given text caption.", "type": "text" } ], "index": 13 } ], "index": 8 } ], "index": 4.5 }, { "type": "text", "bbox": [ 107, 223, 505, 289 ], "lines": [], "index": 16.5, "bbox_fs": [ 105, 223, 505, 291 ], "lines_deleted": true }, { "type": "text", "bbox": [ 108, 295, 504, 340 ], "lines": [ { "bbox": [ 106, 295, 505, 307 ], "spans": [ { "bbox": [ 106, 295, 505, 307 ], "score": 1.0, "content": "In summary, we present the first study on inducing phrase structure from visually-grounded speech", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 306, 506, 319 ], "spans": [ { "bbox": [ 106, 306, 506, 319 ], "score": 1.0, "content": "without relying on text, introducing the AV-NSL model (§3) with comprehensive experiments (§4)", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 317, 506, 331 ], "spans": [ { "bbox": [ 105, 317, 506, 331 ], "score": 1.0, "content": "and analysis (§5). As a by product, we improve over the previous state of the art in unsupervised", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 329, 214, 340 ], "spans": [ { "bbox": [ 106, 329, 214, 340 ], "score": 1.0, "content": "word segmentation (§4.4).", "type": "text" } ], "index": 23 } ], "index": 21.5, "bbox_fs": [ 105, 295, 506, 340 ] }, { "type": "title", "bbox": [ 108, 350, 211, 362 ], "lines": [ { "bbox": [ 105, 349, 213, 365 ], "spans": [ { "bbox": [ 105, 349, 213, 365 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "title", "bbox": [ 107, 374, 428, 386 ], "lines": [ { "bbox": [ 105, 373, 430, 388 ], "spans": [ { "bbox": [ 105, 373, 430, 388 ], "score": 1.0, "content": "2.1 UNSUPERVISED AND DISTANTLY SUPERVISED GRAMMAR INDUCTION", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 395, 505, 527 ], "lines": [ { "bbox": [ 105, 394, 505, 408 ], "spans": [ { "bbox": [ 105, 394, 505, 408 ], "score": 1.0, "content": "Much work has been proposed to induce grammar from different sources of distant supervision,", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 406, 505, 419 ], "spans": [ { "bbox": [ 105, 406, 505, 419 ], "score": 1.0, "content": "including language modeling (Shen et al., 2018; 2019; Kim et al., 2019a;b), masked language mod-", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 417, 505, 430 ], "spans": [ { "bbox": [ 106, 417, 505, 430 ], "score": 1.0, "content": "eling (Drozdov et al., 2019), natural language inference (Li et al., 2019), and, more recently, visual", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 428, 506, 441 ], "spans": [ { "bbox": [ 105, 428, 506, 441 ], "score": 1.0, "content": "grounding via image-caption matching (Shi et al., 2019; Zhao & Titov, 2020; Hong et al., 2021;", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 439, 505, 452 ], "spans": [ { "bbox": [ 106, 439, 505, 452 ], "score": 1.0, "content": "Wan et al., 2022, inter alia). There has also been extensive study directly targeting unsupervised", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 450, 505, 462 ], "spans": [ { "bbox": [ 105, 450, 505, 462 ], "score": 1.0, "content": "constituency parsing (Klein & Manning, 2002; 2004; Bod, 2006; Spitkovsky et al., 2013, inter alia).", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 461, 506, 474 ], "spans": [ { "bbox": [ 105, 461, 506, 474 ], "score": 1.0, "content": "To the best of our knowledge, existing work on grammar induction from distant supervision has", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 470, 505, 486 ], "spans": [ { "bbox": [ 105, 470, 505, 486 ], "score": 1.0, "content": "been based almost exclusively on text input. The most relevant work to ours is MMC-PCFG (Zhang", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 482, 505, 496 ], "spans": [ { "bbox": [ 105, 482, 505, 496 ], "score": 1.0, "content": "et al., 2021), where speech features are treated as an auxiliary input for video-text grammar in-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 494, 505, 506 ], "spans": [ { "bbox": [ 105, 494, 505, 506 ], "score": 1.0, "content": "duction. However, text data and an off-the-shelf automatic speech recognition (ASR) model are", "type": "text" } ], "index": 35 }, { "bbox": [ 104, 504, 505, 519 ], "spans": [ { "bbox": [ 104, 504, 505, 519 ], "score": 1.0, "content": "required. In contrast to them, AV-NSL induces constituency parse trees from raw speech bypassing", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 516, 352, 528 ], "spans": [ { "bbox": [ 105, 516, 352, 528 ], "score": 1.0, "content": "text, with distant supervision from parallel audio-visual data.", "type": "text" } ], "index": 37 } ], "index": 31.5, "bbox_fs": [ 104, 394, 506, 528 ] }, { "type": "title", "bbox": [ 107, 541, 369, 552 ], "lines": [ { "bbox": [ 106, 540, 370, 553 ], "spans": [ { "bbox": [ 106, 540, 370, 553 ], "score": 1.0, "content": "2.2 UNSUPERVISED LANGUAGE ACQUISITION FROM SPEECH", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 561, 505, 693 ], "lines": [ { "bbox": [ 105, 560, 505, 574 ], "spans": [ { "bbox": [ 105, 560, 505, 574 ], "score": 1.0, "content": "The earliest work (de Sa, 1994; De Marcken, 1996; Roy & Pentland, 2002) on language acquisition", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 572, 505, 585 ], "spans": [ { "bbox": [ 105, 572, 505, 585 ], "score": 1.0, "content": "from speech required phonetic lexicon/labels in the process. The idea of spoken term discovery, i.e.,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 583, 505, 595 ], "spans": [ { "bbox": [ 106, 583, 505, 595 ], "score": 1.0, "content": "discovering repetitive patterns or keywords from unannotated speech, was first addressed by Park", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 594, 505, 606 ], "spans": [ { "bbox": [ 105, 594, 505, 606 ], "score": 1.0, "content": "& Glass (2007). Thereafter, subsequent work improved upon the original (Zhang & Glass, 2009;", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 604, 505, 617 ], "spans": [ { "bbox": [ 105, 604, 505, 617 ], "score": 1.0, "content": "Jansen & Van Durme, 2011; McInnes & Goldwater, 2011; Zhang, 2013, inter alia). Other related", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 616, 506, 630 ], "spans": [ { "bbox": [ 105, 616, 506, 630 ], "score": 1.0, "content": "work has considered tasks like unsupervised word segmentation and unsupervised ASR, sometimes", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 626, 506, 640 ], "spans": [ { "bbox": [ 105, 626, 506, 640 ], "score": 1.0, "content": "jointly with spoken term discovery (Lee & Glass, 2012; Lee et al., 2015; Kamper et al., 2015; 2017;", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 636, 506, 651 ], "spans": [ { "bbox": [ 105, 636, 506, 651 ], "score": 1.0, "content": "Kamper & van Niekerk, 2021; Chorowski et al., 2021; Bhati et al., 2021; Kamper, 2022; Algayres", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "et al., 2022) The discovery of lexical units was applied to text-free language modeling (Nguyen", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 660, 506, 672 ], "spans": [ { "bbox": [ 105, 660, 506, 672 ], "score": 1.0, "content": "et al., 2020; Peng & Harwath, 2022a) and speech generation (Lakhotia et al., 2021; Polyak et al.,", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 670, 506, 684 ], "spans": [ { "bbox": [ 105, 670, 506, 684 ], "score": 1.0, "content": "2021; Kharitonov et al., 2022). The ZeroSpeech challenges (Versteegh et al., 2015; Dunbar et al.,", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 681, 443, 695 ], "spans": [ { "bbox": [ 105, 681, 443, 695 ], "score": 1.0, "content": "2017; 2019; 2020; Nguyen et al., 2020) have been a major driving force in the field.", "type": "text" } ], "index": 50 } ], "index": 44.5, "bbox_fs": [ 105, 560, 506, 695 ] }, { "type": "text", "bbox": [ 108, 699, 504, 731 ], "lines": [ { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Harwath (2018) opened up a new direction in visually grounded language acquisition, showing", "type": "text" } ], "index": 51 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "word-like (Harwath & Glass, 2017) and phone-like (Harwath et al., 2020) units are acquired from", "type": "text" } ], "index": 52 }, { "bbox": [ 106, 720, 506, 733 ], "spans": [ { "bbox": [ 106, 720, 506, 733 ], "score": 1.0, "content": "speech by analyzing audio-visual retrieval models. Numerous works have studied the character-", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "istics of the linguistic information acquired in visually grounded speech models (Havard et al.,", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "2019; Khorrami & Ras¨ anen, 2021; Olaleye & Kamper, 2021; Wang & Hasegawa-Johnson, 2021; ¨", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "Mitja Nikolaus, 2022). Peng & Harwath (2022b) shows that clear word segmentation and identifica-", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 106, 115, 505, 128 ], "spans": [ { "bbox": [ 106, 115, 505, 128 ], "score": 1.0, "content": "tion naturally emerge from a visually grounded, self-supervised speech model named VG-HuBERT,", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 105, 125, 506, 141 ], "spans": [ { "bbox": [ 105, 125, 506, 141 ], "score": 1.0, "content": "by analyzing the model’s self-attention heads. Unlike the above, AV-NSL acquires phrase structure,", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 105, 137, 407, 150 ], "spans": [ { "bbox": [ 105, 137, 407, 150 ], "score": 1.0, "content": "in the form of constituency parsing on top of unsupervised word segments.", "type": "text", "cross_page": true } ], "index": 5 } ], "index": 52, "bbox_fs": [ 105, 698, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 149 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "istics of the linguistic information acquired in visually grounded speech models (Havard et al.,", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "2019; Khorrami & Ras¨ anen, 2021; Olaleye & Kamper, 2021; Wang & Hasegawa-Johnson, 2021; ¨", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "Mitja Nikolaus, 2022). Peng & Harwath (2022b) shows that clear word segmentation and identifica-", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 115, 505, 128 ], "spans": [ { "bbox": [ 106, 115, 505, 128 ], "score": 1.0, "content": "tion naturally emerge from a visually grounded, self-supervised speech model named VG-HuBERT,", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 125, 506, 141 ], "spans": [ { "bbox": [ 105, 125, 506, 141 ], "score": 1.0, "content": "by analyzing the model’s self-attention heads. Unlike the above, AV-NSL acquires phrase structure,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 137, 407, 150 ], "spans": [ { "bbox": [ 105, 137, 407, 150 ], "score": 1.0, "content": "in the form of constituency parsing on top of unsupervised word segments.", "type": "text" } ], "index": 5 } ], "index": 2.5 }, { "type": "title", "bbox": [ 108, 163, 305, 173 ], "lines": [ { "bbox": [ 106, 163, 308, 175 ], "spans": [ { "bbox": [ 106, 163, 308, 175 ], "score": 1.0, "content": "2.3 SPEECH PARSING AND ITS APPLICATIONS", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 106, 183, 505, 293 ], "lines": [ { "bbox": [ 105, 183, 506, 195 ], "spans": [ { "bbox": [ 105, 183, 506, 195 ], "score": 1.0, "content": "Early work on speech parsing can be traced back to the SParseval toolkit (Roark et al., 2006), for", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 195, 506, 207 ], "spans": [ { "bbox": [ 105, 195, 506, 207 ], "score": 1.0, "content": "evaluating text parsers given (errorful) ASR output. Tran et al. (2018; 2019); Tran & Ostendorf", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 205, 505, 218 ], "spans": [ { "bbox": [ 106, 205, 505, 218 ], "score": 1.0, "content": "(2021) explored the use of acoustic-prosodic features for text parsing with auxiliary speech input.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 216, 505, 229 ], "spans": [ { "bbox": [ 105, 216, 505, 229 ], "score": 1.0, "content": "Lou et al. (2019) trained a text parser (Kitaev & Klein, 2018) to detect speech disfluencies. In the", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 227, 505, 240 ], "spans": [ { "bbox": [ 104, 227, 505, 240 ], "score": 1.0, "content": "past, syntax has also been studied in the context of speech prosody (Wagner & Watson, 2010; Kohn ¨", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 238, 506, 251 ], "spans": [ { "bbox": [ 105, 238, 506, 251 ], "score": 1.0, "content": "et al., 2018). The most relevant work to ours is Pupier et al. (2022), where a text dependency parser", "type": "text" } ], "index": 12 }, { "bbox": [ 104, 249, 506, 263 ], "spans": [ { "bbox": [ 104, 249, 506, 263 ], "score": 1.0, "content": "is trained from speech jointly with an ASR model. Moreover, text syntax parsing has been applied", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 260, 505, 272 ], "spans": [ { "bbox": [ 105, 260, 505, 272 ], "score": 1.0, "content": "to prosody modeling in end-to-end text-to-speech (TTS; Guo et al., 2019; Tyagi et al., 2020; Kaiki", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 271, 505, 284 ], "spans": [ { "bbox": [ 105, 271, 505, 284 ], "score": 1.0, "content": "et al., 2021). This work builds on top of pre-existing text parsing algorithms or pre-existing phrase", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 282, 464, 294 ], "spans": [ { "bbox": [ 106, 282, 464, 294 ], "score": 1.0, "content": "structures from text, whereas we study phrase structure acquisition in the absence of text.", "type": "text" } ], "index": 16 } ], "index": 11.5 }, { "type": "title", "bbox": [ 107, 310, 173, 322 ], "lines": [ { "bbox": [ 104, 308, 175, 326 ], "spans": [ { "bbox": [ 104, 308, 175, 326 ], "score": 1.0, "content": "3 METHOD", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "image", "bbox": [ 109, 342, 501, 508 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 342, 501, 508 ], "group_id": 0, "lines": [ { "bbox": [ 109, 342, 501, 508 ], "spans": [ { "bbox": [ 109, 342, 501, 508 ], "score": 0.968, "type": "image", "image_path": "1022ab95dbbcbbc5e4f5d65c9b2b1a42eaaa69555b3c232f911b502e0c1cc2d5.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 109, 342, 501, 397.3333333333333 ], "spans": [], "index": 18 }, { "bbox": [ 109, 397.3333333333333, 501, 452.66666666666663 ], "spans": [], "index": 19 }, { "bbox": [ 109, 452.66666666666663, 501, 507.99999999999994 ], "spans": [], "index": 20 } ] }, { "type": "image_caption", "bbox": [ 106, 516, 502, 528 ], "group_id": 0, "lines": [ { "bbox": [ 107, 515, 504, 530 ], "spans": [ { "bbox": [ 107, 515, 504, 530 ], "score": 1.0, "content": "Figure 2: Illustration of AV-NSL, which extends VG-NSL (Shi et al., 2019) to audio-visual inputs.", "type": "text" } ], "index": 21 } ], "index": 21 } ], "index": 20.0 }, { "type": "text", "bbox": [ 107, 542, 505, 642 ], "lines": [ { "bbox": [ 105, 542, 505, 555 ], "spans": [ { "bbox": [ 105, 542, 505, 555 ], "score": 1.0, "content": "Given a set of paired spoken captions and images, the Audio-Visual Neural Syntax Learner (AV-", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 553, 505, 567 ], "spans": [ { "bbox": [ 105, 553, 505, 567 ], "score": 1.0, "content": "NSL) infers phrase structures from subsequences of raw speech segments without relying on text.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 564, 505, 578 ], "spans": [ { "bbox": [ 105, 564, 505, 578 ], "score": 1.0, "content": "The basis of AV-NSL is the Visually-Grounded Neural Syntax Learner (VG-NSL) (Shi et al., 2019).", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 574, 506, 590 ], "spans": [ { "bbox": [ 105, 574, 506, 590 ], "score": 1.0, "content": "VG-NSL learns constituency parse trees by guiding a sequential tree sampling process with text-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 586, 506, 601 ], "spans": [ { "bbox": [ 105, 586, 506, 601 ], "score": 1.0, "content": "image matching. To extend VG-NSL to audio-visual inputs, the central challenge is extracting", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 597, 505, 611 ], "spans": [ { "bbox": [ 105, 597, 505, 611 ], "score": 1.0, "content": "semantically-meaningful word segments from unannotated speech. We break down the problem into", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 609, 505, 622 ], "spans": [ { "bbox": [ 105, 609, 505, 622 ], "score": 1.0, "content": "a two-step process: (1) obtaining sequences of word segments, and (2) extracting segment-level", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 618, 506, 633 ], "spans": [ { "bbox": [ 105, 618, 506, 633 ], "score": 1.0, "content": "self-supervised representations. With these simple modifications, AV-NSL learns non-trivial phrase", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 631, 461, 644 ], "spans": [ { "bbox": [ 105, 631, 461, 644 ], "score": 1.0, "content": "structure without ever reading text, instead by listening to speech and looking at images.", "type": "text" } ], "index": 30 } ], "index": 26 }, { "type": "title", "bbox": [ 107, 655, 417, 667 ], "lines": [ { "bbox": [ 105, 654, 419, 669 ], "spans": [ { "bbox": [ 105, 654, 419, 669 ], "score": 1.0, "content": "3.1 BACKGROUND: VISUALLY-GROUNDED NEURAL SYNTAX LEARNER", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 107, 677, 504, 732 ], "lines": [ { "bbox": [ 105, 675, 506, 691 ], "spans": [ { "bbox": [ 105, 675, 506, 691 ], "score": 1.0, "content": "VG-NSL (Shi et al., 2019) is composed of a bottom-up text parser and a text-image embedding", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "matching module. The parser consists of an embedding similarity scoring function score and an", "type": "text" } ], "index": 33 }, { "bbox": [ 101, 699, 509, 734 ], "spans": [ { "bbox": [ 101, 699, 158, 734 ], "score": 1.0, "content": "embedding cembeddings sively scorin", "type": "text" }, { "bbox": [ 158, 709, 219, 722 ], "score": 0.93, "content": "{ \\cal { W } } = \\{ w _ { i } ^ { 0 } \\} _ { i = 1 } ^ { N }", "type": "inline_equation" }, { "bbox": [ 219, 699, 259, 734 ], "score": 1.0, "content": "nction comof length g adjacent", "type": "text" }, { "bbox": [ 260, 712, 270, 720 ], "score": 0.83, "content": "N", "type": "inline_equation" }, { "bbox": [ 270, 699, 392, 734 ], "score": 1.0, "content": "ne. Given a text caption, den, the parser synthesizes a consmbeddings at each step. At step", "type": "text" }, { "bbox": [ 398, 699, 509, 734 ], "score": 1.0, "content": "ed by a sequence of wordtuency parse tree by recur-, VG-NSL (1) evaluates all", "type": "text" } ], "index": 34 }, { "bbox": [ 393, 721, 398, 731 ], "spans": [ { "bbox": [ 393, 721, 398, 731 ], "score": 0.58, "content": "t", "type": "inline_equation" } ], "index": 35 } ], "index": 33.5 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 303, 751, 309, 759 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "", "type": "text", "height": 12, "width": 9 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 149 ], "lines": [], "index": 2.5, "bbox_fs": [ 105, 82, 506, 150 ], "lines_deleted": true }, { "type": "title", "bbox": [ 108, 163, 305, 173 ], "lines": [ { "bbox": [ 106, 163, 308, 175 ], "spans": [ { "bbox": [ 106, 163, 308, 175 ], "score": 1.0, "content": "2.3 SPEECH PARSING AND ITS APPLICATIONS", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 106, 183, 505, 293 ], "lines": [ { "bbox": [ 105, 183, 506, 195 ], "spans": [ { "bbox": [ 105, 183, 506, 195 ], "score": 1.0, "content": "Early work on speech parsing can be traced back to the SParseval toolkit (Roark et al., 2006), for", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 195, 506, 207 ], "spans": [ { "bbox": [ 105, 195, 506, 207 ], "score": 1.0, "content": "evaluating text parsers given (errorful) ASR output. Tran et al. (2018; 2019); Tran & Ostendorf", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 205, 505, 218 ], "spans": [ { "bbox": [ 106, 205, 505, 218 ], "score": 1.0, "content": "(2021) explored the use of acoustic-prosodic features for text parsing with auxiliary speech input.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 216, 505, 229 ], "spans": [ { "bbox": [ 105, 216, 505, 229 ], "score": 1.0, "content": "Lou et al. (2019) trained a text parser (Kitaev & Klein, 2018) to detect speech disfluencies. In the", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 227, 505, 240 ], "spans": [ { "bbox": [ 104, 227, 505, 240 ], "score": 1.0, "content": "past, syntax has also been studied in the context of speech prosody (Wagner & Watson, 2010; Kohn ¨", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 238, 506, 251 ], "spans": [ { "bbox": [ 105, 238, 506, 251 ], "score": 1.0, "content": "et al., 2018). The most relevant work to ours is Pupier et al. (2022), where a text dependency parser", "type": "text" } ], "index": 12 }, { "bbox": [ 104, 249, 506, 263 ], "spans": [ { "bbox": [ 104, 249, 506, 263 ], "score": 1.0, "content": "is trained from speech jointly with an ASR model. Moreover, text syntax parsing has been applied", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 260, 505, 272 ], "spans": [ { "bbox": [ 105, 260, 505, 272 ], "score": 1.0, "content": "to prosody modeling in end-to-end text-to-speech (TTS; Guo et al., 2019; Tyagi et al., 2020; Kaiki", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 271, 505, 284 ], "spans": [ { "bbox": [ 105, 271, 505, 284 ], "score": 1.0, "content": "et al., 2021). This work builds on top of pre-existing text parsing algorithms or pre-existing phrase", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 282, 464, 294 ], "spans": [ { "bbox": [ 106, 282, 464, 294 ], "score": 1.0, "content": "structures from text, whereas we study phrase structure acquisition in the absence of text.", "type": "text" } ], "index": 16 } ], "index": 11.5, "bbox_fs": [ 104, 183, 506, 294 ] }, { "type": "title", "bbox": [ 107, 310, 173, 322 ], "lines": [ { "bbox": [ 104, 308, 175, 326 ], "spans": [ { "bbox": [ 104, 308, 175, 326 ], "score": 1.0, "content": "3 METHOD", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "image", "bbox": [ 109, 342, 501, 508 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 342, 501, 508 ], "group_id": 0, "lines": [ { "bbox": [ 109, 342, 501, 508 ], "spans": [ { "bbox": [ 109, 342, 501, 508 ], "score": 0.968, "type": "image", "image_path": "1022ab95dbbcbbc5e4f5d65c9b2b1a42eaaa69555b3c232f911b502e0c1cc2d5.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 109, 342, 501, 397.3333333333333 ], "spans": [], "index": 18 }, { "bbox": [ 109, 397.3333333333333, 501, 452.66666666666663 ], "spans": [], "index": 19 }, { "bbox": [ 109, 452.66666666666663, 501, 507.99999999999994 ], "spans": [], "index": 20 } ] }, { "type": "image_caption", "bbox": [ 106, 516, 502, 528 ], "group_id": 0, "lines": [ { "bbox": [ 107, 515, 504, 530 ], "spans": [ { "bbox": [ 107, 515, 504, 530 ], "score": 1.0, "content": "Figure 2: Illustration of AV-NSL, which extends VG-NSL (Shi et al., 2019) to audio-visual inputs.", "type": "text" } ], "index": 21 } ], "index": 21 } ], "index": 20.0 }, { "type": "text", "bbox": [ 107, 542, 505, 642 ], "lines": [ { "bbox": [ 105, 542, 505, 555 ], "spans": [ { "bbox": [ 105, 542, 505, 555 ], "score": 1.0, "content": "Given a set of paired spoken captions and images, the Audio-Visual Neural Syntax Learner (AV-", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 553, 505, 567 ], "spans": [ { "bbox": [ 105, 553, 505, 567 ], "score": 1.0, "content": "NSL) infers phrase structures from subsequences of raw speech segments without relying on text.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 564, 505, 578 ], "spans": [ { "bbox": [ 105, 564, 505, 578 ], "score": 1.0, "content": "The basis of AV-NSL is the Visually-Grounded Neural Syntax Learner (VG-NSL) (Shi et al., 2019).", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 574, 506, 590 ], "spans": [ { "bbox": [ 105, 574, 506, 590 ], "score": 1.0, "content": "VG-NSL learns constituency parse trees by guiding a sequential tree sampling process with text-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 586, 506, 601 ], "spans": [ { "bbox": [ 105, 586, 506, 601 ], "score": 1.0, "content": "image matching. To extend VG-NSL to audio-visual inputs, the central challenge is extracting", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 597, 505, 611 ], "spans": [ { "bbox": [ 105, 597, 505, 611 ], "score": 1.0, "content": "semantically-meaningful word segments from unannotated speech. We break down the problem into", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 609, 505, 622 ], "spans": [ { "bbox": [ 105, 609, 505, 622 ], "score": 1.0, "content": "a two-step process: (1) obtaining sequences of word segments, and (2) extracting segment-level", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 618, 506, 633 ], "spans": [ { "bbox": [ 105, 618, 506, 633 ], "score": 1.0, "content": "self-supervised representations. With these simple modifications, AV-NSL learns non-trivial phrase", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 631, 461, 644 ], "spans": [ { "bbox": [ 105, 631, 461, 644 ], "score": 1.0, "content": "structure without ever reading text, instead by listening to speech and looking at images.", "type": "text" } ], "index": 30 } ], "index": 26, "bbox_fs": [ 105, 542, 506, 644 ] }, { "type": "title", "bbox": [ 107, 655, 417, 667 ], "lines": [ { "bbox": [ 105, 654, 419, 669 ], "spans": [ { "bbox": [ 105, 654, 419, 669 ], "score": 1.0, "content": "3.1 BACKGROUND: VISUALLY-GROUNDED NEURAL SYNTAX LEARNER", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 107, 677, 504, 732 ], "lines": [ { "bbox": [ 105, 675, 506, 691 ], "spans": [ { "bbox": [ 105, 675, 506, 691 ], "score": 1.0, "content": "VG-NSL (Shi et al., 2019) is composed of a bottom-up text parser and a text-image embedding", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "matching module. The parser consists of an embedding similarity scoring function score and an", "type": "text" } ], "index": 33 }, { "bbox": [ 101, 699, 509, 734 ], "spans": [ { "bbox": [ 101, 699, 158, 734 ], "score": 1.0, "content": "embedding cembeddings sively scorin", "type": "text" }, { "bbox": [ 158, 709, 219, 722 ], "score": 0.93, "content": "{ \\cal { W } } = \\{ w _ { i } ^ { 0 } \\} _ { i = 1 } ^ { N }", "type": "inline_equation" }, { "bbox": [ 219, 699, 259, 734 ], "score": 1.0, "content": "nction comof length g adjacent", "type": "text" }, { "bbox": [ 260, 712, 270, 720 ], "score": 0.83, "content": "N", "type": "inline_equation" }, { "bbox": [ 270, 699, 392, 734 ], "score": 1.0, "content": "ne. Given a text caption, den, the parser synthesizes a consmbeddings at each step. At step", "type": "text" }, { "bbox": [ 398, 699, 509, 734 ], "score": 1.0, "content": "ed by a sequence of wordtuency parse tree by recur-, VG-NSL (1) evaluates all", "type": "text" } ], "index": 34 }, { "bbox": [ 393, 721, 398, 731 ], "spans": [ { "bbox": [ 393, 721, 398, 731 ], "score": 0.58, "content": "t", "type": "inline_equation" } ], "index": 35 } ], "index": 33.5, "bbox_fs": [ 101, 675, 509, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 150 ], "lines": [ { "bbox": [ 105, 81, 506, 96 ], "spans": [ { "bbox": [ 105, 81, 238, 96 ], "score": 1.0, "content": "consecutive pairs of embeddings", "type": "text" }, { "bbox": [ 239, 82, 281, 95 ], "score": 0.93, "content": "\\langle w _ { i } ^ { t } , w _ { i + 1 } ^ { t } \\rangle", "type": "inline_equation" }, { "bbox": [ 281, 81, 506, 96 ], "score": 1.0, "content": "and assigns a scalar score to each with score, (2) selects", "type": "text" } ], "index": 0 }, { "bbox": [ 104, 93, 506, 108 ], "spans": [ { "bbox": [ 104, 93, 131, 108 ], "score": 1.0, "content": "a pair", "type": "text" }, { "bbox": [ 132, 94, 179, 107 ], "score": 0.92, "content": "\\langle w _ { i ^ { \\prime } } ^ { t } , w _ { i ^ { \\prime } + 1 } ^ { t } \\rangle", "type": "inline_equation" }, { "bbox": [ 179, 93, 506, 108 ], "score": 1.0, "content": "based on the corresponding scores,3 and (3) combines the selected pair of embed-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 505, 119 ], "spans": [ { "bbox": [ 105, 105, 505, 119 ], "score": 1.0, "content": "dings via combine to form a new phrase embedding for the next step, copying the remaining ones to", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 116, 506, 129 ], "spans": [ { "bbox": [ 106, 116, 506, 129 ], "score": 1.0, "content": "the next step. In VG-NSL, score is parameterized by a 2-layer ReLU-activated MLP, and combine is", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 127, 505, 141 ], "spans": [ { "bbox": [ 105, 127, 505, 141 ], "score": 1.0, "content": "defined by the L2-normalized sum of the input embeddings. The resulting tree is inherently binary", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 138, 505, 151 ], "spans": [ { "bbox": [ 105, 138, 160, 151 ], "score": 1.0, "content": "and there are", "type": "text" }, { "bbox": [ 160, 138, 182, 149 ], "score": 0.85, "content": "N - 1", "type": "inline_equation" }, { "bbox": [ 182, 138, 505, 151 ], "score": 1.0, "content": "combining steps in total, as the tree parser must combine two nodes in each step.", "type": "text" } ], "index": 5 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 155, 505, 210 ], "lines": [ { "bbox": [ 106, 155, 505, 167 ], "spans": [ { "bbox": [ 106, 155, 505, 167 ], "score": 1.0, "content": "The text-image embedding matching module of VG-NSL is based on the standard hinge-based triplet", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 167, 504, 178 ], "spans": [ { "bbox": [ 105, 167, 504, 178 ], "score": 1.0, "content": "loss (Kiros et al., 2014), where the sentence-based loss is modified to a phrase-based one. Addition-", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 177, 505, 189 ], "spans": [ { "bbox": [ 106, 177, 505, 189 ], "score": 1.0, "content": "ally, the loss function is adapted to estimate the visual concreteness of a text span: intuitively, the", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 187, 505, 201 ], "spans": [ { "bbox": [ 105, 187, 305, 201 ], "score": 1.0, "content": "smaller the loss related to a candidate constituent", "type": "text" }, { "bbox": [ 305, 190, 311, 198 ], "score": 0.74, "content": "c", "type": "inline_equation" }, { "bbox": [ 311, 187, 434, 201 ], "score": 1.0, "content": ", the larger the concreteness of", "type": "text" }, { "bbox": [ 434, 190, 439, 198 ], "score": 0.73, "content": "c", "type": "inline_equation" }, { "bbox": [ 439, 187, 505, 201 ], "score": 1.0, "content": ", and vice versa.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 199, 301, 211 ], "spans": [ { "bbox": [ 105, 199, 241, 211 ], "score": 1.0, "content": "The concreteness of a constituent", "type": "text" }, { "bbox": [ 242, 201, 248, 209 ], "score": 0.77, "content": "c", "type": "inline_equation" }, { "bbox": [ 248, 199, 301, 211 ], "score": 1.0, "content": "is defined as", "type": "text" } ], "index": 10 } ], "index": 8 }, { "type": "interline_equation", "bbox": [ 159, 211, 482, 238 ], "lines": [ { "bbox": [ 159, 211, 482, 238 ], "spans": [ { "bbox": [ 159, 211, 482, 238 ], "score": 0.91, "content": "\\mathbf { \\nabla } \\cdot e \\left( \\mathbf { c } ; \\mathbf { i } \\right) = \\sum _ { \\mathbf { c } ^ { \\prime } } \\left[ \\cos \\left( \\mathbf { i } , \\mathbf { c } \\right) - \\cos \\left( \\mathbf { i } , \\mathbf { c } ^ { \\prime } \\right) - \\delta \\right] _ { + } + \\sum _ { \\mathbf { i } ^ { \\prime } } \\left[ \\cos \\left( \\mathbf { i } ^ { \\prime } , \\mathbf { c } \\right) - \\cos \\left( \\mathbf { i } ^ { \\prime } , \\mathbf { c } \\right) - \\delta \\right] _ { + } ,", "type": "interline_equation", "image_path": "6815656dfcb654c315392bbb2c36e1bb9d70834ef62cd242118ff9413cb90c91.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 159, 211, 482, 220.0 ], "spans": [], "index": 11 }, { "bbox": [ 159, 220.0, 482, 229.0 ], "spans": [], "index": 12 }, { "bbox": [ 159, 229.0, 482, 238.0 ], "spans": [], "index": 13 } ] }, { "type": "text", "bbox": [ 106, 240, 505, 296 ], "lines": [ { "bbox": [ 105, 240, 504, 253 ], "spans": [ { "bbox": [ 105, 240, 262, 253 ], "score": 1.0, "content": "where c is the vector representation of", "type": "text" }, { "bbox": [ 263, 243, 268, 251 ], "score": 0.58, "content": "c", "type": "inline_equation" }, { "bbox": [ 269, 240, 484, 253 ], "score": 1.0, "content": "; i is the corresponding vector of the parallel image of", "type": "text" }, { "bbox": [ 485, 241, 504, 252 ], "score": 0.45, "content": "c ; \\mathbf { c } ^ { \\prime }", "type": "inline_equation" } ], "index": 14 }, { "bbox": [ 105, 250, 506, 264 ], "spans": [ { "bbox": [ 105, 250, 393, 264 ], "score": 1.0, "content": "is a candidate constituent from a sentence that is not in parallel with i;", "type": "text" }, { "bbox": [ 393, 252, 401, 262 ], "score": 0.77, "content": "\\mathbf { i } ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 401, 250, 506, 264 ], "score": 1.0, "content": "is an image that is not in", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 506, 276 ], "spans": [ { "bbox": [ 105, 263, 159, 276 ], "score": 1.0, "content": "parallel with", "type": "text" }, { "bbox": [ 159, 263, 176, 273 ], "score": 0.27, "content": "c ; \\delta", "type": "inline_equation" }, { "bbox": [ 176, 263, 287, 276 ], "score": 1.0, "content": "is a constant margin. Here,", "type": "text" }, { "bbox": [ 288, 263, 358, 275 ], "score": 0.92, "content": "[ \\cdot ] _ { + } : = \\operatorname* { m a x } ( \\cdot , 0 )", "type": "inline_equation" }, { "bbox": [ 359, 263, 506, 276 ], "score": 1.0, "content": ". Finally, the estimated concreteness", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 285 ], "spans": [ { "bbox": [ 105, 274, 505, 285 ], "score": 1.0, "content": "scores are passed back to the parser as rewards to the constituents. VG-NSL jointly optimizes the", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 285, 471, 297 ], "spans": [ { "bbox": [ 106, 285, 471, 297 ], "score": 1.0, "content": "visual-semantic embedding loss, and trains the parser with REINFORCE (Williams, 1992).", "type": "text" } ], "index": 18 } ], "index": 16 }, { "type": "title", "bbox": [ 107, 309, 319, 320 ], "lines": [ { "bbox": [ 106, 309, 320, 322 ], "spans": [ { "bbox": [ 106, 309, 320, 322 ], "score": 1.0, "content": "3.2 AUDIO-VISUAL NEURAL SYNTAX LEARNER", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 329, 296, 538 ], "lines": [ { "bbox": [ 106, 329, 297, 342 ], "spans": [ { "bbox": [ 106, 329, 297, 342 ], "score": 1.0, "content": "AV-NSL extends VG-NSL by: (1) incorporat-", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 341, 297, 353 ], "spans": [ { "bbox": [ 106, 341, 297, 353 ], "score": 1.0, "content": "ing an audio-visual word segmentation model", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 352, 297, 364 ], "spans": [ { "bbox": [ 106, 352, 297, 364 ], "score": 1.0, "content": "for obtaining sequences of word segments from", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 361, 298, 376 ], "spans": [ { "bbox": [ 105, 361, 298, 376 ], "score": 1.0, "content": "unannotated speech, (2) jointly optimizing", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 373, 297, 387 ], "spans": [ { "bbox": [ 105, 373, 297, 387 ], "score": 1.0, "content": "segment-level embeddings along with phrase", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 384, 297, 397 ], "spans": [ { "bbox": [ 105, 384, 297, 397 ], "score": 1.0, "content": "structure induction, and (3) employing deeper", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 396, 297, 407 ], "spans": [ { "bbox": [ 106, 396, 297, 407 ], "score": 1.0, "content": "score and combine function parameterization in", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 407, 297, 419 ], "spans": [ { "bbox": [ 105, 407, 297, 419 ], "score": 1.0, "content": "the parsing module. We empirically found (3)", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 418, 297, 430 ], "spans": [ { "bbox": [ 105, 418, 297, 430 ], "score": 1.0, "content": "necessary, mainly because speech embeddings", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 429, 297, 440 ], "spans": [ { "bbox": [ 106, 429, 297, 440 ], "score": 1.0, "content": "are inherently richer, less clean, and semanti-", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 440, 297, 451 ], "spans": [ { "bbox": [ 106, 440, 297, 451 ], "score": 1.0, "content": "cally more ambiguous than word embeddings.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 450, 298, 463 ], "spans": [ { "bbox": [ 105, 450, 298, 463 ], "score": 1.0, "content": "In AV-NSL, score is parameterized by a 4-layer", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 461, 297, 473 ], "spans": [ { "bbox": [ 106, 461, 297, 473 ], "score": 1.0, "content": "MLP with GELU nonlinearities (Hendrycks &", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 472, 297, 484 ], "spans": [ { "bbox": [ 106, 472, 297, 484 ], "score": 1.0, "content": "Gimpel, 2016), and combine is a 5-layer MLP", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 483, 297, 496 ], "spans": [ { "bbox": [ 106, 483, 297, 496 ], "score": 1.0, "content": "with GELUs. On the other hand, such param-", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 495, 297, 506 ], "spans": [ { "bbox": [ 106, 495, 297, 506 ], "score": 1.0, "content": "eterization may cause the text-based sampling", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 506, 297, 517 ], "spans": [ { "bbox": [ 106, 506, 297, 517 ], "score": 1.0, "content": "procedure to favor sampling the visually-salient", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 516, 297, 529 ], "spans": [ { "bbox": [ 106, 516, 297, 529 ], "score": 1.0, "content": "words (Shi et al., 2019; Kojima et al., 2020).", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 527, 283, 539 ], "spans": [ { "bbox": [ 106, 527, 283, 539 ], "score": 1.0, "content": "We describe (1) and (2) in detail as follows.", "type": "text" } ], "index": 38 } ], "index": 29 }, { "type": "image", "bbox": [ 305, 327, 504, 402 ], "blocks": [ { "type": "image_body", "bbox": [ 305, 327, 504, 402 ], "group_id": 0, "lines": [ { "bbox": [ 305, 327, 504, 402 ], "spans": [ { "bbox": [ 305, 327, 504, 402 ], "score": 0.967, "type": "image", "image_path": "6556f77930feb888985877de3fee66012f8b8c7a64af3426d7862217fe5324ef.jpg" } ] } ], "index": 41.5, "virtual_lines": [ { "bbox": [ 305, 327, 504, 339.5 ], "spans": [], "index": 39 }, { "bbox": [ 305, 339.5, 504, 352.0 ], "spans": [], "index": 40 }, { "bbox": [ 305, 352.0, 504, 364.5 ], "spans": [], "index": 41 }, { "bbox": [ 305, 364.5, 504, 377.0 ], "spans": [], "index": 42 }, { "bbox": [ 305, 377.0, 504, 389.5 ], "spans": [], "index": 43 }, { "bbox": [ 305, 389.5, 504, 402.0 ], "spans": [], "index": 44 } ] }, { "type": "image_caption", "bbox": [ 304, 407, 504, 516 ], "group_id": 0, "lines": [ { "bbox": [ 304, 406, 505, 418 ], "spans": [ { "bbox": [ 304, 406, 505, 418 ], "score": 1.0, "content": "Figure 3: Example of word segmentation from", "type": "text" } ], "index": 45 }, { "bbox": [ 303, 417, 505, 429 ], "spans": [ { "bbox": [ 303, 417, 505, 429 ], "score": 1.0, "content": "VG-HuBERT (top). We use the midpoints of ad-", "type": "text" } ], "index": 46 }, { "bbox": [ 303, 429, 505, 439 ], "spans": [ { "bbox": [ 303, 429, 505, 439 ], "score": 1.0, "content": "jacent attention boundaries (vertical blue dashed", "type": "text" } ], "index": 47 }, { "bbox": [ 303, 439, 505, 451 ], "spans": [ { "bbox": [ 303, 439, 505, 451 ], "score": 1.0, "content": "lines) as the word boundaries. We observe that", "type": "text" } ], "index": 48 }, { "bbox": [ 303, 450, 505, 462 ], "spans": [ { "bbox": [ 303, 450, 505, 462 ], "score": 1.0, "content": "function words are ignored by VG-HuBERT; to", "type": "text" } ], "index": 49 }, { "bbox": [ 303, 461, 505, 473 ], "spans": [ { "bbox": [ 303, 461, 505, 473 ], "score": 1.0, "content": "account for this, we introduce segment inser-", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 472, 505, 485 ], "spans": [ { "bbox": [ 303, 472, 505, 485 ], "score": 1.0, "content": "tion (bottom): short segments are placed in long", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 483, 505, 495 ], "spans": [ { "bbox": [ 303, 483, 505, 495 ], "score": 1.0, "content": "enough gaps between existing segments, such that", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 494, 505, 506 ], "spans": [ { "bbox": [ 304, 494, 505, 506 ], "score": 1.0, "content": "function words are recovered. Inserted segments", "type": "text" } ], "index": 53 }, { "bbox": [ 303, 505, 478, 517 ], "spans": [ { "bbox": [ 303, 505, 374, 517 ], "score": 1.0, "content": "are marked with", "type": "text" }, { "bbox": [ 375, 506, 387, 516 ], "score": 0.52, "content": "\\cdot _ { + } \\cdot", "type": "inline_equation" }, { "bbox": [ 387, 505, 478, 517 ], "score": 1.0, "content": ". Best viewed in color.", "type": "text" } ], "index": 54 } ], "index": 49.5 } ], "index": 45.5 }, { "type": "text", "bbox": [ 107, 544, 505, 632 ], "lines": [ { "bbox": [ 105, 543, 506, 557 ], "spans": [ { "bbox": [ 105, 543, 506, 557 ], "score": 1.0, "content": "Audio-visual word segmentation: AV-NSL leverages VG-HuBERT Peng & Harwath (2022b) for", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 555, 505, 568 ], "spans": [ { "bbox": [ 105, 555, 505, 568 ], "score": 1.0, "content": "word segmentation (Figure 2; bottom). VG-HuBERT is trained to associate spoken captions with", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 567, 505, 578 ], "spans": [ { "bbox": [ 105, 567, 505, 578 ], "score": 1.0, "content": "natural images via retrieval training, without any textual supervision. After training, spoken word", "type": "text" } ], "index": 57 }, { "bbox": [ 105, 576, 506, 590 ], "spans": [ { "bbox": [ 105, 576, 506, 590 ], "score": 1.0, "content": "segmentation emerges via magnitude thresholding the self-attention heads of the model’s audio en-", "type": "text" } ], "index": 58 }, { "bbox": [ 105, 587, 506, 600 ], "spans": [ { "bbox": [ 105, 587, 168, 600 ], "score": 1.0, "content": "coder: at layer", "type": "text" }, { "bbox": [ 168, 588, 172, 598 ], "score": 0.42, "content": "l", "type": "inline_equation" }, { "bbox": [ 173, 587, 506, 600 ], "score": 1.0, "content": ", we threshold each CLS token attention weights over each temporal speech frame", "type": "text" } ], "index": 59 }, { "bbox": [ 105, 599, 506, 611 ], "spans": [ { "bbox": [ 105, 599, 203, 611 ], "score": 1.0, "content": "token to only show top", "type": "text" }, { "bbox": [ 203, 599, 218, 610 ], "score": 0.89, "content": "p \\%", "type": "inline_equation" }, { "bbox": [ 218, 599, 506, 611 ], "score": 1.0, "content": "of the magnitude. In Figure 3, we visualize the attention weights that", "type": "text" } ], "index": 60 }, { "bbox": [ 105, 610, 505, 622 ], "spans": [ { "bbox": [ 105, 610, 505, 622 ], "score": 1.0, "content": "each speech frame receives from the CLS token. Weights from different attention heads are plotted", "type": "text" } ], "index": 61 }, { "bbox": [ 105, 620, 475, 633 ], "spans": [ { "bbox": [ 105, 620, 475, 633 ], "score": 1.0, "content": "in different colors, and color transparency represents the magnitude of the attention weights.", "type": "text" } ], "index": 62 } ], "index": 58.5 }, { "type": "text", "bbox": [ 107, 637, 505, 704 ], "lines": [ { "bbox": [ 105, 637, 505, 650 ], "spans": [ { "bbox": [ 105, 637, 505, 650 ], "score": 1.0, "content": "However, an issue we observed with VG-HuBERT is that they tend to ignore function words such", "type": "text" } ], "index": 63 }, { "bbox": [ 105, 649, 506, 661 ], "spans": [ { "bbox": [ 105, 649, 117, 661 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 118, 649, 132, 659 ], "score": 0.3, "content": "\\mathbf { \\ddot { a } } ^ { , , }", "type": "inline_equation" }, { "bbox": [ 133, 649, 506, 661 ], "score": 1.0, "content": ", “the”, and “of”. While this is less of an issue for word segmentation and identification, it", "type": "text" } ], "index": 64 }, { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "is problematic for our purpose, as the function words are critical for phrase induction. Therefore,", "type": "text" } ], "index": 65 }, { "bbox": [ 105, 671, 505, 684 ], "spans": [ { "bbox": [ 105, 671, 505, 684 ], "score": 1.0, "content": "we devise a simple heuristic to pick up function words’ segments – segment insertion. We insert", "type": "text" } ], "index": 66 }, { "bbox": [ 104, 681, 506, 694 ], "spans": [ { "bbox": [ 104, 681, 419, 694 ], "score": 1.0, "content": "a short word segment whenever there is a sufficiently long enough gap of", "type": "text" }, { "bbox": [ 420, 684, 426, 691 ], "score": 0.62, "content": "s", "type": "inline_equation" }, { "bbox": [ 426, 681, 506, 694 ], "score": 1.0, "content": "seconds, and VG-", "type": "text" } ], "index": 67 }, { "bbox": [ 105, 692, 505, 705 ], "spans": [ { "bbox": [ 105, 692, 505, 705 ], "score": 1.0, "content": "HuBERT fails to place an attention segment. See bottom of Figure 3. Since this could introduce", "type": "text" } ], "index": 68 } ], "index": 65.5 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 712, 504, 732 ], "lines": [ { "bbox": [ 118, 709, 506, 724 ], "spans": [ { "bbox": [ 118, 709, 506, 724 ], "score": 1.0, "content": "3In the training stage, the pair is sampled from a distribution where the probability of a pair is proportional", "type": "text" } ] }, { "bbox": [ 105, 721, 330, 732 ], "spans": [ { "bbox": [ 105, 721, 116, 732 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 116, 721, 157, 732 ], "score": 0.75, "content": "\\exp ( s c o r e )", "type": "inline_equation" }, { "bbox": [ 157, 721, 330, 732 ], "score": 1.0, "content": "; in the inference stage, the arg max is selected.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 25, 308, 38 ], "spans": [ { "bbox": [ 107, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 759 ], "lines": [] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 150 ], "lines": [ { "bbox": [ 105, 81, 506, 96 ], "spans": [ { "bbox": [ 105, 81, 238, 96 ], "score": 1.0, "content": "consecutive pairs of embeddings", "type": "text" }, { "bbox": [ 239, 82, 281, 95 ], "score": 0.93, "content": "\\langle w _ { i } ^ { t } , w _ { i + 1 } ^ { t } \\rangle", "type": "inline_equation" }, { "bbox": [ 281, 81, 506, 96 ], "score": 1.0, "content": "and assigns a scalar score to each with score, (2) selects", "type": "text" } ], "index": 0 }, { "bbox": [ 104, 93, 506, 108 ], "spans": [ { "bbox": [ 104, 93, 131, 108 ], "score": 1.0, "content": "a pair", "type": "text" }, { "bbox": [ 132, 94, 179, 107 ], "score": 0.92, "content": "\\langle w _ { i ^ { \\prime } } ^ { t } , w _ { i ^ { \\prime } + 1 } ^ { t } \\rangle", "type": "inline_equation" }, { "bbox": [ 179, 93, 506, 108 ], "score": 1.0, "content": "based on the corresponding scores,3 and (3) combines the selected pair of embed-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 505, 119 ], "spans": [ { "bbox": [ 105, 105, 505, 119 ], "score": 1.0, "content": "dings via combine to form a new phrase embedding for the next step, copying the remaining ones to", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 116, 506, 129 ], "spans": [ { "bbox": [ 106, 116, 506, 129 ], "score": 1.0, "content": "the next step. In VG-NSL, score is parameterized by a 2-layer ReLU-activated MLP, and combine is", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 127, 505, 141 ], "spans": [ { "bbox": [ 105, 127, 505, 141 ], "score": 1.0, "content": "defined by the L2-normalized sum of the input embeddings. The resulting tree is inherently binary", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 138, 505, 151 ], "spans": [ { "bbox": [ 105, 138, 160, 151 ], "score": 1.0, "content": "and there are", "type": "text" }, { "bbox": [ 160, 138, 182, 149 ], "score": 0.85, "content": "N - 1", "type": "inline_equation" }, { "bbox": [ 182, 138, 505, 151 ], "score": 1.0, "content": "combining steps in total, as the tree parser must combine two nodes in each step.", "type": "text" } ], "index": 5 } ], "index": 2.5, "bbox_fs": [ 104, 81, 506, 151 ] }, { "type": "text", "bbox": [ 107, 155, 505, 210 ], "lines": [ { "bbox": [ 106, 155, 505, 167 ], "spans": [ { "bbox": [ 106, 155, 505, 167 ], "score": 1.0, "content": "The text-image embedding matching module of VG-NSL is based on the standard hinge-based triplet", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 167, 504, 178 ], "spans": [ { "bbox": [ 105, 167, 504, 178 ], "score": 1.0, "content": "loss (Kiros et al., 2014), where the sentence-based loss is modified to a phrase-based one. Addition-", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 177, 505, 189 ], "spans": [ { "bbox": [ 106, 177, 505, 189 ], "score": 1.0, "content": "ally, the loss function is adapted to estimate the visual concreteness of a text span: intuitively, the", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 187, 505, 201 ], "spans": [ { "bbox": [ 105, 187, 305, 201 ], "score": 1.0, "content": "smaller the loss related to a candidate constituent", "type": "text" }, { "bbox": [ 305, 190, 311, 198 ], "score": 0.74, "content": "c", "type": "inline_equation" }, { "bbox": [ 311, 187, 434, 201 ], "score": 1.0, "content": ", the larger the concreteness of", "type": "text" }, { "bbox": [ 434, 190, 439, 198 ], "score": 0.73, "content": "c", "type": "inline_equation" }, { "bbox": [ 439, 187, 505, 201 ], "score": 1.0, "content": ", and vice versa.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 199, 301, 211 ], "spans": [ { "bbox": [ 105, 199, 241, 211 ], "score": 1.0, "content": "The concreteness of a constituent", "type": "text" }, { "bbox": [ 242, 201, 248, 209 ], "score": 0.77, "content": "c", "type": "inline_equation" }, { "bbox": [ 248, 199, 301, 211 ], "score": 1.0, "content": "is defined as", "type": "text" } ], "index": 10 } ], "index": 8, "bbox_fs": [ 105, 155, 505, 211 ] }, { "type": "interline_equation", "bbox": [ 159, 211, 482, 238 ], "lines": [ { "bbox": [ 159, 211, 482, 238 ], "spans": [ { "bbox": [ 159, 211, 482, 238 ], "score": 0.91, "content": "\\mathbf { \\nabla } \\cdot e \\left( \\mathbf { c } ; \\mathbf { i } \\right) = \\sum _ { \\mathbf { c } ^ { \\prime } } \\left[ \\cos \\left( \\mathbf { i } , \\mathbf { c } \\right) - \\cos \\left( \\mathbf { i } , \\mathbf { c } ^ { \\prime } \\right) - \\delta \\right] _ { + } + \\sum _ { \\mathbf { i } ^ { \\prime } } \\left[ \\cos \\left( \\mathbf { i } ^ { \\prime } , \\mathbf { c } \\right) - \\cos \\left( \\mathbf { i } ^ { \\prime } , \\mathbf { c } \\right) - \\delta \\right] _ { + } ,", "type": "interline_equation", "image_path": "6815656dfcb654c315392bbb2c36e1bb9d70834ef62cd242118ff9413cb90c91.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 159, 211, 482, 220.0 ], "spans": [], "index": 11 }, { "bbox": [ 159, 220.0, 482, 229.0 ], "spans": [], "index": 12 }, { "bbox": [ 159, 229.0, 482, 238.0 ], "spans": [], "index": 13 } ] }, { "type": "text", "bbox": [ 106, 240, 505, 296 ], "lines": [ { "bbox": [ 105, 240, 504, 253 ], "spans": [ { "bbox": [ 105, 240, 262, 253 ], "score": 1.0, "content": "where c is the vector representation of", "type": "text" }, { "bbox": [ 263, 243, 268, 251 ], "score": 0.58, "content": "c", "type": "inline_equation" }, { "bbox": [ 269, 240, 484, 253 ], "score": 1.0, "content": "; i is the corresponding vector of the parallel image of", "type": "text" }, { "bbox": [ 485, 241, 504, 252 ], "score": 0.45, "content": "c ; \\mathbf { c } ^ { \\prime }", "type": "inline_equation" } ], "index": 14 }, { "bbox": [ 105, 250, 506, 264 ], "spans": [ { "bbox": [ 105, 250, 393, 264 ], "score": 1.0, "content": "is a candidate constituent from a sentence that is not in parallel with i;", "type": "text" }, { "bbox": [ 393, 252, 401, 262 ], "score": 0.77, "content": "\\mathbf { i } ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 401, 250, 506, 264 ], "score": 1.0, "content": "is an image that is not in", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 506, 276 ], "spans": [ { "bbox": [ 105, 263, 159, 276 ], "score": 1.0, "content": "parallel with", "type": "text" }, { "bbox": [ 159, 263, 176, 273 ], "score": 0.27, "content": "c ; \\delta", "type": "inline_equation" }, { "bbox": [ 176, 263, 287, 276 ], "score": 1.0, "content": "is a constant margin. Here,", "type": "text" }, { "bbox": [ 288, 263, 358, 275 ], "score": 0.92, "content": "[ \\cdot ] _ { + } : = \\operatorname* { m a x } ( \\cdot , 0 )", "type": "inline_equation" }, { "bbox": [ 359, 263, 506, 276 ], "score": 1.0, "content": ". Finally, the estimated concreteness", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 285 ], "spans": [ { "bbox": [ 105, 274, 505, 285 ], "score": 1.0, "content": "scores are passed back to the parser as rewards to the constituents. VG-NSL jointly optimizes the", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 285, 471, 297 ], "spans": [ { "bbox": [ 106, 285, 471, 297 ], "score": 1.0, "content": "visual-semantic embedding loss, and trains the parser with REINFORCE (Williams, 1992).", "type": "text" } ], "index": 18 } ], "index": 16, "bbox_fs": [ 105, 240, 506, 297 ] }, { "type": "title", "bbox": [ 107, 309, 319, 320 ], "lines": [ { "bbox": [ 106, 309, 320, 322 ], "spans": [ { "bbox": [ 106, 309, 320, 322 ], "score": 1.0, "content": "3.2 AUDIO-VISUAL NEURAL SYNTAX LEARNER", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 329, 296, 538 ], "lines": [ { "bbox": [ 106, 329, 297, 342 ], "spans": [ { "bbox": [ 106, 329, 297, 342 ], "score": 1.0, "content": "AV-NSL extends VG-NSL by: (1) incorporat-", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 341, 297, 353 ], "spans": [ { "bbox": [ 106, 341, 297, 353 ], "score": 1.0, "content": "ing an audio-visual word segmentation model", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 352, 297, 364 ], "spans": [ { "bbox": [ 106, 352, 297, 364 ], "score": 1.0, "content": "for obtaining sequences of word segments from", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 361, 298, 376 ], "spans": [ { "bbox": [ 105, 361, 298, 376 ], "score": 1.0, "content": "unannotated speech, (2) jointly optimizing", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 373, 297, 387 ], "spans": [ { "bbox": [ 105, 373, 297, 387 ], "score": 1.0, "content": "segment-level embeddings along with phrase", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 384, 297, 397 ], "spans": [ { "bbox": [ 105, 384, 297, 397 ], "score": 1.0, "content": "structure induction, and (3) employing deeper", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 396, 297, 407 ], "spans": [ { "bbox": [ 106, 396, 297, 407 ], "score": 1.0, "content": "score and combine function parameterization in", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 407, 297, 419 ], "spans": [ { "bbox": [ 105, 407, 297, 419 ], "score": 1.0, "content": "the parsing module. We empirically found (3)", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 418, 297, 430 ], "spans": [ { "bbox": [ 105, 418, 297, 430 ], "score": 1.0, "content": "necessary, mainly because speech embeddings", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 429, 297, 440 ], "spans": [ { "bbox": [ 106, 429, 297, 440 ], "score": 1.0, "content": "are inherently richer, less clean, and semanti-", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 440, 297, 451 ], "spans": [ { "bbox": [ 106, 440, 297, 451 ], "score": 1.0, "content": "cally more ambiguous than word embeddings.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 450, 298, 463 ], "spans": [ { "bbox": [ 105, 450, 298, 463 ], "score": 1.0, "content": "In AV-NSL, score is parameterized by a 4-layer", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 461, 297, 473 ], "spans": [ { "bbox": [ 106, 461, 297, 473 ], "score": 1.0, "content": "MLP with GELU nonlinearities (Hendrycks &", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 472, 297, 484 ], "spans": [ { "bbox": [ 106, 472, 297, 484 ], "score": 1.0, "content": "Gimpel, 2016), and combine is a 5-layer MLP", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 483, 297, 496 ], "spans": [ { "bbox": [ 106, 483, 297, 496 ], "score": 1.0, "content": "with GELUs. On the other hand, such param-", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 495, 297, 506 ], "spans": [ { "bbox": [ 106, 495, 297, 506 ], "score": 1.0, "content": "eterization may cause the text-based sampling", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 506, 297, 517 ], "spans": [ { "bbox": [ 106, 506, 297, 517 ], "score": 1.0, "content": "procedure to favor sampling the visually-salient", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 516, 297, 529 ], "spans": [ { "bbox": [ 106, 516, 297, 529 ], "score": 1.0, "content": "words (Shi et al., 2019; Kojima et al., 2020).", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 527, 283, 539 ], "spans": [ { "bbox": [ 106, 527, 283, 539 ], "score": 1.0, "content": "We describe (1) and (2) in detail as follows.", "type": "text" } ], "index": 38 } ], "index": 29, "bbox_fs": [ 105, 329, 298, 539 ] }, { "type": "image", "bbox": [ 305, 327, 504, 402 ], "blocks": [ { "type": "image_body", "bbox": [ 305, 327, 504, 402 ], "group_id": 0, "lines": [ { "bbox": [ 305, 327, 504, 402 ], "spans": [ { "bbox": [ 305, 327, 504, 402 ], "score": 0.967, "type": "image", "image_path": "6556f77930feb888985877de3fee66012f8b8c7a64af3426d7862217fe5324ef.jpg" } ] } ], "index": 41.5, "virtual_lines": [ { "bbox": [ 305, 327, 504, 339.5 ], "spans": [], "index": 39 }, { "bbox": [ 305, 339.5, 504, 352.0 ], "spans": [], "index": 40 }, { "bbox": [ 305, 352.0, 504, 364.5 ], "spans": [], "index": 41 }, { "bbox": [ 305, 364.5, 504, 377.0 ], "spans": [], "index": 42 }, { "bbox": [ 305, 377.0, 504, 389.5 ], "spans": [], "index": 43 }, { "bbox": [ 305, 389.5, 504, 402.0 ], "spans": [], "index": 44 } ] }, { "type": "image_caption", "bbox": [ 304, 407, 504, 516 ], "group_id": 0, "lines": [ { "bbox": [ 304, 406, 505, 418 ], "spans": [ { "bbox": [ 304, 406, 505, 418 ], "score": 1.0, "content": "Figure 3: Example of word segmentation from", "type": "text" } ], "index": 45 }, { "bbox": [ 303, 417, 505, 429 ], "spans": [ { "bbox": [ 303, 417, 505, 429 ], "score": 1.0, "content": "VG-HuBERT (top). We use the midpoints of ad-", "type": "text" } ], "index": 46 }, { "bbox": [ 303, 429, 505, 439 ], "spans": [ { "bbox": [ 303, 429, 505, 439 ], "score": 1.0, "content": "jacent attention boundaries (vertical blue dashed", "type": "text" } ], "index": 47 }, { "bbox": [ 303, 439, 505, 451 ], "spans": [ { "bbox": [ 303, 439, 505, 451 ], "score": 1.0, "content": "lines) as the word boundaries. We observe that", "type": "text" } ], "index": 48 }, { "bbox": [ 303, 450, 505, 462 ], "spans": [ { "bbox": [ 303, 450, 505, 462 ], "score": 1.0, "content": "function words are ignored by VG-HuBERT; to", "type": "text" } ], "index": 49 }, { "bbox": [ 303, 461, 505, 473 ], "spans": [ { "bbox": [ 303, 461, 505, 473 ], "score": 1.0, "content": "account for this, we introduce segment inser-", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 472, 505, 485 ], "spans": [ { "bbox": [ 303, 472, 505, 485 ], "score": 1.0, "content": "tion (bottom): short segments are placed in long", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 483, 505, 495 ], "spans": [ { "bbox": [ 303, 483, 505, 495 ], "score": 1.0, "content": "enough gaps between existing segments, such that", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 494, 505, 506 ], "spans": [ { "bbox": [ 304, 494, 505, 506 ], "score": 1.0, "content": "function words are recovered. Inserted segments", "type": "text" } ], "index": 53 }, { "bbox": [ 303, 505, 478, 517 ], "spans": [ { "bbox": [ 303, 505, 374, 517 ], "score": 1.0, "content": "are marked with", "type": "text" }, { "bbox": [ 375, 506, 387, 516 ], "score": 0.52, "content": "\\cdot _ { + } \\cdot", "type": "inline_equation" }, { "bbox": [ 387, 505, 478, 517 ], "score": 1.0, "content": ". Best viewed in color.", "type": "text" } ], "index": 54 } ], "index": 49.5 } ], "index": 45.5 }, { "type": "text", "bbox": [ 107, 544, 505, 632 ], "lines": [ { "bbox": [ 105, 543, 506, 557 ], "spans": [ { "bbox": [ 105, 543, 506, 557 ], "score": 1.0, "content": "Audio-visual word segmentation: AV-NSL leverages VG-HuBERT Peng & Harwath (2022b) for", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 555, 505, 568 ], "spans": [ { "bbox": [ 105, 555, 505, 568 ], "score": 1.0, "content": "word segmentation (Figure 2; bottom). VG-HuBERT is trained to associate spoken captions with", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 567, 505, 578 ], "spans": [ { "bbox": [ 105, 567, 505, 578 ], "score": 1.0, "content": "natural images via retrieval training, without any textual supervision. After training, spoken word", "type": "text" } ], "index": 57 }, { "bbox": [ 105, 576, 506, 590 ], "spans": [ { "bbox": [ 105, 576, 506, 590 ], "score": 1.0, "content": "segmentation emerges via magnitude thresholding the self-attention heads of the model’s audio en-", "type": "text" } ], "index": 58 }, { "bbox": [ 105, 587, 506, 600 ], "spans": [ { "bbox": [ 105, 587, 168, 600 ], "score": 1.0, "content": "coder: at layer", "type": "text" }, { "bbox": [ 168, 588, 172, 598 ], "score": 0.42, "content": "l", "type": "inline_equation" }, { "bbox": [ 173, 587, 506, 600 ], "score": 1.0, "content": ", we threshold each CLS token attention weights over each temporal speech frame", "type": "text" } ], "index": 59 }, { "bbox": [ 105, 599, 506, 611 ], "spans": [ { "bbox": [ 105, 599, 203, 611 ], "score": 1.0, "content": "token to only show top", "type": "text" }, { "bbox": [ 203, 599, 218, 610 ], "score": 0.89, "content": "p \\%", "type": "inline_equation" }, { "bbox": [ 218, 599, 506, 611 ], "score": 1.0, "content": "of the magnitude. In Figure 3, we visualize the attention weights that", "type": "text" } ], "index": 60 }, { "bbox": [ 105, 610, 505, 622 ], "spans": [ { "bbox": [ 105, 610, 505, 622 ], "score": 1.0, "content": "each speech frame receives from the CLS token. Weights from different attention heads are plotted", "type": "text" } ], "index": 61 }, { "bbox": [ 105, 620, 475, 633 ], "spans": [ { "bbox": [ 105, 620, 475, 633 ], "score": 1.0, "content": "in different colors, and color transparency represents the magnitude of the attention weights.", "type": "text" } ], "index": 62 } ], "index": 58.5, "bbox_fs": [ 105, 543, 506, 633 ] }, { "type": "text", "bbox": [ 107, 637, 505, 704 ], "lines": [ { "bbox": [ 105, 637, 505, 650 ], "spans": [ { "bbox": [ 105, 637, 505, 650 ], "score": 1.0, "content": "However, an issue we observed with VG-HuBERT is that they tend to ignore function words such", "type": "text" } ], "index": 63 }, { "bbox": [ 105, 649, 506, 661 ], "spans": [ { "bbox": [ 105, 649, 117, 661 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 118, 649, 132, 659 ], "score": 0.3, "content": "\\mathbf { \\ddot { a } } ^ { , , }", "type": "inline_equation" }, { "bbox": [ 133, 649, 506, 661 ], "score": 1.0, "content": ", “the”, and “of”. While this is less of an issue for word segmentation and identification, it", "type": "text" } ], "index": 64 }, { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "is problematic for our purpose, as the function words are critical for phrase induction. Therefore,", "type": "text" } ], "index": 65 }, { "bbox": [ 105, 671, 505, 684 ], "spans": [ { "bbox": [ 105, 671, 505, 684 ], "score": 1.0, "content": "we devise a simple heuristic to pick up function words’ segments – segment insertion. We insert", "type": "text" } ], "index": 66 }, { "bbox": [ 104, 681, 506, 694 ], "spans": [ { "bbox": [ 104, 681, 419, 694 ], "score": 1.0, "content": "a short word segment whenever there is a sufficiently long enough gap of", "type": "text" }, { "bbox": [ 420, 684, 426, 691 ], "score": 0.62, "content": "s", "type": "inline_equation" }, { "bbox": [ 426, 681, 506, 694 ], "score": 1.0, "content": "seconds, and VG-", "type": "text" } ], "index": 67 }, { "bbox": [ 105, 692, 505, 705 ], "spans": [ { "bbox": [ 105, 692, 505, 705 ], "score": 1.0, "content": "HuBERT fails to place an attention segment. See bottom of Figure 3. Since this could introduce", "type": "text" } ], "index": 68 }, { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "false positives (inserting segments where there is no word spoken), we apply unsupervised voice", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 94, 505, 106 ], "spans": [ { "bbox": [ 106, 94, 505, 106 ], "score": 1.0, "content": "activity detection (Tan et al., 2020) to further restrict segment insertion only in voiced regions. The", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 104, 505, 118 ], "spans": [ { "bbox": [ 105, 104, 214, 118 ], "score": 1.0, "content": "length of the insertion gap", "type": "text", "cross_page": true }, { "bbox": [ 214, 107, 220, 115 ], "score": 0.33, "content": "s", "type": "inline_equation", "cross_page": true }, { "bbox": [ 221, 104, 374, 118 ], "score": 1.0, "content": ", the VG-HuBERT segmentation layer", "type": "text", "cross_page": true }, { "bbox": [ 375, 105, 379, 115 ], "score": 0.43, "content": "l", "type": "inline_equation", "cross_page": true }, { "bbox": [ 380, 104, 505, 118 ], "score": 1.0, "content": ", attention magnitude threshold", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 106, 115, 505, 128 ], "spans": [ { "bbox": [ 106, 115, 121, 127 ], "score": 0.88, "content": "p \\%", "type": "inline_equation", "cross_page": true }, { "bbox": [ 121, 115, 505, 128 ], "score": 1.0, "content": ", and model training snapshots over different random seeds and training steps, are all determined", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 106, 127, 466, 138 ], "spans": [ { "bbox": [ 106, 127, 466, 138 ], "score": 1.0, "content": "in an unsupervised fashion with minimal Bayes’ risk decoding, introduced in Section 3.4.", "type": "text", "cross_page": true } ], "index": 4 } ], "index": 65.5, "bbox_fs": [ 104, 637, 506, 705 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 138 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "false positives (inserting segments where there is no word spoken), we apply unsupervised voice", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 94, 505, 106 ], "spans": [ { "bbox": [ 106, 94, 505, 106 ], "score": 1.0, "content": "activity detection (Tan et al., 2020) to further restrict segment insertion only in voiced regions. The", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 118 ], "spans": [ { "bbox": [ 105, 104, 214, 118 ], "score": 1.0, "content": "length of the insertion gap", "type": "text" }, { "bbox": [ 214, 107, 220, 115 ], "score": 0.33, "content": "s", "type": "inline_equation" }, { "bbox": [ 221, 104, 374, 118 ], "score": 1.0, "content": ", the VG-HuBERT segmentation layer", "type": "text" }, { "bbox": [ 375, 105, 379, 115 ], "score": 0.43, "content": "l", "type": "inline_equation" }, { "bbox": [ 380, 104, 505, 118 ], "score": 1.0, "content": ", attention magnitude threshold", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 115, 505, 128 ], "spans": [ { "bbox": [ 106, 115, 121, 127 ], "score": 0.88, "content": "p \\%", "type": "inline_equation" }, { "bbox": [ 121, 115, 505, 128 ], "score": 1.0, "content": ", and model training snapshots over different random seeds and training steps, are all determined", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 127, 466, 138 ], "spans": [ { "bbox": [ 106, 127, 466, 138 ], "score": 1.0, "content": "in an unsupervised fashion with minimal Bayes’ risk decoding, introduced in Section 3.4.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "text", "bbox": [ 106, 143, 505, 255 ], "lines": [ { "bbox": [ 106, 144, 505, 155 ], "spans": [ { "bbox": [ 106, 144, 505, 155 ], "score": 1.0, "content": "Speech segment representations: Given the word segments from the audio-visual segmentation", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 154, 505, 167 ], "spans": [ { "bbox": [ 105, 154, 505, 167 ], "score": 1.0, "content": "model, segment representations are extracted as inputs for the tree sampling module. Ideally, these", "type": "text" } ], "index": 6 }, { "bbox": [ 101, 162, 510, 188 ], "spans": [ { "bbox": [ 101, 162, 407, 188 ], "score": 1.0, "content": "segments should be semantically-meaningful and mimic word embeddings method is speech discretization that converts the inputs into sequences of d", "type": "text" }, { "bbox": [ 407, 164, 467, 177 ], "score": 0.93, "content": "\\mathbf { \\bar { \\mathit { W } } } = \\{ w _ { i } ^ { 0 } \\} _ { i = 1 } ^ { N }", "type": "inline_equation" }, { "bbox": [ 468, 162, 510, 188 ], "score": 1.0, "content": ". A naive(Lakhotia", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 187, 505, 200 ], "spans": [ { "bbox": [ 105, 187, 505, 200 ], "score": 1.0, "content": "et al., 2021). Yet, we are targeting word-level phrase structures, while speech discretization, namely", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 198, 505, 210 ], "spans": [ { "bbox": [ 105, 198, 505, 210 ], "score": 1.0, "content": "acoustic unit discovery, are sub-phone level, which does not fit into our setup. Different from it, AV-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 210, 504, 221 ], "spans": [ { "bbox": [ 106, 210, 504, 221 ], "score": 1.0, "content": "NSL is based on continuous segment-level self-supervised representations. Let’s denote the frame-", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 219, 507, 236 ], "spans": [ { "bbox": [ 104, 219, 235, 236 ], "score": 1.0, "content": "level representation sequence as", "type": "text" }, { "bbox": [ 235, 219, 290, 233 ], "score": 0.95, "content": "R = \\{ r _ { j } \\} _ { j = 1 } ^ { T }", "type": "inline_equation" }, { "bbox": [ 290, 219, 320, 236 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 320, 221, 329, 230 ], "score": 0.81, "content": "T", "type": "inline_equation" }, { "bbox": [ 329, 219, 507, 236 ], "score": 1.0, "content": "is the speech sequence length. Audio-visual", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 231, 507, 247 ], "spans": [ { "bbox": [ 105, 231, 271, 247 ], "score": 1.0, "content": "word segmentation returns an alignment", "type": "text" }, { "bbox": [ 272, 232, 322, 245 ], "score": 0.92, "content": "\\bar { \\boldsymbol { A } } ( i ) = \\boldsymbol { r } _ { p : q }", "type": "inline_equation" }, { "bbox": [ 322, 231, 479, 247 ], "score": 1.0, "content": "that maps the ith word segment to the", "type": "text" }, { "bbox": [ 479, 234, 485, 244 ], "score": 0.28, "content": "p", "type": "inline_equation" }, { "bbox": [ 486, 231, 507, 247 ], "score": 1.0, "content": "th to", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 243, 477, 257 ], "spans": [ { "bbox": [ 106, 245, 112, 255 ], "score": 0.3, "content": "q", "type": "inline_equation" }, { "bbox": [ 113, 243, 477, 257 ], "score": 1.0, "content": "th acoustic frames. The segment-level continuous representation for the ith word is simply,", "type": "text" } ], "index": 13 } ], "index": 9 }, { "type": "interline_equation", "bbox": [ 268, 254, 343, 282 ], "lines": [ { "bbox": [ 268, 254, 343, 282 ], "spans": [ { "bbox": [ 268, 254, 343, 282 ], "score": 0.93, "content": "w _ { i } ^ { 0 } = \\sum _ { t \\in A ( i ) } \\stackrel { } { a _ { i t } } r _ { i t }", "type": "interline_equation", "image_path": "09041c7a29476e99070cf5ddd2b6899c44f676091dc62483b1414ad4d764d04c.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 268, 254, 343, 268.0 ], "spans": [], "index": 14 }, { "bbox": [ 268, 268.0, 343, 282.0 ], "spans": [], "index": 15 } ] }, { "type": "text", "bbox": [ 107, 283, 505, 361 ], "lines": [ { "bbox": [ 105, 283, 505, 296 ], "spans": [ { "bbox": [ 105, 283, 134, 296 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 286, 147, 295 ], "score": 0.85, "content": "a _ { i t }", "type": "inline_equation" }, { "bbox": [ 148, 283, 380, 296 ], "score": 1.0, "content": "is the attention weights over the segments specified by", "type": "text" }, { "bbox": [ 380, 284, 399, 296 ], "score": 0.91, "content": "A ( i )", "type": "inline_equation" }, { "bbox": [ 400, 283, 505, 296 ], "score": 1.0, "content": ". By default in AV-NSL,", "type": "text" } ], "index": 16 }, { "bbox": [ 107, 295, 506, 307 ], "spans": [ { "bbox": [ 107, 295, 115, 305 ], "score": 0.8, "content": "R", "type": "inline_equation" }, { "bbox": [ 116, 295, 325, 307 ], "score": 1.0, "content": "is the layer representation from VG-HuBERT, and", "type": "text" }, { "bbox": [ 326, 296, 339, 306 ], "score": 0.85, "content": "a _ { i t }", "type": "inline_equation" }, { "bbox": [ 339, 295, 506, 307 ], "score": 1.0, "content": "is the CLS token attention weights over", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 306, 505, 318 ], "spans": [ { "bbox": [ 106, 306, 505, 318 ], "score": 1.0, "content": "frames within each segment. In some cases, visual grounding is not available in AV-NSL’s word", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 316, 506, 330 ], "spans": [ { "bbox": [ 105, 316, 384, 330 ], "score": 1.0, "content": "segmentation, e.g. VG-HuBERT is not available. We instead take", "type": "text" }, { "bbox": [ 384, 317, 393, 327 ], "score": 0.78, "content": "R", "type": "inline_equation" }, { "bbox": [ 393, 316, 506, 330 ], "score": 1.0, "content": "as the layer representation", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 326, 505, 341 ], "spans": [ { "bbox": [ 105, 326, 298, 341 ], "score": 1.0, "content": "from a vanilla HuBERT (Hsu et al., 2021a), and", "type": "text" }, { "bbox": [ 298, 329, 311, 339 ], "score": 0.88, "content": "a _ { i t }", "type": "inline_equation" }, { "bbox": [ 311, 326, 505, 341 ], "score": 1.0, "content": "is parameterized by a hidden layer that is jointly", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 338, 506, 352 ], "spans": [ { "bbox": [ 106, 338, 506, 352 ], "score": 1.0, "content": "optimized with the tree sampling module. Despite its simplicity, AV-NSL learns meaningful phrase", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 350, 322, 362 ], "spans": [ { "bbox": [ 106, 350, 322, 362 ], "score": 1.0, "content": "structures on these segment representation sequences.", "type": "text" } ], "index": 22 } ], "index": 19 }, { "type": "title", "bbox": [ 108, 374, 201, 385 ], "lines": [ { "bbox": [ 105, 372, 202, 387 ], "spans": [ { "bbox": [ 105, 372, 202, 387 ], "score": 1.0, "content": "3.3 SELF-TRAINING", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 393, 505, 471 ], "lines": [ { "bbox": [ 105, 394, 505, 407 ], "spans": [ { "bbox": [ 105, 394, 505, 407 ], "score": 1.0, "content": "A self-training procedure is introduced for AV-NSL to further improve its parsing capability. Previ-", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 406, 505, 418 ], "spans": [ { "bbox": [ 105, 406, 505, 418 ], "score": 1.0, "content": "ously, it has been shown that self-training consistently improves the performance of text-based un-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 416, 506, 430 ], "spans": [ { "bbox": [ 105, 416, 506, 430 ], "score": 1.0, "content": "supervised constituency parsing. In Shi et al. (2020), the self-training model was based on Benepar", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 427, 505, 441 ], "spans": [ { "bbox": [ 105, 427, 505, 441 ], "score": 1.0, "content": "(Kitaev & Klein, 2018), a supervised neural constituency parser, which (1) takes a sentence as the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 438, 505, 452 ], "spans": [ { "bbox": [ 105, 438, 505, 452 ], "score": 1.0, "content": "input, (2) maps it to word representations, and (3) predicts a score for any constituency parse tree. In", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 448, 505, 463 ], "spans": [ { "bbox": [ 105, 448, 505, 463 ], "score": 1.0, "content": "the inference stage, the model evaluates all possible tree structures and outputs the highest-scoring", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 461, 414, 473 ], "spans": [ { "bbox": [ 105, 461, 414, 473 ], "score": 1.0, "content": "one using the CKY algorithm (Kasami, 1966; Younger, 1967; Cocke, 1969).", "type": "text" } ], "index": 30 } ], "index": 27 }, { "type": "text", "bbox": [ 108, 477, 505, 543 ], "lines": [ { "bbox": [ 105, 477, 505, 490 ], "spans": [ { "bbox": [ 105, 477, 505, 490 ], "score": 1.0, "content": "In this work, we introduce s-Benepar, which is based on the original Benepar, except the model in-", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 488, 505, 500 ], "spans": [ { "bbox": [ 105, 488, 505, 500 ], "score": 1.0, "content": "put is the segment-level continuous HuBERT representations mean-pooled over unsupervised word", "type": "text" } ], "index": 32 }, { "bbox": [ 104, 498, 506, 512 ], "spans": [ { "bbox": [ 104, 498, 506, 512 ], "score": 1.0, "content": "segmentation from VG-HuBERT with segment insertion, and model output is AV-NSL’s inferred", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 509, 505, 524 ], "spans": [ { "bbox": [ 105, 509, 505, 524 ], "score": 1.0, "content": "constituency parse from Section 3.2. We also removed part-of-speech tag prediction as in Benepar,", "type": "text" } ], "index": 34 }, { "bbox": [ 104, 519, 506, 535 ], "spans": [ { "bbox": [ 104, 519, 443, 535 ], "score": 1.0, "content": "as there is no textual supervision in our setting. To summarize, with paired speech", "type": "text" }, { "bbox": [ 443, 522, 459, 532 ], "score": 0.9, "content": "D _ { A }", "type": "inline_equation" }, { "bbox": [ 459, 519, 506, 535 ], "score": 1.0, "content": "and image", "type": "text" } ], "index": 35 }, { "bbox": [ 107, 532, 403, 544 ], "spans": [ { "bbox": [ 107, 532, 123, 543 ], "score": 0.91, "content": "D _ { V }", "type": "inline_equation" }, { "bbox": [ 123, 532, 403, 544 ], "score": 1.0, "content": "data, the training scheme for AV-NSL with self-training is as follows:", "type": "text" } ], "index": 36 } ], "index": 33.5 }, { "type": "text", "bbox": [ 130, 545, 502, 603 ], "lines": [ { "bbox": [ 128, 543, 493, 560 ], "spans": [ { "bbox": [ 128, 543, 308, 560 ], "score": 1.0, "content": "1. Train an AV-NSL from audio-visual data", "type": "text" }, { "bbox": [ 308, 545, 350, 558 ], "score": 0.93, "content": "( D _ { A } , D _ { V } )", "type": "inline_equation" }, { "bbox": [ 351, 543, 468, 560 ], "score": 1.0, "content": "and obtain the trained model", "type": "text" }, { "bbox": [ 469, 546, 488, 556 ], "score": 0.89, "content": "M _ { a v }", "type": "inline_equation" }, { "bbox": [ 488, 543, 493, 560 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 37 }, { "bbox": [ 129, 556, 499, 570 ], "spans": [ { "bbox": [ 129, 556, 221, 570 ], "score": 1.0, "content": "2. Generate parse tree", "type": "text" }, { "bbox": [ 222, 557, 233, 568 ], "score": 0.88, "content": "T _ { 0 }", "type": "inline_equation" }, { "bbox": [ 234, 556, 255, 570 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 255, 557, 274, 568 ], "score": 0.9, "content": "M _ { a v }", "type": "inline_equation" }, { "bbox": [ 275, 556, 290, 570 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 290, 558, 306, 568 ], "score": 0.89, "content": "D _ { A }", "type": "inline_equation" }, { "bbox": [ 306, 556, 405, 570 ], "score": 1.0, "content": ". Obtain audio-tree pairs", "type": "text" }, { "bbox": [ 406, 557, 443, 569 ], "score": 0.92, "content": "( D _ { A } , T _ { 0 } )", "type": "inline_equation" }, { "bbox": [ 443, 556, 462, 570 ], "score": 1.0, "content": ". Set", "type": "text" }, { "bbox": [ 463, 557, 495, 568 ], "score": 0.91, "content": "T = T _ { 0 }", "type": "inline_equation" }, { "bbox": [ 495, 556, 499, 570 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 38 }, { "bbox": [ 128, 567, 415, 581 ], "spans": [ { "bbox": [ 128, 567, 243, 581 ], "score": 1.0, "content": "3. Train an s-Benepar from", "type": "text" }, { "bbox": [ 243, 569, 278, 580 ], "score": 0.91, "content": "( D _ { A } , T )", "type": "inline_equation" }, { "bbox": [ 278, 567, 396, 581 ], "score": 1.0, "content": "and obtain the trained model", "type": "text" }, { "bbox": [ 396, 569, 411, 580 ], "score": 0.89, "content": "M _ { s } ^ { i }", "type": "inline_equation" }, { "bbox": [ 411, 567, 415, 581 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 39 }, { "bbox": [ 130, 579, 491, 592 ], "spans": [ { "bbox": [ 130, 579, 221, 592 ], "score": 1.0, "content": "4. Generate parse tree", "type": "text" }, { "bbox": [ 222, 580, 232, 591 ], "score": 0.86, "content": "T _ { i }", "type": "inline_equation" }, { "bbox": [ 232, 579, 253, 592 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 254, 580, 269, 591 ], "score": 0.9, "content": "M _ { s } ^ { i }", "type": "inline_equation" }, { "bbox": [ 269, 579, 284, 592 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 284, 580, 300, 591 ], "score": 0.89, "content": "D _ { A }", "type": "inline_equation" }, { "bbox": [ 300, 579, 399, 592 ], "score": 1.0, "content": ". Obtain audio-tree pairs", "type": "text" }, { "bbox": [ 399, 580, 436, 591 ], "score": 0.92, "content": "( D _ { A } , T _ { i } )", "type": "inline_equation" }, { "bbox": [ 436, 579, 455, 592 ], "score": 1.0, "content": ". Set", "type": "text" }, { "bbox": [ 456, 580, 486, 591 ], "score": 0.91, "content": "T = T _ { i }", "type": "inline_equation" }, { "bbox": [ 487, 579, 491, 592 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 40 }, { "bbox": [ 128, 590, 505, 603 ], "spans": [ { "bbox": [ 128, 590, 451, 603 ], "score": 1.0, "content": "5. Go to Step 3 if we have not reached the desirable number of iterations; return", "type": "text" }, { "bbox": [ 452, 591, 460, 601 ], "score": 0.83, "content": "T", "type": "inline_equation" }, { "bbox": [ 461, 590, 505, 603 ], "score": 1.0, "content": "otherwise.", "type": "text" } ], "index": 41 } ], "index": 39 }, { "type": "text", "bbox": [ 109, 615, 486, 627 ], "lines": [ { "bbox": [ 106, 614, 487, 628 ], "spans": [ { "bbox": [ 106, 614, 319, 628 ], "score": 1.0, "content": "We find it helpful to iterate s-Benepar training twice", "type": "text" }, { "bbox": [ 319, 616, 344, 626 ], "score": 0.85, "content": "( i = 2", "type": "inline_equation" }, { "bbox": [ 344, 614, 487, 628 ], "score": 1.0, "content": "), but the results plateau afterwards.", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "title", "bbox": [ 108, 639, 248, 651 ], "lines": [ { "bbox": [ 106, 639, 249, 652 ], "spans": [ { "bbox": [ 106, 639, 249, 652 ], "score": 1.0, "content": "3.4 UNSUPERVISED DECODING", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "text", "bbox": [ 107, 660, 504, 704 ], "lines": [ { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "One key ingredient of AV-NSL is applying minimum Bayes risk (MBR) decoding (Bickel &", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 671, 505, 684 ], "spans": [ { "bbox": [ 105, 671, 505, 684 ], "score": 1.0, "content": "Li, 1977) as the selection criterion for fully-unsupervised spoken word segmentation and phrase-", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 682, 506, 695 ], "spans": [ { "bbox": [ 105, 682, 506, 695 ], "score": 1.0, "content": "structure induction.4 Specifically, this is in contrast to all prior unsupervised word segmentation", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 693, 485, 707 ], "spans": [ { "bbox": [ 105, 693, 485, 707 ], "score": 1.0, "content": "work, in which ground truth word segments from a development set are required for decoding.", "type": "text" } ], "index": 47 } ], "index": 45.5 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 711, 506, 732 ], "lines": [ { "bbox": [ 118, 709, 506, 724 ], "spans": [ { "bbox": [ 118, 709, 506, 724 ], "score": 1.0, "content": "4MBR decoding is widely adopted in machine translation (Kumar & Byrne, 2004; Zhang & Gildea, 2008;", "type": "text" } ] }, { "bbox": [ 106, 720, 206, 732 ], "spans": [ { "bbox": [ 106, 720, 206, 732 ], "score": 1.0, "content": "Shi et al., 2022, inter alia).", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 25, 308, 38 ], "spans": [ { "bbox": [ 107, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "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": "text", "bbox": [ 107, 82, 505, 138 ], "lines": [], "index": 2, "bbox_fs": [ 105, 82, 505, 138 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 143, 505, 255 ], "lines": [ { "bbox": [ 106, 144, 505, 155 ], "spans": [ { "bbox": [ 106, 144, 505, 155 ], "score": 1.0, "content": "Speech segment representations: Given the word segments from the audio-visual segmentation", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 154, 505, 167 ], "spans": [ { "bbox": [ 105, 154, 505, 167 ], "score": 1.0, "content": "model, segment representations are extracted as inputs for the tree sampling module. Ideally, these", "type": "text" } ], "index": 6 }, { "bbox": [ 101, 162, 510, 188 ], "spans": [ { "bbox": [ 101, 162, 407, 188 ], "score": 1.0, "content": "segments should be semantically-meaningful and mimic word embeddings method is speech discretization that converts the inputs into sequences of d", "type": "text" }, { "bbox": [ 407, 164, 467, 177 ], "score": 0.93, "content": "\\mathbf { \\bar { \\mathit { W } } } = \\{ w _ { i } ^ { 0 } \\} _ { i = 1 } ^ { N }", "type": "inline_equation" }, { "bbox": [ 468, 162, 510, 188 ], "score": 1.0, "content": ". A naive(Lakhotia", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 187, 505, 200 ], "spans": [ { "bbox": [ 105, 187, 505, 200 ], "score": 1.0, "content": "et al., 2021). Yet, we are targeting word-level phrase structures, while speech discretization, namely", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 198, 505, 210 ], "spans": [ { "bbox": [ 105, 198, 505, 210 ], "score": 1.0, "content": "acoustic unit discovery, are sub-phone level, which does not fit into our setup. Different from it, AV-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 210, 504, 221 ], "spans": [ { "bbox": [ 106, 210, 504, 221 ], "score": 1.0, "content": "NSL is based on continuous segment-level self-supervised representations. Let’s denote the frame-", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 219, 507, 236 ], "spans": [ { "bbox": [ 104, 219, 235, 236 ], "score": 1.0, "content": "level representation sequence as", "type": "text" }, { "bbox": [ 235, 219, 290, 233 ], "score": 0.95, "content": "R = \\{ r _ { j } \\} _ { j = 1 } ^ { T }", "type": "inline_equation" }, { "bbox": [ 290, 219, 320, 236 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 320, 221, 329, 230 ], "score": 0.81, "content": "T", "type": "inline_equation" }, { "bbox": [ 329, 219, 507, 236 ], "score": 1.0, "content": "is the speech sequence length. Audio-visual", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 231, 507, 247 ], "spans": [ { "bbox": [ 105, 231, 271, 247 ], "score": 1.0, "content": "word segmentation returns an alignment", "type": "text" }, { "bbox": [ 272, 232, 322, 245 ], "score": 0.92, "content": "\\bar { \\boldsymbol { A } } ( i ) = \\boldsymbol { r } _ { p : q }", "type": "inline_equation" }, { "bbox": [ 322, 231, 479, 247 ], "score": 1.0, "content": "that maps the ith word segment to the", "type": "text" }, { "bbox": [ 479, 234, 485, 244 ], "score": 0.28, "content": "p", "type": "inline_equation" }, { "bbox": [ 486, 231, 507, 247 ], "score": 1.0, "content": "th to", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 243, 477, 257 ], "spans": [ { "bbox": [ 106, 245, 112, 255 ], "score": 0.3, "content": "q", "type": "inline_equation" }, { "bbox": [ 113, 243, 477, 257 ], "score": 1.0, "content": "th acoustic frames. The segment-level continuous representation for the ith word is simply,", "type": "text" } ], "index": 13 } ], "index": 9, "bbox_fs": [ 101, 144, 510, 257 ] }, { "type": "interline_equation", "bbox": [ 268, 254, 343, 282 ], "lines": [ { "bbox": [ 268, 254, 343, 282 ], "spans": [ { "bbox": [ 268, 254, 343, 282 ], "score": 0.93, "content": "w _ { i } ^ { 0 } = \\sum _ { t \\in A ( i ) } \\stackrel { } { a _ { i t } } r _ { i t }", "type": "interline_equation", "image_path": "09041c7a29476e99070cf5ddd2b6899c44f676091dc62483b1414ad4d764d04c.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 268, 254, 343, 268.0 ], "spans": [], "index": 14 }, { "bbox": [ 268, 268.0, 343, 282.0 ], "spans": [], "index": 15 } ] }, { "type": "text", "bbox": [ 107, 283, 505, 361 ], "lines": [ { "bbox": [ 105, 283, 505, 296 ], "spans": [ { "bbox": [ 105, 283, 134, 296 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 286, 147, 295 ], "score": 0.85, "content": "a _ { i t }", "type": "inline_equation" }, { "bbox": [ 148, 283, 380, 296 ], "score": 1.0, "content": "is the attention weights over the segments specified by", "type": "text" }, { "bbox": [ 380, 284, 399, 296 ], "score": 0.91, "content": "A ( i )", "type": "inline_equation" }, { "bbox": [ 400, 283, 505, 296 ], "score": 1.0, "content": ". By default in AV-NSL,", "type": "text" } ], "index": 16 }, { "bbox": [ 107, 295, 506, 307 ], "spans": [ { "bbox": [ 107, 295, 115, 305 ], "score": 0.8, "content": "R", "type": "inline_equation" }, { "bbox": [ 116, 295, 325, 307 ], "score": 1.0, "content": "is the layer representation from VG-HuBERT, and", "type": "text" }, { "bbox": [ 326, 296, 339, 306 ], "score": 0.85, "content": "a _ { i t }", "type": "inline_equation" }, { "bbox": [ 339, 295, 506, 307 ], "score": 1.0, "content": "is the CLS token attention weights over", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 306, 505, 318 ], "spans": [ { "bbox": [ 106, 306, 505, 318 ], "score": 1.0, "content": "frames within each segment. In some cases, visual grounding is not available in AV-NSL’s word", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 316, 506, 330 ], "spans": [ { "bbox": [ 105, 316, 384, 330 ], "score": 1.0, "content": "segmentation, e.g. VG-HuBERT is not available. We instead take", "type": "text" }, { "bbox": [ 384, 317, 393, 327 ], "score": 0.78, "content": "R", "type": "inline_equation" }, { "bbox": [ 393, 316, 506, 330 ], "score": 1.0, "content": "as the layer representation", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 326, 505, 341 ], "spans": [ { "bbox": [ 105, 326, 298, 341 ], "score": 1.0, "content": "from a vanilla HuBERT (Hsu et al., 2021a), and", "type": "text" }, { "bbox": [ 298, 329, 311, 339 ], "score": 0.88, "content": "a _ { i t }", "type": "inline_equation" }, { "bbox": [ 311, 326, 505, 341 ], "score": 1.0, "content": "is parameterized by a hidden layer that is jointly", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 338, 506, 352 ], "spans": [ { "bbox": [ 106, 338, 506, 352 ], "score": 1.0, "content": "optimized with the tree sampling module. Despite its simplicity, AV-NSL learns meaningful phrase", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 350, 322, 362 ], "spans": [ { "bbox": [ 106, 350, 322, 362 ], "score": 1.0, "content": "structures on these segment representation sequences.", "type": "text" } ], "index": 22 } ], "index": 19, "bbox_fs": [ 105, 283, 506, 362 ] }, { "type": "title", "bbox": [ 108, 374, 201, 385 ], "lines": [ { "bbox": [ 105, 372, 202, 387 ], "spans": [ { "bbox": [ 105, 372, 202, 387 ], "score": 1.0, "content": "3.3 SELF-TRAINING", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 393, 505, 471 ], "lines": [ { "bbox": [ 105, 394, 505, 407 ], "spans": [ { "bbox": [ 105, 394, 505, 407 ], "score": 1.0, "content": "A self-training procedure is introduced for AV-NSL to further improve its parsing capability. Previ-", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 406, 505, 418 ], "spans": [ { "bbox": [ 105, 406, 505, 418 ], "score": 1.0, "content": "ously, it has been shown that self-training consistently improves the performance of text-based un-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 416, 506, 430 ], "spans": [ { "bbox": [ 105, 416, 506, 430 ], "score": 1.0, "content": "supervised constituency parsing. In Shi et al. (2020), the self-training model was based on Benepar", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 427, 505, 441 ], "spans": [ { "bbox": [ 105, 427, 505, 441 ], "score": 1.0, "content": "(Kitaev & Klein, 2018), a supervised neural constituency parser, which (1) takes a sentence as the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 438, 505, 452 ], "spans": [ { "bbox": [ 105, 438, 505, 452 ], "score": 1.0, "content": "input, (2) maps it to word representations, and (3) predicts a score for any constituency parse tree. In", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 448, 505, 463 ], "spans": [ { "bbox": [ 105, 448, 505, 463 ], "score": 1.0, "content": "the inference stage, the model evaluates all possible tree structures and outputs the highest-scoring", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 461, 414, 473 ], "spans": [ { "bbox": [ 105, 461, 414, 473 ], "score": 1.0, "content": "one using the CKY algorithm (Kasami, 1966; Younger, 1967; Cocke, 1969).", "type": "text" } ], "index": 30 } ], "index": 27, "bbox_fs": [ 105, 394, 506, 473 ] }, { "type": "text", "bbox": [ 108, 477, 505, 543 ], "lines": [ { "bbox": [ 105, 477, 505, 490 ], "spans": [ { "bbox": [ 105, 477, 505, 490 ], "score": 1.0, "content": "In this work, we introduce s-Benepar, which is based on the original Benepar, except the model in-", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 488, 505, 500 ], "spans": [ { "bbox": [ 105, 488, 505, 500 ], "score": 1.0, "content": "put is the segment-level continuous HuBERT representations mean-pooled over unsupervised word", "type": "text" } ], "index": 32 }, { "bbox": [ 104, 498, 506, 512 ], "spans": [ { "bbox": [ 104, 498, 506, 512 ], "score": 1.0, "content": "segmentation from VG-HuBERT with segment insertion, and model output is AV-NSL’s inferred", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 509, 505, 524 ], "spans": [ { "bbox": [ 105, 509, 505, 524 ], "score": 1.0, "content": "constituency parse from Section 3.2. We also removed part-of-speech tag prediction as in Benepar,", "type": "text" } ], "index": 34 }, { "bbox": [ 104, 519, 506, 535 ], "spans": [ { "bbox": [ 104, 519, 443, 535 ], "score": 1.0, "content": "as there is no textual supervision in our setting. To summarize, with paired speech", "type": "text" }, { "bbox": [ 443, 522, 459, 532 ], "score": 0.9, "content": "D _ { A }", "type": "inline_equation" }, { "bbox": [ 459, 519, 506, 535 ], "score": 1.0, "content": "and image", "type": "text" } ], "index": 35 }, { "bbox": [ 107, 532, 403, 544 ], "spans": [ { "bbox": [ 107, 532, 123, 543 ], "score": 0.91, "content": "D _ { V }", "type": "inline_equation" }, { "bbox": [ 123, 532, 403, 544 ], "score": 1.0, "content": "data, the training scheme for AV-NSL with self-training is as follows:", "type": "text" } ], "index": 36 } ], "index": 33.5, "bbox_fs": [ 104, 477, 506, 544 ] }, { "type": "index", "bbox": [ 130, 545, 502, 603 ], "lines": [ { "bbox": [ 128, 543, 493, 560 ], "spans": [ { "bbox": [ 128, 543, 308, 560 ], "score": 1.0, "content": "1. Train an AV-NSL from audio-visual data", "type": "text" }, { "bbox": [ 308, 545, 350, 558 ], "score": 0.93, "content": "( D _ { A } , D _ { V } )", "type": "inline_equation" }, { "bbox": [ 351, 543, 468, 560 ], "score": 1.0, "content": "and obtain the trained model", "type": "text" }, { "bbox": [ 469, 546, 488, 556 ], "score": 0.89, "content": "M _ { a v }", "type": "inline_equation" }, { "bbox": [ 488, 543, 493, 560 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 37, "is_list_start_line": true }, { "bbox": [ 129, 556, 499, 570 ], "spans": [ { "bbox": [ 129, 556, 221, 570 ], "score": 1.0, "content": "2. Generate parse tree", "type": "text" }, { "bbox": [ 222, 557, 233, 568 ], "score": 0.88, "content": "T _ { 0 }", "type": "inline_equation" }, { "bbox": [ 234, 556, 255, 570 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 255, 557, 274, 568 ], "score": 0.9, "content": "M _ { a v }", "type": "inline_equation" }, { "bbox": [ 275, 556, 290, 570 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 290, 558, 306, 568 ], "score": 0.89, "content": "D _ { A }", "type": "inline_equation" }, { "bbox": [ 306, 556, 405, 570 ], "score": 1.0, "content": ". Obtain audio-tree pairs", "type": "text" }, { "bbox": [ 406, 557, 443, 569 ], "score": 0.92, "content": "( D _ { A } , T _ { 0 } )", "type": "inline_equation" }, { "bbox": [ 443, 556, 462, 570 ], "score": 1.0, "content": ". Set", "type": "text" }, { "bbox": [ 463, 557, 495, 568 ], "score": 0.91, "content": "T = T _ { 0 }", "type": "inline_equation" }, { "bbox": [ 495, 556, 499, 570 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 128, 567, 415, 581 ], "spans": [ { "bbox": [ 128, 567, 243, 581 ], "score": 1.0, "content": "3. Train an s-Benepar from", "type": "text" }, { "bbox": [ 243, 569, 278, 580 ], "score": 0.91, "content": "( D _ { A } , T )", "type": "inline_equation" }, { "bbox": [ 278, 567, 396, 581 ], "score": 1.0, "content": "and obtain the trained model", "type": "text" }, { "bbox": [ 396, 569, 411, 580 ], "score": 0.89, "content": "M _ { s } ^ { i }", "type": "inline_equation" }, { "bbox": [ 411, 567, 415, 581 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 39, "is_list_start_line": true }, { "bbox": [ 130, 579, 491, 592 ], "spans": [ { "bbox": [ 130, 579, 221, 592 ], "score": 1.0, "content": "4. Generate parse tree", "type": "text" }, { "bbox": [ 222, 580, 232, 591 ], "score": 0.86, "content": "T _ { i }", "type": "inline_equation" }, { "bbox": [ 232, 579, 253, 592 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 254, 580, 269, 591 ], "score": 0.9, "content": "M _ { s } ^ { i }", "type": "inline_equation" }, { "bbox": [ 269, 579, 284, 592 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 284, 580, 300, 591 ], "score": 0.89, "content": "D _ { A }", "type": "inline_equation" }, { "bbox": [ 300, 579, 399, 592 ], "score": 1.0, "content": ". Obtain audio-tree pairs", "type": "text" }, { "bbox": [ 399, 580, 436, 591 ], "score": 0.92, "content": "( D _ { A } , T _ { i } )", "type": "inline_equation" }, { "bbox": [ 436, 579, 455, 592 ], "score": 1.0, "content": ". Set", "type": "text" }, { "bbox": [ 456, 580, 486, 591 ], "score": 0.91, "content": "T = T _ { i }", "type": "inline_equation" }, { "bbox": [ 487, 579, 491, 592 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 128, 590, 505, 603 ], "spans": [ { "bbox": [ 128, 590, 451, 603 ], "score": 1.0, "content": "5. Go to Step 3 if we have not reached the desirable number of iterations; return", "type": "text" }, { "bbox": [ 452, 591, 460, 601 ], "score": 0.83, "content": "T", "type": "inline_equation" }, { "bbox": [ 461, 590, 505, 603 ], "score": 1.0, "content": "otherwise.", "type": "text" } ], "index": 41, "is_list_start_line": true } ], "index": 39, "bbox_fs": [ 128, 543, 505, 603 ] }, { "type": "text", "bbox": [ 109, 615, 486, 627 ], "lines": [ { "bbox": [ 106, 614, 487, 628 ], "spans": [ { "bbox": [ 106, 614, 319, 628 ], "score": 1.0, "content": "We find it helpful to iterate s-Benepar training twice", "type": "text" }, { "bbox": [ 319, 616, 344, 626 ], "score": 0.85, "content": "( i = 2", "type": "inline_equation" }, { "bbox": [ 344, 614, 487, 628 ], "score": 1.0, "content": "), but the results plateau afterwards.", "type": "text" } ], "index": 42 } ], "index": 42, "bbox_fs": [ 106, 614, 487, 628 ] }, { "type": "title", "bbox": [ 108, 639, 248, 651 ], "lines": [ { "bbox": [ 106, 639, 249, 652 ], "spans": [ { "bbox": [ 106, 639, 249, 652 ], "score": 1.0, "content": "3.4 UNSUPERVISED DECODING", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "text", "bbox": [ 107, 660, 504, 704 ], "lines": [ { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "One key ingredient of AV-NSL is applying minimum Bayes risk (MBR) decoding (Bickel &", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 671, 505, 684 ], "spans": [ { "bbox": [ 105, 671, 505, 684 ], "score": 1.0, "content": "Li, 1977) as the selection criterion for fully-unsupervised spoken word segmentation and phrase-", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 682, 506, 695 ], "spans": [ { "bbox": [ 105, 682, 506, 695 ], "score": 1.0, "content": "structure induction.4 Specifically, this is in contrast to all prior unsupervised word segmentation", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 693, 485, 707 ], "spans": [ { "bbox": [ 105, 693, 485, 707 ], "score": 1.0, "content": "work, in which ground truth word segments from a development set are required for decoding.", "type": "text" } ], "index": 47 } ], "index": 45.5, "bbox_fs": [ 105, 659, 506, 707 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 82, 504, 105 ], "lines": [ { "bbox": [ 105, 81, 504, 96 ], "spans": [ { "bbox": [ 105, 81, 256, 96 ], "score": 1.0, "content": "At a high level, given a loss function", "type": "text" }, { "bbox": [ 256, 82, 312, 95 ], "score": 0.93, "content": "\\ell _ { M B R } ( O _ { 1 } , O _ { 2 } )", "type": "inline_equation" }, { "bbox": [ 313, 81, 399, 96 ], "score": 1.0, "content": "between two outputs", "type": "text" }, { "bbox": [ 399, 83, 412, 93 ], "score": 0.89, "content": "O _ { 1 }", "type": "inline_equation" }, { "bbox": [ 413, 81, 431, 96 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 431, 83, 444, 93 ], "score": 0.89, "content": "O _ { 2 }", "type": "inline_equation" }, { "bbox": [ 444, 81, 497, 96 ], "score": 1.0, "content": ", and a set of", "type": "text" }, { "bbox": [ 497, 83, 504, 92 ], "score": 0.79, "content": "k", "type": "inline_equation" } ], "index": 0 }, { "bbox": [ 106, 93, 336, 106 ], "spans": [ { "bbox": [ 106, 93, 138, 106 ], "score": 1.0, "content": "outputs", "type": "text" }, { "bbox": [ 139, 93, 218, 106 ], "score": 0.94, "content": "\\mathcal { O } = \\{ O _ { 1 } , \\ldots , O _ { k } \\}", "type": "inline_equation" }, { "bbox": [ 218, 93, 336, 106 ], "score": 1.0, "content": ", we select the optimal output", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "interline_equation", "bbox": [ 233, 106, 378, 133 ], "lines": [ { "bbox": [ 233, 106, 378, 133 ], "spans": [ { "bbox": [ 233, 106, 378, 133 ], "score": 0.94, "content": "\\hat { O } = \\arg \\operatorname* { m i n } _ { O ^ { \\prime } \\in { \\mathcal O } } \\sum _ { O ^ { \\prime \\prime } \\in { \\mathcal O } } \\ell _ { M B R } ( O ^ { \\prime } , O ^ { \\prime \\prime } ) .", "type": "interline_equation", "image_path": "05f4cb5da860d44c1c6f81a2e82c85670940b072bcd0f89b4e5cb02bef871169.jpg" } ] } ], "index": 2.5, "virtual_lines": [ { "bbox": [ 233, 106, 378, 119.5 ], "spans": [], "index": 2 }, { "bbox": [ 233, 119.5, 378, 133.0 ], "spans": [], "index": 3 } ] }, { "type": "text", "bbox": [ 106, 140, 505, 195 ], "lines": [ { "bbox": [ 105, 139, 505, 153 ], "spans": [ { "bbox": [ 105, 139, 444, 153 ], "score": 1.0, "content": "For word segmentation, we define the loss between two segmentation proposals", "type": "text" }, { "bbox": [ 444, 141, 456, 151 ], "score": 0.87, "content": " { \\boldsymbol { S } } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 457, 139, 478, 153 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 478, 140, 489, 151 ], "score": 0.87, "content": "S _ { 2 }", "type": "inline_equation" }, { "bbox": [ 490, 139, 505, 153 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 151, 505, 164 ], "spans": [ { "bbox": [ 106, 151, 242, 163 ], "score": 0.92, "content": "\\ell _ { M B R } ( S _ { 1 } , S _ { 2 } ) ^ { - } = - \\mathrm { M I O U } ( S _ { 1 } , S _ { 2 } )", "type": "inline_equation" }, { "bbox": [ 243, 151, 275, 164 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 275, 151, 319, 163 ], "score": 0.73, "content": "\\mathrm { { M I O U } } ( \\cdot , \\cdot )", "type": "inline_equation" }, { "bbox": [ 319, 151, 505, 164 ], "score": 1.0, "content": "denotes the mean intersection over union ra-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 162, 505, 174 ], "spans": [ { "bbox": [ 106, 162, 338, 174 ], "score": 1.0, "content": "tio across all matched pairs of predicted word spans from", "type": "text" }, { "bbox": [ 338, 162, 350, 173 ], "score": 0.88, "content": "S _ { 1 }", "type": "inline_equation" }, { "bbox": [ 351, 162, 368, 174 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 369, 162, 380, 173 ], "score": 0.87, "content": "S _ { 2 }", "type": "inline_equation" }, { "bbox": [ 380, 162, 505, 174 ], "score": 1.0, "content": ". We match the predicted word", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 172, 506, 186 ], "spans": [ { "bbox": [ 105, 172, 506, 186 ], "score": 1.0, "content": "spans using the maximum weight matching algorithm (Galil, 1986), where word spans correspond", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 184, 499, 197 ], "spans": [ { "bbox": [ 105, 184, 499, 197 ], "score": 1.0, "content": "to vertices, and we define edge weights by the temporal overlap between the corresponding spans.", "type": "text" } ], "index": 8 } ], "index": 6 }, { "type": "text", "bbox": [ 105, 200, 504, 223 ], "lines": [ { "bbox": [ 105, 199, 505, 214 ], "spans": [ { "bbox": [ 105, 199, 448, 214 ], "score": 1.0, "content": "For phrase structure induction, we define the loss function between two parse trees", "type": "text" }, { "bbox": [ 448, 201, 460, 212 ], "score": 0.86, "content": "\\mathcal { T } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 460, 199, 479, 214 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 479, 201, 491, 212 ], "score": 0.86, "content": "\\mathcal { T } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 491, 199, 505, 214 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 9 }, { "bbox": [ 107, 211, 456, 224 ], "spans": [ { "bbox": [ 107, 212, 232, 224 ], "score": 0.92, "content": "\\ell _ { M B R } ( \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 2 } ) = 1 - F _ { 1 } ( \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 2 } )", "type": "inline_equation" }, { "bbox": [ 232, 211, 263, 224 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 263, 212, 293, 224 ], "score": 0.93, "content": "F _ { 1 } ( \\cdot , \\cdot )", "type": "inline_equation" }, { "bbox": [ 293, 211, 341, 224 ], "score": 1.0, "content": "denotes the", "type": "text" }, { "bbox": [ 342, 212, 353, 223 ], "score": 0.89, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 354, 211, 456, 224 ], "score": 1.0, "content": "score between two trees.", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "title", "bbox": [ 108, 238, 200, 251 ], "lines": [ { "bbox": [ 105, 237, 201, 253 ], "spans": [ { "bbox": [ 105, 237, 201, 253 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "title", "bbox": [ 107, 262, 169, 274 ], "lines": [ { "bbox": [ 105, 261, 171, 276 ], "spans": [ { "bbox": [ 105, 261, 171, 276 ], "score": 1.0, "content": "4.1 SETTING", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 283, 505, 339 ], "lines": [ { "bbox": [ 106, 284, 505, 295 ], "spans": [ { "bbox": [ 106, 284, 505, 295 ], "score": 1.0, "content": "Dataset: All models are evaluated on SpokenCOCO, the spoken version of MSCOCO (Lin et al.,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 293, 504, 307 ], "spans": [ { "bbox": [ 105, 293, 462, 307 ], "score": 1.0, "content": "2014) where the text captions are read out by MTurk users (Hsu et al., 2021b). It contains", "type": "text" }, { "bbox": [ 462, 294, 504, 305 ], "score": 0.56, "content": "8 3 \\mathrm { k } / 5 \\mathrm { k } / 5 \\mathrm { k }", "type": "inline_equation" } ], "index": 14 }, { "bbox": [ 105, 306, 505, 318 ], "spans": [ { "bbox": [ 105, 306, 505, 318 ], "score": 1.0, "content": "images for training, validation, and test: each image has 5 corresponding spoken captions. Spoken-", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 315, 506, 329 ], "spans": [ { "bbox": [ 105, 315, 264, 329 ], "score": 1.0, "content": "COCO totals 740h of read speech from", "type": "text" }, { "bbox": [ 264, 316, 283, 327 ], "score": 0.63, "content": "2 . 3 \\mathrm { k }", "type": "inline_equation" }, { "bbox": [ 284, 315, 506, 329 ], "score": 1.0, "content": "speakers, with an average utterance duration of about 4", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 326, 286, 341 ], "spans": [ { "bbox": [ 105, 326, 286, 341 ], "score": 1.0, "content": "seconds, covering 29K different word types.", "type": "text" } ], "index": 17 } ], "index": 15 }, { "type": "text", "bbox": [ 107, 344, 505, 410 ], "lines": [ { "bbox": [ 106, 344, 505, 357 ], "spans": [ { "bbox": [ 106, 344, 505, 357 ], "score": 1.0, "content": "Preprocessing: For oracle word segmentation, we ran an off-the-shelf English ASR from Montreal", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 354, 505, 368 ], "spans": [ { "bbox": [ 105, 354, 505, 368 ], "score": 1.0, "content": "Force Aligner (McAuliffe et al., 2017) that was pre-trained on Librispeech and adapted to Spoken-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 365, 506, 379 ], "spans": [ { "bbox": [ 105, 365, 506, 379 ], "score": 1.0, "content": "COCO. We removed a few utterances that have mismatches in their ASR transcripts and their text", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 376, 506, 390 ], "spans": [ { "bbox": [ 105, 376, 506, 390 ], "score": 1.0, "content": "captions. Following Shi et al. (2019), we included trivial spans in tree evaluation. Additionally, we", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 388, 506, 401 ], "spans": [ { "bbox": [ 105, 388, 506, 401 ], "score": 1.0, "content": "ran an off-the-shelf English parser (Kitaev & Klein, 2018) on the ASR transcript (normalized text", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 399, 402, 412 ], "spans": [ { "bbox": [ 106, 399, 402, 412 ], "score": 1.0, "content": "with punctuation removed) to generate the oracle trees for SpokenCOCO.", "type": "text" } ], "index": 23 } ], "index": 20.5 }, { "type": "title", "bbox": [ 108, 423, 246, 434 ], "lines": [ { "bbox": [ 105, 423, 248, 435 ], "spans": [ { "bbox": [ 105, 423, 248, 435 ], "score": 1.0, "content": "4.2 BASELINES AND TOPLINES", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 108, 443, 505, 488 ], "lines": [ { "bbox": [ 106, 443, 505, 456 ], "spans": [ { "bbox": [ 106, 443, 505, 456 ], "score": 1.0, "content": "AV-NSL segments speech waveforms into word segments, then learns phrase structures on top of the", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 455, 505, 467 ], "spans": [ { "bbox": [ 106, 455, 505, 467 ], "score": 1.0, "content": "learned segments. Both segmentation and structure induction are fully-unsupervised and visually-", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 466, 506, 478 ], "spans": [ { "bbox": [ 105, 466, 506, 478 ], "score": 1.0, "content": "grounded. To help us examine the role of each component in AV-NSL, we therefore further construct", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 477, 432, 489 ], "spans": [ { "bbox": [ 106, 477, 432, 489 ], "score": 1.0, "content": "the following baselines and toplines. Their full descriptions are in Appendix A.1.", "type": "text" } ], "index": 28 } ], "index": 26.5 }, { "type": "text", "bbox": [ 107, 493, 503, 516 ], "lines": [ { "bbox": [ 107, 494, 504, 505 ], "spans": [ { "bbox": [ 107, 494, 504, 505 ], "score": 1.0, "content": "Trivial tree structures: Following (Shi et al., 2019), we include baselines without linguistic infor-", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 505, 468, 517 ], "spans": [ { "bbox": [ 106, 505, 468, 517 ], "score": 1.0, "content": "mation: random binary trees, left-branching binary trees, and right-branching binary trees.", "type": "text" } ], "index": 30 } ], "index": 29.5 }, { "type": "text", "bbox": [ 107, 521, 505, 577 ], "lines": [ { "bbox": [ 106, 520, 505, 534 ], "spans": [ { "bbox": [ 106, 520, 505, 534 ], "score": 1.0, "content": "AV-cPCFG: We train compound probabilistic context free grammar (cPCFG) (Kim et al., 2019a) on", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 533, 505, 545 ], "spans": [ { "bbox": [ 106, 533, 505, 545 ], "score": 1.0, "content": "word-level discrete speech tokens. Similar to AV-NSL, word segments and segment representations", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 543, 506, 556 ], "spans": [ { "bbox": [ 105, 543, 506, 556 ], "score": 1.0, "content": "are based on VG-HuBERT. Different from AV-NSL, the segment representations are discretized via", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 553, 505, 567 ], "spans": [ { "bbox": [ 105, 553, 505, 567 ], "score": 1.0, "content": "kmeans to obtain word-level discrete indices. In short, AV-cPCFG leverages visual cues only for", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 565, 432, 578 ], "spans": [ { "bbox": [ 105, 565, 432, 578 ], "score": 1.0, "content": "segmentation and segment representations, but not for phrase structure induction.", "type": "text" } ], "index": 35 } ], "index": 33 }, { "type": "text", "bbox": [ 107, 582, 504, 626 ], "lines": [ { "bbox": [ 105, 581, 505, 595 ], "spans": [ { "bbox": [ 105, 581, 505, 595 ], "score": 1.0, "content": "DPDP-cPCFG: Instead of training cPCFG on audio-visual word segments and audio-visual seg-", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 593, 505, 605 ], "spans": [ { "bbox": [ 106, 593, 505, 605 ], "score": 1.0, "content": "ment representations, DPDP-cPCFG does not rely on any visual grounding throughout. Instead,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 603, 506, 617 ], "spans": [ { "bbox": [ 105, 603, 506, 617 ], "score": 1.0, "content": "DPDP (Kamper, 2022) and vanilla HuBERT representations are used. As in AV-cPCFG, kmeans is", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 615, 245, 627 ], "spans": [ { "bbox": [ 106, 615, 245, 627 ], "score": 1.0, "content": "used for word-level discretization.", "type": "text" } ], "index": 39 } ], "index": 37.5 }, { "type": "text", "bbox": [ 106, 632, 504, 654 ], "lines": [ { "bbox": [ 106, 632, 505, 644 ], "spans": [ { "bbox": [ 106, 632, 505, 644 ], "score": 1.0, "content": "Oracle AV-NSL: To remove the uncertainty of unsupervised word segmentation, we directly train", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 644, 369, 655 ], "spans": [ { "bbox": [ 106, 644, 369, 655 ], "score": 1.0, "content": "AV-NSL on top of oracle word segmentation via force alignment.", "type": "text" } ], "index": 41 } ], "index": 40.5 }, { "type": "title", "bbox": [ 108, 667, 225, 678 ], "lines": [ { "bbox": [ 105, 666, 226, 680 ], "spans": [ { "bbox": [ 105, 666, 226, 680 ], "score": 1.0, "content": "4.3 EVALUATION METRIC", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 107, 687, 504, 732 ], "lines": [ { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "Word segmentation. We use the standard word boundary prediction metrics (precision, recall and", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "F1), which are calculated by comparing the temporal position between inferred word boundaries and", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 710, 506, 723 ], "spans": [ { "bbox": [ 105, 710, 506, 723 ], "score": 1.0, "content": "force aligned word boundaries. In particular, following Peng & Harwath (2022b), when an inferred", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 721, 501, 732 ], "spans": [ { "bbox": [ 106, 721, 214, 732 ], "score": 1.0, "content": "boundary is located within", "type": "text" }, { "bbox": [ 215, 721, 245, 731 ], "score": 0.81, "content": "\\pm 2 0 m s", "type": "inline_equation" }, { "bbox": [ 245, 721, 501, 732 ], "score": 1.0, "content": "of a force aligned boundary, we declare a successful prediction.", "type": "text" } ], "index": 46 } ], "index": 44.5 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 26, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 309, 760 ], "lines": [ { "bbox": [ 302, 751, 310, 762 ], "spans": [ { "bbox": [ 302, 751, 310, 762 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 82, 504, 105 ], "lines": [ { "bbox": [ 105, 81, 504, 96 ], "spans": [ { "bbox": [ 105, 81, 256, 96 ], "score": 1.0, "content": "At a high level, given a loss function", "type": "text" }, { "bbox": [ 256, 82, 312, 95 ], "score": 0.93, "content": "\\ell _ { M B R } ( O _ { 1 } , O _ { 2 } )", "type": "inline_equation" }, { "bbox": [ 313, 81, 399, 96 ], "score": 1.0, "content": "between two outputs", "type": "text" }, { "bbox": [ 399, 83, 412, 93 ], "score": 0.89, "content": "O _ { 1 }", "type": "inline_equation" }, { "bbox": [ 413, 81, 431, 96 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 431, 83, 444, 93 ], "score": 0.89, "content": "O _ { 2 }", "type": "inline_equation" }, { "bbox": [ 444, 81, 497, 96 ], "score": 1.0, "content": ", and a set of", "type": "text" }, { "bbox": [ 497, 83, 504, 92 ], "score": 0.79, "content": "k", "type": "inline_equation" } ], "index": 0 }, { "bbox": [ 106, 93, 336, 106 ], "spans": [ { "bbox": [ 106, 93, 138, 106 ], "score": 1.0, "content": "outputs", "type": "text" }, { "bbox": [ 139, 93, 218, 106 ], "score": 0.94, "content": "\\mathcal { O } = \\{ O _ { 1 } , \\ldots , O _ { k } \\}", "type": "inline_equation" }, { "bbox": [ 218, 93, 336, 106 ], "score": 1.0, "content": ", we select the optimal output", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 81, 504, 106 ] }, { "type": "interline_equation", "bbox": [ 233, 106, 378, 133 ], "lines": [ { "bbox": [ 233, 106, 378, 133 ], "spans": [ { "bbox": [ 233, 106, 378, 133 ], "score": 0.94, "content": "\\hat { O } = \\arg \\operatorname* { m i n } _ { O ^ { \\prime } \\in { \\mathcal O } } \\sum _ { O ^ { \\prime \\prime } \\in { \\mathcal O } } \\ell _ { M B R } ( O ^ { \\prime } , O ^ { \\prime \\prime } ) .", "type": "interline_equation", "image_path": "05f4cb5da860d44c1c6f81a2e82c85670940b072bcd0f89b4e5cb02bef871169.jpg" } ] } ], "index": 2.5, "virtual_lines": [ { "bbox": [ 233, 106, 378, 119.5 ], "spans": [], "index": 2 }, { "bbox": [ 233, 119.5, 378, 133.0 ], "spans": [], "index": 3 } ] }, { "type": "text", "bbox": [ 106, 140, 505, 195 ], "lines": [ { "bbox": [ 105, 139, 505, 153 ], "spans": [ { "bbox": [ 105, 139, 444, 153 ], "score": 1.0, "content": "For word segmentation, we define the loss between two segmentation proposals", "type": "text" }, { "bbox": [ 444, 141, 456, 151 ], "score": 0.87, "content": " { \\boldsymbol { S } } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 457, 139, 478, 153 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 478, 140, 489, 151 ], "score": 0.87, "content": "S _ { 2 }", "type": "inline_equation" }, { "bbox": [ 490, 139, 505, 153 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 151, 505, 164 ], "spans": [ { "bbox": [ 106, 151, 242, 163 ], "score": 0.92, "content": "\\ell _ { M B R } ( S _ { 1 } , S _ { 2 } ) ^ { - } = - \\mathrm { M I O U } ( S _ { 1 } , S _ { 2 } )", "type": "inline_equation" }, { "bbox": [ 243, 151, 275, 164 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 275, 151, 319, 163 ], "score": 0.73, "content": "\\mathrm { { M I O U } } ( \\cdot , \\cdot )", "type": "inline_equation" }, { "bbox": [ 319, 151, 505, 164 ], "score": 1.0, "content": "denotes the mean intersection over union ra-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 162, 505, 174 ], "spans": [ { "bbox": [ 106, 162, 338, 174 ], "score": 1.0, "content": "tio across all matched pairs of predicted word spans from", "type": "text" }, { "bbox": [ 338, 162, 350, 173 ], "score": 0.88, "content": "S _ { 1 }", "type": "inline_equation" }, { "bbox": [ 351, 162, 368, 174 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 369, 162, 380, 173 ], "score": 0.87, "content": "S _ { 2 }", "type": "inline_equation" }, { "bbox": [ 380, 162, 505, 174 ], "score": 1.0, "content": ". We match the predicted word", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 172, 506, 186 ], "spans": [ { "bbox": [ 105, 172, 506, 186 ], "score": 1.0, "content": "spans using the maximum weight matching algorithm (Galil, 1986), where word spans correspond", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 184, 499, 197 ], "spans": [ { "bbox": [ 105, 184, 499, 197 ], "score": 1.0, "content": "to vertices, and we define edge weights by the temporal overlap between the corresponding spans.", "type": "text" } ], "index": 8 } ], "index": 6, "bbox_fs": [ 105, 139, 506, 197 ] }, { "type": "text", "bbox": [ 105, 200, 504, 223 ], "lines": [ { "bbox": [ 105, 199, 505, 214 ], "spans": [ { "bbox": [ 105, 199, 448, 214 ], "score": 1.0, "content": "For phrase structure induction, we define the loss function between two parse trees", "type": "text" }, { "bbox": [ 448, 201, 460, 212 ], "score": 0.86, "content": "\\mathcal { T } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 460, 199, 479, 214 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 479, 201, 491, 212 ], "score": 0.86, "content": "\\mathcal { T } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 491, 199, 505, 214 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 9 }, { "bbox": [ 107, 211, 456, 224 ], "spans": [ { "bbox": [ 107, 212, 232, 224 ], "score": 0.92, "content": "\\ell _ { M B R } ( \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 2 } ) = 1 - F _ { 1 } ( \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 2 } )", "type": "inline_equation" }, { "bbox": [ 232, 211, 263, 224 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 263, 212, 293, 224 ], "score": 0.93, "content": "F _ { 1 } ( \\cdot , \\cdot )", "type": "inline_equation" }, { "bbox": [ 293, 211, 341, 224 ], "score": 1.0, "content": "denotes the", "type": "text" }, { "bbox": [ 342, 212, 353, 223 ], "score": 0.89, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 354, 211, 456, 224 ], "score": 1.0, "content": "score between two trees.", "type": "text" } ], "index": 10 } ], "index": 9.5, "bbox_fs": [ 105, 199, 505, 224 ] }, { "type": "title", "bbox": [ 108, 238, 200, 251 ], "lines": [ { "bbox": [ 105, 237, 201, 253 ], "spans": [ { "bbox": [ 105, 237, 201, 253 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "title", "bbox": [ 107, 262, 169, 274 ], "lines": [ { "bbox": [ 105, 261, 171, 276 ], "spans": [ { "bbox": [ 105, 261, 171, 276 ], "score": 1.0, "content": "4.1 SETTING", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 283, 505, 339 ], "lines": [ { "bbox": [ 106, 284, 505, 295 ], "spans": [ { "bbox": [ 106, 284, 505, 295 ], "score": 1.0, "content": "Dataset: All models are evaluated on SpokenCOCO, the spoken version of MSCOCO (Lin et al.,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 293, 504, 307 ], "spans": [ { "bbox": [ 105, 293, 462, 307 ], "score": 1.0, "content": "2014) where the text captions are read out by MTurk users (Hsu et al., 2021b). It contains", "type": "text" }, { "bbox": [ 462, 294, 504, 305 ], "score": 0.56, "content": "8 3 \\mathrm { k } / 5 \\mathrm { k } / 5 \\mathrm { k }", "type": "inline_equation" } ], "index": 14 }, { "bbox": [ 105, 306, 505, 318 ], "spans": [ { "bbox": [ 105, 306, 505, 318 ], "score": 1.0, "content": "images for training, validation, and test: each image has 5 corresponding spoken captions. Spoken-", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 315, 506, 329 ], "spans": [ { "bbox": [ 105, 315, 264, 329 ], "score": 1.0, "content": "COCO totals 740h of read speech from", "type": "text" }, { "bbox": [ 264, 316, 283, 327 ], "score": 0.63, "content": "2 . 3 \\mathrm { k }", "type": "inline_equation" }, { "bbox": [ 284, 315, 506, 329 ], "score": 1.0, "content": "speakers, with an average utterance duration of about 4", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 326, 286, 341 ], "spans": [ { "bbox": [ 105, 326, 286, 341 ], "score": 1.0, "content": "seconds, covering 29K different word types.", "type": "text" } ], "index": 17 } ], "index": 15, "bbox_fs": [ 105, 284, 506, 341 ] }, { "type": "text", "bbox": [ 107, 344, 505, 410 ], "lines": [ { "bbox": [ 106, 344, 505, 357 ], "spans": [ { "bbox": [ 106, 344, 505, 357 ], "score": 1.0, "content": "Preprocessing: For oracle word segmentation, we ran an off-the-shelf English ASR from Montreal", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 354, 505, 368 ], "spans": [ { "bbox": [ 105, 354, 505, 368 ], "score": 1.0, "content": "Force Aligner (McAuliffe et al., 2017) that was pre-trained on Librispeech and adapted to Spoken-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 365, 506, 379 ], "spans": [ { "bbox": [ 105, 365, 506, 379 ], "score": 1.0, "content": "COCO. We removed a few utterances that have mismatches in their ASR transcripts and their text", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 376, 506, 390 ], "spans": [ { "bbox": [ 105, 376, 506, 390 ], "score": 1.0, "content": "captions. Following Shi et al. (2019), we included trivial spans in tree evaluation. Additionally, we", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 388, 506, 401 ], "spans": [ { "bbox": [ 105, 388, 506, 401 ], "score": 1.0, "content": "ran an off-the-shelf English parser (Kitaev & Klein, 2018) on the ASR transcript (normalized text", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 399, 402, 412 ], "spans": [ { "bbox": [ 106, 399, 402, 412 ], "score": 1.0, "content": "with punctuation removed) to generate the oracle trees for SpokenCOCO.", "type": "text" } ], "index": 23 } ], "index": 20.5, "bbox_fs": [ 105, 344, 506, 412 ] }, { "type": "title", "bbox": [ 108, 423, 246, 434 ], "lines": [ { "bbox": [ 105, 423, 248, 435 ], "spans": [ { "bbox": [ 105, 423, 248, 435 ], "score": 1.0, "content": "4.2 BASELINES AND TOPLINES", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 108, 443, 505, 488 ], "lines": [ { "bbox": [ 106, 443, 505, 456 ], "spans": [ { "bbox": [ 106, 443, 505, 456 ], "score": 1.0, "content": "AV-NSL segments speech waveforms into word segments, then learns phrase structures on top of the", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 455, 505, 467 ], "spans": [ { "bbox": [ 106, 455, 505, 467 ], "score": 1.0, "content": "learned segments. Both segmentation and structure induction are fully-unsupervised and visually-", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 466, 506, 478 ], "spans": [ { "bbox": [ 105, 466, 506, 478 ], "score": 1.0, "content": "grounded. To help us examine the role of each component in AV-NSL, we therefore further construct", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 477, 432, 489 ], "spans": [ { "bbox": [ 106, 477, 432, 489 ], "score": 1.0, "content": "the following baselines and toplines. Their full descriptions are in Appendix A.1.", "type": "text" } ], "index": 28 } ], "index": 26.5, "bbox_fs": [ 105, 443, 506, 489 ] }, { "type": "text", "bbox": [ 107, 493, 503, 516 ], "lines": [ { "bbox": [ 107, 494, 504, 505 ], "spans": [ { "bbox": [ 107, 494, 504, 505 ], "score": 1.0, "content": "Trivial tree structures: Following (Shi et al., 2019), we include baselines without linguistic infor-", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 505, 468, 517 ], "spans": [ { "bbox": [ 106, 505, 468, 517 ], "score": 1.0, "content": "mation: random binary trees, left-branching binary trees, and right-branching binary trees.", "type": "text" } ], "index": 30 } ], "index": 29.5, "bbox_fs": [ 106, 494, 504, 517 ] }, { "type": "text", "bbox": [ 107, 521, 505, 577 ], "lines": [ { "bbox": [ 106, 520, 505, 534 ], "spans": [ { "bbox": [ 106, 520, 505, 534 ], "score": 1.0, "content": "AV-cPCFG: We train compound probabilistic context free grammar (cPCFG) (Kim et al., 2019a) on", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 533, 505, 545 ], "spans": [ { "bbox": [ 106, 533, 505, 545 ], "score": 1.0, "content": "word-level discrete speech tokens. Similar to AV-NSL, word segments and segment representations", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 543, 506, 556 ], "spans": [ { "bbox": [ 105, 543, 506, 556 ], "score": 1.0, "content": "are based on VG-HuBERT. Different from AV-NSL, the segment representations are discretized via", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 553, 505, 567 ], "spans": [ { "bbox": [ 105, 553, 505, 567 ], "score": 1.0, "content": "kmeans to obtain word-level discrete indices. In short, AV-cPCFG leverages visual cues only for", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 565, 432, 578 ], "spans": [ { "bbox": [ 105, 565, 432, 578 ], "score": 1.0, "content": "segmentation and segment representations, but not for phrase structure induction.", "type": "text" } ], "index": 35 } ], "index": 33, "bbox_fs": [ 105, 520, 506, 578 ] }, { "type": "text", "bbox": [ 107, 582, 504, 626 ], "lines": [ { "bbox": [ 105, 581, 505, 595 ], "spans": [ { "bbox": [ 105, 581, 505, 595 ], "score": 1.0, "content": "DPDP-cPCFG: Instead of training cPCFG on audio-visual word segments and audio-visual seg-", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 593, 505, 605 ], "spans": [ { "bbox": [ 106, 593, 505, 605 ], "score": 1.0, "content": "ment representations, DPDP-cPCFG does not rely on any visual grounding throughout. Instead,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 603, 506, 617 ], "spans": [ { "bbox": [ 105, 603, 506, 617 ], "score": 1.0, "content": "DPDP (Kamper, 2022) and vanilla HuBERT representations are used. As in AV-cPCFG, kmeans is", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 615, 245, 627 ], "spans": [ { "bbox": [ 106, 615, 245, 627 ], "score": 1.0, "content": "used for word-level discretization.", "type": "text" } ], "index": 39 } ], "index": 37.5, "bbox_fs": [ 105, 581, 506, 627 ] }, { "type": "text", "bbox": [ 106, 632, 504, 654 ], "lines": [ { "bbox": [ 106, 632, 505, 644 ], "spans": [ { "bbox": [ 106, 632, 505, 644 ], "score": 1.0, "content": "Oracle AV-NSL: To remove the uncertainty of unsupervised word segmentation, we directly train", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 644, 369, 655 ], "spans": [ { "bbox": [ 106, 644, 369, 655 ], "score": 1.0, "content": "AV-NSL on top of oracle word segmentation via force alignment.", "type": "text" } ], "index": 41 } ], "index": 40.5, "bbox_fs": [ 106, 632, 505, 655 ] }, { "type": "title", "bbox": [ 108, 667, 225, 678 ], "lines": [ { "bbox": [ 105, 666, 226, 680 ], "spans": [ { "bbox": [ 105, 666, 226, 680 ], "score": 1.0, "content": "4.3 EVALUATION METRIC", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 107, 687, 504, 732 ], "lines": [ { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "Word segmentation. We use the standard word boundary prediction metrics (precision, recall and", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "F1), which are calculated by comparing the temporal position between inferred word boundaries and", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 710, 506, 723 ], "spans": [ { "bbox": [ 105, 710, 506, 723 ], "score": 1.0, "content": "force aligned word boundaries. In particular, following Peng & Harwath (2022b), when an inferred", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 721, 501, 732 ], "spans": [ { "bbox": [ 106, 721, 214, 732 ], "score": 1.0, "content": "boundary is located within", "type": "text" }, { "bbox": [ 215, 721, 245, 731 ], "score": 0.81, "content": "\\pm 2 0 m s", "type": "inline_equation" }, { "bbox": [ 245, 721, 501, 732 ], "score": 1.0, "content": "of a force aligned boundary, we declare a successful prediction.", "type": "text" } ], "index": 46 } ], "index": 44.5, "bbox_fs": [ 105, 687, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 126 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 466, 95 ], "score": 1.0, "content": "Parsing. For parsing with oracle word segmentation, we use EVALB to calculate the", "type": "text" }, { "bbox": [ 467, 83, 479, 93 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 479, 82, 505, 95 ], "score": 1.0, "content": "score", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 92, 506, 106 ], "spans": [ { "bbox": [ 105, 92, 506, 106 ], "score": 1.0, "content": "between the predicted and ground-truth parse trees.5 For parsing with inferred word segmentation,", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 505, 117 ], "spans": [ { "bbox": [ 105, 105, 505, 117 ], "score": 1.0, "content": "due to the mismatch in the number of nodes between the predicted and ground-truth parse trees, we", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 116, 489, 128 ], "spans": [ { "bbox": [ 105, 116, 489, 128 ], "score": 1.0, "content": "introduce the structured average intersection-over-union ratio (SAIOU) as an additional metric.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 106, 132, 505, 215 ], "lines": [ { "bbox": [ 106, 131, 506, 145 ], "spans": [ { "bbox": [ 106, 131, 506, 145 ], "score": 1.0, "content": "SAIOU takes both word segmentation quality and temporal overlap between induced constituents", "type": "text" } ], "index": 4 }, { "bbox": [ 102, 142, 509, 175 ], "spans": [ { "bbox": [ 102, 142, 146, 175 ], "score": 1.0, "content": "into consutterance", "type": "text" }, { "bbox": [ 146, 154, 264, 167 ], "score": 0.91, "content": "\\mathcal { T } _ { 1 } = \\{ c _ { 1 , i } = ( \\ell _ { 1 , i } , \\dot { r } _ { 1 , i } ) \\} _ { i = 1 } ^ { n _ { 1 } }", "type": "inline_equation" }, { "bbox": [ 264, 142, 270, 175 ], "score": 1.0, "content": "pa", "type": "text" }, { "bbox": [ 276, 142, 284, 175 ], "score": 1.0, "content": "t id", "type": "text" }, { "bbox": [ 284, 154, 406, 168 ], "score": 0.92, "content": "\\mathcal { T } _ { 2 } = \\{ c _ { 2 , j } = ( \\ell _ { 2 , j } , r _ { 2 , j } ) \\} _ { j = 1 } ^ { n _ { 2 } }", "type": "inline_equation" }, { "bbox": [ 406, 142, 509, 175 ], "score": 1.0, "content": "s over the same speech, represented by a set of", "type": "text" } ], "index": 5 }, { "bbox": [ 101, 166, 466, 199 ], "spans": [ { "bbox": [ 101, 166, 181, 199 ], "score": 1.0, "content": "constituency tempthe constituents in", "type": "text" }, { "bbox": [ 193, 166, 211, 199 ], "score": 1.0, "content": "l boand", "type": "text" }, { "bbox": [ 222, 166, 226, 199 ], "score": 1.0, "content": "d,", "type": "text" }, { "bbox": [ 245, 168, 251, 177 ], "score": 0.78, "content": "\\ell", "type": "inline_equation" }, { "bbox": [ 270, 169, 276, 177 ], "score": 0.73, "content": "r", "type": "inline_equation" }, { "bbox": [ 436, 166, 466, 199 ], "score": 1.0, "content": "ignmen, where", "type": "text" } ], "index": 6 }, { "bbox": [ 182, 178, 504, 193 ], "spans": [ { "bbox": [ 182, 180, 193, 190 ], "score": 0.87, "content": "\\mathcal { T } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 211, 178, 222, 190 ], "score": 0.81, "content": "\\mathcal { T } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 226, 179, 436, 193 ], "score": 0.81, "content": "\\begin{array} { r } { \\hat { \\mathcal { A } } = \\arg \\operatorname* { m a x } _ { \\nu a l i d \\mathcal { A } } \\sum _ { i = 1 } ^ { n _ { 1 } } \\sum _ { j = 1 } ^ { n _ { 2 } } \\mathcal { A } _ { i , j } \\mathrm { I o U } ( c _ { 1 , i } , c _ { 2 , j } ) } \\end{array}", "type": "inline_equation" }, { "bbox": [ 467, 179, 504, 192 ], "score": 0.92, "content": "A _ { i , j } = 1", "type": "inline_equation" } ], "index": 7 }, { "bbox": [ 106, 191, 505, 205 ], "spans": [ { "bbox": [ 106, 191, 139, 204 ], "score": 1.0, "content": "denotes", "type": "text" }, { "bbox": [ 140, 194, 155, 204 ], "score": 0.88, "content": "c _ { 1 , i }", "type": "inline_equation" }, { "bbox": [ 155, 191, 203, 204 ], "score": 1.0, "content": "aligns with", "type": "text" }, { "bbox": [ 203, 194, 219, 204 ], "score": 0.88, "content": "c _ { 2 , j }", "type": "inline_equation" }, { "bbox": [ 219, 191, 240, 204 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 240, 192, 278, 205 ], "score": 0.92, "content": "A _ { i , j } = 0", "type": "inline_equation" }, { "bbox": [ 278, 191, 324, 204 ], "score": 1.0, "content": "otherwise;", "type": "text" }, { "bbox": [ 324, 193, 360, 204 ], "score": 0.84, "content": "\\operatorname { I o U } ( \\cdot , \\cdot )", "type": "inline_equation" }, { "bbox": [ 360, 191, 505, 204 ], "score": 1.0, "content": "denotes the intersection-over-union", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 203, 472, 216 ], "spans": [ { "bbox": [ 105, 203, 282, 216 ], "score": 1.0, "content": "ratio between two spans. A valid alignment", "type": "text" }, { "bbox": [ 282, 204, 291, 213 ], "score": 0.81, "content": "\\mathcal { A }", "type": "inline_equation" }, { "bbox": [ 292, 203, 472, 216 ], "score": 1.0, "content": "is one that satisfies the following conditions:", "type": "text" } ], "index": 9 } ], "index": 6.5 }, { "type": "text", "bbox": [ 107, 219, 411, 231 ], "lines": [ { "bbox": [ 106, 217, 412, 233 ], "spans": [ { "bbox": [ 106, 217, 412, 233 ], "score": 1.0, "content": "1. Any constituent may be aligned with up to 1 constituent in the other tree;", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 108, 235, 275, 247 ], "lines": [ { "bbox": [ 105, 233, 276, 249 ], "spans": [ { "bbox": [ 105, 233, 178, 249 ], "score": 1.0, "content": "2. For any pair of", "type": "text" }, { "bbox": [ 179, 236, 183, 245 ], "score": 0.77, "content": "i", "type": "inline_equation" }, { "bbox": [ 184, 233, 201, 249 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 201, 236, 208, 247 ], "score": 0.83, "content": "j", "type": "inline_equation" }, { "bbox": [ 208, 233, 235, 249 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 236, 235, 272, 248 ], "score": 0.92, "content": "A _ { i , j } = 1", "type": "inline_equation" }, { "bbox": [ 272, 233, 276, 249 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 106, 250, 474, 304 ], "lines": [ { "bbox": [ 105, 250, 474, 264 ], "spans": [ { "bbox": [ 105, 250, 192, 264 ], "score": 1.0, "content": "• Any descendant of", "type": "text" }, { "bbox": [ 192, 253, 228, 263 ], "score": 0.29, "content": "c _ { 1 , i } , c _ { 1 , k }", "type": "inline_equation" }, { "bbox": [ 228, 250, 373, 264 ], "score": 1.0, "content": ", may either align to a descendant of", "type": "text" }, { "bbox": [ 374, 253, 389, 263 ], "score": 0.89, "content": "c _ { 2 , j }", "type": "inline_equation" }, { "bbox": [ 390, 250, 474, 264 ], "score": 1.0, "content": "or be left unaligned;", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 263, 455, 278 ], "spans": [ { "bbox": [ 105, 263, 181, 278 ], "score": 1.0, "content": "• Any ancestor of", "type": "text" }, { "bbox": [ 181, 266, 219, 277 ], "score": 0.32, "content": "c _ { 1 , i } , c _ { 1 , k ^ { \\prime } }", "type": "inline_equation" }, { "bbox": [ 219, 263, 354, 278 ], "score": 1.0, "content": ", may either align to a ancestor of", "type": "text" }, { "bbox": [ 354, 266, 370, 277 ], "score": 0.89, "content": "c _ { 2 , j }", "type": "inline_equation" }, { "bbox": [ 370, 263, 455, 278 ], "score": 1.0, "content": "or be left unaligned;", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 277, 474, 292 ], "spans": [ { "bbox": [ 105, 277, 192, 292 ], "score": 1.0, "content": "• Any descendant of", "type": "text" }, { "bbox": [ 192, 280, 228, 291 ], "score": 0.75, "content": "c _ { 2 , j } , c _ { 2 , p }", "type": "inline_equation" }, { "bbox": [ 228, 277, 374, 292 ], "score": 1.0, "content": ", may either align to a descendant of", "type": "text" }, { "bbox": [ 374, 280, 389, 291 ], "score": 0.88, "content": "c _ { 1 , i }", "type": "inline_equation" }, { "bbox": [ 389, 277, 474, 292 ], "score": 1.0, "content": "or be left unaligned;", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 291, 454, 305 ], "spans": [ { "bbox": [ 105, 291, 181, 305 ], "score": 1.0, "content": "• Any ancestor of", "type": "text" }, { "bbox": [ 181, 294, 220, 304 ], "score": 0.48, "content": "c _ { 2 , j } , c _ { 2 , p ^ { \\prime } }", "type": "inline_equation" }, { "bbox": [ 220, 291, 354, 305 ], "score": 1.0, "content": ", may either align to a ancestor of", "type": "text" }, { "bbox": [ 355, 294, 370, 304 ], "score": 0.89, "content": "c _ { 1 , i }", "type": "inline_equation" }, { "bbox": [ 370, 291, 454, 305 ], "score": 1.0, "content": "or be left unaligned.", "type": "text" } ], "index": 15 } ], "index": 13.5 }, { "type": "text", "bbox": [ 105, 313, 487, 326 ], "lines": [ { "bbox": [ 106, 312, 488, 329 ], "spans": [ { "bbox": [ 106, 312, 223, 329 ], "score": 1.0, "content": "Given the optimal alignment", "type": "text" }, { "bbox": [ 223, 312, 232, 324 ], "score": 0.84, "content": "\\hat { A }", "type": "inline_equation" }, { "bbox": [ 232, 312, 434, 329 ], "score": 1.0, "content": ", we calculate the structured average IOU between", "type": "text" }, { "bbox": [ 434, 315, 445, 325 ], "score": 0.88, "content": "\\mathcal { T } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 445, 312, 463, 329 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 463, 315, 474, 325 ], "score": 0.87, "content": "\\mathcal { T } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 474, 312, 488, 329 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "interline_equation", "bbox": [ 182, 328, 428, 368 ], "lines": [ { "bbox": [ 182, 328, 428, 368 ], "spans": [ { "bbox": [ 182, 328, 428, 368 ], "score": 0.95, "content": "\\operatorname { S A I o U } ( \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 2 } ) = \\frac { 2 } { n _ { 1 } + n _ { 2 } } \\left( \\sum _ { i = 1 } ^ { n _ { 1 } } \\sum _ { j = 1 } ^ { n _ { 2 } } \\hat { A } _ { i , j } \\mathrm { I o U } ( c _ { 1 , i } , c _ { 2 , j } ) \\right) .", "type": "interline_equation", "image_path": "44b46d1aab90ab2d5d4ff14b21f09616a5265b08cb4be359e83470e364277370.jpg" } ] } ], "index": 18, "virtual_lines": [ { "bbox": [ 182, 328, 428, 341.3333333333333 ], "spans": [], "index": 17 }, { "bbox": [ 182, 341.3333333333333, 428, 354.66666666666663 ], "spans": [], "index": 18 }, { "bbox": [ 182, 354.66666666666663, 428, 367.99999999999994 ], "spans": [], "index": 19 } ] }, { "type": "title", "bbox": [ 108, 379, 299, 390 ], "lines": [ { "bbox": [ 105, 378, 301, 392 ], "spans": [ { "bbox": [ 105, 378, 301, 392 ], "score": 1.0, "content": "4.4 UNSUPERVISED WORD SEGMENTATION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 399, 505, 443 ], "lines": [ { "bbox": [ 105, 399, 506, 413 ], "spans": [ { "bbox": [ 105, 399, 506, 413 ], "score": 1.0, "content": "We validate our decision of adopting VG-HuBERT to extract word-like units from raw speech wave-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 410, 506, 423 ], "spans": [ { "bbox": [ 105, 410, 506, 423 ], "score": 1.0, "content": "forms for later phrase structure parsing. In particular, we investigate two questions: (1) How does", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 421, 506, 434 ], "spans": [ { "bbox": [ 105, 421, 506, 434 ], "score": 1.0, "content": "segment insertion affect word segmentation performance? (2) how does MBR-based VG-HuBERT", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 433, 294, 444 ], "spans": [ { "bbox": [ 105, 433, 294, 444 ], "score": 1.0, "content": "compare to supervised selected VG-HuBERT?", "type": "text" } ], "index": 24 } ], "index": 22.5 }, { "type": "text", "bbox": [ 106, 449, 505, 592 ], "lines": [ { "bbox": [ 105, 448, 505, 462 ], "spans": [ { "bbox": [ 105, 448, 505, 462 ], "score": 1.0, "content": "In Table 1, in addition to VG-HuBERT, we also list a speech-only word segmentation algorithm", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 461, 504, 472 ], "spans": [ { "bbox": [ 106, 461, 504, 472 ], "score": 1.0, "content": "DPDP (Kamper, 2022). Note that audio-visual model VG-HuBERT significantly outperform DPDP.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 469, 506, 486 ], "spans": [ { "bbox": [ 105, 469, 506, 486 ], "score": 1.0, "content": "For question (1), by comparing the third row and the fourth row, as expected we see that performing", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 482, 505, 496 ], "spans": [ { "bbox": [ 105, 482, 505, 496 ], "score": 1.0, "content": "segment insertion improves recall and hurts precision, and slightly improves F1. For question (2), by", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 493, 505, 506 ], "spans": [ { "bbox": [ 106, 493, 505, 506 ], "score": 1.0, "content": "comparing the fourth row and the fifth row (second to last row), we see that MBR selection actually", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 504, 505, 516 ], "spans": [ { "bbox": [ 106, 504, 505, 516 ], "score": 1.0, "content": "leads to better performance than supervised selection. The final MBR selection we adopted is based", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 516, 504, 527 ], "spans": [ { "bbox": [ 106, 516, 504, 527 ], "score": 1.0, "content": "on the last row, where we first performed MBR selection on SpokenCOCO val set on all 405 candi-", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 526, 505, 539 ], "spans": [ { "bbox": [ 106, 526, 505, 539 ], "score": 1.0, "content": "dates, and subsequently chose the 10 most selected combinations to perform another round of MBR", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 536, 506, 550 ], "spans": [ { "bbox": [ 105, 536, 506, 550 ], "score": 1.0, "content": "decoding. Getting the top 10 most selected combinations does not require knowing the performance", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 549, 505, 560 ], "spans": [ { "bbox": [ 106, 549, 505, 560 ], "score": 1.0, "content": "on segmentation, and therefore this process is still completely unsupervised. The reason for doing 2", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 559, 505, 571 ], "spans": [ { "bbox": [ 106, 559, 505, 571 ], "score": 1.0, "content": "iterations of MBR is because performing MBR on 405 candidates on SpokenCOCO training set is", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 569, 506, 583 ], "spans": [ { "bbox": [ 105, 569, 506, 583 ], "score": 1.0, "content": "estimated to take 2 months, and MBR on 10 candidates can be done in 5 days. Comparing the last", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 581, 426, 593 ], "spans": [ { "bbox": [ 106, 581, 426, 593 ], "score": 1.0, "content": "two rows, we observe that two iterations of MBR does not lead to worse results.", "type": "text" } ], "index": 37 } ], "index": 31 }, { "type": "title", "bbox": [ 107, 605, 342, 616 ], "lines": [ { "bbox": [ 105, 604, 343, 618 ], "spans": [ { "bbox": [ 105, 604, 343, 618 ], "score": 1.0, "content": "4.5 UNSUPERVISED PHRASE STRUCTURE INDUCTION", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 625, 505, 714 ], "lines": [ { "bbox": [ 106, 625, 505, 639 ], "spans": [ { "bbox": [ 106, 625, 505, 639 ], "score": 1.0, "content": "We quantitatively show that AV-NSL learns meaningful phrase structure given word segments.", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 637, 505, 649 ], "spans": [ { "bbox": [ 106, 637, 505, 649 ], "score": 1.0, "content": "First, Table 2 is the main result of the fully-unsupervised AV-NSL on SpokenCOCO, evaluated", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 648, 506, 661 ], "spans": [ { "bbox": [ 105, 648, 506, 661 ], "score": 1.0, "content": "with SAIOU. The best performing AV-NSL is based on our improved VG-HuBERT with MBR", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 659, 506, 672 ], "spans": [ { "bbox": [ 105, 659, 506, 672 ], "score": 1.0, "content": "top 10 selection for word segmentation, attention-weighted mean-pool over VG-HuBERT layers as", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 669, 505, 683 ], "spans": [ { "bbox": [ 105, 669, 505, 683 ], "score": 1.0, "content": "the segment representations, and another MBR decoding over all phrase structure induction hyper-", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 680, 506, 694 ], "spans": [ { "bbox": [ 105, 680, 506, 694 ], "score": 1.0, "content": "parameters. Comparing AV-NSL against AV-cPCFG and AV-cPCFG against DPDP-cPCFG, we", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 692, 506, 704 ], "spans": [ { "bbox": [ 105, 692, 506, 704 ], "score": 1.0, "content": "empirically show the necessity of training AV-NSL on continuous segment representation instead of", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 702, 494, 716 ], "spans": [ { "bbox": [ 105, 702, 494, 716 ], "score": 1.0, "content": "discretized speech tokens, and the effectiveness of visual-grounding in our overall model design.", "type": "text" } ], "index": 46 } ], "index": 42.5 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 119, 721, 280, 732 ], "lines": [ { "bbox": [ 118, 720, 281, 734 ], "spans": [ { "bbox": [ 118, 720, 281, 734 ], "score": 1.0, "content": "5https://nlp.cs.nyu.edu/evalb/", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 25, 308, 38 ], "spans": [ { "bbox": [ 107, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "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": "text", "bbox": [ 107, 82, 505, 126 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 466, 95 ], "score": 1.0, "content": "Parsing. For parsing with oracle word segmentation, we use EVALB to calculate the", "type": "text" }, { "bbox": [ 467, 83, 479, 93 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 479, 82, 505, 95 ], "score": 1.0, "content": "score", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 92, 506, 106 ], "spans": [ { "bbox": [ 105, 92, 506, 106 ], "score": 1.0, "content": "between the predicted and ground-truth parse trees.5 For parsing with inferred word segmentation,", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 505, 117 ], "spans": [ { "bbox": [ 105, 105, 505, 117 ], "score": 1.0, "content": "due to the mismatch in the number of nodes between the predicted and ground-truth parse trees, we", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 116, 489, 128 ], "spans": [ { "bbox": [ 105, 116, 489, 128 ], "score": 1.0, "content": "introduce the structured average intersection-over-union ratio (SAIOU) as an additional metric.", "type": "text" } ], "index": 3 } ], "index": 1.5, "bbox_fs": [ 105, 82, 506, 128 ] }, { "type": "text", "bbox": [ 106, 132, 505, 215 ], "lines": [ { "bbox": [ 106, 131, 506, 145 ], "spans": [ { "bbox": [ 106, 131, 506, 145 ], "score": 1.0, "content": "SAIOU takes both word segmentation quality and temporal overlap between induced constituents", "type": "text" } ], "index": 4 }, { "bbox": [ 102, 142, 509, 175 ], "spans": [ { "bbox": [ 102, 142, 146, 175 ], "score": 1.0, "content": "into consutterance", "type": "text" }, { "bbox": [ 146, 154, 264, 167 ], "score": 0.91, "content": "\\mathcal { T } _ { 1 } = \\{ c _ { 1 , i } = ( \\ell _ { 1 , i } , \\dot { r } _ { 1 , i } ) \\} _ { i = 1 } ^ { n _ { 1 } }", "type": "inline_equation" }, { "bbox": [ 264, 142, 270, 175 ], "score": 1.0, "content": "pa", "type": "text" }, { "bbox": [ 276, 142, 284, 175 ], "score": 1.0, "content": "t id", "type": "text" }, { "bbox": [ 284, 154, 406, 168 ], "score": 0.92, "content": "\\mathcal { T } _ { 2 } = \\{ c _ { 2 , j } = ( \\ell _ { 2 , j } , r _ { 2 , j } ) \\} _ { j = 1 } ^ { n _ { 2 } }", "type": "inline_equation" }, { "bbox": [ 406, 142, 509, 175 ], "score": 1.0, "content": "s over the same speech, represented by a set of", "type": "text" } ], "index": 5 }, { "bbox": [ 101, 166, 466, 199 ], "spans": [ { "bbox": [ 101, 166, 181, 199 ], "score": 1.0, "content": "constituency tempthe constituents in", "type": "text" }, { "bbox": [ 193, 166, 211, 199 ], "score": 1.0, "content": "l boand", "type": "text" }, { "bbox": [ 222, 166, 226, 199 ], "score": 1.0, "content": "d,", "type": "text" }, { "bbox": [ 245, 168, 251, 177 ], "score": 0.78, "content": "\\ell", "type": "inline_equation" }, { "bbox": [ 270, 169, 276, 177 ], "score": 0.73, "content": "r", "type": "inline_equation" }, { "bbox": [ 436, 166, 466, 199 ], "score": 1.0, "content": "ignmen, where", "type": "text" } ], "index": 6 }, { "bbox": [ 182, 178, 504, 193 ], "spans": [ { "bbox": [ 182, 180, 193, 190 ], "score": 0.87, "content": "\\mathcal { T } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 211, 178, 222, 190 ], "score": 0.81, "content": "\\mathcal { T } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 226, 179, 436, 193 ], "score": 0.81, "content": "\\begin{array} { r } { \\hat { \\mathcal { A } } = \\arg \\operatorname* { m a x } _ { \\nu a l i d \\mathcal { A } } \\sum _ { i = 1 } ^ { n _ { 1 } } \\sum _ { j = 1 } ^ { n _ { 2 } } \\mathcal { A } _ { i , j } \\mathrm { I o U } ( c _ { 1 , i } , c _ { 2 , j } ) } \\end{array}", "type": "inline_equation" }, { "bbox": [ 467, 179, 504, 192 ], "score": 0.92, "content": "A _ { i , j } = 1", "type": "inline_equation" } ], "index": 7 }, { "bbox": [ 106, 191, 505, 205 ], "spans": [ { "bbox": [ 106, 191, 139, 204 ], "score": 1.0, "content": "denotes", "type": "text" }, { "bbox": [ 140, 194, 155, 204 ], "score": 0.88, "content": "c _ { 1 , i }", "type": "inline_equation" }, { "bbox": [ 155, 191, 203, 204 ], "score": 1.0, "content": "aligns with", "type": "text" }, { "bbox": [ 203, 194, 219, 204 ], "score": 0.88, "content": "c _ { 2 , j }", "type": "inline_equation" }, { "bbox": [ 219, 191, 240, 204 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 240, 192, 278, 205 ], "score": 0.92, "content": "A _ { i , j } = 0", "type": "inline_equation" }, { "bbox": [ 278, 191, 324, 204 ], "score": 1.0, "content": "otherwise;", "type": "text" }, { "bbox": [ 324, 193, 360, 204 ], "score": 0.84, "content": "\\operatorname { I o U } ( \\cdot , \\cdot )", "type": "inline_equation" }, { "bbox": [ 360, 191, 505, 204 ], "score": 1.0, "content": "denotes the intersection-over-union", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 203, 472, 216 ], "spans": [ { "bbox": [ 105, 203, 282, 216 ], "score": 1.0, "content": "ratio between two spans. A valid alignment", "type": "text" }, { "bbox": [ 282, 204, 291, 213 ], "score": 0.81, "content": "\\mathcal { A }", "type": "inline_equation" }, { "bbox": [ 292, 203, 472, 216 ], "score": 1.0, "content": "is one that satisfies the following conditions:", "type": "text" } ], "index": 9 } ], "index": 6.5, "bbox_fs": [ 101, 131, 509, 216 ] }, { "type": "text", "bbox": [ 107, 219, 411, 231 ], "lines": [ { "bbox": [ 106, 217, 412, 233 ], "spans": [ { "bbox": [ 106, 217, 412, 233 ], "score": 1.0, "content": "1. Any constituent may be aligned with up to 1 constituent in the other tree;", "type": "text" } ], "index": 10 } ], "index": 10, "bbox_fs": [ 106, 217, 412, 233 ] }, { "type": "text", "bbox": [ 108, 235, 275, 247 ], "lines": [ { "bbox": [ 105, 233, 276, 249 ], "spans": [ { "bbox": [ 105, 233, 178, 249 ], "score": 1.0, "content": "2. For any pair of", "type": "text" }, { "bbox": [ 179, 236, 183, 245 ], "score": 0.77, "content": "i", "type": "inline_equation" }, { "bbox": [ 184, 233, 201, 249 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 201, 236, 208, 247 ], "score": 0.83, "content": "j", "type": "inline_equation" }, { "bbox": [ 208, 233, 235, 249 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 236, 235, 272, 248 ], "score": 0.92, "content": "A _ { i , j } = 1", "type": "inline_equation" }, { "bbox": [ 272, 233, 276, 249 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 11 } ], "index": 11, "bbox_fs": [ 105, 233, 276, 249 ] }, { "type": "list", "bbox": [ 106, 250, 474, 304 ], "lines": [ { "bbox": [ 105, 250, 474, 264 ], "spans": [ { "bbox": [ 105, 250, 192, 264 ], "score": 1.0, "content": "• Any descendant of", "type": "text" }, { "bbox": [ 192, 253, 228, 263 ], "score": 0.29, "content": "c _ { 1 , i } , c _ { 1 , k }", "type": "inline_equation" }, { "bbox": [ 228, 250, 373, 264 ], "score": 1.0, "content": ", may either align to a descendant of", "type": "text" }, { "bbox": [ 374, 253, 389, 263 ], "score": 0.89, "content": "c _ { 2 , j }", "type": "inline_equation" }, { "bbox": [ 390, 250, 474, 264 ], "score": 1.0, "content": "or be left unaligned;", "type": "text" } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 105, 263, 455, 278 ], "spans": [ { "bbox": [ 105, 263, 181, 278 ], "score": 1.0, "content": "• Any ancestor of", "type": "text" }, { "bbox": [ 181, 266, 219, 277 ], "score": 0.32, "content": "c _ { 1 , i } , c _ { 1 , k ^ { \\prime } }", "type": "inline_equation" }, { "bbox": [ 219, 263, 354, 278 ], "score": 1.0, "content": ", may either align to a ancestor of", "type": "text" }, { "bbox": [ 354, 266, 370, 277 ], "score": 0.89, "content": "c _ { 2 , j }", "type": "inline_equation" }, { "bbox": [ 370, 263, 455, 278 ], "score": 1.0, "content": "or be left unaligned;", "type": "text" } ], "index": 13, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 277, 474, 292 ], "spans": [ { "bbox": [ 105, 277, 192, 292 ], "score": 1.0, "content": "• Any descendant of", "type": "text" }, { "bbox": [ 192, 280, 228, 291 ], "score": 0.75, "content": "c _ { 2 , j } , c _ { 2 , p }", "type": "inline_equation" }, { "bbox": [ 228, 277, 374, 292 ], "score": 1.0, "content": ", may either align to a descendant of", "type": "text" }, { "bbox": [ 374, 280, 389, 291 ], "score": 0.88, "content": "c _ { 1 , i }", "type": "inline_equation" }, { "bbox": [ 389, 277, 474, 292 ], "score": 1.0, "content": "or be left unaligned;", "type": "text" } ], "index": 14, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 291, 454, 305 ], "spans": [ { "bbox": [ 105, 291, 181, 305 ], "score": 1.0, "content": "• Any ancestor of", "type": "text" }, { "bbox": [ 181, 294, 220, 304 ], "score": 0.48, "content": "c _ { 2 , j } , c _ { 2 , p ^ { \\prime } }", "type": "inline_equation" }, { "bbox": [ 220, 291, 354, 305 ], "score": 1.0, "content": ", may either align to a ancestor of", "type": "text" }, { "bbox": [ 355, 294, 370, 304 ], "score": 0.89, "content": "c _ { 1 , i }", "type": "inline_equation" }, { "bbox": [ 370, 291, 454, 305 ], "score": 1.0, "content": "or be left unaligned.", "type": "text" } ], "index": 15, "is_list_start_line": true, "is_list_end_line": true } ], "index": 13.5, "bbox_fs": [ 105, 250, 474, 305 ] }, { "type": "text", "bbox": [ 105, 313, 487, 326 ], "lines": [ { "bbox": [ 106, 312, 488, 329 ], "spans": [ { "bbox": [ 106, 312, 223, 329 ], "score": 1.0, "content": "Given the optimal alignment", "type": "text" }, { "bbox": [ 223, 312, 232, 324 ], "score": 0.84, "content": "\\hat { A }", "type": "inline_equation" }, { "bbox": [ 232, 312, 434, 329 ], "score": 1.0, "content": ", we calculate the structured average IOU between", "type": "text" }, { "bbox": [ 434, 315, 445, 325 ], "score": 0.88, "content": "\\mathcal { T } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 445, 312, 463, 329 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 463, 315, 474, 325 ], "score": 0.87, "content": "\\mathcal { T } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 474, 312, 488, 329 ], "score": 1.0, "content": "by", "type": "text" } ], "index": 16 } ], "index": 16, "bbox_fs": [ 106, 312, 488, 329 ] }, { "type": "interline_equation", "bbox": [ 182, 328, 428, 368 ], "lines": [ { "bbox": [ 182, 328, 428, 368 ], "spans": [ { "bbox": [ 182, 328, 428, 368 ], "score": 0.95, "content": "\\operatorname { S A I o U } ( \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 2 } ) = \\frac { 2 } { n _ { 1 } + n _ { 2 } } \\left( \\sum _ { i = 1 } ^ { n _ { 1 } } \\sum _ { j = 1 } ^ { n _ { 2 } } \\hat { A } _ { i , j } \\mathrm { I o U } ( c _ { 1 , i } , c _ { 2 , j } ) \\right) .", "type": "interline_equation", "image_path": "44b46d1aab90ab2d5d4ff14b21f09616a5265b08cb4be359e83470e364277370.jpg" } ] } ], "index": 18, "virtual_lines": [ { "bbox": [ 182, 328, 428, 341.3333333333333 ], "spans": [], "index": 17 }, { "bbox": [ 182, 341.3333333333333, 428, 354.66666666666663 ], "spans": [], "index": 18 }, { "bbox": [ 182, 354.66666666666663, 428, 367.99999999999994 ], "spans": [], "index": 19 } ] }, { "type": "title", "bbox": [ 108, 379, 299, 390 ], "lines": [ { "bbox": [ 105, 378, 301, 392 ], "spans": [ { "bbox": [ 105, 378, 301, 392 ], "score": 1.0, "content": "4.4 UNSUPERVISED WORD SEGMENTATION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 399, 505, 443 ], "lines": [ { "bbox": [ 105, 399, 506, 413 ], "spans": [ { "bbox": [ 105, 399, 506, 413 ], "score": 1.0, "content": "We validate our decision of adopting VG-HuBERT to extract word-like units from raw speech wave-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 410, 506, 423 ], "spans": [ { "bbox": [ 105, 410, 506, 423 ], "score": 1.0, "content": "forms for later phrase structure parsing. In particular, we investigate two questions: (1) How does", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 421, 506, 434 ], "spans": [ { "bbox": [ 105, 421, 506, 434 ], "score": 1.0, "content": "segment insertion affect word segmentation performance? (2) how does MBR-based VG-HuBERT", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 433, 294, 444 ], "spans": [ { "bbox": [ 105, 433, 294, 444 ], "score": 1.0, "content": "compare to supervised selected VG-HuBERT?", "type": "text" } ], "index": 24 } ], "index": 22.5, "bbox_fs": [ 105, 399, 506, 444 ] }, { "type": "text", "bbox": [ 106, 449, 505, 592 ], "lines": [ { "bbox": [ 105, 448, 505, 462 ], "spans": [ { "bbox": [ 105, 448, 505, 462 ], "score": 1.0, "content": "In Table 1, in addition to VG-HuBERT, we also list a speech-only word segmentation algorithm", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 461, 504, 472 ], "spans": [ { "bbox": [ 106, 461, 504, 472 ], "score": 1.0, "content": "DPDP (Kamper, 2022). Note that audio-visual model VG-HuBERT significantly outperform DPDP.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 469, 506, 486 ], "spans": [ { "bbox": [ 105, 469, 506, 486 ], "score": 1.0, "content": "For question (1), by comparing the third row and the fourth row, as expected we see that performing", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 482, 505, 496 ], "spans": [ { "bbox": [ 105, 482, 505, 496 ], "score": 1.0, "content": "segment insertion improves recall and hurts precision, and slightly improves F1. For question (2), by", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 493, 505, 506 ], "spans": [ { "bbox": [ 106, 493, 505, 506 ], "score": 1.0, "content": "comparing the fourth row and the fifth row (second to last row), we see that MBR selection actually", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 504, 505, 516 ], "spans": [ { "bbox": [ 106, 504, 505, 516 ], "score": 1.0, "content": "leads to better performance than supervised selection. The final MBR selection we adopted is based", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 516, 504, 527 ], "spans": [ { "bbox": [ 106, 516, 504, 527 ], "score": 1.0, "content": "on the last row, where we first performed MBR selection on SpokenCOCO val set on all 405 candi-", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 526, 505, 539 ], "spans": [ { "bbox": [ 106, 526, 505, 539 ], "score": 1.0, "content": "dates, and subsequently chose the 10 most selected combinations to perform another round of MBR", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 536, 506, 550 ], "spans": [ { "bbox": [ 105, 536, 506, 550 ], "score": 1.0, "content": "decoding. Getting the top 10 most selected combinations does not require knowing the performance", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 549, 505, 560 ], "spans": [ { "bbox": [ 106, 549, 505, 560 ], "score": 1.0, "content": "on segmentation, and therefore this process is still completely unsupervised. The reason for doing 2", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 559, 505, 571 ], "spans": [ { "bbox": [ 106, 559, 505, 571 ], "score": 1.0, "content": "iterations of MBR is because performing MBR on 405 candidates on SpokenCOCO training set is", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 569, 506, 583 ], "spans": [ { "bbox": [ 105, 569, 506, 583 ], "score": 1.0, "content": "estimated to take 2 months, and MBR on 10 candidates can be done in 5 days. Comparing the last", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 581, 426, 593 ], "spans": [ { "bbox": [ 106, 581, 426, 593 ], "score": 1.0, "content": "two rows, we observe that two iterations of MBR does not lead to worse results.", "type": "text" } ], "index": 37 } ], "index": 31, "bbox_fs": [ 105, 448, 506, 593 ] }, { "type": "title", "bbox": [ 107, 605, 342, 616 ], "lines": [ { "bbox": [ 105, 604, 343, 618 ], "spans": [ { "bbox": [ 105, 604, 343, 618 ], "score": 1.0, "content": "4.5 UNSUPERVISED PHRASE STRUCTURE INDUCTION", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 625, 505, 714 ], "lines": [ { "bbox": [ 106, 625, 505, 639 ], "spans": [ { "bbox": [ 106, 625, 505, 639 ], "score": 1.0, "content": "We quantitatively show that AV-NSL learns meaningful phrase structure given word segments.", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 637, 505, 649 ], "spans": [ { "bbox": [ 106, 637, 505, 649 ], "score": 1.0, "content": "First, Table 2 is the main result of the fully-unsupervised AV-NSL on SpokenCOCO, evaluated", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 648, 506, 661 ], "spans": [ { "bbox": [ 105, 648, 506, 661 ], "score": 1.0, "content": "with SAIOU. The best performing AV-NSL is based on our improved VG-HuBERT with MBR", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 659, 506, 672 ], "spans": [ { "bbox": [ 105, 659, 506, 672 ], "score": 1.0, "content": "top 10 selection for word segmentation, attention-weighted mean-pool over VG-HuBERT layers as", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 669, 505, 683 ], "spans": [ { "bbox": [ 105, 669, 505, 683 ], "score": 1.0, "content": "the segment representations, and another MBR decoding over all phrase structure induction hyper-", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 680, 506, 694 ], "spans": [ { "bbox": [ 105, 680, 506, 694 ], "score": 1.0, "content": "parameters. Comparing AV-NSL against AV-cPCFG and AV-cPCFG against DPDP-cPCFG, we", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 692, 506, 704 ], "spans": [ { "bbox": [ 105, 692, 506, 704 ], "score": 1.0, "content": "empirically show the necessity of training AV-NSL on continuous segment representation instead of", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 702, 494, 716 ], "spans": [ { "bbox": [ 105, 702, 494, 716 ], "score": 1.0, "content": "discretized speech tokens, and the effectiveness of visual-grounding in our overall model design.", "type": "text" } ], "index": 46 } ], "index": 42.5, "bbox_fs": [ 105, 625, 506, 716 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 118, 80, 491, 151 ], "blocks": [ { "type": "table_body", "bbox": [ 118, 80, 491, 151 ], "group_id": 0, "lines": [ { "bbox": [ 118, 80, 491, 151 ], "spans": [ { "bbox": [ 118, 80, 491, 151 ], "score": 0.978, "html": "
MethodInsertionOut. Sel.#Sel. Cand.PrecisionRecallF1
DPDP (Kamper,2022)supervised17.379.0011.85
VG-HuBERT (Peng & Harwath,2022b)supervised36.1927.2231.07
Improved VG-HuBERT (Ours)supervised34.3429.8531.94
MBR40533.8334.3734.10
√MBR (2iter)405→1033.3134.9034.09
", "type": "table", "image_path": "d1a014a1d922b1e22175a66b65ca0a9aae6bfe302357e5514cc4322c4dae2407.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 118, 80, 491, 103.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 118, 103.66666666666667, 491, 127.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 118, 127.33333333333334, 491, 151.0 ], "spans": [], "index": 2 } ] }, { "type": "table_caption", "bbox": [ 107, 155, 505, 200 ], "group_id": 0, "lines": [ { "bbox": [ 105, 154, 506, 168 ], "spans": [ { "bbox": [ 105, 154, 506, 168 ], "score": 1.0, "content": "Table 1: Word Segmentation Performance on SpokenCOCO validation set. Out. Sel. denotes output", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 166, 505, 178 ], "spans": [ { "bbox": [ 106, 166, 505, 178 ], "score": 1.0, "content": "selection methods, and #Sel. Cand. denotes the number of candidate models to be selected. MBR", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 177, 504, 190 ], "spans": [ { "bbox": [ 106, 177, 470, 190 ], "score": 1.0, "content": "(2iter) means we first run MBR on all 405 candidates, and then run MBR again on the", "type": "text" }, { "bbox": [ 470, 177, 504, 188 ], "score": 0.36, "content": "1 0 \\ \\mathrm { m o s t }", "type": "inline_equation" } ], "index": 5 }, { "bbox": [ 105, 188, 475, 201 ], "spans": [ { "bbox": [ 105, 188, 459, 201 ], "score": 1.0, "content": "selected candidates. Our improved VG-HuBERT with MBR achieves the best boundary", "type": "text" }, { "bbox": [ 459, 189, 470, 200 ], "score": 0.87, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 471, 188, 475, 201 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "table", "bbox": [ 121, 218, 487, 332 ], "blocks": [ { "type": "table_body", "bbox": [ 121, 218, 487, 332 ], "group_id": 1, "lines": [ { "bbox": [ 121, 218, 487, 332 ], "spans": [ { "bbox": [ 121, 218, 487, 332 ], "score": 0.982, "html": "
ModelOutput SelectionSAIoU
Syntax InductionSegmentationSeg.Representation (continuous/discrete)
Right-BranchingVG-HuBERT+MBR100.546
Right-BranchingDPDP0.478
AV-NSLVG-HuBERT+MBR10VG-HuBERT1o (continuous)MBR0.516
AV-NSLVG-HuBERT+MBR10VG-HuBERT10,11,12 (continuous)MBR0.521
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+4k km (discrete)last ckpt.0.499
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+8k km (discrete)last ckpt.0.481
DPDP-cPCFGDPDPHuBERT2+2k km (discrete)last ckpt.0.465
DPDP-cPCFGDPDPHuBERT1o+2k km (discrete)last ckpt.0.426
", "type": "table", "image_path": "ed3fe25b38fe86bfed07938db0d37e7c07a8fa0c1eaa7fe377733ffb637f40d5.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 121, 218, 487, 256.0 ], "spans": [], "index": 7 }, { "bbox": [ 121, 256.0, 487, 294.0 ], "spans": [], "index": 8 }, { "bbox": [ 121, 294.0, 487, 332.0 ], "spans": [], "index": 9 } ] }, { "type": "table_caption", "bbox": [ 108, 339, 504, 362 ], "group_id": 1, "lines": [ { "bbox": [ 107, 339, 505, 351 ], "spans": [ { "bbox": [ 107, 339, 505, 351 ], "score": 1.0, "content": "Table 2: Fully-unsupervised phrase structure induction results on SpokenCOCO. The best overall", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 349, 502, 362 ], "spans": [ { "bbox": [ 106, 349, 502, 362 ], "score": 1.0, "content": "number and the best number produced by neural models are in boldface. Full table in Appendix 7.", "type": "text" } ], "index": 11 } ], "index": 10.5 } ], "index": 9.25 }, { "type": "table", "bbox": [ 114, 380, 494, 442 ], "blocks": [ { "type": "table_body", "bbox": [ 114, 380, 494, 442 ], "group_id": 2, "lines": [ { "bbox": [ 114, 380, 494, 442 ], "spans": [ { "bbox": [ 114, 380, 494, 442 ], "score": 0.975, "html": "
ModelSegmentationSeg. Representationtree targetOutput SelectionSAIoU
trainvaltest
s-BeneparVG-HuBERT+MBR10HuBERT2AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT6AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT2,4,6.8,10,12AV-NSLAV-NSLoracleMBR0.536
", "type": "table", "image_path": "4eb398d91fd631b2dd846d3a1856a6bcea7a37bddddaec82c4b9d009c5997b9e.jpg" } ] } ], "index": 13, "virtual_lines": [ { "bbox": [ 114, 380, 494, 400.6666666666667 ], "spans": [], "index": 12 }, { "bbox": [ 114, 400.6666666666667, 494, 421.33333333333337 ], "spans": [], "index": 13 }, { "bbox": [ 114, 421.33333333333337, 494, 442.00000000000006 ], "spans": [], "index": 14 } ] }, { "type": "table_footnote", "bbox": [ 107, 450, 505, 473 ], "group_id": 2, "lines": [ { "bbox": [ 105, 448, 505, 462 ], "spans": [ { "bbox": [ 105, 448, 505, 462 ], "score": 1.0, "content": "Table 3: Single round self-training in Section 3.3 improves the best AV-NSL from Table 2. We train", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 460, 438, 473 ], "spans": [ { "bbox": [ 105, 460, 438, 473 ], "score": 1.0, "content": "s-Benepar on the trees from fully-unsupervised AV-NSL. Full table in Appendix 8.", "type": "text" } ], "index": 16 } ], "index": 15.5 } ], "index": 14.25 }, { "type": "text", "bbox": [ 107, 501, 505, 633 ], "lines": [ { "bbox": [ 105, 500, 505, 514 ], "spans": [ { "bbox": [ 105, 500, 505, 514 ], "score": 1.0, "content": "Secondly, Table 3 shows that our proposed self-training with s-Benepar complements AV-NSL.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 512, 505, 525 ], "spans": [ { "bbox": [ 106, 512, 505, 525 ], "score": 1.0, "content": "Generally, a single round of self-training improves the SAIOU, and our best s-Benepar improves", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 523, 505, 536 ], "spans": [ { "bbox": [ 105, 523, 505, 536 ], "score": 1.0, "content": "the best AV-NSL from 0.521 to 0.538. Thirdly, Table 4 isolates phrase structure induction from", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 534, 505, 547 ], "spans": [ { "bbox": [ 105, 534, 505, 547 ], "score": 1.0, "content": "word segmentation quality with oracle AV-NSL. Different from Table 2, since there is no mismatch", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 545, 505, 558 ], "spans": [ { "bbox": [ 105, 545, 273, 558 ], "score": 1.0, "content": "in the number of tree nodes, we can adopt", "type": "text" }, { "bbox": [ 273, 545, 285, 556 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 285, 545, 505, 558 ], "score": 1.0, "content": "evaluation. With proper segment-level representations,", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 556, 505, 568 ], "spans": [ { "bbox": [ 105, 556, 505, 568 ], "score": 1.0, "content": "unsupervised oracle AV-NSL matches or out-performs text-based VG-NSL. Similar to Tabel 3, self-", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 568, 505, 579 ], "spans": [ { "bbox": [ 106, 568, 505, 579 ], "score": 1.0, "content": "training with s-Benepar on oracle AV-NSL trees further improves the syntax induction results, almost", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 577, 505, 591 ], "spans": [ { "bbox": [ 105, 577, 505, 591 ], "score": 1.0, "content": "matching that of right-branching tree. Last but not least, perhaps surprisingly, right-branching trees", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 370, 602 ], "score": 1.0, "content": "(RBT) on the given word segmentation reach the best SAIOU and", "type": "text" }, { "bbox": [ 370, 590, 383, 600 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 383, 588, 505, 602 ], "score": 1.0, "content": "scores. We note that the right-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "branching approach highly aligns with the head-initial property of English (Baker, 2001), especially", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 611, 505, 623 ], "spans": [ { "bbox": [ 105, 611, 505, 623 ], "score": 1.0, "content": "in our setting where all punctuation marks were removed; thus, it is nontrivial for AV-NSL to reach", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 622, 482, 635 ], "spans": [ { "bbox": [ 106, 622, 482, 635 ], "score": 1.0, "content": "the performance on par with RBT without inductive biases favoring any specific type of trees.", "type": "text" } ], "index": 28 } ], "index": 22.5 }, { "type": "title", "bbox": [ 107, 658, 181, 670 ], "lines": [ { "bbox": [ 105, 656, 183, 672 ], "spans": [ { "bbox": [ 105, 656, 183, 672 ], "score": 1.0, "content": "5 ANALYSES", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 107, 687, 504, 732 ], "lines": [ { "bbox": [ 105, 686, 506, 701 ], "spans": [ { "bbox": [ 105, 686, 506, 701 ], "score": 1.0, "content": "Unsupervised Constituent Recall: Following Shi et al. (2019), we show the recall of specific types", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 699, 506, 711 ], "spans": [ { "bbox": [ 106, 699, 506, 711 ], "score": 1.0, "content": "of constituents (Table 5). While VG-NSL benefits from the head-initial (HI) bias, where abstract", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 710, 506, 722 ], "spans": [ { "bbox": [ 105, 710, 506, 722 ], "score": 1.0, "content": "words are encouraged to appear in the beginning of a constituent, it is worth noting that AV-NSL", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 721, 506, 734 ], "spans": [ { "bbox": [ 105, 721, 506, 734 ], "score": 1.0, "content": "outperforms all variations of VG-NSL, without inductive biases favoring any specific types of trees.", "type": "text" } ], "index": 33 } ], "index": 31.5 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 759 ], "lines": [ { "bbox": [ 302, 750, 309, 761 ], "spans": [ { "bbox": [ 302, 750, 309, 761 ], "score": 1.0, "content": "8", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 118, 80, 491, 151 ], "blocks": [ { "type": "table_body", "bbox": [ 118, 80, 491, 151 ], "group_id": 0, "lines": [ { "bbox": [ 118, 80, 491, 151 ], "spans": [ { "bbox": [ 118, 80, 491, 151 ], "score": 0.978, "html": "
MethodInsertionOut. Sel.#Sel. Cand.PrecisionRecallF1
DPDP (Kamper,2022)supervised17.379.0011.85
VG-HuBERT (Peng & Harwath,2022b)supervised36.1927.2231.07
Improved VG-HuBERT (Ours)supervised34.3429.8531.94
MBR40533.8334.3734.10
√MBR (2iter)405→1033.3134.9034.09
", "type": "table", "image_path": "d1a014a1d922b1e22175a66b65ca0a9aae6bfe302357e5514cc4322c4dae2407.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 118, 80, 491, 103.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 118, 103.66666666666667, 491, 127.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 118, 127.33333333333334, 491, 151.0 ], "spans": [], "index": 2 } ] }, { "type": "table_caption", "bbox": [ 107, 155, 505, 200 ], "group_id": 0, "lines": [ { "bbox": [ 105, 154, 506, 168 ], "spans": [ { "bbox": [ 105, 154, 506, 168 ], "score": 1.0, "content": "Table 1: Word Segmentation Performance on SpokenCOCO validation set. Out. Sel. denotes output", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 166, 505, 178 ], "spans": [ { "bbox": [ 106, 166, 505, 178 ], "score": 1.0, "content": "selection methods, and #Sel. Cand. denotes the number of candidate models to be selected. MBR", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 177, 504, 190 ], "spans": [ { "bbox": [ 106, 177, 470, 190 ], "score": 1.0, "content": "(2iter) means we first run MBR on all 405 candidates, and then run MBR again on the", "type": "text" }, { "bbox": [ 470, 177, 504, 188 ], "score": 0.36, "content": "1 0 \\ \\mathrm { m o s t }", "type": "inline_equation" } ], "index": 5 }, { "bbox": [ 105, 188, 475, 201 ], "spans": [ { "bbox": [ 105, 188, 459, 201 ], "score": 1.0, "content": "selected candidates. Our improved VG-HuBERT with MBR achieves the best boundary", "type": "text" }, { "bbox": [ 459, 189, 470, 200 ], "score": 0.87, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 471, 188, 475, 201 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "table", "bbox": [ 121, 218, 487, 332 ], "blocks": [ { "type": "table_body", "bbox": [ 121, 218, 487, 332 ], "group_id": 1, "lines": [ { "bbox": [ 121, 218, 487, 332 ], "spans": [ { "bbox": [ 121, 218, 487, 332 ], "score": 0.982, "html": "
ModelOutput SelectionSAIoU
Syntax InductionSegmentationSeg.Representation (continuous/discrete)
Right-BranchingVG-HuBERT+MBR100.546
Right-BranchingDPDP0.478
AV-NSLVG-HuBERT+MBR10VG-HuBERT1o (continuous)MBR0.516
AV-NSLVG-HuBERT+MBR10VG-HuBERT10,11,12 (continuous)MBR0.521
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+4k km (discrete)last ckpt.0.499
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+8k km (discrete)last ckpt.0.481
DPDP-cPCFGDPDPHuBERT2+2k km (discrete)last ckpt.0.465
DPDP-cPCFGDPDPHuBERT1o+2k km (discrete)last ckpt.0.426
", "type": "table", "image_path": "ed3fe25b38fe86bfed07938db0d37e7c07a8fa0c1eaa7fe377733ffb637f40d5.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 121, 218, 487, 256.0 ], "spans": [], "index": 7 }, { "bbox": [ 121, 256.0, 487, 294.0 ], "spans": [], "index": 8 }, { "bbox": [ 121, 294.0, 487, 332.0 ], "spans": [], "index": 9 } ] }, { "type": "table_caption", "bbox": [ 108, 339, 504, 362 ], "group_id": 1, "lines": [ { "bbox": [ 107, 339, 505, 351 ], "spans": [ { "bbox": [ 107, 339, 505, 351 ], "score": 1.0, "content": "Table 2: Fully-unsupervised phrase structure induction results on SpokenCOCO. The best overall", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 349, 502, 362 ], "spans": [ { "bbox": [ 106, 349, 502, 362 ], "score": 1.0, "content": "number and the best number produced by neural models are in boldface. Full table in Appendix 7.", "type": "text" } ], "index": 11 } ], "index": 10.5 } ], "index": 9.25 }, { "type": "table", "bbox": [ 114, 380, 494, 442 ], "blocks": [ { "type": "table_body", "bbox": [ 114, 380, 494, 442 ], "group_id": 2, "lines": [ { "bbox": [ 114, 380, 494, 442 ], "spans": [ { "bbox": [ 114, 380, 494, 442 ], "score": 0.975, "html": "
ModelSegmentationSeg. Representationtree targetOutput SelectionSAIoU
trainvaltest
s-BeneparVG-HuBERT+MBR10HuBERT2AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT6AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT2,4,6.8,10,12AV-NSLAV-NSLoracleMBR0.536
", "type": "table", "image_path": "4eb398d91fd631b2dd846d3a1856a6bcea7a37bddddaec82c4b9d009c5997b9e.jpg" } ] } ], "index": 13, "virtual_lines": [ { "bbox": [ 114, 380, 494, 400.6666666666667 ], "spans": [], "index": 12 }, { "bbox": [ 114, 400.6666666666667, 494, 421.33333333333337 ], "spans": [], "index": 13 }, { "bbox": [ 114, 421.33333333333337, 494, 442.00000000000006 ], "spans": [], "index": 14 } ] }, { "type": "table_footnote", "bbox": [ 107, 450, 505, 473 ], "group_id": 2, "lines": [ { "bbox": [ 105, 448, 505, 462 ], "spans": [ { "bbox": [ 105, 448, 505, 462 ], "score": 1.0, "content": "Table 3: Single round self-training in Section 3.3 improves the best AV-NSL from Table 2. We train", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 460, 438, 473 ], "spans": [ { "bbox": [ 105, 460, 438, 473 ], "score": 1.0, "content": "s-Benepar on the trees from fully-unsupervised AV-NSL. Full table in Appendix 8.", "type": "text" } ], "index": 16 } ], "index": 15.5 } ], "index": 14.25 }, { "type": "text", "bbox": [ 107, 501, 505, 633 ], "lines": [ { "bbox": [ 105, 500, 505, 514 ], "spans": [ { "bbox": [ 105, 500, 505, 514 ], "score": 1.0, "content": "Secondly, Table 3 shows that our proposed self-training with s-Benepar complements AV-NSL.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 512, 505, 525 ], "spans": [ { "bbox": [ 106, 512, 505, 525 ], "score": 1.0, "content": "Generally, a single round of self-training improves the SAIOU, and our best s-Benepar improves", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 523, 505, 536 ], "spans": [ { "bbox": [ 105, 523, 505, 536 ], "score": 1.0, "content": "the best AV-NSL from 0.521 to 0.538. Thirdly, Table 4 isolates phrase structure induction from", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 534, 505, 547 ], "spans": [ { "bbox": [ 105, 534, 505, 547 ], "score": 1.0, "content": "word segmentation quality with oracle AV-NSL. Different from Table 2, since there is no mismatch", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 545, 505, 558 ], "spans": [ { "bbox": [ 105, 545, 273, 558 ], "score": 1.0, "content": "in the number of tree nodes, we can adopt", "type": "text" }, { "bbox": [ 273, 545, 285, 556 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 285, 545, 505, 558 ], "score": 1.0, "content": "evaluation. With proper segment-level representations,", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 556, 505, 568 ], "spans": [ { "bbox": [ 105, 556, 505, 568 ], "score": 1.0, "content": "unsupervised oracle AV-NSL matches or out-performs text-based VG-NSL. Similar to Tabel 3, self-", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 568, 505, 579 ], "spans": [ { "bbox": [ 106, 568, 505, 579 ], "score": 1.0, "content": "training with s-Benepar on oracle AV-NSL trees further improves the syntax induction results, almost", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 577, 505, 591 ], "spans": [ { "bbox": [ 105, 577, 505, 591 ], "score": 1.0, "content": "matching that of right-branching tree. Last but not least, perhaps surprisingly, right-branching trees", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 370, 602 ], "score": 1.0, "content": "(RBT) on the given word segmentation reach the best SAIOU and", "type": "text" }, { "bbox": [ 370, 590, 383, 600 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 383, 588, 505, 602 ], "score": 1.0, "content": "scores. We note that the right-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 599, 505, 613 ], "spans": [ { "bbox": [ 105, 599, 505, 613 ], "score": 1.0, "content": "branching approach highly aligns with the head-initial property of English (Baker, 2001), especially", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 611, 505, 623 ], "spans": [ { "bbox": [ 105, 611, 505, 623 ], "score": 1.0, "content": "in our setting where all punctuation marks were removed; thus, it is nontrivial for AV-NSL to reach", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 622, 482, 635 ], "spans": [ { "bbox": [ 106, 622, 482, 635 ], "score": 1.0, "content": "the performance on par with RBT without inductive biases favoring any specific type of trees.", "type": "text" } ], "index": 28 } ], "index": 22.5, "bbox_fs": [ 105, 500, 505, 635 ] }, { "type": "title", "bbox": [ 107, 658, 181, 670 ], "lines": [ { "bbox": [ 105, 656, 183, 672 ], "spans": [ { "bbox": [ 105, 656, 183, 672 ], "score": 1.0, "content": "5 ANALYSES", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 107, 687, 504, 732 ], "lines": [ { "bbox": [ 105, 686, 506, 701 ], "spans": [ { "bbox": [ 105, 686, 506, 701 ], "score": 1.0, "content": "Unsupervised Constituent Recall: Following Shi et al. (2019), we show the recall of specific types", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 699, 506, 711 ], "spans": [ { "bbox": [ 106, 699, 506, 711 ], "score": 1.0, "content": "of constituents (Table 5). While VG-NSL benefits from the head-initial (HI) bias, where abstract", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 710, 506, 722 ], "spans": [ { "bbox": [ 105, 710, 506, 722 ], "score": 1.0, "content": "words are encouraged to appear in the beginning of a constituent, it is worth noting that AV-NSL", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 721, 506, 734 ], "spans": [ { "bbox": [ 105, 721, 506, 734 ], "score": 1.0, "content": "outperforms all variations of VG-NSL, without inductive biases favoring any specific types of trees.", "type": "text" } ], "index": 33 } ], "index": 31.5, "bbox_fs": [ 105, 686, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 175, 80, 434, 214 ], "blocks": [ { "type": "table_body", "bbox": [ 175, 80, 434, 214 ], "group_id": 0, "lines": [ { "bbox": [ 175, 80, 434, 214 ], "spans": [ { "bbox": [ 175, 80, 434, 214 ], "score": 0.98, "html": "
ModelOutput SelectionF1
Syntax InductionSeg.Representation
Random32.77
Left-Branching24.56
Right-Branching VG-NSLSupervised57.39 53.11
word embeddings
oracle AV-NSLlog-Mel spectrogramSupervised42.01
oracle AV-NSLHuBERT2Supervised55.51
oracle AV-NSLHuBERT2MBR54.99
oracle AV-NSLHuBERT2,4,6,8,10,12,24MBR55.96
oracle AV-NSL →s-BeneparHuBERT2MBR57.24
oracle AV-NSL →s-BeneparHuBERT12MBR57.33
", "type": "table", "image_path": "2900bc078940adfa7559dc5dc2c462de7b955f3c7bbd057c4263e7022455d359.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 175, 80, 434, 124.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 175, 124.66666666666666, 434, 169.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 175, 169.33333333333331, 434, 213.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "table_caption", "bbox": [ 113, 222, 492, 234 ], "group_id": 0, "lines": [ { "bbox": [ 118, 221, 492, 235 ], "spans": [ { "bbox": [ 118, 221, 492, 235 ], "score": 1.0, "content": "Table 4: Phrase structure induction with oracle segmentation given. Full table in Appendix 9.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "text", "bbox": [ 107, 257, 503, 280 ], "lines": [ { "bbox": [ 105, 256, 505, 272 ], "spans": [ { "bbox": [ 105, 256, 505, 272 ], "score": 1.0, "content": "Ablation Study: We present two ablations to examine the effectiveness of high-quality word seg-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 269, 505, 282 ], "spans": [ { "bbox": [ 105, 269, 505, 282 ], "score": 1.0, "content": "mentation and visual representation (Table 6). We train AV-NSL with the following modifications:", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "text", "bbox": [ 107, 296, 505, 345 ], "lines": [ { "bbox": [ 106, 297, 504, 309 ], "spans": [ { "bbox": [ 106, 297, 504, 309 ], "score": 1.0, "content": "1. Fix the visual representations, but replace oracle segmentation with naive uniform word segmen-", "type": "text" } ], "index": 6 }, { "bbox": [ 118, 307, 435, 320 ], "spans": [ { "bbox": [ 118, 307, 435, 320 ], "score": 1.0, "content": "tation, where the number of words in each caption is given (uniform AV-NSL).", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 321, 505, 335 ], "spans": [ { "bbox": [ 105, 321, 505, 335 ], "score": 1.0, "content": "2. Fix the oracle word segmentation, but replace visual embeddings with random images, where", "type": "text" } ], "index": 8 }, { "bbox": [ 119, 334, 379, 346 ], "spans": [ { "bbox": [ 119, 334, 379, 346 ], "score": 1.0, "content": "each pixel is independently sampled from a uniform distribution.", "type": "text" } ], "index": 9 } ], "index": 7.5 }, { "type": "text", "bbox": [ 107, 353, 505, 408 ], "lines": [ { "bbox": [ 106, 352, 505, 366 ], "spans": [ { "bbox": [ 106, 352, 505, 366 ], "score": 1.0, "content": "We observe that there are significant performance drops in both settings, comparing to the AV-NSL", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 363, 505, 377 ], "spans": [ { "bbox": [ 105, 363, 505, 377 ], "score": 1.0, "content": "trained with oracle segmentation and high-quality visual representation. This set of results comple-", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 375, 505, 387 ], "spans": [ { "bbox": [ 105, 375, 505, 387 ], "score": 1.0, "content": "ment Table 2, stressing that precise word segmentation and high-quality visual representations are", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 385, 505, 399 ], "spans": [ { "bbox": [ 105, 385, 505, 399 ], "score": 1.0, "content": "both necessary for phrase structure induction from speech. Furthermore, we provide tree structure", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 397, 421, 410 ], "spans": [ { "bbox": [ 105, 397, 421, 410 ], "score": 1.0, "content": "and word segmentation visualizations for qualitative analysis in the Appendix.", "type": "text" } ], "index": 14 } ], "index": 12 }, { "type": "text", "bbox": [ 318, 431, 501, 520 ], "lines": [ { "bbox": [ 317, 431, 502, 443 ], "spans": [ { "bbox": [ 317, 431, 502, 443 ], "score": 1.0, "content": "Table 6: Top rows: performance of AV-NSL", "type": "text" } ], "index": 16 }, { "bbox": [ 317, 442, 502, 455 ], "spans": [ { "bbox": [ 317, 442, 502, 455 ], "score": 1.0, "content": "with word segmentation in various quality", "type": "text" } ], "index": 18 }, { "bbox": [ 317, 453, 503, 465 ], "spans": [ { "bbox": [ 317, 453, 503, 465 ], "score": 1.0, "content": "and high-quality visual embeddings. Bot-", "type": "text" } ], "index": 20 }, { "bbox": [ 316, 465, 503, 475 ], "spans": [ { "bbox": [ 316, 465, 503, 475 ], "score": 1.0, "content": "tom rows: performance of AV-NSL with vi-", "type": "text" } ], "index": 22 }, { "bbox": [ 317, 475, 502, 487 ], "spans": [ { "bbox": [ 317, 475, 502, 487 ], "score": 1.0, "content": "sual embeddings in various quality and high-", "type": "text" } ], "index": 24 }, { "bbox": [ 317, 486, 503, 497 ], "spans": [ { "bbox": [ 317, 486, 503, 497 ], "score": 1.0, "content": "quality word segmentation. DINO: a self-", "type": "text" } ], "index": 26 }, { "bbox": [ 317, 496, 503, 509 ], "spans": [ { "bbox": [ 317, 496, 503, 509 ], "score": 1.0, "content": "supervised model that produces high-quality", "type": "text" } ], "index": 29 }, { "bbox": [ 317, 508, 489, 520 ], "spans": [ { "bbox": [ 317, 508, 489, 520 ], "score": 1.0, "content": "visual representations (Caron et al., 2021).", "type": "text" } ], "index": 30 } ], "index": 23.0 }, { "type": "table", "bbox": [ 108, 524, 311, 590 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 430, 313, 519 ], "group_id": 1, "lines": [ { "bbox": [ 105, 430, 315, 443 ], "spans": [ { "bbox": [ 105, 430, 315, 443 ], "score": 1.0, "content": "Table 5: Recall of specific typed phrases, includ-", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 442, 315, 454 ], "spans": [ { "bbox": [ 106, 442, 315, 454 ], "score": 1.0, "content": "ing noun phrases (NP), verb phrases (VP), preposi-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 453, 315, 465 ], "spans": [ { "bbox": [ 106, 453, 315, 465 ], "score": 1.0, "content": "tional phrases (PP) and adjective phrases (ADJP),", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 464, 315, 476 ], "spans": [ { "bbox": [ 105, 464, 158, 476 ], "score": 1.0, "content": "and overall", "type": "text" }, { "bbox": [ 159, 465, 171, 475 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 171, 464, 315, 476 ], "score": 1.0, "content": "score, evaluated on the Spoken-", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 475, 315, 486 ], "spans": [ { "bbox": [ 106, 475, 315, 486 ], "score": 1.0, "content": "COCO test split. The VG-NSL numbers are taken", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 486, 315, 497 ], "spans": [ { "bbox": [ 106, 486, 315, 497 ], "score": 1.0, "content": "from (Shi et al., 2019). AV-NSL here are trained", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 497, 316, 508 ], "spans": [ { "bbox": [ 105, 497, 316, 508 ], "score": 1.0, "content": "on oracle segmentation with vanilla HuBERT as the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 509, 195, 520 ], "spans": [ { "bbox": [ 105, 509, 195, 520 ], "score": 1.0, "content": "layer representations.", "type": "text" } ], "index": 28 } ], "index": 22.0 }, { "type": "table_body", "bbox": [ 108, 524, 311, 590 ], "group_id": 1, "lines": [ { "bbox": [ 108, 524, 311, 590 ], "spans": [ { "bbox": [ 108, 524, 311, 590 ], "score": 0.981, "html": "
ModelF1Constituent Recall
NPVPPPADJP
VG-NSL (Shi et al.,2019)50.479.626.242.022.0
VG-NSL + HI53.374.632.566.521.7
VG-NSL + HI+ FastText54.478.824.465.622.0
oracle AV-NSL55.655.568.166.622.1
", "type": "table", "image_path": "38d137326beea01d78ca72e0e3591ad3cdcf2650a0372d5e5d022ca0e875d59d.jpg" } ] } ], "index": 35, "virtual_lines": [ { "bbox": [ 108, 524, 311, 537.2 ], "spans": [], "index": 31 }, { "bbox": [ 108, 537.2, 311, 550.4000000000001 ], "spans": [], "index": 33 }, { "bbox": [ 108, 550.4000000000001, 311, 563.6000000000001 ], "spans": [], "index": 35 }, { "bbox": [ 108, 563.6000000000001, 311, 576.8000000000002 ], "spans": [], "index": 37 }, { "bbox": [ 108, 576.8000000000002, 311, 590.0000000000002 ], "spans": [], "index": 39 } ] } ], "index": 28.5 }, { "type": "table", "bbox": [ 323, 525, 496, 591 ], "blocks": [ { "type": "table_body", "bbox": [ 323, 525, 496, 591 ], "group_id": 2, "lines": [ { "bbox": [ 323, 525, 496, 591 ], "spans": [ { "bbox": [ 323, 525, 496, 591 ], "score": 0.976, "html": "
ModelVisualF1
Syntax InductionSeg.Repre.
oracle AV-NSLHuBERT10ResNet10150.50
uniform AV-NSLHuBERT10ResNet10136.62
oracle AV-NSLHuBERT255.71
DINO
oracle AV-NSLHuBERT2random31.23
", "type": "table", "image_path": "02dfd66c2176ae4821ac6dbb8aebf51317ed8a95a7f0d2294aa6dae357fad75d.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 323, 525, 496, 538.2 ], "spans": [], "index": 32 }, { "bbox": [ 323, 538.2, 496, 551.4000000000001 ], "spans": [], "index": 34 }, { "bbox": [ 323, 551.4000000000001, 496, 564.6000000000001 ], "spans": [], "index": 36 }, { "bbox": [ 323, 564.6000000000001, 496, 577.8000000000002 ], "spans": [], "index": 38 }, { "bbox": [ 323, 577.8000000000002, 496, 591.0000000000002 ], "spans": [], "index": 40 } ] } ], "index": 36 }, { "type": "title", "bbox": [ 107, 612, 196, 624 ], "lines": [ { "bbox": [ 105, 609, 198, 628 ], "spans": [ { "bbox": [ 105, 609, 198, 628 ], "score": 1.0, "content": "6 CONCLUSION", "type": "text" } ], "index": 41 } ], "index": 41 }, { "type": "text", "bbox": [ 106, 632, 505, 732 ], "lines": [ { "bbox": [ 106, 633, 504, 645 ], "spans": [ { "bbox": [ 106, 633, 504, 645 ], "score": 1.0, "content": "In recent years, there have been fruitful progresses in multi-modal induction for zero-resource speech", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "processing and grammar induction respectively. The idea of leveraging the visual modality to learn", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 655, 506, 668 ], "spans": [ { "bbox": [ 105, 655, 506, 668 ], "score": 1.0, "content": "language competence, either lexicon units from speech or syntactic structure from text, is an at-", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 666, 506, 679 ], "spans": [ { "bbox": [ 106, 666, 506, 679 ], "score": 1.0, "content": "tractive approach for modeling human language acquisition. Our study contributes to both lines of", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 676, 505, 690 ], "spans": [ { "bbox": [ 105, 676, 505, 690 ], "score": 1.0, "content": "research, by presenting an unifying framework that learns phrase structure from visually-grounded", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 687, 506, 702 ], "spans": [ { "bbox": [ 105, 687, 506, 702 ], "score": 1.0, "content": "speech, without any text. We show that our proposed model, AV-NSL, infers meaningful con-", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 698, 506, 713 ], "spans": [ { "bbox": [ 105, 698, 506, 713 ], "score": 1.0, "content": "stituency parse trees on top of continuous word segment representations, both quantitatively and", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 710, 506, 722 ], "spans": [ { "bbox": [ 105, 710, 506, 722 ], "score": 1.0, "content": "qualitatively. To justify our modeling design choices, we construct several baselines and introduce a", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 720, 504, 734 ], "spans": [ { "bbox": [ 105, 720, 504, 734 ], "score": 1.0, "content": "novel evaluation metric. We envision our research as the first of many in textless structure learning.", "type": "text" } ], "index": 50 } ], "index": 46 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 175, 80, 434, 214 ], "blocks": [ { "type": "table_body", "bbox": [ 175, 80, 434, 214 ], "group_id": 0, "lines": [ { "bbox": [ 175, 80, 434, 214 ], "spans": [ { "bbox": [ 175, 80, 434, 214 ], "score": 0.98, "html": "
ModelOutput SelectionF1
Syntax InductionSeg.Representation
Random32.77
Left-Branching24.56
Right-Branching VG-NSLSupervised57.39 53.11
word embeddings
oracle AV-NSLlog-Mel spectrogramSupervised42.01
oracle AV-NSLHuBERT2Supervised55.51
oracle AV-NSLHuBERT2MBR54.99
oracle AV-NSLHuBERT2,4,6,8,10,12,24MBR55.96
oracle AV-NSL →s-BeneparHuBERT2MBR57.24
oracle AV-NSL →s-BeneparHuBERT12MBR57.33
", "type": "table", "image_path": "2900bc078940adfa7559dc5dc2c462de7b955f3c7bbd057c4263e7022455d359.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 175, 80, 434, 124.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 175, 124.66666666666666, 434, 169.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 175, 169.33333333333331, 434, 213.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "table_caption", "bbox": [ 113, 222, 492, 234 ], "group_id": 0, "lines": [ { "bbox": [ 118, 221, 492, 235 ], "spans": [ { "bbox": [ 118, 221, 492, 235 ], "score": 1.0, "content": "Table 4: Phrase structure induction with oracle segmentation given. Full table in Appendix 9.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "text", "bbox": [ 107, 257, 503, 280 ], "lines": [ { "bbox": [ 105, 256, 505, 272 ], "spans": [ { "bbox": [ 105, 256, 505, 272 ], "score": 1.0, "content": "Ablation Study: We present two ablations to examine the effectiveness of high-quality word seg-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 269, 505, 282 ], "spans": [ { "bbox": [ 105, 269, 505, 282 ], "score": 1.0, "content": "mentation and visual representation (Table 6). We train AV-NSL with the following modifications:", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 105, 256, 505, 282 ] }, { "type": "list", "bbox": [ 107, 296, 505, 345 ], "lines": [ { "bbox": [ 106, 297, 504, 309 ], "spans": [ { "bbox": [ 106, 297, 504, 309 ], "score": 1.0, "content": "1. Fix the visual representations, but replace oracle segmentation with naive uniform word segmen-", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 118, 307, 435, 320 ], "spans": [ { "bbox": [ 118, 307, 435, 320 ], "score": 1.0, "content": "tation, where the number of words in each caption is given (uniform AV-NSL).", "type": "text" } ], "index": 7, "is_list_end_line": true }, { "bbox": [ 105, 321, 505, 335 ], "spans": [ { "bbox": [ 105, 321, 505, 335 ], "score": 1.0, "content": "2. Fix the oracle word segmentation, but replace visual embeddings with random images, where", "type": "text" } ], "index": 8, "is_list_start_line": true }, { "bbox": [ 119, 334, 379, 346 ], "spans": [ { "bbox": [ 119, 334, 379, 346 ], "score": 1.0, "content": "each pixel is independently sampled from a uniform distribution.", "type": "text" } ], "index": 9, "is_list_end_line": true } ], "index": 7.5, "bbox_fs": [ 105, 297, 505, 346 ] }, { "type": "text", "bbox": [ 107, 353, 505, 408 ], "lines": [ { "bbox": [ 106, 352, 505, 366 ], "spans": [ { "bbox": [ 106, 352, 505, 366 ], "score": 1.0, "content": "We observe that there are significant performance drops in both settings, comparing to the AV-NSL", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 363, 505, 377 ], "spans": [ { "bbox": [ 105, 363, 505, 377 ], "score": 1.0, "content": "trained with oracle segmentation and high-quality visual representation. This set of results comple-", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 375, 505, 387 ], "spans": [ { "bbox": [ 105, 375, 505, 387 ], "score": 1.0, "content": "ment Table 2, stressing that precise word segmentation and high-quality visual representations are", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 385, 505, 399 ], "spans": [ { "bbox": [ 105, 385, 505, 399 ], "score": 1.0, "content": "both necessary for phrase structure induction from speech. Furthermore, we provide tree structure", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 397, 421, 410 ], "spans": [ { "bbox": [ 105, 397, 421, 410 ], "score": 1.0, "content": "and word segmentation visualizations for qualitative analysis in the Appendix.", "type": "text" } ], "index": 14 } ], "index": 12, "bbox_fs": [ 105, 352, 505, 410 ] }, { "type": "text", "bbox": [ 318, 431, 501, 520 ], "lines": [ { "bbox": [ 317, 431, 502, 443 ], "spans": [ { "bbox": [ 317, 431, 502, 443 ], "score": 1.0, "content": "Table 6: Top rows: performance of AV-NSL", "type": "text" } ], "index": 16 }, { "bbox": [ 317, 442, 502, 455 ], "spans": [ { "bbox": [ 317, 442, 502, 455 ], "score": 1.0, "content": "with word segmentation in various quality", "type": "text" } ], "index": 18 }, { "bbox": [ 317, 453, 503, 465 ], "spans": [ { "bbox": [ 317, 453, 503, 465 ], "score": 1.0, "content": "and high-quality visual embeddings. Bot-", "type": "text" } ], "index": 20 }, { "bbox": [ 316, 465, 503, 475 ], "spans": [ { "bbox": [ 316, 465, 503, 475 ], "score": 1.0, "content": "tom rows: performance of AV-NSL with vi-", "type": "text" } ], "index": 22 }, { "bbox": [ 317, 475, 502, 487 ], "spans": [ { "bbox": [ 317, 475, 502, 487 ], "score": 1.0, "content": "sual embeddings in various quality and high-", "type": "text" } ], "index": 24 }, { "bbox": [ 317, 486, 503, 497 ], "spans": [ { "bbox": [ 317, 486, 503, 497 ], "score": 1.0, "content": "quality word segmentation. DINO: a self-", "type": "text" } ], "index": 26 }, { "bbox": [ 317, 496, 503, 509 ], "spans": [ { "bbox": [ 317, 496, 503, 509 ], "score": 1.0, "content": "supervised model that produces high-quality", "type": "text" } ], "index": 29 }, { "bbox": [ 317, 508, 489, 520 ], "spans": [ { "bbox": [ 317, 508, 489, 520 ], "score": 1.0, "content": "visual representations (Caron et al., 2021).", "type": "text" } ], "index": 30 } ], "index": 23.0, "bbox_fs": [ 316, 431, 503, 520 ] }, { "type": "table", "bbox": [ 108, 524, 311, 590 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 430, 313, 519 ], "group_id": 1, "lines": [ { "bbox": [ 105, 430, 315, 443 ], "spans": [ { "bbox": [ 105, 430, 315, 443 ], "score": 1.0, "content": "Table 5: Recall of specific typed phrases, includ-", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 442, 315, 454 ], "spans": [ { "bbox": [ 106, 442, 315, 454 ], "score": 1.0, "content": "ing noun phrases (NP), verb phrases (VP), preposi-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 453, 315, 465 ], "spans": [ { "bbox": [ 106, 453, 315, 465 ], "score": 1.0, "content": "tional phrases (PP) and adjective phrases (ADJP),", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 464, 315, 476 ], "spans": [ { "bbox": [ 105, 464, 158, 476 ], "score": 1.0, "content": "and overall", "type": "text" }, { "bbox": [ 159, 465, 171, 475 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 171, 464, 315, 476 ], "score": 1.0, "content": "score, evaluated on the Spoken-", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 475, 315, 486 ], "spans": [ { "bbox": [ 106, 475, 315, 486 ], "score": 1.0, "content": "COCO test split. The VG-NSL numbers are taken", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 486, 315, 497 ], "spans": [ { "bbox": [ 106, 486, 315, 497 ], "score": 1.0, "content": "from (Shi et al., 2019). AV-NSL here are trained", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 497, 316, 508 ], "spans": [ { "bbox": [ 105, 497, 316, 508 ], "score": 1.0, "content": "on oracle segmentation with vanilla HuBERT as the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 509, 195, 520 ], "spans": [ { "bbox": [ 105, 509, 195, 520 ], "score": 1.0, "content": "layer representations.", "type": "text" } ], "index": 28 } ], "index": 22.0 }, { "type": "table_body", "bbox": [ 108, 524, 311, 590 ], "group_id": 1, "lines": [ { "bbox": [ 108, 524, 311, 590 ], "spans": [ { "bbox": [ 108, 524, 311, 590 ], "score": 0.981, "html": "
ModelF1Constituent Recall
NPVPPPADJP
VG-NSL (Shi et al.,2019)50.479.626.242.022.0
VG-NSL + HI53.374.632.566.521.7
VG-NSL + HI+ FastText54.478.824.465.622.0
oracle AV-NSL55.655.568.166.622.1
", "type": "table", "image_path": "38d137326beea01d78ca72e0e3591ad3cdcf2650a0372d5e5d022ca0e875d59d.jpg" } ] } ], "index": 35, "virtual_lines": [ { "bbox": [ 108, 524, 311, 537.2 ], "spans": [], "index": 31 }, { "bbox": [ 108, 537.2, 311, 550.4000000000001 ], "spans": [], "index": 33 }, { "bbox": [ 108, 550.4000000000001, 311, 563.6000000000001 ], "spans": [], "index": 35 }, { "bbox": [ 108, 563.6000000000001, 311, 576.8000000000002 ], "spans": [], "index": 37 }, { "bbox": [ 108, 576.8000000000002, 311, 590.0000000000002 ], "spans": [], "index": 39 } ] } ], "index": 28.5 }, { "type": "table", "bbox": [ 323, 525, 496, 591 ], "blocks": [ { "type": "table_body", "bbox": [ 323, 525, 496, 591 ], "group_id": 2, "lines": [ { "bbox": [ 323, 525, 496, 591 ], "spans": [ { "bbox": [ 323, 525, 496, 591 ], "score": 0.976, "html": "
ModelVisualF1
Syntax InductionSeg.Repre.
oracle AV-NSLHuBERT10ResNet10150.50
uniform AV-NSLHuBERT10ResNet10136.62
oracle AV-NSLHuBERT255.71
DINO
oracle AV-NSLHuBERT2random31.23
", "type": "table", "image_path": "02dfd66c2176ae4821ac6dbb8aebf51317ed8a95a7f0d2294aa6dae357fad75d.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 323, 525, 496, 538.2 ], "spans": [], "index": 32 }, { "bbox": [ 323, 538.2, 496, 551.4000000000001 ], "spans": [], "index": 34 }, { "bbox": [ 323, 551.4000000000001, 496, 564.6000000000001 ], "spans": [], "index": 36 }, { "bbox": [ 323, 564.6000000000001, 496, 577.8000000000002 ], "spans": [], "index": 38 }, { "bbox": [ 323, 577.8000000000002, 496, 591.0000000000002 ], "spans": [], "index": 40 } ] } ], "index": 36 }, { "type": "title", "bbox": [ 107, 612, 196, 624 ], "lines": [ { "bbox": [ 105, 609, 198, 628 ], "spans": [ { "bbox": [ 105, 609, 198, 628 ], "score": 1.0, "content": "6 CONCLUSION", "type": "text" } ], "index": 41 } ], "index": 41 }, { "type": "text", "bbox": [ 106, 632, 505, 732 ], "lines": [ { "bbox": [ 106, 633, 504, 645 ], "spans": [ { "bbox": [ 106, 633, 504, 645 ], "score": 1.0, "content": "In recent years, there have been fruitful progresses in multi-modal induction for zero-resource speech", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "processing and grammar induction respectively. The idea of leveraging the visual modality to learn", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 655, 506, 668 ], "spans": [ { "bbox": [ 105, 655, 506, 668 ], "score": 1.0, "content": "language competence, either lexicon units from speech or syntactic structure from text, is an at-", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 666, 506, 679 ], "spans": [ { "bbox": [ 106, 666, 506, 679 ], "score": 1.0, "content": "tractive approach for modeling human language acquisition. Our study contributes to both lines of", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 676, 505, 690 ], "spans": [ { "bbox": [ 105, 676, 505, 690 ], "score": 1.0, "content": "research, by presenting an unifying framework that learns phrase structure from visually-grounded", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 687, 506, 702 ], "spans": [ { "bbox": [ 105, 687, 506, 702 ], "score": 1.0, "content": "speech, without any text. We show that our proposed model, AV-NSL, infers meaningful con-", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 698, 506, 713 ], "spans": [ { "bbox": [ 105, 698, 506, 713 ], "score": 1.0, "content": "stituency parse trees on top of continuous word segment representations, both quantitatively and", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 710, 506, 722 ], "spans": [ { "bbox": [ 105, 710, 506, 722 ], "score": 1.0, "content": "qualitatively. To justify our modeling design choices, we construct several baselines and introduce a", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 720, 504, 734 ], "spans": [ { "bbox": [ 105, 720, 504, 734 ], "score": 1.0, "content": "novel evaluation metric. We envision our research as the first of many in textless structure learning.", "type": "text" } ], "index": 50 } ], "index": 46, "bbox_fs": [ 105, 633, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 81, 211, 94 ], "lines": [ { "bbox": [ 106, 80, 213, 95 ], "spans": [ { "bbox": [ 106, 80, 213, 95 ], "score": 1.0, "content": "ETHICS STATEMENT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 100, 505, 167 ], "lines": [ { "bbox": [ 105, 100, 505, 113 ], "spans": [ { "bbox": [ 105, 100, 505, 113 ], "score": 1.0, "content": "This work is scientific at its core, as the goal is to study the process of grammar induction from", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 111, 505, 123 ], "spans": [ { "bbox": [ 105, 111, 505, 123 ], "score": 1.0, "content": "speech with visual grounding. The data used in this work is also publicly available. One potential", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 123, 505, 135 ], "spans": [ { "bbox": [ 105, 123, 505, 135 ], "score": 1.0, "content": "concern is that the data and experiments are based on English, which does not represent the global", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 133, 505, 146 ], "spans": [ { "bbox": [ 105, 133, 505, 146 ], "score": 1.0, "content": "human population. However, we believe that our proposed method is general enough to be applied", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 144, 505, 157 ], "spans": [ { "bbox": [ 105, 144, 505, 157 ], "score": 1.0, "content": "to other spoken languages when the data is available, because we do not use any language specific", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 155, 452, 167 ], "spans": [ { "bbox": [ 105, 155, 452, 167 ], "score": 1.0, "content": "speech processing techniques, and we do not have any built-in bias within the models.", "type": "text" } ], "index": 6 } ], "index": 3.5 }, { "type": "title", "bbox": [ 108, 183, 268, 195 ], "lines": [ { "bbox": [ 106, 182, 269, 196 ], "spans": [ { "bbox": [ 106, 182, 269, 196 ], "score": 1.0, "content": "REPRODUCIBILITY STATEMENT", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 202, 505, 246 ], "lines": [ { "bbox": [ 106, 202, 505, 214 ], "spans": [ { "bbox": [ 106, 202, 505, 214 ], "score": 1.0, "content": "AV-NSL code, s-Benepar code, and SAIOU evaluation code will be made publicly available. AV-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 213, 505, 225 ], "spans": [ { "bbox": [ 105, 213, 505, 225 ], "score": 1.0, "content": "NSL code is based on the VG-NSL codebase. s-Benepar code is based on the Benepar codebase.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 224, 505, 237 ], "spans": [ { "bbox": [ 105, 224, 505, 237 ], "score": 1.0, "content": "SpokenCOCO is publicly available to download. All models are trained on a single GPU. We also in-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 235, 503, 247 ], "spans": [ { "bbox": [ 106, 235, 503, 247 ], "score": 1.0, "content": "cluded as many experimental details as we can in the main content of the paper and in Appendix A.2.", "type": "text" } ], "index": 11 } ], "index": 9.5 }, { "type": "title", "bbox": [ 107, 263, 176, 275 ], "lines": [ { "bbox": [ 106, 263, 176, 276 ], "spans": [ { "bbox": [ 106, 263, 176, 276 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 108, 281, 503, 315 ], "lines": [ { "bbox": [ 105, 281, 505, 294 ], "spans": [ { "bbox": [ 105, 281, 505, 294 ], "score": 1.0, "content": "Robin Algayres, Tristan Ricoul, Julien Karadayi, Hugo Laurenc¸on, Salah Zaiem, Abdelrahman Mo-", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 293, 505, 305 ], "spans": [ { "bbox": [ 115, 293, 505, 305 ], "score": 1.0, "content": "hamed, Benoˆıt Sagot, and Emmanuel Dupoux. Dp-parse: Finding word boundaries from raw", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 304, 411, 315 ], "spans": [ { "bbox": [ 115, 304, 411, 315 ], "score": 1.0, "content": "speech with an instance lexicon. arXiv preprint arXiv:2206.11332, 2022.", "type": "text" } ], "index": 15 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 322, 345, 334 ], "lines": [ { "bbox": [ 106, 322, 345, 335 ], "spans": [ { "bbox": [ 106, 322, 345, 335 ], "score": 1.0, "content": "Mark C Baker. The atoms of language. Basic Books, 2001.", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 107, 340, 503, 364 ], "lines": [ { "bbox": [ 106, 340, 505, 353 ], "spans": [ { "bbox": [ 106, 340, 505, 353 ], "score": 1.0, "content": "Saurabhchand Bhati, Jesus Villalba, Piotr ´ Zelasko, Laureano Moro-Velazquez, and Najim Dehak. ˙", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 352, 505, 364 ], "spans": [ { "bbox": [ 116, 352, 505, 364 ], "score": 1.0, "content": "Segmental contrastive predictive coding for unsupervised word segmentation. Interspeech, 2021.", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "text", "bbox": [ 107, 370, 401, 382 ], "lines": [ { "bbox": [ 105, 370, 401, 383 ], "spans": [ { "bbox": [ 105, 370, 401, 383 ], "score": 1.0, "content": "Peter J Bickel and Bo Li. Mathematical statistics. In Test. Citeseer, 1977.", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 389, 441, 401 ], "lines": [ { "bbox": [ 105, 389, 442, 403 ], "spans": [ { "bbox": [ 105, 389, 442, 403 ], "score": 1.0, "content": "Rens Bod. An all-subtrees approach to unsupervised parsing. COLING-ACL, 2006.", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 408, 506, 442 ], "lines": [ { "bbox": [ 105, 408, 506, 421 ], "spans": [ { "bbox": [ 105, 408, 506, 421 ], "score": 1.0, "content": "Mathilde Caron, Hugo Touvron, Ishan Misra, Herve J ´ egou, Julien Mairal, Piotr Bojanowski, and ´", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 420, 506, 432 ], "spans": [ { "bbox": [ 116, 420, 506, 432 ], "score": 1.0, "content": "Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 431, 454, 443 ], "spans": [ { "bbox": [ 116, 431, 454, 443 ], "score": 1.0, "content": "the IEEE/CVF International Conference on Computer Vision, pp. 9650–9660, 2021.", "type": "text" } ], "index": 23 } ], "index": 22 }, { "type": "text", "bbox": [ 106, 448, 505, 483 ], "lines": [ { "bbox": [ 106, 449, 505, 461 ], "spans": [ { "bbox": [ 106, 449, 505, 461 ], "score": 1.0, "content": "Jan Chorowski, Grzegorz Ciesielski, Jarosław Dzikowski, Adrian Łancucki, Ricard Marxer, Ma- ´", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 460, 505, 473 ], "spans": [ { "bbox": [ 116, 460, 505, 473 ], "score": 1.0, "content": "teusz Opala, Piotr Pusz, Paweł Rychlikowski, and Michał Stypułkowski. Aligned contrastive", "type": "text" } ], "index": 25 }, { "bbox": [ 116, 471, 268, 483 ], "spans": [ { "bbox": [ 116, 471, 268, 483 ], "score": 1.0, "content": "predictive coding. Interspeech, 2021.", "type": "text" } ], "index": 26 } ], "index": 25 }, { "type": "text", "bbox": [ 105, 489, 504, 512 ], "lines": [ { "bbox": [ 104, 488, 505, 504 ], "spans": [ { "bbox": [ 104, 488, 505, 504 ], "score": 1.0, "content": "John Cocke. Programming languages and their compilers: Preliminary notes. New York University,", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 500, 143, 514 ], "spans": [ { "bbox": [ 115, 500, 143, 514 ], "score": 1.0, "content": "1969.", "type": "text" } ], "index": 28 } ], "index": 27.5 }, { "type": "text", "bbox": [ 106, 519, 504, 542 ], "lines": [ { "bbox": [ 106, 519, 505, 533 ], "spans": [ { "bbox": [ 106, 519, 505, 533 ], "score": 1.0, "content": "Carl De Marcken. Unsupervised language acquisition. PhD thesis, Massachusetts Institute of Tech-", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 530, 174, 543 ], "spans": [ { "bbox": [ 114, 530, 174, 543 ], "score": 1.0, "content": "nology, 1996.", "type": "text" } ], "index": 30 } ], "index": 29.5 }, { "type": "text", "bbox": [ 107, 549, 420, 561 ], "lines": [ { "bbox": [ 106, 549, 420, 562 ], "spans": [ { "bbox": [ 106, 549, 420, 562 ], "score": 1.0, "content": "Virginia R de Sa. Learning classification with unlabeled data. NeurIPS, 1994.", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 105, 568, 504, 591 ], "lines": [ { "bbox": [ 106, 569, 504, 580 ], "spans": [ { "bbox": [ 106, 569, 504, 580 ], "score": 1.0, "content": "Andrew Drozdov, Pat Verga, Mohit Yadav, Mohit Iyyer, and Andrew McCallum. Unsupervised", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 579, 478, 592 ], "spans": [ { "bbox": [ 115, 579, 478, 592 ], "score": 1.0, "content": "latent tree induction with deep inside-outside recursive autoencoders. NAACL-HLT, 2019.", "type": "text" } ], "index": 33 } ], "index": 32.5 }, { "type": "text", "bbox": [ 106, 598, 504, 621 ], "lines": [ { "bbox": [ 106, 597, 505, 611 ], "spans": [ { "bbox": [ 106, 597, 505, 611 ], "score": 1.0, "content": "Ewan Dunbar, Xuan Nga Cao, Juan Benjumea, Julien Karadayi, Mathieu Bernard, Laurent Besacier,", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 608, 506, 623 ], "spans": [ { "bbox": [ 115, 608, 506, 623 ], "score": 1.0, "content": "Xavier Anguera, and Emmanuel Dupoux. The zero resource speech challenge 2017. ASRU, 2017.", "type": "text" } ], "index": 35 } ], "index": 34.5 }, { "type": "text", "bbox": [ 107, 627, 505, 662 ], "lines": [ { "bbox": [ 105, 627, 505, 641 ], "spans": [ { "bbox": [ 105, 627, 505, 641 ], "score": 1.0, "content": "Ewan Dunbar, Robin Algayres, Julien Karadayi, Mathieu Bernard, Juan Benjumea, Xuan-Nga Cao,", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 639, 505, 651 ], "spans": [ { "bbox": [ 115, 639, 505, 651 ], "score": 1.0, "content": "Lucie Miskic, Charlotte Dugrain, Lucas Ondel, Alan W Black, et al. The zero resource speech", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 649, 317, 663 ], "spans": [ { "bbox": [ 116, 649, 317, 663 ], "score": 1.0, "content": "challenge 2019: Tts without t. Interspeech, 2019.", "type": "text" } ], "index": 38 } ], "index": 37 }, { "type": "text", "bbox": [ 107, 668, 504, 702 ], "lines": [ { "bbox": [ 105, 668, 505, 682 ], "spans": [ { "bbox": [ 105, 668, 505, 682 ], "score": 1.0, "content": "Ewan Dunbar, Julien Karadayi, Mathieu Bernard, Xuan-Nga Cao, Robin Algayres, Lucas Ondel,", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 679, 505, 693 ], "spans": [ { "bbox": [ 114, 679, 505, 693 ], "score": 1.0, "content": "Laurent Besacier, Sakriani Sakti, and Emmanuel Dupoux. The zero resource speech challenge", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 691, 405, 704 ], "spans": [ { "bbox": [ 116, 691, 405, 704 ], "score": 1.0, "content": "2020: Discovering discrete subword and word units. Interspeech, 2020.", "type": "text" } ], "index": 41 } ], "index": 40 }, { "type": "text", "bbox": [ 107, 709, 502, 732 ], "lines": [ { "bbox": [ 106, 708, 504, 722 ], "spans": [ { "bbox": [ 106, 708, 504, 722 ], "score": 1.0, "content": "Emmanuel Dupoux. Cognitive science in the era of artificial intelligence: A roadmap for reverse-", "type": "text" } ], "index": 42 }, { "bbox": [ 116, 721, 396, 732 ], "spans": [ { "bbox": [ 116, 721, 396, 732 ], "score": 1.0, "content": "engineering the infant language-learner. Cognition, 173:43–59, 2018.", "type": "text" } ], "index": 43 } ], "index": 42.5 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 765 ], "spans": [ { "bbox": [ 299, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 81, 211, 94 ], "lines": [ { "bbox": [ 106, 80, 213, 95 ], "spans": [ { "bbox": [ 106, 80, 213, 95 ], "score": 1.0, "content": "ETHICS STATEMENT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 100, 505, 167 ], "lines": [ { "bbox": [ 105, 100, 505, 113 ], "spans": [ { "bbox": [ 105, 100, 505, 113 ], "score": 1.0, "content": "This work is scientific at its core, as the goal is to study the process of grammar induction from", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 111, 505, 123 ], "spans": [ { "bbox": [ 105, 111, 505, 123 ], "score": 1.0, "content": "speech with visual grounding. The data used in this work is also publicly available. One potential", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 123, 505, 135 ], "spans": [ { "bbox": [ 105, 123, 505, 135 ], "score": 1.0, "content": "concern is that the data and experiments are based on English, which does not represent the global", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 133, 505, 146 ], "spans": [ { "bbox": [ 105, 133, 505, 146 ], "score": 1.0, "content": "human population. However, we believe that our proposed method is general enough to be applied", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 144, 505, 157 ], "spans": [ { "bbox": [ 105, 144, 505, 157 ], "score": 1.0, "content": "to other spoken languages when the data is available, because we do not use any language specific", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 155, 452, 167 ], "spans": [ { "bbox": [ 105, 155, 452, 167 ], "score": 1.0, "content": "speech processing techniques, and we do not have any built-in bias within the models.", "type": "text" } ], "index": 6 } ], "index": 3.5, "bbox_fs": [ 105, 100, 505, 167 ] }, { "type": "title", "bbox": [ 108, 183, 268, 195 ], "lines": [ { "bbox": [ 106, 182, 269, 196 ], "spans": [ { "bbox": [ 106, 182, 269, 196 ], "score": 1.0, "content": "REPRODUCIBILITY STATEMENT", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 202, 505, 246 ], "lines": [ { "bbox": [ 106, 202, 505, 214 ], "spans": [ { "bbox": [ 106, 202, 505, 214 ], "score": 1.0, "content": "AV-NSL code, s-Benepar code, and SAIOU evaluation code will be made publicly available. AV-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 213, 505, 225 ], "spans": [ { "bbox": [ 105, 213, 505, 225 ], "score": 1.0, "content": "NSL code is based on the VG-NSL codebase. s-Benepar code is based on the Benepar codebase.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 224, 505, 237 ], "spans": [ { "bbox": [ 105, 224, 505, 237 ], "score": 1.0, "content": "SpokenCOCO is publicly available to download. All models are trained on a single GPU. We also in-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 235, 503, 247 ], "spans": [ { "bbox": [ 106, 235, 503, 247 ], "score": 1.0, "content": "cluded as many experimental details as we can in the main content of the paper and in Appendix A.2.", "type": "text" } ], "index": 11 } ], "index": 9.5, "bbox_fs": [ 105, 202, 505, 247 ] }, { "type": "title", "bbox": [ 107, 263, 176, 275 ], "lines": [ { "bbox": [ 106, 263, 176, 276 ], "spans": [ { "bbox": [ 106, 263, 176, 276 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 108, 281, 503, 315 ], "lines": [ { "bbox": [ 105, 281, 505, 294 ], "spans": [ { "bbox": [ 105, 281, 505, 294 ], "score": 1.0, "content": "Robin Algayres, Tristan Ricoul, Julien Karadayi, Hugo Laurenc¸on, Salah Zaiem, Abdelrahman Mo-", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 293, 505, 305 ], "spans": [ { "bbox": [ 115, 293, 505, 305 ], "score": 1.0, "content": "hamed, Benoˆıt Sagot, and Emmanuel Dupoux. Dp-parse: Finding word boundaries from raw", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 304, 411, 315 ], "spans": [ { "bbox": [ 115, 304, 411, 315 ], "score": 1.0, "content": "speech with an instance lexicon. arXiv preprint arXiv:2206.11332, 2022.", "type": "text" } ], "index": 15 } ], "index": 14, "bbox_fs": [ 105, 281, 505, 315 ] }, { "type": "text", "bbox": [ 106, 322, 345, 334 ], "lines": [ { "bbox": [ 106, 322, 345, 335 ], "spans": [ { "bbox": [ 106, 322, 345, 335 ], "score": 1.0, "content": "Mark C Baker. The atoms of language. Basic Books, 2001.", "type": "text" } ], "index": 16 } ], "index": 16, "bbox_fs": [ 106, 322, 345, 335 ] }, { "type": "text", "bbox": [ 107, 340, 503, 364 ], "lines": [ { "bbox": [ 106, 340, 505, 353 ], "spans": [ { "bbox": [ 106, 340, 505, 353 ], "score": 1.0, "content": "Saurabhchand Bhati, Jesus Villalba, Piotr ´ Zelasko, Laureano Moro-Velazquez, and Najim Dehak. ˙", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 352, 505, 364 ], "spans": [ { "bbox": [ 116, 352, 505, 364 ], "score": 1.0, "content": "Segmental contrastive predictive coding for unsupervised word segmentation. Interspeech, 2021.", "type": "text" } ], "index": 18 } ], "index": 17.5, "bbox_fs": [ 106, 340, 505, 364 ] }, { "type": "text", "bbox": [ 107, 370, 401, 382 ], "lines": [ { "bbox": [ 105, 370, 401, 383 ], "spans": [ { "bbox": [ 105, 370, 401, 383 ], "score": 1.0, "content": "Peter J Bickel and Bo Li. Mathematical statistics. In Test. Citeseer, 1977.", "type": "text" } ], "index": 19 } ], "index": 19, "bbox_fs": [ 105, 370, 401, 383 ] }, { "type": "text", "bbox": [ 107, 389, 441, 401 ], "lines": [ { "bbox": [ 105, 389, 442, 403 ], "spans": [ { "bbox": [ 105, 389, 442, 403 ], "score": 1.0, "content": "Rens Bod. An all-subtrees approach to unsupervised parsing. COLING-ACL, 2006.", "type": "text" } ], "index": 20 } ], "index": 20, "bbox_fs": [ 105, 389, 442, 403 ] }, { "type": "text", "bbox": [ 107, 408, 506, 442 ], "lines": [ { "bbox": [ 105, 408, 506, 421 ], "spans": [ { "bbox": [ 105, 408, 506, 421 ], "score": 1.0, "content": "Mathilde Caron, Hugo Touvron, Ishan Misra, Herve J ´ egou, Julien Mairal, Piotr Bojanowski, and ´", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 420, 506, 432 ], "spans": [ { "bbox": [ 116, 420, 506, 432 ], "score": 1.0, "content": "Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 431, 454, 443 ], "spans": [ { "bbox": [ 116, 431, 454, 443 ], "score": 1.0, "content": "the IEEE/CVF International Conference on Computer Vision, pp. 9650–9660, 2021.", "type": "text" } ], "index": 23 } ], "index": 22, "bbox_fs": [ 105, 408, 506, 443 ] }, { "type": "text", "bbox": [ 106, 448, 505, 483 ], "lines": [ { "bbox": [ 106, 449, 505, 461 ], "spans": [ { "bbox": [ 106, 449, 505, 461 ], "score": 1.0, "content": "Jan Chorowski, Grzegorz Ciesielski, Jarosław Dzikowski, Adrian Łancucki, Ricard Marxer, Ma- ´", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 460, 505, 473 ], "spans": [ { "bbox": [ 116, 460, 505, 473 ], "score": 1.0, "content": "teusz Opala, Piotr Pusz, Paweł Rychlikowski, and Michał Stypułkowski. Aligned contrastive", "type": "text" } ], "index": 25 }, { "bbox": [ 116, 471, 268, 483 ], "spans": [ { "bbox": [ 116, 471, 268, 483 ], "score": 1.0, "content": "predictive coding. Interspeech, 2021.", "type": "text" } ], "index": 26 } ], "index": 25, "bbox_fs": [ 106, 449, 505, 483 ] }, { "type": "text", "bbox": [ 105, 489, 504, 512 ], "lines": [ { "bbox": [ 104, 488, 505, 504 ], "spans": [ { "bbox": [ 104, 488, 505, 504 ], "score": 1.0, "content": "John Cocke. Programming languages and their compilers: Preliminary notes. New York University,", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 500, 143, 514 ], "spans": [ { "bbox": [ 115, 500, 143, 514 ], "score": 1.0, "content": "1969.", "type": "text" } ], "index": 28 } ], "index": 27.5, "bbox_fs": [ 104, 488, 505, 514 ] }, { "type": "text", "bbox": [ 106, 519, 504, 542 ], "lines": [ { "bbox": [ 106, 519, 505, 533 ], "spans": [ { "bbox": [ 106, 519, 505, 533 ], "score": 1.0, "content": "Carl De Marcken. Unsupervised language acquisition. PhD thesis, Massachusetts Institute of Tech-", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 530, 174, 543 ], "spans": [ { "bbox": [ 114, 530, 174, 543 ], "score": 1.0, "content": "nology, 1996.", "type": "text" } ], "index": 30 } ], "index": 29.5, "bbox_fs": [ 106, 519, 505, 543 ] }, { "type": "text", "bbox": [ 107, 549, 420, 561 ], "lines": [ { "bbox": [ 106, 549, 420, 562 ], "spans": [ { "bbox": [ 106, 549, 420, 562 ], "score": 1.0, "content": "Virginia R de Sa. Learning classification with unlabeled data. NeurIPS, 1994.", "type": "text" } ], "index": 31 } ], "index": 31, "bbox_fs": [ 106, 549, 420, 562 ] }, { "type": "text", "bbox": [ 105, 568, 504, 591 ], "lines": [ { "bbox": [ 106, 569, 504, 580 ], "spans": [ { "bbox": [ 106, 569, 504, 580 ], "score": 1.0, "content": "Andrew Drozdov, Pat Verga, Mohit Yadav, Mohit Iyyer, and Andrew McCallum. Unsupervised", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 579, 478, 592 ], "spans": [ { "bbox": [ 115, 579, 478, 592 ], "score": 1.0, "content": "latent tree induction with deep inside-outside recursive autoencoders. NAACL-HLT, 2019.", "type": "text" } ], "index": 33 } ], "index": 32.5, "bbox_fs": [ 106, 569, 504, 592 ] }, { "type": "text", "bbox": [ 106, 598, 504, 621 ], "lines": [ { "bbox": [ 106, 597, 505, 611 ], "spans": [ { "bbox": [ 106, 597, 505, 611 ], "score": 1.0, "content": "Ewan Dunbar, Xuan Nga Cao, Juan Benjumea, Julien Karadayi, Mathieu Bernard, Laurent Besacier,", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 608, 506, 623 ], "spans": [ { "bbox": [ 115, 608, 506, 623 ], "score": 1.0, "content": "Xavier Anguera, and Emmanuel Dupoux. The zero resource speech challenge 2017. ASRU, 2017.", "type": "text" } ], "index": 35 } ], "index": 34.5, "bbox_fs": [ 106, 597, 506, 623 ] }, { "type": "text", "bbox": [ 107, 627, 505, 662 ], "lines": [ { "bbox": [ 105, 627, 505, 641 ], "spans": [ { "bbox": [ 105, 627, 505, 641 ], "score": 1.0, "content": "Ewan Dunbar, Robin Algayres, Julien Karadayi, Mathieu Bernard, Juan Benjumea, Xuan-Nga Cao,", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 639, 505, 651 ], "spans": [ { "bbox": [ 115, 639, 505, 651 ], "score": 1.0, "content": "Lucie Miskic, Charlotte Dugrain, Lucas Ondel, Alan W Black, et al. The zero resource speech", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 649, 317, 663 ], "spans": [ { "bbox": [ 116, 649, 317, 663 ], "score": 1.0, "content": "challenge 2019: Tts without t. Interspeech, 2019.", "type": "text" } ], "index": 38 } ], "index": 37, "bbox_fs": [ 105, 627, 505, 663 ] }, { "type": "text", "bbox": [ 107, 668, 504, 702 ], "lines": [ { "bbox": [ 105, 668, 505, 682 ], "spans": [ { "bbox": [ 105, 668, 505, 682 ], "score": 1.0, "content": "Ewan Dunbar, Julien Karadayi, Mathieu Bernard, Xuan-Nga Cao, Robin Algayres, Lucas Ondel,", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 679, 505, 693 ], "spans": [ { "bbox": [ 114, 679, 505, 693 ], "score": 1.0, "content": "Laurent Besacier, Sakriani Sakti, and Emmanuel Dupoux. The zero resource speech challenge", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 691, 405, 704 ], "spans": [ { "bbox": [ 116, 691, 405, 704 ], "score": 1.0, "content": "2020: Discovering discrete subword and word units. Interspeech, 2020.", "type": "text" } ], "index": 41 } ], "index": 40, "bbox_fs": [ 105, 668, 505, 704 ] }, { "type": "text", "bbox": [ 107, 709, 502, 732 ], "lines": [ { "bbox": [ 106, 708, 504, 722 ], "spans": [ { "bbox": [ 106, 708, 504, 722 ], "score": 1.0, "content": "Emmanuel Dupoux. Cognitive science in the era of artificial intelligence: A roadmap for reverse-", "type": "text" } ], "index": 42 }, { "bbox": [ 116, 721, 396, 732 ], "spans": [ { "bbox": [ 116, 721, 396, 732 ], "score": 1.0, "content": "engineering the infant language-learner. Cognition, 173:43–59, 2018.", "type": "text" } ], "index": 43 } ], "index": 42.5, "bbox_fs": [ 106, 708, 504, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "Zvi Galil. Efficient algorithms for finding maximum matching in graphs. ACM Comput. Surv., 18", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 504, 107 ], "spans": [ { "bbox": [ 115, 93, 504, 107 ], "score": 1.0, "content": "(1):23–38, mar 1986. ISSN 0360-0300. doi: 10.1145/6462.6502. URL https://doi.org/", "type": "text" } ], "index": 1 }, { "bbox": [ 117, 105, 223, 116 ], "spans": [ { "bbox": [ 117, 105, 223, 116 ], "score": 1.0, "content": "10.1145/6462.6502.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 107, 124, 503, 147 ], "lines": [ { "bbox": [ 105, 123, 505, 137 ], "spans": [ { "bbox": [ 105, 123, 505, 137 ], "score": 1.0, "content": "Haohan Guo, Frank K Soong, Lei He, and Lei Xie. Exploiting syntactic features in a parsed tree to", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 136, 289, 147 ], "spans": [ { "bbox": [ 116, 136, 289, 147 ], "score": 1.0, "content": "improve end-to-end tts. Interspeech, 2019.", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "text", "bbox": [ 107, 155, 503, 178 ], "lines": [ { "bbox": [ 105, 154, 505, 168 ], "spans": [ { "bbox": [ 105, 154, 505, 168 ], "score": 1.0, "content": "David Harwath and James R Glass. Learning word-like units from joint audio-visual analysis. ACL,", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 165, 144, 179 ], "spans": [ { "bbox": [ 114, 165, 144, 179 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "text", "bbox": [ 107, 186, 504, 209 ], "lines": [ { "bbox": [ 105, 186, 505, 199 ], "spans": [ { "bbox": [ 105, 186, 505, 199 ], "score": 1.0, "content": "David Harwath, Wei-Ning Hsu, and James Glass. Learning hierarchical discrete linguistic units", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 198, 298, 210 ], "spans": [ { "bbox": [ 116, 198, 298, 210 ], "score": 1.0, "content": "from visually-grounded speech. ICLR, 2020.", "type": "text" } ], "index": 8 } ], "index": 7.5 }, { "type": "text", "bbox": [ 107, 217, 504, 240 ], "lines": [ { "bbox": [ 105, 217, 505, 230 ], "spans": [ { "bbox": [ 105, 217, 505, 230 ], "score": 1.0, "content": "David Frank Harwath. Learning spoken language through vision. PhD thesis, Massachusetts Insti-", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 228, 221, 241 ], "spans": [ { "bbox": [ 116, 228, 221, 241 ], "score": 1.0, "content": "tute of Technology, 2018.", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 248, 505, 272 ], "lines": [ { "bbox": [ 105, 247, 505, 262 ], "spans": [ { "bbox": [ 105, 247, 505, 262 ], "score": 1.0, "content": "William N. Havard, Jean-Pierre Chevrot, and Laurent Besacier. Word recognition, competition, and", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 259, 390, 272 ], "spans": [ { "bbox": [ 116, 259, 390, 272 ], "score": 1.0, "content": "activation in a model of visually grounded speech. In CoNLL, 2019.", "type": "text" } ], "index": 12 } ], "index": 11.5 }, { "type": "text", "bbox": [ 106, 279, 505, 303 ], "lines": [ { "bbox": [ 104, 278, 506, 294 ], "spans": [ { "bbox": [ 104, 278, 506, 294 ], "score": 1.0, "content": "Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 290, 220, 303 ], "spans": [ { "bbox": [ 115, 290, 220, 303 ], "score": 1.0, "content": "arXiv:1606.08415, 2016.", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "text", "bbox": [ 107, 311, 504, 334 ], "lines": [ { "bbox": [ 106, 311, 505, 324 ], "spans": [ { "bbox": [ 106, 311, 505, 324 ], "score": 1.0, "content": "Yining Hong, Qing Li, Song-Chun Zhu, and Siyuan Huang. Vlgrammar: Grounded grammar in-", "type": "text" } ], "index": 15 }, { "bbox": [ 117, 322, 299, 334 ], "spans": [ { "bbox": [ 117, 322, 299, 334 ], "score": 1.0, "content": "duction of vision and language. ICCV, 2021.", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "text", "bbox": [ 107, 341, 505, 387 ], "lines": [ { "bbox": [ 105, 342, 505, 354 ], "spans": [ { "bbox": [ 105, 342, 505, 354 ], "score": 1.0, "content": "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov,", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 353, 506, 366 ], "spans": [ { "bbox": [ 115, 353, 506, 366 ], "score": 1.0, "content": "and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 363, 506, 378 ], "spans": [ { "bbox": [ 115, 363, 506, 378 ], "score": 1.0, "content": "prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing,", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 374, 209, 387 ], "spans": [ { "bbox": [ 115, 374, 209, 387 ], "score": 1.0, "content": "29:3451–3460, 2021a.", "type": "text" } ], "index": 20 } ], "index": 18.5 }, { "type": "text", "bbox": [ 106, 395, 504, 418 ], "lines": [ { "bbox": [ 105, 393, 505, 408 ], "spans": [ { "bbox": [ 105, 393, 505, 408 ], "score": 1.0, "content": "Wei-Ning Hsu, David Harwath, Christopher Song, and James Glass. Text-free image-to-speech", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 406, 333, 418 ], "spans": [ { "bbox": [ 115, 406, 333, 418 ], "score": 1.0, "content": "synthesis using learned segmental units. ACL, 2021b.", "type": "text" } ], "index": 22 } ], "index": 21.5 }, { "type": "text", "bbox": [ 106, 426, 503, 449 ], "lines": [ { "bbox": [ 105, 424, 505, 439 ], "spans": [ { "bbox": [ 105, 424, 505, 439 ], "score": 1.0, "content": "Aren Jansen and Benjamin Van Durme. Efficient spoken term discovery using randomized algo-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 437, 202, 448 ], "spans": [ { "bbox": [ 115, 437, 202, 448 ], "score": 1.0, "content": "rithms. ASRU, 2011.", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "text", "bbox": [ 106, 456, 505, 502 ], "lines": [ { "bbox": [ 106, 457, 504, 469 ], "spans": [ { "bbox": [ 106, 457, 504, 469 ], "score": 1.0, "content": "Aren Jansen, Emmanuel Dupoux, Sharon Goldwater, Mark Johnson, Sanjeev Khudanpur, Kenneth", "type": "text" } ], "index": 25 }, { "bbox": [ 116, 468, 505, 480 ], "spans": [ { "bbox": [ 116, 468, 505, 480 ], "score": 1.0, "content": "Church, Naomi Feldman, Hynek Hermansky, Florian Metze, Richard Rose, et al. A summary of", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 479, 505, 493 ], "spans": [ { "bbox": [ 115, 479, 505, 493 ], "score": 1.0, "content": "the 2012 jhu clsp workshop on zero resource speech technologies and models of early language", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 491, 229, 502 ], "spans": [ { "bbox": [ 115, 491, 229, 502 ], "score": 1.0, "content": "acquisition. ICASSP, 2013.", "type": "text" } ], "index": 28 } ], "index": 26.5 }, { "type": "text", "bbox": [ 106, 509, 506, 555 ], "lines": [ { "bbox": [ 104, 509, 506, 523 ], "spans": [ { "bbox": [ 104, 509, 506, 523 ], "score": 1.0, "content": "Nobuyoshi Kaiki, Sakriani Sakti, and Satoshi Nakamura. Using local phrase dependency struc-", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 521, 507, 534 ], "spans": [ { "bbox": [ 115, 521, 507, 534 ], "score": 1.0, "content": "ture information in neural sequence-to-sequence speech synthesis. In 2021 24th Conference of", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 531, 507, 545 ], "spans": [ { "bbox": [ 115, 531, 507, 545 ], "score": 1.0, "content": "the Oriental COCOSDA International Committee for the Co-ordination and Standardisation of", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 543, 478, 555 ], "spans": [ { "bbox": [ 115, 543, 478, 555 ], "score": 1.0, "content": "Speech Databases and Assessment Techniques (O-COCOSDA), pp. 206–211. IEEE, 2021.", "type": "text" } ], "index": 32 } ], "index": 30.5 }, { "type": "text", "bbox": [ 107, 563, 504, 586 ], "lines": [ { "bbox": [ 105, 562, 505, 577 ], "spans": [ { "bbox": [ 105, 562, 505, 577 ], "score": 1.0, "content": "Herman Kamper. Word segmentation on discovered phone units with dynamic programming and", "type": "text" } ], "index": 33 }, { "bbox": [ 116, 574, 376, 586 ], "spans": [ { "bbox": [ 116, 574, 376, 586 ], "score": 1.0, "content": "self-supervised scoring. arXiv preprint arXiv:2202.11929, 2022.", "type": "text" } ], "index": 34 } ], "index": 33.5 }, { "type": "text", "bbox": [ 105, 594, 504, 617 ], "lines": [ { "bbox": [ 106, 594, 505, 607 ], "spans": [ { "bbox": [ 106, 594, 505, 607 ], "score": 1.0, "content": "Herman Kamper and Benjamin van Niekerk. Towards unsupervised phone and word segmentation", "type": "text" } ], "index": 35 }, { "bbox": [ 116, 605, 419, 618 ], "spans": [ { "bbox": [ 116, 605, 419, 618 ], "score": 1.0, "content": "using self-supervised vector-quantized neural networks. Interspeech, 2021.", "type": "text" } ], "index": 36 } ], "index": 35.5 }, { "type": "text", "bbox": [ 107, 625, 504, 659 ], "lines": [ { "bbox": [ 105, 625, 505, 638 ], "spans": [ { "bbox": [ 105, 625, 505, 638 ], "score": 1.0, "content": "Herman Kamper, Aren Jansen, and Sharon Goldwater. Fully unsupervised small-vocabulary speech", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 637, 505, 649 ], "spans": [ { "bbox": [ 115, 637, 505, 649 ], "score": 1.0, "content": "recognition using a segmental bayesian model. In Sixteenth Annual Conference of the Interna-", "type": "text" } ], "index": 38 }, { "bbox": [ 116, 647, 316, 659 ], "spans": [ { "bbox": [ 116, 647, 316, 659 ], "score": 1.0, "content": "tional Speech Communication Association, 2015.", "type": "text" } ], "index": 39 } ], "index": 38 }, { "type": "text", "bbox": [ 108, 667, 504, 701 ], "lines": [ { "bbox": [ 105, 666, 505, 680 ], "spans": [ { "bbox": [ 105, 666, 505, 680 ], "score": 1.0, "content": "Herman Kamper, Aren Jansen, and Sharon Goldwater. A segmental framework for fully-", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 678, 505, 691 ], "spans": [ { "bbox": [ 115, 678, 505, 691 ], "score": 1.0, "content": "unsupervised large-vocabulary speech recognition. Computer Speech & Language, 46:154–174,", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 688, 143, 702 ], "spans": [ { "bbox": [ 115, 688, 143, 702 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 42 } ], "index": 41 }, { "type": "text", "bbox": [ 108, 709, 502, 732 ], "lines": [ { "bbox": [ 106, 707, 504, 723 ], "spans": [ { "bbox": [ 106, 707, 504, 723 ], "score": 1.0, "content": "Tadao Kasami. An efficient recognition and syntax-analysis algorithm for context-free languages.", "type": "text" } ], "index": 43 }, { "bbox": [ 117, 720, 348, 733 ], "spans": [ { "bbox": [ 117, 720, 348, 733 ], "score": 1.0, "content": "Coordinated Science Laboratory Report no. R-257, 1966.", "type": "text" } ], "index": 44 } ], "index": 43.5 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "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 } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "Zvi Galil. Efficient algorithms for finding maximum matching in graphs. ACM Comput. Surv., 18", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 504, 107 ], "spans": [ { "bbox": [ 115, 93, 504, 107 ], "score": 1.0, "content": "(1):23–38, mar 1986. ISSN 0360-0300. doi: 10.1145/6462.6502. URL https://doi.org/", "type": "text" } ], "index": 1 }, { "bbox": [ 117, 105, 223, 116 ], "spans": [ { "bbox": [ 117, 105, 223, 116 ], "score": 1.0, "content": "10.1145/6462.6502.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 105, 82, 505, 116 ] }, { "type": "text", "bbox": [ 107, 124, 503, 147 ], "lines": [ { "bbox": [ 105, 123, 505, 137 ], "spans": [ { "bbox": [ 105, 123, 505, 137 ], "score": 1.0, "content": "Haohan Guo, Frank K Soong, Lei He, and Lei Xie. Exploiting syntactic features in a parsed tree to", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 136, 289, 147 ], "spans": [ { "bbox": [ 116, 136, 289, 147 ], "score": 1.0, "content": "improve end-to-end tts. Interspeech, 2019.", "type": "text" } ], "index": 4 } ], "index": 3.5, "bbox_fs": [ 105, 123, 505, 147 ] }, { "type": "text", "bbox": [ 107, 155, 503, 178 ], "lines": [ { "bbox": [ 105, 154, 505, 168 ], "spans": [ { "bbox": [ 105, 154, 505, 168 ], "score": 1.0, "content": "David Harwath and James R Glass. Learning word-like units from joint audio-visual analysis. ACL,", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 165, 144, 179 ], "spans": [ { "bbox": [ 114, 165, 144, 179 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 105, 154, 505, 179 ] }, { "type": "text", "bbox": [ 107, 186, 504, 209 ], "lines": [ { "bbox": [ 105, 186, 505, 199 ], "spans": [ { "bbox": [ 105, 186, 505, 199 ], "score": 1.0, "content": "David Harwath, Wei-Ning Hsu, and James Glass. Learning hierarchical discrete linguistic units", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 198, 298, 210 ], "spans": [ { "bbox": [ 116, 198, 298, 210 ], "score": 1.0, "content": "from visually-grounded speech. ICLR, 2020.", "type": "text" } ], "index": 8 } ], "index": 7.5, "bbox_fs": [ 105, 186, 505, 210 ] }, { "type": "text", "bbox": [ 107, 217, 504, 240 ], "lines": [ { "bbox": [ 105, 217, 505, 230 ], "spans": [ { "bbox": [ 105, 217, 505, 230 ], "score": 1.0, "content": "David Frank Harwath. Learning spoken language through vision. PhD thesis, Massachusetts Insti-", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 228, 221, 241 ], "spans": [ { "bbox": [ 116, 228, 221, 241 ], "score": 1.0, "content": "tute of Technology, 2018.", "type": "text" } ], "index": 10 } ], "index": 9.5, "bbox_fs": [ 105, 217, 505, 241 ] }, { "type": "text", "bbox": [ 107, 248, 505, 272 ], "lines": [ { "bbox": [ 105, 247, 505, 262 ], "spans": [ { "bbox": [ 105, 247, 505, 262 ], "score": 1.0, "content": "William N. Havard, Jean-Pierre Chevrot, and Laurent Besacier. Word recognition, competition, and", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 259, 390, 272 ], "spans": [ { "bbox": [ 116, 259, 390, 272 ], "score": 1.0, "content": "activation in a model of visually grounded speech. In CoNLL, 2019.", "type": "text" } ], "index": 12 } ], "index": 11.5, "bbox_fs": [ 105, 247, 505, 272 ] }, { "type": "text", "bbox": [ 106, 279, 505, 303 ], "lines": [ { "bbox": [ 104, 278, 506, 294 ], "spans": [ { "bbox": [ 104, 278, 506, 294 ], "score": 1.0, "content": "Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 290, 220, 303 ], "spans": [ { "bbox": [ 115, 290, 220, 303 ], "score": 1.0, "content": "arXiv:1606.08415, 2016.", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 104, 278, 506, 303 ] }, { "type": "text", "bbox": [ 107, 311, 504, 334 ], "lines": [ { "bbox": [ 106, 311, 505, 324 ], "spans": [ { "bbox": [ 106, 311, 505, 324 ], "score": 1.0, "content": "Yining Hong, Qing Li, Song-Chun Zhu, and Siyuan Huang. Vlgrammar: Grounded grammar in-", "type": "text" } ], "index": 15 }, { "bbox": [ 117, 322, 299, 334 ], "spans": [ { "bbox": [ 117, 322, 299, 334 ], "score": 1.0, "content": "duction of vision and language. ICCV, 2021.", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 106, 311, 505, 334 ] }, { "type": "text", "bbox": [ 107, 341, 505, 387 ], "lines": [ { "bbox": [ 105, 342, 505, 354 ], "spans": [ { "bbox": [ 105, 342, 505, 354 ], "score": 1.0, "content": "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov,", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 353, 506, 366 ], "spans": [ { "bbox": [ 115, 353, 506, 366 ], "score": 1.0, "content": "and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 363, 506, 378 ], "spans": [ { "bbox": [ 115, 363, 506, 378 ], "score": 1.0, "content": "prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing,", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 374, 209, 387 ], "spans": [ { "bbox": [ 115, 374, 209, 387 ], "score": 1.0, "content": "29:3451–3460, 2021a.", "type": "text" } ], "index": 20 } ], "index": 18.5, "bbox_fs": [ 105, 342, 506, 387 ] }, { "type": "text", "bbox": [ 106, 395, 504, 418 ], "lines": [ { "bbox": [ 105, 393, 505, 408 ], "spans": [ { "bbox": [ 105, 393, 505, 408 ], "score": 1.0, "content": "Wei-Ning Hsu, David Harwath, Christopher Song, and James Glass. Text-free image-to-speech", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 406, 333, 418 ], "spans": [ { "bbox": [ 115, 406, 333, 418 ], "score": 1.0, "content": "synthesis using learned segmental units. ACL, 2021b.", "type": "text" } ], "index": 22 } ], "index": 21.5, "bbox_fs": [ 105, 393, 505, 418 ] }, { "type": "text", "bbox": [ 106, 426, 503, 449 ], "lines": [ { "bbox": [ 105, 424, 505, 439 ], "spans": [ { "bbox": [ 105, 424, 505, 439 ], "score": 1.0, "content": "Aren Jansen and Benjamin Van Durme. Efficient spoken term discovery using randomized algo-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 437, 202, 448 ], "spans": [ { "bbox": [ 115, 437, 202, 448 ], "score": 1.0, "content": "rithms. ASRU, 2011.", "type": "text" } ], "index": 24 } ], "index": 23.5, "bbox_fs": [ 105, 424, 505, 448 ] }, { "type": "text", "bbox": [ 106, 456, 505, 502 ], "lines": [ { "bbox": [ 106, 457, 504, 469 ], "spans": [ { "bbox": [ 106, 457, 504, 469 ], "score": 1.0, "content": "Aren Jansen, Emmanuel Dupoux, Sharon Goldwater, Mark Johnson, Sanjeev Khudanpur, Kenneth", "type": "text" } ], "index": 25 }, { "bbox": [ 116, 468, 505, 480 ], "spans": [ { "bbox": [ 116, 468, 505, 480 ], "score": 1.0, "content": "Church, Naomi Feldman, Hynek Hermansky, Florian Metze, Richard Rose, et al. A summary of", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 479, 505, 493 ], "spans": [ { "bbox": [ 115, 479, 505, 493 ], "score": 1.0, "content": "the 2012 jhu clsp workshop on zero resource speech technologies and models of early language", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 491, 229, 502 ], "spans": [ { "bbox": [ 115, 491, 229, 502 ], "score": 1.0, "content": "acquisition. ICASSP, 2013.", "type": "text" } ], "index": 28 } ], "index": 26.5, "bbox_fs": [ 106, 457, 505, 502 ] }, { "type": "text", "bbox": [ 106, 509, 506, 555 ], "lines": [ { "bbox": [ 104, 509, 506, 523 ], "spans": [ { "bbox": [ 104, 509, 506, 523 ], "score": 1.0, "content": "Nobuyoshi Kaiki, Sakriani Sakti, and Satoshi Nakamura. Using local phrase dependency struc-", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 521, 507, 534 ], "spans": [ { "bbox": [ 115, 521, 507, 534 ], "score": 1.0, "content": "ture information in neural sequence-to-sequence speech synthesis. In 2021 24th Conference of", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 531, 507, 545 ], "spans": [ { "bbox": [ 115, 531, 507, 545 ], "score": 1.0, "content": "the Oriental COCOSDA International Committee for the Co-ordination and Standardisation of", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 543, 478, 555 ], "spans": [ { "bbox": [ 115, 543, 478, 555 ], "score": 1.0, "content": "Speech Databases and Assessment Techniques (O-COCOSDA), pp. 206–211. IEEE, 2021.", "type": "text" } ], "index": 32 } ], "index": 30.5, "bbox_fs": [ 104, 509, 507, 555 ] }, { "type": "text", "bbox": [ 107, 563, 504, 586 ], "lines": [ { "bbox": [ 105, 562, 505, 577 ], "spans": [ { "bbox": [ 105, 562, 505, 577 ], "score": 1.0, "content": "Herman Kamper. Word segmentation on discovered phone units with dynamic programming and", "type": "text" } ], "index": 33 }, { "bbox": [ 116, 574, 376, 586 ], "spans": [ { "bbox": [ 116, 574, 376, 586 ], "score": 1.0, "content": "self-supervised scoring. arXiv preprint arXiv:2202.11929, 2022.", "type": "text" } ], "index": 34 } ], "index": 33.5, "bbox_fs": [ 105, 562, 505, 586 ] }, { "type": "text", "bbox": [ 105, 594, 504, 617 ], "lines": [ { "bbox": [ 106, 594, 505, 607 ], "spans": [ { "bbox": [ 106, 594, 505, 607 ], "score": 1.0, "content": "Herman Kamper and Benjamin van Niekerk. Towards unsupervised phone and word segmentation", "type": "text" } ], "index": 35 }, { "bbox": [ 116, 605, 419, 618 ], "spans": [ { "bbox": [ 116, 605, 419, 618 ], "score": 1.0, "content": "using self-supervised vector-quantized neural networks. Interspeech, 2021.", "type": "text" } ], "index": 36 } ], "index": 35.5, "bbox_fs": [ 106, 594, 505, 618 ] }, { "type": "text", "bbox": [ 107, 625, 504, 659 ], "lines": [ { "bbox": [ 105, 625, 505, 638 ], "spans": [ { "bbox": [ 105, 625, 505, 638 ], "score": 1.0, "content": "Herman Kamper, Aren Jansen, and Sharon Goldwater. Fully unsupervised small-vocabulary speech", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 637, 505, 649 ], "spans": [ { "bbox": [ 115, 637, 505, 649 ], "score": 1.0, "content": "recognition using a segmental bayesian model. In Sixteenth Annual Conference of the Interna-", "type": "text" } ], "index": 38 }, { "bbox": [ 116, 647, 316, 659 ], "spans": [ { "bbox": [ 116, 647, 316, 659 ], "score": 1.0, "content": "tional Speech Communication Association, 2015.", "type": "text" } ], "index": 39 } ], "index": 38, "bbox_fs": [ 105, 625, 505, 659 ] }, { "type": "text", "bbox": [ 108, 667, 504, 701 ], "lines": [ { "bbox": [ 105, 666, 505, 680 ], "spans": [ { "bbox": [ 105, 666, 505, 680 ], "score": 1.0, "content": "Herman Kamper, Aren Jansen, and Sharon Goldwater. A segmental framework for fully-", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 678, 505, 691 ], "spans": [ { "bbox": [ 115, 678, 505, 691 ], "score": 1.0, "content": "unsupervised large-vocabulary speech recognition. Computer Speech & Language, 46:154–174,", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 688, 143, 702 ], "spans": [ { "bbox": [ 115, 688, 143, 702 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 42 } ], "index": 41, "bbox_fs": [ 105, 666, 505, 702 ] }, { "type": "text", "bbox": [ 108, 709, 502, 732 ], "lines": [ { "bbox": [ 106, 707, 504, 723 ], "spans": [ { "bbox": [ 106, 707, 504, 723 ], "score": 1.0, "content": "Tadao Kasami. An efficient recognition and syntax-analysis algorithm for context-free languages.", "type": "text" } ], "index": 43 }, { "bbox": [ 117, 720, 348, 733 ], "spans": [ { "bbox": [ 117, 720, 348, 733 ], "score": 1.0, "content": "Coordinated Science Laboratory Report no. R-257, 1966.", "type": "text" } ], "index": 44 } ], "index": 43.5, "bbox_fs": [ 106, 707, 504, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 106, 82, 505, 95 ], "spans": [ { "bbox": [ 106, 82, 505, 95 ], "score": 1.0, "content": "Eugene Kharitonov, Ann Lee, Adam Polyak, Yossi Adi, Jade Copet, Kushal Lakhotia, Tu-Anh", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 504, 107 ], "spans": [ { "bbox": [ 115, 93, 504, 107 ], "score": 1.0, "content": "Nguyen, Morgane Riviere, Abdelrahman Mohamed, Emmanuel Dupoux, et al. Text-free prosody- `", "type": "text" } ], "index": 1 }, { "bbox": [ 116, 105, 346, 117 ], "spans": [ { "bbox": [ 116, 105, 346, 117 ], "score": 1.0, "content": "aware generative spoken language modeling. ACL, 2022.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 108, 123, 504, 157 ], "lines": [ { "bbox": [ 105, 122, 506, 137 ], "spans": [ { "bbox": [ 105, 122, 506, 137 ], "score": 1.0, "content": "Khazar Khorrami and Okko Johannes Ras¨ anen. Can phones, syllables, and words emerge as ¨", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 135, 505, 147 ], "spans": [ { "bbox": [ 116, 135, 505, 147 ], "score": 1.0, "content": "side-products of cross-situational audiovisual learning? - a computational investigation. ArXiv,", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 146, 209, 157 ], "spans": [ { "bbox": [ 116, 146, 209, 157 ], "score": 1.0, "content": "abs/2109.14200, 2021.", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "text", "bbox": [ 107, 165, 503, 188 ], "lines": [ { "bbox": [ 106, 164, 505, 179 ], "spans": [ { "bbox": [ 106, 164, 505, 179 ], "score": 1.0, "content": "Yoon Kim, Chris Dyer, and Alexander M Rush. Compound probabilistic context-free grammars for", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 177, 250, 188 ], "spans": [ { "bbox": [ 115, 177, 250, 188 ], "score": 1.0, "content": "grammar induction. ACL, 2019a.", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "text", "bbox": [ 107, 195, 503, 219 ], "lines": [ { "bbox": [ 106, 195, 504, 209 ], "spans": [ { "bbox": [ 106, 195, 504, 209 ], "score": 1.0, "content": "Yoon Kim, Alexander M Rush, Lei Yu, Adhiguna Kuncoro, Chris Dyer, and Gabor Melis. Unsuper- ´", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 207, 374, 219 ], "spans": [ { "bbox": [ 116, 207, 374, 219 ], "score": 1.0, "content": "vised recurrent neural network grammars. NAACL-HLT, 2019b.", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 106, 226, 502, 249 ], "lines": [ { "bbox": [ 105, 226, 504, 240 ], "spans": [ { "bbox": [ 105, 226, 504, 240 ], "score": 1.0, "content": "Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 237, 443, 250 ], "spans": [ { "bbox": [ 116, 237, 443, 250 ], "score": 1.0, "content": "with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014.", "type": "text" } ], "index": 11 } ], "index": 10.5 }, { "type": "text", "bbox": [ 106, 257, 491, 269 ], "lines": [ { "bbox": [ 105, 255, 491, 271 ], "spans": [ { "bbox": [ 105, 255, 491, 271 ], "score": 1.0, "content": "Nikita Kitaev and Dan Klein. Constituency parsing with a self-attentive encoder. In ACL, 2018.", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 276, 503, 299 ], "lines": [ { "bbox": [ 105, 275, 505, 290 ], "spans": [ { "bbox": [ 105, 275, 505, 290 ], "score": 1.0, "content": "Dan Klein and Christopher D Manning. A generative constituent-context model for improved gram-", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 288, 226, 299 ], "spans": [ { "bbox": [ 116, 288, 226, 299 ], "score": 1.0, "content": "mar induction. ACL, 2002.", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "text", "bbox": [ 107, 307, 504, 330 ], "lines": [ { "bbox": [ 105, 307, 506, 320 ], "spans": [ { "bbox": [ 105, 307, 506, 320 ], "score": 1.0, "content": "Dan Klein and Christopher D Manning. Corpus-based induction of syntactic structure: Models of", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 318, 286, 330 ], "spans": [ { "bbox": [ 116, 318, 286, 330 ], "score": 1.0, "content": "dependency and constituency. ACL, 2004.", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "text", "bbox": [ 106, 337, 505, 361 ], "lines": [ { "bbox": [ 106, 336, 505, 351 ], "spans": [ { "bbox": [ 106, 336, 505, 351 ], "score": 1.0, "content": "Arne Kohn, Timo Baumann, and Oskar D ¨ orfler. An empirical analysis of the correlation of syntax ¨", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 349, 247, 361 ], "spans": [ { "bbox": [ 115, 349, 247, 361 ], "score": 1.0, "content": "and prosody. Interspeech, 2018.", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "text", "bbox": [ 105, 368, 505, 391 ], "lines": [ { "bbox": [ 106, 368, 505, 381 ], "spans": [ { "bbox": [ 106, 368, 505, 381 ], "score": 1.0, "content": "Noriyuki Kojima, Hadar Averbuch-Elor, Alexander M Rush, and Yoav Artzi. What is learned in", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 379, 344, 392 ], "spans": [ { "bbox": [ 116, 379, 344, 392 ], "score": 1.0, "content": "visually grounded neural syntax acquisition. ACL, 2020.", "type": "text" } ], "index": 20 } ], "index": 19.5 }, { "type": "text", "bbox": [ 107, 398, 505, 455 ], "lines": [ { "bbox": [ 106, 399, 505, 411 ], "spans": [ { "bbox": [ 106, 399, 505, 411 ], "score": 1.0, "content": "Shankar Kumar and William Byrne. Minimum Bayes-risk decoding for statistical machine transla-", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 410, 505, 422 ], "spans": [ { "bbox": [ 116, 410, 505, 422 ], "score": 1.0, "content": "tion. In Proceedings of the Human Language Technology Conference of the North American", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 420, 505, 434 ], "spans": [ { "bbox": [ 116, 420, 505, 434 ], "score": 1.0, "content": "Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pp. 169–176,", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 430, 505, 446 ], "spans": [ { "bbox": [ 115, 430, 505, 446 ], "score": 1.0, "content": "Boston, Massachusetts, USA, May 2 - May 7 2004. Association for Computational Linguistics.", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 442, 342, 457 ], "spans": [ { "bbox": [ 115, 442, 342, 457 ], "score": 1.0, "content": "URL https://aclanthology.org/N04-1022.", "type": "text" } ], "index": 25 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 462, 505, 507 ], "lines": [ { "bbox": [ 106, 463, 505, 475 ], "spans": [ { "bbox": [ 106, 463, 505, 475 ], "score": 1.0, "content": "Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte,", "type": "text" } ], "index": 26 }, { "bbox": [ 117, 474, 505, 486 ], "spans": [ { "bbox": [ 117, 474, 505, 486 ], "score": 1.0, "content": "Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, et al. On generative", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 485, 505, 498 ], "spans": [ { "bbox": [ 116, 485, 505, 498 ], "score": 1.0, "content": "spoken language modeling from raw audio. Transactions of the Association for Computational", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 496, 248, 507 ], "spans": [ { "bbox": [ 116, 496, 248, 507 ], "score": 1.0, "content": "Linguistics, 9:1336–1354, 2021.", "type": "text" } ], "index": 29 } ], "index": 27.5 }, { "type": "text", "bbox": [ 106, 515, 503, 537 ], "lines": [ { "bbox": [ 105, 514, 505, 529 ], "spans": [ { "bbox": [ 105, 514, 505, 529 ], "score": 1.0, "content": "Chia-ying Lee and James Glass. A nonparametric bayesian approach to acoustic model discovery.", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 525, 165, 538 ], "spans": [ { "bbox": [ 115, 525, 165, 538 ], "score": 1.0, "content": "ACL, 2012.", "type": "text" } ], "index": 31 } ], "index": 30.5 }, { "type": "text", "bbox": [ 106, 545, 504, 569 ], "lines": [ { "bbox": [ 106, 545, 505, 559 ], "spans": [ { "bbox": [ 106, 545, 505, 559 ], "score": 1.0, "content": "Chia-ying Lee, Timothy J O’donnell, and James Glass. Unsupervised lexicon discovery from acous-", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 556, 481, 569 ], "spans": [ { "bbox": [ 115, 556, 481, 569 ], "score": 1.0, "content": "tic input. Transactions of the Association for Computational Linguistics, 3:389–403, 2015.", "type": "text" } ], "index": 33 } ], "index": 32.5 }, { "type": "text", "bbox": [ 105, 576, 503, 599 ], "lines": [ { "bbox": [ 105, 574, 505, 591 ], "spans": [ { "bbox": [ 105, 574, 505, 591 ], "score": 1.0, "content": "Bowen Li, Lili Mou, and Frank Keller. An imitation learning approach to unsupervised parsing.", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 587, 165, 599 ], "spans": [ { "bbox": [ 115, 587, 165, 599 ], "score": 1.0, "content": "ACL, 2019.", "type": "text" } ], "index": 35 } ], "index": 34.5 }, { "type": "text", "bbox": [ 107, 606, 505, 641 ], "lines": [ { "bbox": [ 106, 607, 505, 619 ], "spans": [ { "bbox": [ 106, 607, 505, 619 ], "score": 1.0, "content": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 617, 505, 630 ], "spans": [ { "bbox": [ 115, 617, 505, 630 ], "score": 1.0, "content": "Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ´ European", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 628, 362, 642 ], "spans": [ { "bbox": [ 116, 628, 362, 642 ], "score": 1.0, "content": "conference on computer vision, pp. 740–755. Springer, 2014.", "type": "text" } ], "index": 38 } ], "index": 37 }, { "type": "text", "bbox": [ 105, 648, 504, 671 ], "lines": [ { "bbox": [ 105, 647, 504, 662 ], "spans": [ { "bbox": [ 105, 647, 504, 662 ], "score": 1.0, "content": "Paria Jamshid Lou, Yufei Wang, and Mark Johnson. Neural constituency parsing of speech tran-", "type": "text" } ], "index": 39 }, { "bbox": [ 115, 660, 231, 671 ], "spans": [ { "bbox": [ 115, 660, 231, 671 ], "score": 1.0, "content": "scripts. NAACL-HLT, 2019.", "type": "text" } ], "index": 40 } ], "index": 39.5 }, { "type": "text", "bbox": [ 106, 678, 503, 702 ], "lines": [ { "bbox": [ 106, 678, 505, 691 ], "spans": [ { "bbox": [ 106, 678, 505, 691 ], "score": 1.0, "content": "Jana M Mason. When do children begin to read: An exploration of four year old children’s letter", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 689, 448, 702 ], "spans": [ { "bbox": [ 115, 689, 448, 702 ], "score": 1.0, "content": "and word reading competencies. Reading Research Quarterly, pp. 203–227, 1980.", "type": "text" } ], "index": 42 } ], "index": 41.5 }, { "type": "text", "bbox": [ 108, 709, 502, 732 ], "lines": [ { "bbox": [ 105, 707, 504, 723 ], "spans": [ { "bbox": [ 105, 707, 504, 723 ], "score": 1.0, "content": "Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger.", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 720, 473, 732 ], "spans": [ { "bbox": [ 116, 720, 473, 732 ], "score": 1.0, "content": "Montreal forced aligner: Trainable text-speech alignment using kaldi. Interspeech, 2017.", "type": "text" } ], "index": 44 } ], "index": 43.5 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "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" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 106, 82, 505, 95 ], "spans": [ { "bbox": [ 106, 82, 505, 95 ], "score": 1.0, "content": "Eugene Kharitonov, Ann Lee, Adam Polyak, Yossi Adi, Jade Copet, Kushal Lakhotia, Tu-Anh", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 504, 107 ], "spans": [ { "bbox": [ 115, 93, 504, 107 ], "score": 1.0, "content": "Nguyen, Morgane Riviere, Abdelrahman Mohamed, Emmanuel Dupoux, et al. Text-free prosody- `", "type": "text" } ], "index": 1 }, { "bbox": [ 116, 105, 346, 117 ], "spans": [ { "bbox": [ 116, 105, 346, 117 ], "score": 1.0, "content": "aware generative spoken language modeling. ACL, 2022.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 82, 505, 117 ] }, { "type": "text", "bbox": [ 108, 123, 504, 157 ], "lines": [ { "bbox": [ 105, 122, 506, 137 ], "spans": [ { "bbox": [ 105, 122, 506, 137 ], "score": 1.0, "content": "Khazar Khorrami and Okko Johannes Ras¨ anen. Can phones, syllables, and words emerge as ¨", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 135, 505, 147 ], "spans": [ { "bbox": [ 116, 135, 505, 147 ], "score": 1.0, "content": "side-products of cross-situational audiovisual learning? - a computational investigation. ArXiv,", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 146, 209, 157 ], "spans": [ { "bbox": [ 116, 146, 209, 157 ], "score": 1.0, "content": "abs/2109.14200, 2021.", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 105, 122, 506, 157 ] }, { "type": "text", "bbox": [ 107, 165, 503, 188 ], "lines": [ { "bbox": [ 106, 164, 505, 179 ], "spans": [ { "bbox": [ 106, 164, 505, 179 ], "score": 1.0, "content": "Yoon Kim, Chris Dyer, and Alexander M Rush. Compound probabilistic context-free grammars for", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 177, 250, 188 ], "spans": [ { "bbox": [ 115, 177, 250, 188 ], "score": 1.0, "content": "grammar induction. ACL, 2019a.", "type": "text" } ], "index": 7 } ], "index": 6.5, "bbox_fs": [ 106, 164, 505, 188 ] }, { "type": "text", "bbox": [ 107, 195, 503, 219 ], "lines": [ { "bbox": [ 106, 195, 504, 209 ], "spans": [ { "bbox": [ 106, 195, 504, 209 ], "score": 1.0, "content": "Yoon Kim, Alexander M Rush, Lei Yu, Adhiguna Kuncoro, Chris Dyer, and Gabor Melis. Unsuper- ´", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 207, 374, 219 ], "spans": [ { "bbox": [ 116, 207, 374, 219 ], "score": 1.0, "content": "vised recurrent neural network grammars. NAACL-HLT, 2019b.", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 106, 195, 504, 219 ] }, { "type": "text", "bbox": [ 106, 226, 502, 249 ], "lines": [ { "bbox": [ 105, 226, 504, 240 ], "spans": [ { "bbox": [ 105, 226, 504, 240 ], "score": 1.0, "content": "Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 237, 443, 250 ], "spans": [ { "bbox": [ 116, 237, 443, 250 ], "score": 1.0, "content": "with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014.", "type": "text" } ], "index": 11 } ], "index": 10.5, "bbox_fs": [ 105, 226, 504, 250 ] }, { "type": "text", "bbox": [ 106, 257, 491, 269 ], "lines": [ { "bbox": [ 105, 255, 491, 271 ], "spans": [ { "bbox": [ 105, 255, 491, 271 ], "score": 1.0, "content": "Nikita Kitaev and Dan Klein. Constituency parsing with a self-attentive encoder. In ACL, 2018.", "type": "text" } ], "index": 12 } ], "index": 12, "bbox_fs": [ 105, 255, 491, 271 ] }, { "type": "text", "bbox": [ 106, 276, 503, 299 ], "lines": [ { "bbox": [ 105, 275, 505, 290 ], "spans": [ { "bbox": [ 105, 275, 505, 290 ], "score": 1.0, "content": "Dan Klein and Christopher D Manning. A generative constituent-context model for improved gram-", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 288, 226, 299 ], "spans": [ { "bbox": [ 116, 288, 226, 299 ], "score": 1.0, "content": "mar induction. ACL, 2002.", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 105, 275, 505, 299 ] }, { "type": "text", "bbox": [ 107, 307, 504, 330 ], "lines": [ { "bbox": [ 105, 307, 506, 320 ], "spans": [ { "bbox": [ 105, 307, 506, 320 ], "score": 1.0, "content": "Dan Klein and Christopher D Manning. Corpus-based induction of syntactic structure: Models of", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 318, 286, 330 ], "spans": [ { "bbox": [ 116, 318, 286, 330 ], "score": 1.0, "content": "dependency and constituency. ACL, 2004.", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 105, 307, 506, 330 ] }, { "type": "text", "bbox": [ 106, 337, 505, 361 ], "lines": [ { "bbox": [ 106, 336, 505, 351 ], "spans": [ { "bbox": [ 106, 336, 505, 351 ], "score": 1.0, "content": "Arne Kohn, Timo Baumann, and Oskar D ¨ orfler. An empirical analysis of the correlation of syntax ¨", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 349, 247, 361 ], "spans": [ { "bbox": [ 115, 349, 247, 361 ], "score": 1.0, "content": "and prosody. Interspeech, 2018.", "type": "text" } ], "index": 18 } ], "index": 17.5, "bbox_fs": [ 106, 336, 505, 361 ] }, { "type": "text", "bbox": [ 105, 368, 505, 391 ], "lines": [ { "bbox": [ 106, 368, 505, 381 ], "spans": [ { "bbox": [ 106, 368, 505, 381 ], "score": 1.0, "content": "Noriyuki Kojima, Hadar Averbuch-Elor, Alexander M Rush, and Yoav Artzi. What is learned in", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 379, 344, 392 ], "spans": [ { "bbox": [ 116, 379, 344, 392 ], "score": 1.0, "content": "visually grounded neural syntax acquisition. ACL, 2020.", "type": "text" } ], "index": 20 } ], "index": 19.5, "bbox_fs": [ 106, 368, 505, 392 ] }, { "type": "text", "bbox": [ 107, 398, 505, 455 ], "lines": [ { "bbox": [ 106, 399, 505, 411 ], "spans": [ { "bbox": [ 106, 399, 505, 411 ], "score": 1.0, "content": "Shankar Kumar and William Byrne. Minimum Bayes-risk decoding for statistical machine transla-", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 410, 505, 422 ], "spans": [ { "bbox": [ 116, 410, 505, 422 ], "score": 1.0, "content": "tion. In Proceedings of the Human Language Technology Conference of the North American", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 420, 505, 434 ], "spans": [ { "bbox": [ 116, 420, 505, 434 ], "score": 1.0, "content": "Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pp. 169–176,", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 430, 505, 446 ], "spans": [ { "bbox": [ 115, 430, 505, 446 ], "score": 1.0, "content": "Boston, Massachusetts, USA, May 2 - May 7 2004. Association for Computational Linguistics.", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 442, 342, 457 ], "spans": [ { "bbox": [ 115, 442, 342, 457 ], "score": 1.0, "content": "URL https://aclanthology.org/N04-1022.", "type": "text" } ], "index": 25 } ], "index": 23, "bbox_fs": [ 106, 399, 505, 457 ] }, { "type": "text", "bbox": [ 107, 462, 505, 507 ], "lines": [ { "bbox": [ 106, 463, 505, 475 ], "spans": [ { "bbox": [ 106, 463, 505, 475 ], "score": 1.0, "content": "Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte,", "type": "text" } ], "index": 26 }, { "bbox": [ 117, 474, 505, 486 ], "spans": [ { "bbox": [ 117, 474, 505, 486 ], "score": 1.0, "content": "Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, et al. On generative", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 485, 505, 498 ], "spans": [ { "bbox": [ 116, 485, 505, 498 ], "score": 1.0, "content": "spoken language modeling from raw audio. Transactions of the Association for Computational", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 496, 248, 507 ], "spans": [ { "bbox": [ 116, 496, 248, 507 ], "score": 1.0, "content": "Linguistics, 9:1336–1354, 2021.", "type": "text" } ], "index": 29 } ], "index": 27.5, "bbox_fs": [ 106, 463, 505, 507 ] }, { "type": "text", "bbox": [ 106, 515, 503, 537 ], "lines": [ { "bbox": [ 105, 514, 505, 529 ], "spans": [ { "bbox": [ 105, 514, 505, 529 ], "score": 1.0, "content": "Chia-ying Lee and James Glass. A nonparametric bayesian approach to acoustic model discovery.", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 525, 165, 538 ], "spans": [ { "bbox": [ 115, 525, 165, 538 ], "score": 1.0, "content": "ACL, 2012.", "type": "text" } ], "index": 31 } ], "index": 30.5, "bbox_fs": [ 105, 514, 505, 538 ] }, { "type": "text", "bbox": [ 106, 545, 504, 569 ], "lines": [ { "bbox": [ 106, 545, 505, 559 ], "spans": [ { "bbox": [ 106, 545, 505, 559 ], "score": 1.0, "content": "Chia-ying Lee, Timothy J O’donnell, and James Glass. Unsupervised lexicon discovery from acous-", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 556, 481, 569 ], "spans": [ { "bbox": [ 115, 556, 481, 569 ], "score": 1.0, "content": "tic input. Transactions of the Association for Computational Linguistics, 3:389–403, 2015.", "type": "text" } ], "index": 33 } ], "index": 32.5, "bbox_fs": [ 106, 545, 505, 569 ] }, { "type": "text", "bbox": [ 105, 576, 503, 599 ], "lines": [ { "bbox": [ 105, 574, 505, 591 ], "spans": [ { "bbox": [ 105, 574, 505, 591 ], "score": 1.0, "content": "Bowen Li, Lili Mou, and Frank Keller. An imitation learning approach to unsupervised parsing.", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 587, 165, 599 ], "spans": [ { "bbox": [ 115, 587, 165, 599 ], "score": 1.0, "content": "ACL, 2019.", "type": "text" } ], "index": 35 } ], "index": 34.5, "bbox_fs": [ 105, 574, 505, 599 ] }, { "type": "text", "bbox": [ 107, 606, 505, 641 ], "lines": [ { "bbox": [ 106, 607, 505, 619 ], "spans": [ { "bbox": [ 106, 607, 505, 619 ], "score": 1.0, "content": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 617, 505, 630 ], "spans": [ { "bbox": [ 115, 617, 505, 630 ], "score": 1.0, "content": "Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ´ European", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 628, 362, 642 ], "spans": [ { "bbox": [ 116, 628, 362, 642 ], "score": 1.0, "content": "conference on computer vision, pp. 740–755. Springer, 2014.", "type": "text" } ], "index": 38 } ], "index": 37, "bbox_fs": [ 106, 607, 505, 642 ] }, { "type": "text", "bbox": [ 105, 648, 504, 671 ], "lines": [ { "bbox": [ 105, 647, 504, 662 ], "spans": [ { "bbox": [ 105, 647, 504, 662 ], "score": 1.0, "content": "Paria Jamshid Lou, Yufei Wang, and Mark Johnson. Neural constituency parsing of speech tran-", "type": "text" } ], "index": 39 }, { "bbox": [ 115, 660, 231, 671 ], "spans": [ { "bbox": [ 115, 660, 231, 671 ], "score": 1.0, "content": "scripts. NAACL-HLT, 2019.", "type": "text" } ], "index": 40 } ], "index": 39.5, "bbox_fs": [ 105, 647, 504, 671 ] }, { "type": "text", "bbox": [ 106, 678, 503, 702 ], "lines": [ { "bbox": [ 106, 678, 505, 691 ], "spans": [ { "bbox": [ 106, 678, 505, 691 ], "score": 1.0, "content": "Jana M Mason. When do children begin to read: An exploration of four year old children’s letter", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 689, 448, 702 ], "spans": [ { "bbox": [ 115, 689, 448, 702 ], "score": 1.0, "content": "and word reading competencies. Reading Research Quarterly, pp. 203–227, 1980.", "type": "text" } ], "index": 42 } ], "index": 41.5, "bbox_fs": [ 106, 678, 505, 702 ] }, { "type": "text", "bbox": [ 108, 709, 502, 732 ], "lines": [ { "bbox": [ 105, 707, 504, 723 ], "spans": [ { "bbox": [ 105, 707, 504, 723 ], "score": 1.0, "content": "Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger.", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 720, 473, 732 ], "spans": [ { "bbox": [ 116, 720, 473, 732 ], "score": 1.0, "content": "Montreal forced aligner: Trainable text-speech alignment using kaldi. Interspeech, 2017.", "type": "text" } ], "index": 44 } ], "index": 43.5, "bbox_fs": [ 105, 707, 504, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 503, 116 ], "lines": [ { "bbox": [ 106, 82, 504, 95 ], "spans": [ { "bbox": [ 106, 82, 504, 95 ], "score": 1.0, "content": "Fergus McInnes and Sharon Goldwater. Unsupervised extraction of recurring words from infant-", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 93, 504, 106 ], "spans": [ { "bbox": [ 116, 93, 504, 106 ], "score": 1.0, "content": "directed speech. In Proceedings of the Annual Meeting of the Cognitive Science Society, vol-", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 104, 177, 116 ], "spans": [ { "bbox": [ 115, 104, 177, 116 ], "score": 1.0, "content": "ume 33, 2011.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 106, 123, 496, 135 ], "lines": [ { "bbox": [ 105, 121, 498, 138 ], "spans": [ { "bbox": [ 105, 121, 498, 138 ], "score": 1.0, "content": "Grzegorz Chrupała Mitja Nikolaus, Afra Alishahi. Learning english with peppa pig. TACL, 2022.", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 106, 142, 504, 165 ], "lines": [ { "bbox": [ 105, 141, 505, 156 ], "spans": [ { "bbox": [ 105, 141, 505, 156 ], "score": 1.0, "content": "Letitia Naigles. Children use syntax to learn verb meanings. Journal of child language, 17(2):", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 153, 181, 165 ], "spans": [ { "bbox": [ 116, 153, 181, 165 ], "score": 1.0, "content": "357–374, 1990.", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "text", "bbox": [ 107, 172, 505, 217 ], "lines": [ { "bbox": [ 106, 172, 506, 185 ], "spans": [ { "bbox": [ 106, 172, 506, 185 ], "score": 1.0, "content": "Tu Nguyen, Maureen de Seyssel, Patricia Roz’e, Morgane Riviere, Evgeny Kharitonov, Alexei `", "type": "text" } ], "index": 6 }, { "bbox": [ 116, 183, 505, 195 ], "spans": [ { "bbox": [ 116, 183, 505, 195 ], "score": 1.0, "content": "Baevski, Ewan Dunbar, and Emmanuel Dupoux. The zero resource speech benchmark 2021:", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 194, 505, 207 ], "spans": [ { "bbox": [ 115, 194, 505, 207 ], "score": 1.0, "content": "Metrics and baselines for unsupervised spoken language modeling. Self-Supervised Learning for", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 205, 343, 217 ], "spans": [ { "bbox": [ 115, 205, 343, 217 ], "score": 1.0, "content": "Speech and Audio Processing NeurIPS Workshop, 2020.", "type": "text" } ], "index": 9 } ], "index": 7.5 }, { "type": "text", "bbox": [ 107, 224, 503, 247 ], "lines": [ { "bbox": [ 106, 224, 505, 237 ], "spans": [ { "bbox": [ 106, 224, 505, 237 ], "score": 1.0, "content": "Kayode Olaleye and Herman Kamper. Attention-based keyword localisation in speech using visual", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 235, 250, 247 ], "spans": [ { "bbox": [ 116, 235, 250, 247 ], "score": 1.0, "content": "grounding. In Interspeech, 2021.", "type": "text" } ], "index": 11 } ], "index": 10.5 }, { "type": "text", "bbox": [ 107, 254, 503, 277 ], "lines": [ { "bbox": [ 106, 254, 505, 267 ], "spans": [ { "bbox": [ 106, 254, 505, 267 ], "score": 1.0, "content": "Alex S Park and James R Glass. Unsupervised pattern discovery in speech. IEEE Transactions on", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 265, 376, 278 ], "spans": [ { "bbox": [ 116, 265, 376, 278 ], "score": 1.0, "content": "Audio, Speech, and Language Processing, 16(1):186–197, 2007.", "type": "text" } ], "index": 13 } ], "index": 12.5 }, { "type": "text", "bbox": [ 107, 283, 505, 318 ], "lines": [ { "bbox": [ 105, 283, 505, 297 ], "spans": [ { "bbox": [ 105, 283, 505, 297 ], "score": 1.0, "content": "Puyuan Peng and David Harwath. Self-supervised representation learning for speech using visual", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 295, 505, 308 ], "spans": [ { "bbox": [ 115, 295, 505, 308 ], "score": 1.0, "content": "grounding and masked language modeling. Self-Supervised Learning for Speech and Audio Pro-", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 306, 257, 318 ], "spans": [ { "bbox": [ 116, 306, 257, 318 ], "score": 1.0, "content": "cessing Workshop at AAAI, 2022a.", "type": "text" } ], "index": 16 } ], "index": 15 }, { "type": "text", "bbox": [ 106, 324, 504, 348 ], "lines": [ { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "Puyuan Peng and David Harwath. Word discovery in visually grounded, self-supervised speech", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 336, 232, 348 ], "spans": [ { "bbox": [ 115, 336, 232, 348 ], "score": 1.0, "content": "models. Interspeech, 2022b.", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "text", "bbox": [ 107, 354, 505, 388 ], "lines": [ { "bbox": [ 106, 354, 505, 368 ], "spans": [ { "bbox": [ 106, 354, 505, 368 ], "score": 1.0, "content": "Adam Polyak, Yossi Adi, Jade Copet, Eugene Kharitonov, Kushal Lakhotia, Wei-Ning Hsu, Ab-", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 366, 505, 379 ], "spans": [ { "bbox": [ 116, 366, 505, 379 ], "score": 1.0, "content": "delrahman Mohamed, and Emmanuel Dupoux. Speech resynthesis from discrete disentangled", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 377, 322, 389 ], "spans": [ { "bbox": [ 117, 377, 322, 389 ], "score": 1.0, "content": "self-supervised representations. Interspeech, 2021.", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 396, 504, 419 ], "lines": [ { "bbox": [ 105, 394, 505, 409 ], "spans": [ { "bbox": [ 105, 394, 505, 409 ], "score": 1.0, "content": "Adrien Pupier, Maximin Coavoux, Benjamin Lecouteux, and Jer´ ome Goulian. End-to-end depen- ˆ", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 407, 335, 419 ], "spans": [ { "bbox": [ 115, 407, 335, 419 ], "score": 1.0, "content": "dency parsing of spoken french. In Interspeech, 2022.", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "text", "bbox": [ 107, 425, 505, 459 ], "lines": [ { "bbox": [ 105, 424, 506, 440 ], "spans": [ { "bbox": [ 105, 424, 506, 440 ], "score": 1.0, "content": "Brian Roark, Mary Harper, Eugene Charniak, Bonnie Dorr, Mark Johnson, Jeremy G Kahn, Yang", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 437, 505, 450 ], "spans": [ { "bbox": [ 115, 437, 505, 450 ], "score": 1.0, "content": "Liu, Mari Ostendorf, John Hale, Anna Krasnyanskaya, et al. Sparseval: Evaluation metrics for", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 448, 236, 460 ], "spans": [ { "bbox": [ 115, 448, 236, 460 ], "score": 1.0, "content": "parsing speech. LREC, 2006.", "type": "text" } ], "index": 26 } ], "index": 25 }, { "type": "text", "bbox": [ 106, 466, 504, 489 ], "lines": [ { "bbox": [ 106, 466, 504, 479 ], "spans": [ { "bbox": [ 106, 466, 504, 479 ], "score": 1.0, "content": "Deb K Roy and Alex P Pentland. Learning words from sights and sounds: A computational model.", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 478, 281, 489 ], "spans": [ { "bbox": [ 116, 478, 281, 489 ], "score": 1.0, "content": "Cognitive science, 26(1):113–146, 2002.", "type": "text" } ], "index": 28 } ], "index": 27.5 }, { "type": "text", "bbox": [ 106, 496, 504, 520 ], "lines": [ { "bbox": [ 105, 494, 505, 511 ], "spans": [ { "bbox": [ 105, 494, 505, 511 ], "score": 1.0, "content": "Yikang Shen, Zhouhan Lin, Chin-Wei Huang, and Aaron Courville. Neural language modeling by", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 508, 312, 520 ], "spans": [ { "bbox": [ 114, 508, 312, 520 ], "score": 1.0, "content": "jointly learning syntax and lexicon. ICLR, 2018.", "type": "text" } ], "index": 30 } ], "index": 29.5 }, { "type": "text", "bbox": [ 105, 526, 504, 550 ], "lines": [ { "bbox": [ 105, 524, 505, 541 ], "spans": [ { "bbox": [ 105, 524, 505, 541 ], "score": 1.0, "content": "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. Ordered neurons: Integrating", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 538, 353, 549 ], "spans": [ { "bbox": [ 115, 538, 353, 549 ], "score": 1.0, "content": "tree structures into recurrent neural networks. ICLR, 2019.", "type": "text" } ], "index": 32 } ], "index": 31.5 }, { "type": "text", "bbox": [ 105, 556, 503, 580 ], "lines": [ { "bbox": [ 105, 556, 505, 569 ], "spans": [ { "bbox": [ 105, 556, 505, 569 ], "score": 1.0, "content": "Freda Shi, Daniel Fried, Marjan Ghazvininejad, Luke Zettlemoyer, and Sida I Wang. Natural lan-", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 567, 447, 580 ], "spans": [ { "bbox": [ 115, 567, 447, 580 ], "score": 1.0, "content": "guage to code translation with execution. arXiv preprint arXiv:2204.11454, 2022.", "type": "text" } ], "index": 34 } ], "index": 33.5 }, { "type": "text", "bbox": [ 105, 586, 504, 609 ], "lines": [ { "bbox": [ 105, 585, 505, 600 ], "spans": [ { "bbox": [ 105, 585, 505, 600 ], "score": 1.0, "content": "Haoyue Shi, Jiayuan Mao, Kevin Gimpel, and Karen Livescu. Visually grounded neural syntax", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 598, 215, 609 ], "spans": [ { "bbox": [ 115, 598, 215, 609 ], "score": 1.0, "content": "acquisition. ACL, 2019.", "type": "text" } ], "index": 36 } ], "index": 35.5 }, { "type": "text", "bbox": [ 104, 616, 504, 640 ], "lines": [ { "bbox": [ 105, 615, 504, 630 ], "spans": [ { "bbox": [ 105, 615, 504, 630 ], "score": 1.0, "content": "Haoyue Shi, Karen Livescu, and Kevin Gimpel. On the role of supervision in unsupervised con-", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 627, 438, 640 ], "spans": [ { "bbox": [ 115, 627, 438, 640 ], "score": 1.0, "content": "stituency parsing. In EMNLP. Association for Computational Linguistics, 2020.", "type": "text" } ], "index": 38 } ], "index": 37.5 }, { "type": "text", "bbox": [ 107, 646, 505, 702 ], "lines": [ { "bbox": [ 106, 646, 505, 660 ], "spans": [ { "bbox": [ 106, 646, 505, 660 ], "score": 1.0, "content": "Valentin I. Spitkovsky, Hiyan Alshawi, and Daniel Jurafsky. Breaking out of local optima with", "type": "text" } ], "index": 39 }, { "bbox": [ 116, 657, 505, 671 ], "spans": [ { "bbox": [ 116, 657, 505, 671 ], "score": 1.0, "content": "count transforms and model recombination: A study in grammar induction. In Proceedings of the", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 667, 505, 682 ], "spans": [ { "bbox": [ 115, 667, 505, 682 ], "score": 1.0, "content": "2013 Conference on Empirical Methods in Natural Language Processing, pp. 1983–1995, Seattle,", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 677, 505, 694 ], "spans": [ { "bbox": [ 115, 677, 505, 694 ], "score": 1.0, "content": "Washington, USA, October 2013. Association for Computational Linguistics. URL https:", "type": "text" } ], "index": 42 }, { "bbox": [ 117, 690, 283, 703 ], "spans": [ { "bbox": [ 117, 690, 283, 703 ], "score": 1.0, "content": "//aclanthology.org/D13-1204.", "type": "text" } ], "index": 43 } ], "index": 41 }, { "type": "text", "bbox": [ 106, 709, 502, 732 ], "lines": [ { "bbox": [ 105, 708, 504, 723 ], "spans": [ { "bbox": [ 105, 708, 504, 723 ], "score": 1.0, "content": "Zheng-Hua Tan, Najim Dehak, et al. rvad: An unsupervised segment-based robust voice activity", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 720, 378, 732 ], "spans": [ { "bbox": [ 115, 720, 378, 732 ], "score": 1.0, "content": "detection method. Computer speech & language, 59:1–21, 2020.", "type": "text" } ], "index": 45 } ], "index": 44.5 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "13", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 503, 116 ], "lines": [ { "bbox": [ 106, 82, 504, 95 ], "spans": [ { "bbox": [ 106, 82, 504, 95 ], "score": 1.0, "content": "Fergus McInnes and Sharon Goldwater. Unsupervised extraction of recurring words from infant-", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 93, 504, 106 ], "spans": [ { "bbox": [ 116, 93, 504, 106 ], "score": 1.0, "content": "directed speech. In Proceedings of the Annual Meeting of the Cognitive Science Society, vol-", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 104, 177, 116 ], "spans": [ { "bbox": [ 115, 104, 177, 116 ], "score": 1.0, "content": "ume 33, 2011.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 82, 504, 116 ] }, { "type": "text", "bbox": [ 106, 123, 496, 135 ], "lines": [ { "bbox": [ 105, 121, 498, 138 ], "spans": [ { "bbox": [ 105, 121, 498, 138 ], "score": 1.0, "content": "Grzegorz Chrupała Mitja Nikolaus, Afra Alishahi. Learning english with peppa pig. TACL, 2022.", "type": "text" } ], "index": 3 } ], "index": 3, "bbox_fs": [ 105, 121, 498, 138 ] }, { "type": "text", "bbox": [ 106, 142, 504, 165 ], "lines": [ { "bbox": [ 105, 141, 505, 156 ], "spans": [ { "bbox": [ 105, 141, 505, 156 ], "score": 1.0, "content": "Letitia Naigles. Children use syntax to learn verb meanings. Journal of child language, 17(2):", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 153, 181, 165 ], "spans": [ { "bbox": [ 116, 153, 181, 165 ], "score": 1.0, "content": "357–374, 1990.", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 105, 141, 505, 165 ] }, { "type": "text", "bbox": [ 107, 172, 505, 217 ], "lines": [ { "bbox": [ 106, 172, 506, 185 ], "spans": [ { "bbox": [ 106, 172, 506, 185 ], "score": 1.0, "content": "Tu Nguyen, Maureen de Seyssel, Patricia Roz’e, Morgane Riviere, Evgeny Kharitonov, Alexei `", "type": "text" } ], "index": 6 }, { "bbox": [ 116, 183, 505, 195 ], "spans": [ { "bbox": [ 116, 183, 505, 195 ], "score": 1.0, "content": "Baevski, Ewan Dunbar, and Emmanuel Dupoux. The zero resource speech benchmark 2021:", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 194, 505, 207 ], "spans": [ { "bbox": [ 115, 194, 505, 207 ], "score": 1.0, "content": "Metrics and baselines for unsupervised spoken language modeling. Self-Supervised Learning for", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 205, 343, 217 ], "spans": [ { "bbox": [ 115, 205, 343, 217 ], "score": 1.0, "content": "Speech and Audio Processing NeurIPS Workshop, 2020.", "type": "text" } ], "index": 9 } ], "index": 7.5, "bbox_fs": [ 106, 172, 506, 217 ] }, { "type": "text", "bbox": [ 107, 224, 503, 247 ], "lines": [ { "bbox": [ 106, 224, 505, 237 ], "spans": [ { "bbox": [ 106, 224, 505, 237 ], "score": 1.0, "content": "Kayode Olaleye and Herman Kamper. Attention-based keyword localisation in speech using visual", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 235, 250, 247 ], "spans": [ { "bbox": [ 116, 235, 250, 247 ], "score": 1.0, "content": "grounding. In Interspeech, 2021.", "type": "text" } ], "index": 11 } ], "index": 10.5, "bbox_fs": [ 106, 224, 505, 247 ] }, { "type": "text", "bbox": [ 107, 254, 503, 277 ], "lines": [ { "bbox": [ 106, 254, 505, 267 ], "spans": [ { "bbox": [ 106, 254, 505, 267 ], "score": 1.0, "content": "Alex S Park and James R Glass. Unsupervised pattern discovery in speech. IEEE Transactions on", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 265, 376, 278 ], "spans": [ { "bbox": [ 116, 265, 376, 278 ], "score": 1.0, "content": "Audio, Speech, and Language Processing, 16(1):186–197, 2007.", "type": "text" } ], "index": 13 } ], "index": 12.5, "bbox_fs": [ 106, 254, 505, 278 ] }, { "type": "text", "bbox": [ 107, 283, 505, 318 ], "lines": [ { "bbox": [ 105, 283, 505, 297 ], "spans": [ { "bbox": [ 105, 283, 505, 297 ], "score": 1.0, "content": "Puyuan Peng and David Harwath. Self-supervised representation learning for speech using visual", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 295, 505, 308 ], "spans": [ { "bbox": [ 115, 295, 505, 308 ], "score": 1.0, "content": "grounding and masked language modeling. Self-Supervised Learning for Speech and Audio Pro-", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 306, 257, 318 ], "spans": [ { "bbox": [ 116, 306, 257, 318 ], "score": 1.0, "content": "cessing Workshop at AAAI, 2022a.", "type": "text" } ], "index": 16 } ], "index": 15, "bbox_fs": [ 105, 283, 505, 318 ] }, { "type": "text", "bbox": [ 106, 324, 504, 348 ], "lines": [ { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "Puyuan Peng and David Harwath. Word discovery in visually grounded, self-supervised speech", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 336, 232, 348 ], "spans": [ { "bbox": [ 115, 336, 232, 348 ], "score": 1.0, "content": "models. Interspeech, 2022b.", "type": "text" } ], "index": 18 } ], "index": 17.5, "bbox_fs": [ 105, 324, 505, 348 ] }, { "type": "text", "bbox": [ 107, 354, 505, 388 ], "lines": [ { "bbox": [ 106, 354, 505, 368 ], "spans": [ { "bbox": [ 106, 354, 505, 368 ], "score": 1.0, "content": "Adam Polyak, Yossi Adi, Jade Copet, Eugene Kharitonov, Kushal Lakhotia, Wei-Ning Hsu, Ab-", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 366, 505, 379 ], "spans": [ { "bbox": [ 116, 366, 505, 379 ], "score": 1.0, "content": "delrahman Mohamed, and Emmanuel Dupoux. Speech resynthesis from discrete disentangled", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 377, 322, 389 ], "spans": [ { "bbox": [ 117, 377, 322, 389 ], "score": 1.0, "content": "self-supervised representations. Interspeech, 2021.", "type": "text" } ], "index": 21 } ], "index": 20, "bbox_fs": [ 106, 354, 505, 389 ] }, { "type": "text", "bbox": [ 106, 396, 504, 419 ], "lines": [ { "bbox": [ 105, 394, 505, 409 ], "spans": [ { "bbox": [ 105, 394, 505, 409 ], "score": 1.0, "content": "Adrien Pupier, Maximin Coavoux, Benjamin Lecouteux, and Jer´ ome Goulian. End-to-end depen- ˆ", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 407, 335, 419 ], "spans": [ { "bbox": [ 115, 407, 335, 419 ], "score": 1.0, "content": "dency parsing of spoken french. In Interspeech, 2022.", "type": "text" } ], "index": 23 } ], "index": 22.5, "bbox_fs": [ 105, 394, 505, 419 ] }, { "type": "text", "bbox": [ 107, 425, 505, 459 ], "lines": [ { "bbox": [ 105, 424, 506, 440 ], "spans": [ { "bbox": [ 105, 424, 506, 440 ], "score": 1.0, "content": "Brian Roark, Mary Harper, Eugene Charniak, Bonnie Dorr, Mark Johnson, Jeremy G Kahn, Yang", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 437, 505, 450 ], "spans": [ { "bbox": [ 115, 437, 505, 450 ], "score": 1.0, "content": "Liu, Mari Ostendorf, John Hale, Anna Krasnyanskaya, et al. Sparseval: Evaluation metrics for", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 448, 236, 460 ], "spans": [ { "bbox": [ 115, 448, 236, 460 ], "score": 1.0, "content": "parsing speech. LREC, 2006.", "type": "text" } ], "index": 26 } ], "index": 25, "bbox_fs": [ 105, 424, 506, 460 ] }, { "type": "text", "bbox": [ 106, 466, 504, 489 ], "lines": [ { "bbox": [ 106, 466, 504, 479 ], "spans": [ { "bbox": [ 106, 466, 504, 479 ], "score": 1.0, "content": "Deb K Roy and Alex P Pentland. Learning words from sights and sounds: A computational model.", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 478, 281, 489 ], "spans": [ { "bbox": [ 116, 478, 281, 489 ], "score": 1.0, "content": "Cognitive science, 26(1):113–146, 2002.", "type": "text" } ], "index": 28 } ], "index": 27.5, "bbox_fs": [ 106, 466, 504, 489 ] }, { "type": "text", "bbox": [ 106, 496, 504, 520 ], "lines": [ { "bbox": [ 105, 494, 505, 511 ], "spans": [ { "bbox": [ 105, 494, 505, 511 ], "score": 1.0, "content": "Yikang Shen, Zhouhan Lin, Chin-Wei Huang, and Aaron Courville. Neural language modeling by", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 508, 312, 520 ], "spans": [ { "bbox": [ 114, 508, 312, 520 ], "score": 1.0, "content": "jointly learning syntax and lexicon. ICLR, 2018.", "type": "text" } ], "index": 30 } ], "index": 29.5, "bbox_fs": [ 105, 494, 505, 520 ] }, { "type": "text", "bbox": [ 105, 526, 504, 550 ], "lines": [ { "bbox": [ 105, 524, 505, 541 ], "spans": [ { "bbox": [ 105, 524, 505, 541 ], "score": 1.0, "content": "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. Ordered neurons: Integrating", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 538, 353, 549 ], "spans": [ { "bbox": [ 115, 538, 353, 549 ], "score": 1.0, "content": "tree structures into recurrent neural networks. ICLR, 2019.", "type": "text" } ], "index": 32 } ], "index": 31.5, "bbox_fs": [ 105, 524, 505, 549 ] }, { "type": "text", "bbox": [ 105, 556, 503, 580 ], "lines": [ { "bbox": [ 105, 556, 505, 569 ], "spans": [ { "bbox": [ 105, 556, 505, 569 ], "score": 1.0, "content": "Freda Shi, Daniel Fried, Marjan Ghazvininejad, Luke Zettlemoyer, and Sida I Wang. Natural lan-", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 567, 447, 580 ], "spans": [ { "bbox": [ 115, 567, 447, 580 ], "score": 1.0, "content": "guage to code translation with execution. arXiv preprint arXiv:2204.11454, 2022.", "type": "text" } ], "index": 34 } ], "index": 33.5, "bbox_fs": [ 105, 556, 505, 580 ] }, { "type": "text", "bbox": [ 105, 586, 504, 609 ], "lines": [ { "bbox": [ 105, 585, 505, 600 ], "spans": [ { "bbox": [ 105, 585, 505, 600 ], "score": 1.0, "content": "Haoyue Shi, Jiayuan Mao, Kevin Gimpel, and Karen Livescu. Visually grounded neural syntax", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 598, 215, 609 ], "spans": [ { "bbox": [ 115, 598, 215, 609 ], "score": 1.0, "content": "acquisition. ACL, 2019.", "type": "text" } ], "index": 36 } ], "index": 35.5, "bbox_fs": [ 105, 585, 505, 609 ] }, { "type": "text", "bbox": [ 104, 616, 504, 640 ], "lines": [ { "bbox": [ 105, 615, 504, 630 ], "spans": [ { "bbox": [ 105, 615, 504, 630 ], "score": 1.0, "content": "Haoyue Shi, Karen Livescu, and Kevin Gimpel. On the role of supervision in unsupervised con-", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 627, 438, 640 ], "spans": [ { "bbox": [ 115, 627, 438, 640 ], "score": 1.0, "content": "stituency parsing. In EMNLP. Association for Computational Linguistics, 2020.", "type": "text" } ], "index": 38 } ], "index": 37.5, "bbox_fs": [ 105, 615, 504, 640 ] }, { "type": "text", "bbox": [ 107, 646, 505, 702 ], "lines": [ { "bbox": [ 106, 646, 505, 660 ], "spans": [ { "bbox": [ 106, 646, 505, 660 ], "score": 1.0, "content": "Valentin I. Spitkovsky, Hiyan Alshawi, and Daniel Jurafsky. Breaking out of local optima with", "type": "text" } ], "index": 39 }, { "bbox": [ 116, 657, 505, 671 ], "spans": [ { "bbox": [ 116, 657, 505, 671 ], "score": 1.0, "content": "count transforms and model recombination: A study in grammar induction. In Proceedings of the", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 667, 505, 682 ], "spans": [ { "bbox": [ 115, 667, 505, 682 ], "score": 1.0, "content": "2013 Conference on Empirical Methods in Natural Language Processing, pp. 1983–1995, Seattle,", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 677, 505, 694 ], "spans": [ { "bbox": [ 115, 677, 505, 694 ], "score": 1.0, "content": "Washington, USA, October 2013. Association for Computational Linguistics. URL https:", "type": "text" } ], "index": 42 }, { "bbox": [ 117, 690, 283, 703 ], "spans": [ { "bbox": [ 117, 690, 283, 703 ], "score": 1.0, "content": "//aclanthology.org/D13-1204.", "type": "text" } ], "index": 43 } ], "index": 41, "bbox_fs": [ 106, 646, 505, 703 ] }, { "type": "text", "bbox": [ 106, 709, 502, 732 ], "lines": [ { "bbox": [ 105, 708, 504, 723 ], "spans": [ { "bbox": [ 105, 708, 504, 723 ], "score": 1.0, "content": "Zheng-Hua Tan, Najim Dehak, et al. rvad: An unsupervised segment-based robust voice activity", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 720, 378, 732 ], "spans": [ { "bbox": [ 115, 720, 378, 732 ], "score": 1.0, "content": "detection method. Computer speech & language, 59:1–21, 2020.", "type": "text" } ], "index": 45 } ], "index": 44.5, "bbox_fs": [ 105, 708, 504, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Trang Tran and Mari Ostendorf. Assessing the use of prosody in constituency parsing of imperfect", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 94, 240, 106 ], "spans": [ { "bbox": [ 115, 94, 240, 106 ], "score": 1.0, "content": "transcripts. Interspeech, 2021.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 108, 111, 504, 146 ], "lines": [ { "bbox": [ 106, 111, 505, 125 ], "spans": [ { "bbox": [ 106, 111, 505, 125 ], "score": 1.0, "content": "Trang Tran, Shubham Toshniwal, Mohit Bansal, Kevin Gimpel, Karen Livescu, and Mari Osten-", "type": "text" } ], "index": 2 }, { "bbox": [ 116, 122, 505, 136 ], "spans": [ { "bbox": [ 116, 122, 505, 136 ], "score": 1.0, "content": "dorf. Parsing speech: a neural approach to integrating lexical and acoustic-prosodic information.", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 134, 198, 146 ], "spans": [ { "bbox": [ 115, 134, 198, 146 ], "score": 1.0, "content": "NAACL-HLT, 2018.", "type": "text" } ], "index": 4 } ], "index": 3 }, { "type": "text", "bbox": [ 105, 153, 504, 176 ], "lines": [ { "bbox": [ 105, 152, 505, 167 ], "spans": [ { "bbox": [ 105, 152, 505, 167 ], "score": 1.0, "content": "Trang Tran, Jiahong Yuan, Yang Liu, and Mari Ostendorf. On the role of style in parsing speech", "type": "text" } ], "index": 5 }, { "bbox": [ 116, 164, 275, 176 ], "spans": [ { "bbox": [ 116, 164, 275, 176 ], "score": 1.0, "content": "with neural models. Interspeech, 2019.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "text", "bbox": [ 107, 183, 504, 217 ], "lines": [ { "bbox": [ 105, 182, 504, 196 ], "spans": [ { "bbox": [ 105, 182, 504, 196 ], "score": 1.0, "content": "Shubhi Tyagi, Marco Nicolis, Jonas Rohnke, Thomas Drugman, and Jaime Lorenzo-Trueba. Dy-", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 194, 504, 207 ], "spans": [ { "bbox": [ 116, 194, 504, 207 ], "score": 1.0, "content": "namic prosody generation for speech synthesis using linguistics-driven acoustic embedding se-", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 205, 225, 217 ], "spans": [ { "bbox": [ 116, 205, 225, 217 ], "score": 1.0, "content": "lection. Interspeech, 2020.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "text", "bbox": [ 106, 223, 503, 247 ], "lines": [ { "bbox": [ 106, 223, 505, 236 ], "spans": [ { "bbox": [ 106, 223, 505, 236 ], "score": 1.0, "content": "Maarten Versteegh, Roland Thiolliere, Thomas Schatz, Xuan Nga Cao, Xavier Anguera, Aren", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 235, 470, 248 ], "spans": [ { "bbox": [ 116, 235, 470, 248 ], "score": 1.0, "content": "Jansen, and Emmanuel Dupoux. The zero resource speech challenge 2015. ISCA, 2015.", "type": "text" } ], "index": 11 } ], "index": 10.5 }, { "type": "text", "bbox": [ 107, 253, 502, 277 ], "lines": [ { "bbox": [ 105, 252, 505, 267 ], "spans": [ { "bbox": [ 105, 252, 505, 267 ], "score": 1.0, "content": "Michael Wagner and Duane G Watson. Experimental and theoretical advances in prosody: A review.", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 265, 356, 277 ], "spans": [ { "bbox": [ 116, 265, 356, 277 ], "score": 1.0, "content": "Language and cognitive processes, 25(7-9):905–945, 2010.", "type": "text" } ], "index": 13 } ], "index": 12.5 }, { "type": "text", "bbox": [ 107, 283, 503, 307 ], "lines": [ { "bbox": [ 105, 282, 504, 297 ], "spans": [ { "bbox": [ 105, 282, 504, 297 ], "score": 1.0, "content": "Bo Wan, Wenjuan Han, Zilong Zheng, and Tinne Tuytelaars. Unsupervised vision-language gram-", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 294, 356, 307 ], "spans": [ { "bbox": [ 115, 294, 356, 307 ], "score": 1.0, "content": "mar induction with shared structure modeling. ICLR, 2022.", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 106, 313, 504, 336 ], "lines": [ { "bbox": [ 105, 312, 505, 327 ], "spans": [ { "bbox": [ 105, 312, 505, 327 ], "score": 1.0, "content": "Liming Wang and Mark Hasegawa-Johnson. A translation framework for visually grounded spoken", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 324, 248, 336 ], "spans": [ { "bbox": [ 115, 324, 248, 336 ], "score": 1.0, "content": "unit discovery. In ACSSC, 2021.", "type": "text" } ], "index": 17 } ], "index": 16.5 }, { "type": "text", "bbox": [ 106, 343, 506, 378 ], "lines": [ { "bbox": [ 106, 344, 505, 356 ], "spans": [ { "bbox": [ 106, 344, 505, 356 ], "score": 1.0, "content": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 354, 504, 367 ], "spans": [ { "bbox": [ 115, 354, 504, 367 ], "score": 1.0, "content": "learning. Machine Learning, 8(3-4):229–256, 1992. URL https://link.springer.com/", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 366, 325, 378 ], "spans": [ { "bbox": [ 116, 366, 325, 378 ], "score": 1.0, "content": "content/pdf/10.1007/BF00992696.pdf.", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "text", "bbox": [ 106, 384, 504, 407 ], "lines": [ { "bbox": [ 105, 383, 506, 397 ], "spans": [ { "bbox": [ 105, 383, 506, 397 ], "score": 1.0, "content": "Daniel H Younger. Recognition and parsing of context-free languages in time n3. Information and", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 396, 239, 406 ], "spans": [ { "bbox": [ 116, 396, 239, 406 ], "score": 1.0, "content": "control, 10(2):189–208, 1967.", "type": "text" } ], "index": 22 } ], "index": 21.5 }, { "type": "text", "bbox": [ 106, 414, 505, 448 ], "lines": [ { "bbox": [ 105, 414, 505, 428 ], "spans": [ { "bbox": [ 105, 414, 505, 428 ], "score": 1.0, "content": "Hao Zhang and Daniel Gildea. Efficient multi-pass decoding for synchronous context free gram-", "type": "text" } ], "index": 23 }, { "bbox": [ 116, 425, 505, 438 ], "spans": [ { "bbox": [ 116, 425, 505, 438 ], "score": 1.0, "content": "mars. In Proceedings of ACL-08: HLT, pp. 209–217, Columbus, Ohio, June 2008. Association", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 435, 469, 450 ], "spans": [ { "bbox": [ 115, 435, 469, 450 ], "score": 1.0, "content": "for Computational Linguistics. URL https://aclanthology.org/P08-1025.", "type": "text" } ], "index": 25 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 454, 504, 478 ], "lines": [ { "bbox": [ 106, 455, 504, 468 ], "spans": [ { "bbox": [ 106, 455, 504, 468 ], "score": 1.0, "content": "Songyang Zhang, Linfeng Song, Lifeng Jin, Kun Xu, Dong Yu, and Jiebo Luo. Video-aided unsu-", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 466, 317, 478 ], "spans": [ { "bbox": [ 115, 466, 317, 478 ], "score": 1.0, "content": "pervised grammar induction. NAACL-HLT, 2021.", "type": "text" } ], "index": 27 } ], "index": 26.5 }, { "type": "text", "bbox": [ 107, 484, 504, 508 ], "lines": [ { "bbox": [ 105, 483, 505, 498 ], "spans": [ { "bbox": [ 105, 483, 505, 498 ], "score": 1.0, "content": "Yaodong Zhang. Unsupervised speech processing with applications to query-by-example spoken", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 496, 411, 508 ], "spans": [ { "bbox": [ 115, 496, 411, 508 ], "score": 1.0, "content": "term detection. PhD thesis, Massachusetts Institute of Technology, 2013.", "type": "text" } ], "index": 29 } ], "index": 28.5 }, { "type": "text", "bbox": [ 106, 514, 505, 538 ], "lines": [ { "bbox": [ 106, 513, 505, 529 ], "spans": [ { "bbox": [ 106, 513, 505, 529 ], "score": 1.0, "content": "Yaodong Zhang and James R Glass. Unsupervised spoken keyword spotting via segmental dtw on", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 527, 272, 538 ], "spans": [ { "bbox": [ 115, 527, 272, 538 ], "score": 1.0, "content": "gaussian posteriorgrams. ASRU, 2009.", "type": "text" } ], "index": 31 } ], "index": 30.5 }, { "type": "text", "bbox": [ 106, 544, 441, 557 ], "lines": [ { "bbox": [ 106, 543, 441, 558 ], "spans": [ { "bbox": [ 106, 543, 441, 558 ], "score": 1.0, "content": "Yanpeng Zhao and Ivan Titov. Visually grounded compound pcfgs. EMNLP, 2020.", "type": "text" } ], "index": 32 } ], "index": 32 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "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": "14", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Trang Tran and Mari Ostendorf. Assessing the use of prosody in constituency parsing of imperfect", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 94, 240, 106 ], "spans": [ { "bbox": [ 115, 94, 240, 106 ], "score": 1.0, "content": "transcripts. Interspeech, 2021.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 82, 505, 106 ] }, { "type": "text", "bbox": [ 108, 111, 504, 146 ], "lines": [ { "bbox": [ 106, 111, 505, 125 ], "spans": [ { "bbox": [ 106, 111, 505, 125 ], "score": 1.0, "content": "Trang Tran, Shubham Toshniwal, Mohit Bansal, Kevin Gimpel, Karen Livescu, and Mari Osten-", "type": "text" } ], "index": 2 }, { "bbox": [ 116, 122, 505, 136 ], "spans": [ { "bbox": [ 116, 122, 505, 136 ], "score": 1.0, "content": "dorf. Parsing speech: a neural approach to integrating lexical and acoustic-prosodic information.", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 134, 198, 146 ], "spans": [ { "bbox": [ 115, 134, 198, 146 ], "score": 1.0, "content": "NAACL-HLT, 2018.", "type": "text" } ], "index": 4 } ], "index": 3, "bbox_fs": [ 106, 111, 505, 146 ] }, { "type": "text", "bbox": [ 105, 153, 504, 176 ], "lines": [ { "bbox": [ 105, 152, 505, 167 ], "spans": [ { "bbox": [ 105, 152, 505, 167 ], "score": 1.0, "content": "Trang Tran, Jiahong Yuan, Yang Liu, and Mari Ostendorf. On the role of style in parsing speech", "type": "text" } ], "index": 5 }, { "bbox": [ 116, 164, 275, 176 ], "spans": [ { "bbox": [ 116, 164, 275, 176 ], "score": 1.0, "content": "with neural models. Interspeech, 2019.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 105, 152, 505, 176 ] }, { "type": "text", "bbox": [ 107, 183, 504, 217 ], "lines": [ { "bbox": [ 105, 182, 504, 196 ], "spans": [ { "bbox": [ 105, 182, 504, 196 ], "score": 1.0, "content": "Shubhi Tyagi, Marco Nicolis, Jonas Rohnke, Thomas Drugman, and Jaime Lorenzo-Trueba. Dy-", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 194, 504, 207 ], "spans": [ { "bbox": [ 116, 194, 504, 207 ], "score": 1.0, "content": "namic prosody generation for speech synthesis using linguistics-driven acoustic embedding se-", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 205, 225, 217 ], "spans": [ { "bbox": [ 116, 205, 225, 217 ], "score": 1.0, "content": "lection. Interspeech, 2020.", "type": "text" } ], "index": 9 } ], "index": 8, "bbox_fs": [ 105, 182, 504, 217 ] }, { "type": "text", "bbox": [ 106, 223, 503, 247 ], "lines": [ { "bbox": [ 106, 223, 505, 236 ], "spans": [ { "bbox": [ 106, 223, 505, 236 ], "score": 1.0, "content": "Maarten Versteegh, Roland Thiolliere, Thomas Schatz, Xuan Nga Cao, Xavier Anguera, Aren", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 235, 470, 248 ], "spans": [ { "bbox": [ 116, 235, 470, 248 ], "score": 1.0, "content": "Jansen, and Emmanuel Dupoux. The zero resource speech challenge 2015. ISCA, 2015.", "type": "text" } ], "index": 11 } ], "index": 10.5, "bbox_fs": [ 106, 223, 505, 248 ] }, { "type": "text", "bbox": [ 107, 253, 502, 277 ], "lines": [ { "bbox": [ 105, 252, 505, 267 ], "spans": [ { "bbox": [ 105, 252, 505, 267 ], "score": 1.0, "content": "Michael Wagner and Duane G Watson. Experimental and theoretical advances in prosody: A review.", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 265, 356, 277 ], "spans": [ { "bbox": [ 116, 265, 356, 277 ], "score": 1.0, "content": "Language and cognitive processes, 25(7-9):905–945, 2010.", "type": "text" } ], "index": 13 } ], "index": 12.5, "bbox_fs": [ 105, 252, 505, 277 ] }, { "type": "text", "bbox": [ 107, 283, 503, 307 ], "lines": [ { "bbox": [ 105, 282, 504, 297 ], "spans": [ { "bbox": [ 105, 282, 504, 297 ], "score": 1.0, "content": "Bo Wan, Wenjuan Han, Zilong Zheng, and Tinne Tuytelaars. Unsupervised vision-language gram-", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 294, 356, 307 ], "spans": [ { "bbox": [ 115, 294, 356, 307 ], "score": 1.0, "content": "mar induction with shared structure modeling. ICLR, 2022.", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 105, 282, 504, 307 ] }, { "type": "text", "bbox": [ 106, 313, 504, 336 ], "lines": [ { "bbox": [ 105, 312, 505, 327 ], "spans": [ { "bbox": [ 105, 312, 505, 327 ], "score": 1.0, "content": "Liming Wang and Mark Hasegawa-Johnson. A translation framework for visually grounded spoken", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 324, 248, 336 ], "spans": [ { "bbox": [ 115, 324, 248, 336 ], "score": 1.0, "content": "unit discovery. In ACSSC, 2021.", "type": "text" } ], "index": 17 } ], "index": 16.5, "bbox_fs": [ 105, 312, 505, 336 ] }, { "type": "text", "bbox": [ 106, 343, 506, 378 ], "lines": [ { "bbox": [ 106, 344, 505, 356 ], "spans": [ { "bbox": [ 106, 344, 505, 356 ], "score": 1.0, "content": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 354, 504, 367 ], "spans": [ { "bbox": [ 115, 354, 504, 367 ], "score": 1.0, "content": "learning. Machine Learning, 8(3-4):229–256, 1992. URL https://link.springer.com/", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 366, 325, 378 ], "spans": [ { "bbox": [ 116, 366, 325, 378 ], "score": 1.0, "content": "content/pdf/10.1007/BF00992696.pdf.", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 106, 344, 505, 378 ] }, { "type": "text", "bbox": [ 106, 384, 504, 407 ], "lines": [ { "bbox": [ 105, 383, 506, 397 ], "spans": [ { "bbox": [ 105, 383, 506, 397 ], "score": 1.0, "content": "Daniel H Younger. Recognition and parsing of context-free languages in time n3. Information and", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 396, 239, 406 ], "spans": [ { "bbox": [ 116, 396, 239, 406 ], "score": 1.0, "content": "control, 10(2):189–208, 1967.", "type": "text" } ], "index": 22 } ], "index": 21.5, "bbox_fs": [ 105, 383, 506, 406 ] }, { "type": "text", "bbox": [ 106, 414, 505, 448 ], "lines": [ { "bbox": [ 105, 414, 505, 428 ], "spans": [ { "bbox": [ 105, 414, 505, 428 ], "score": 1.0, "content": "Hao Zhang and Daniel Gildea. Efficient multi-pass decoding for synchronous context free gram-", "type": "text" } ], "index": 23 }, { "bbox": [ 116, 425, 505, 438 ], "spans": [ { "bbox": [ 116, 425, 505, 438 ], "score": 1.0, "content": "mars. In Proceedings of ACL-08: HLT, pp. 209–217, Columbus, Ohio, June 2008. Association", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 435, 469, 450 ], "spans": [ { "bbox": [ 115, 435, 469, 450 ], "score": 1.0, "content": "for Computational Linguistics. URL https://aclanthology.org/P08-1025.", "type": "text" } ], "index": 25 } ], "index": 24, "bbox_fs": [ 105, 414, 505, 450 ] }, { "type": "text", "bbox": [ 107, 454, 504, 478 ], "lines": [ { "bbox": [ 106, 455, 504, 468 ], "spans": [ { "bbox": [ 106, 455, 504, 468 ], "score": 1.0, "content": "Songyang Zhang, Linfeng Song, Lifeng Jin, Kun Xu, Dong Yu, and Jiebo Luo. Video-aided unsu-", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 466, 317, 478 ], "spans": [ { "bbox": [ 115, 466, 317, 478 ], "score": 1.0, "content": "pervised grammar induction. NAACL-HLT, 2021.", "type": "text" } ], "index": 27 } ], "index": 26.5, "bbox_fs": [ 106, 455, 504, 478 ] }, { "type": "text", "bbox": [ 107, 484, 504, 508 ], "lines": [ { "bbox": [ 105, 483, 505, 498 ], "spans": [ { "bbox": [ 105, 483, 505, 498 ], "score": 1.0, "content": "Yaodong Zhang. Unsupervised speech processing with applications to query-by-example spoken", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 496, 411, 508 ], "spans": [ { "bbox": [ 115, 496, 411, 508 ], "score": 1.0, "content": "term detection. PhD thesis, Massachusetts Institute of Technology, 2013.", "type": "text" } ], "index": 29 } ], "index": 28.5, "bbox_fs": [ 105, 483, 505, 508 ] }, { "type": "text", "bbox": [ 106, 514, 505, 538 ], "lines": [ { "bbox": [ 106, 513, 505, 529 ], "spans": [ { "bbox": [ 106, 513, 505, 529 ], "score": 1.0, "content": "Yaodong Zhang and James R Glass. Unsupervised spoken keyword spotting via segmental dtw on", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 527, 272, 538 ], "spans": [ { "bbox": [ 115, 527, 272, 538 ], "score": 1.0, "content": "gaussian posteriorgrams. ASRU, 2009.", "type": "text" } ], "index": 31 } ], "index": 30.5, "bbox_fs": [ 106, 513, 505, 538 ] }, { "type": "text", "bbox": [ 106, 544, 441, 557 ], "lines": [ { "bbox": [ 106, 543, 441, 558 ], "spans": [ { "bbox": [ 106, 543, 441, 558 ], "score": 1.0, "content": "Yanpeng Zhao and Ivan Titov. Visually grounded compound pcfgs. EMNLP, 2020.", "type": "text" } ], "index": 32 } ], "index": 32, "bbox_fs": [ 106, 543, 441, 558 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 182, 93 ], "lines": [ { "bbox": [ 105, 79, 185, 97 ], "spans": [ { "bbox": [ 105, 79, 185, 97 ], "score": 1.0, "content": "A APPENDIX", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 108, 106, 182, 118 ], "lines": [ { "bbox": [ 105, 105, 183, 119 ], "spans": [ { "bbox": [ 105, 105, 183, 119 ], "score": 1.0, "content": "A.1 BASELINES", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 107, 127, 505, 215 ], "lines": [ { "bbox": [ 106, 127, 504, 139 ], "spans": [ { "bbox": [ 106, 127, 504, 139 ], "score": 1.0, "content": "AV-cPCFG: We train compound probabilistic context free grammar (cPCFG) (Kim et al., 2019a)", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 137, 505, 150 ], "spans": [ { "bbox": [ 105, 137, 505, 150 ], "score": 1.0, "content": "on word-level discrete speech tokens. Similar to AV-NSL, word segments are obtained from VG-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 148, 505, 161 ], "spans": [ { "bbox": [ 105, 148, 505, 161 ], "score": 1.0, "content": "HuBERT with segment insertion, and segment representations are extracted from VG-Hubert layer", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 159, 505, 172 ], "spans": [ { "bbox": [ 106, 159, 505, 172 ], "score": 1.0, "content": "10 with CLS attention weighted mean-pool. Different from AV-NSL, the segment representations", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 171, 505, 182 ], "spans": [ { "bbox": [ 105, 171, 505, 182 ], "score": 1.0, "content": "are discretized via kmeans to obtain word-level discrete indices. Because the discretization is word-", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 181, 504, 193 ], "spans": [ { "bbox": [ 105, 181, 401, 193 ], "score": 1.0, "content": "level instead of phone-level, we swept the number of kmeans cluster over", "type": "text" }, { "bbox": [ 401, 181, 502, 193 ], "score": 0.55, "content": "\\left\\{ 1 \\mathrm { k } , 2 \\mathrm { k } , 4 \\mathrm { k } , 8 \\mathrm { k } , 1 2 \\mathrm { k } , 1 6 \\mathrm { k } , \\right.", "type": "inline_equation" }, { "bbox": [ 503, 181, 504, 193 ], "score": 0.583, "content": "，", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 191, 506, 205 ], "spans": [ { "bbox": [ 106, 192, 127, 204 ], "score": 0.8, "content": "2 0 \\mathrm { k } \\}", "type": "inline_equation" }, { "bbox": [ 128, 191, 506, 205 ], "score": 1.0, "content": ", which corresponds to the dictionary size in cPCFG. In summary, AV-cPCFG leverages visual", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 204, 487, 216 ], "spans": [ { "bbox": [ 105, 204, 487, 216 ], "score": 1.0, "content": "cues only for segmentation and segment representations, but not for phrase structure induction.", "type": "text" } ], "index": 9 } ], "index": 5.5 }, { "type": "text", "bbox": [ 107, 220, 505, 275 ], "lines": [ { "bbox": [ 105, 219, 505, 234 ], "spans": [ { "bbox": [ 105, 219, 505, 234 ], "score": 1.0, "content": "DPDP-cPCFG: Instead of training cPCFG on audio-visual word segments and audio-visual seg-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 231, 505, 244 ], "spans": [ { "bbox": [ 105, 231, 505, 244 ], "score": 1.0, "content": "ment representations, DPDP-cPCFG does not rely on any visual grounding throughout. Instead,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 242, 505, 255 ], "spans": [ { "bbox": [ 105, 242, 505, 255 ], "score": 1.0, "content": "DPDP (Kamper, 2022), a recent speech-only word segmentation algorithm, and vanilla HuBERT", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 252, 505, 267 ], "spans": [ { "bbox": [ 105, 252, 505, 267 ], "score": 1.0, "content": "representations mean-pooled over DPDP segments are used. We swept through HuBERT layer {2,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 264, 425, 276 ], "spans": [ { "bbox": [ 105, 264, 425, 276 ], "score": 1.0, "content": "4, 6, 8, 10, 12}. As in AV-cPCFG, kmeans is used for word-level discretization.", "type": "text" } ], "index": 14 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 281, 505, 336 ], "lines": [ { "bbox": [ 105, 280, 505, 294 ], "spans": [ { "bbox": [ 105, 280, 505, 294 ], "score": 1.0, "content": "oracle AV-NSL: To remove the uncertainty of unsupervised word segmentation, we directly train", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 291, 505, 305 ], "spans": [ { "bbox": [ 106, 291, 505, 305 ], "score": 1.0, "content": "AV-NSL on top of oracle word segmentation via force alignment. The segment representations are", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 303, 505, 316 ], "spans": [ { "bbox": [ 105, 303, 363, 316 ], "score": 1.0, "content": "based on learnable attention pooling over vanilla HuBERT layer", "type": "text" }, { "bbox": [ 364, 303, 438, 315 ], "score": 0.84, "content": "\\{ 2 , 4 , 6 , 8 , 1 0 , 1 2 \\}", "type": "inline_equation" }, { "bbox": [ 439, 303, 505, 316 ], "score": 1.0, "content": "representations.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 312, 506, 327 ], "spans": [ { "bbox": [ 106, 312, 506, 327 ], "score": 1.0, "content": "We also tried log Mel spectrograms and HuBERT-L 300M to examine the effectiveness of different", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 325, 494, 338 ], "spans": [ { "bbox": [ 105, 325, 494, 338 ], "score": 1.0, "content": "input representations. One note is that simpler score and combine parametrization suffices here6.", "type": "text" } ], "index": 19 } ], "index": 17 }, { "type": "title", "bbox": [ 108, 349, 221, 361 ], "lines": [ { "bbox": [ 106, 350, 222, 362 ], "spans": [ { "bbox": [ 106, 350, 222, 362 ], "score": 1.0, "content": "A.2 HYPERPARAMETERS", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 370, 505, 414 ], "lines": [ { "bbox": [ 105, 370, 505, 383 ], "spans": [ { "bbox": [ 105, 370, 431, 383 ], "score": 1.0, "content": "For VG-HuBERT, we run MBR selection on the combination of insertion gap", "type": "text" }, { "bbox": [ 431, 370, 484, 382 ], "score": 0.85, "content": "\\{ 0 . 1 , 0 . 2 , 0 . 3 \\}", "type": "inline_equation" }, { "bbox": [ 484, 370, 505, 383 ], "score": 1.0, "content": "sec-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 381, 505, 393 ], "spans": [ { "bbox": [ 105, 381, 211, 393 ], "score": 1.0, "content": "onds, segmentation layer", "type": "text" }, { "bbox": [ 211, 381, 252, 393 ], "score": 0.7, "content": "\\{ 9 , 1 0 , 1 1 \\}", "type": "inline_equation" }, { "bbox": [ 252, 381, 407, 393 ], "score": 1.0, "content": ", attention magnitude threshold at top", "type": "text" }, { "bbox": [ 407, 381, 478, 393 ], "score": 0.9, "content": "\\{ 3 0 \\% , 2 0 \\% , 1 0 \\% \\}", "type": "inline_equation" }, { "bbox": [ 478, 381, 505, 393 ], "score": 1.0, "content": ", three", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 392, 505, 406 ], "spans": [ { "bbox": [ 105, 392, 505, 406 ], "score": 1.0, "content": "training random seeds, and model snapshots at training step 20k, 30k, 40k, 50k, 60k. This gives 405", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 404, 196, 414 ], "spans": [ { "bbox": [ 106, 404, 196, 414 ], "score": 1.0, "content": "combinations in total.", "type": "text" } ], "index": 24 } ], "index": 22.5 }, { "type": "title", "bbox": [ 108, 428, 227, 439 ], "lines": [ { "bbox": [ 107, 428, 228, 440 ], "spans": [ { "bbox": [ 107, 428, 228, 440 ], "score": 1.0, "content": "A.3 FULL RESULTS TABLE", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "title", "bbox": [ 108, 448, 252, 460 ], "lines": [ { "bbox": [ 106, 447, 252, 461 ], "spans": [ { "bbox": [ 106, 447, 252, 461 ], "score": 1.0, "content": "A.4 WORD SEGMENTATION VIZ", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 109, 469, 505, 502 ], "lines": [ { "bbox": [ 106, 468, 505, 482 ], "spans": [ { "bbox": [ 106, 468, 505, 482 ], "score": 1.0, "content": "We show more examples of word segmentation generated by our improved VG-HuBERT in Figure 4.", "type": "text" } ], "index": 27 }, { "bbox": [ 107, 480, 505, 492 ], "spans": [ { "bbox": [ 107, 480, 201, 492 ], "score": 1.0, "content": "Segments marked with", "type": "text" }, { "bbox": [ 201, 480, 218, 491 ], "score": 0.73, "content": "\" + \"", "type": "inline_equation" }, { "bbox": [ 218, 480, 505, 492 ], "score": 1.0, "content": "are inserted segments, and vertical blue dotted lines are inferred word", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 490, 156, 504 ], "spans": [ { "bbox": [ 106, 490, 156, 504 ], "score": 1.0, "content": "boundaries.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "title", "bbox": [ 108, 516, 289, 527 ], "lines": [ { "bbox": [ 106, 516, 290, 529 ], "spans": [ { "bbox": [ 106, 516, 290, 529 ], "score": 1.0, "content": "A.5 VISUALIZATION OF INDUCED TREES", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 108, 536, 276, 548 ], "lines": [ { "bbox": [ 106, 535, 279, 550 ], "spans": [ { "bbox": [ 106, 535, 279, 550 ], "score": 1.0, "content": "We visualize the induced trees in Figure 5.", "type": "text" } ], "index": 31 } ], "index": 31 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 114, 721, 503, 732 ], "lines": [ { "bbox": [ 118, 718, 505, 734 ], "spans": [ { "bbox": [ 118, 718, 505, 734 ], "score": 1.0, "content": "6We found that for oracle AV-NSL, the original score and combine parametrization in VG-NSL works better.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "15", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 182, 93 ], "lines": [ { "bbox": [ 105, 79, 185, 97 ], "spans": [ { "bbox": [ 105, 79, 185, 97 ], "score": 1.0, "content": "A APPENDIX", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 108, 106, 182, 118 ], "lines": [ { "bbox": [ 105, 105, 183, 119 ], "spans": [ { "bbox": [ 105, 105, 183, 119 ], "score": 1.0, "content": "A.1 BASELINES", "type": "text" } ], "index": 1 } ], "index": 1, "bbox_fs": [ 105, 105, 183, 119 ] }, { "type": "text", "bbox": [ 107, 127, 505, 215 ], "lines": [ { "bbox": [ 106, 127, 504, 139 ], "spans": [ { "bbox": [ 106, 127, 504, 139 ], "score": 1.0, "content": "AV-cPCFG: We train compound probabilistic context free grammar (cPCFG) (Kim et al., 2019a)", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 137, 505, 150 ], "spans": [ { "bbox": [ 105, 137, 505, 150 ], "score": 1.0, "content": "on word-level discrete speech tokens. Similar to AV-NSL, word segments are obtained from VG-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 148, 505, 161 ], "spans": [ { "bbox": [ 105, 148, 505, 161 ], "score": 1.0, "content": "HuBERT with segment insertion, and segment representations are extracted from VG-Hubert layer", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 159, 505, 172 ], "spans": [ { "bbox": [ 106, 159, 505, 172 ], "score": 1.0, "content": "10 with CLS attention weighted mean-pool. Different from AV-NSL, the segment representations", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 171, 505, 182 ], "spans": [ { "bbox": [ 105, 171, 505, 182 ], "score": 1.0, "content": "are discretized via kmeans to obtain word-level discrete indices. Because the discretization is word-", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 181, 504, 193 ], "spans": [ { "bbox": [ 105, 181, 401, 193 ], "score": 1.0, "content": "level instead of phone-level, we swept the number of kmeans cluster over", "type": "text" }, { "bbox": [ 401, 181, 502, 193 ], "score": 0.55, "content": "\\left\\{ 1 \\mathrm { k } , 2 \\mathrm { k } , 4 \\mathrm { k } , 8 \\mathrm { k } , 1 2 \\mathrm { k } , 1 6 \\mathrm { k } , \\right.", "type": "inline_equation" }, { "bbox": [ 503, 181, 504, 193 ], "score": 0.583, "content": "，", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 191, 506, 205 ], "spans": [ { "bbox": [ 106, 192, 127, 204 ], "score": 0.8, "content": "2 0 \\mathrm { k } \\}", "type": "inline_equation" }, { "bbox": [ 128, 191, 506, 205 ], "score": 1.0, "content": ", which corresponds to the dictionary size in cPCFG. In summary, AV-cPCFG leverages visual", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 204, 487, 216 ], "spans": [ { "bbox": [ 105, 204, 487, 216 ], "score": 1.0, "content": "cues only for segmentation and segment representations, but not for phrase structure induction.", "type": "text" } ], "index": 9 } ], "index": 5.5, "bbox_fs": [ 105, 127, 506, 216 ] }, { "type": "text", "bbox": [ 107, 220, 505, 275 ], "lines": [ { "bbox": [ 105, 219, 505, 234 ], "spans": [ { "bbox": [ 105, 219, 505, 234 ], "score": 1.0, "content": "DPDP-cPCFG: Instead of training cPCFG on audio-visual word segments and audio-visual seg-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 231, 505, 244 ], "spans": [ { "bbox": [ 105, 231, 505, 244 ], "score": 1.0, "content": "ment representations, DPDP-cPCFG does not rely on any visual grounding throughout. Instead,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 242, 505, 255 ], "spans": [ { "bbox": [ 105, 242, 505, 255 ], "score": 1.0, "content": "DPDP (Kamper, 2022), a recent speech-only word segmentation algorithm, and vanilla HuBERT", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 252, 505, 267 ], "spans": [ { "bbox": [ 105, 252, 505, 267 ], "score": 1.0, "content": "representations mean-pooled over DPDP segments are used. We swept through HuBERT layer {2,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 264, 425, 276 ], "spans": [ { "bbox": [ 105, 264, 425, 276 ], "score": 1.0, "content": "4, 6, 8, 10, 12}. As in AV-cPCFG, kmeans is used for word-level discretization.", "type": "text" } ], "index": 14 } ], "index": 12, "bbox_fs": [ 105, 219, 505, 276 ] }, { "type": "text", "bbox": [ 107, 281, 505, 336 ], "lines": [ { "bbox": [ 105, 280, 505, 294 ], "spans": [ { "bbox": [ 105, 280, 505, 294 ], "score": 1.0, "content": "oracle AV-NSL: To remove the uncertainty of unsupervised word segmentation, we directly train", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 291, 505, 305 ], "spans": [ { "bbox": [ 106, 291, 505, 305 ], "score": 1.0, "content": "AV-NSL on top of oracle word segmentation via force alignment. The segment representations are", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 303, 505, 316 ], "spans": [ { "bbox": [ 105, 303, 363, 316 ], "score": 1.0, "content": "based on learnable attention pooling over vanilla HuBERT layer", "type": "text" }, { "bbox": [ 364, 303, 438, 315 ], "score": 0.84, "content": "\\{ 2 , 4 , 6 , 8 , 1 0 , 1 2 \\}", "type": "inline_equation" }, { "bbox": [ 439, 303, 505, 316 ], "score": 1.0, "content": "representations.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 312, 506, 327 ], "spans": [ { "bbox": [ 106, 312, 506, 327 ], "score": 1.0, "content": "We also tried log Mel spectrograms and HuBERT-L 300M to examine the effectiveness of different", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 325, 494, 338 ], "spans": [ { "bbox": [ 105, 325, 494, 338 ], "score": 1.0, "content": "input representations. One note is that simpler score and combine parametrization suffices here6.", "type": "text" } ], "index": 19 } ], "index": 17, "bbox_fs": [ 105, 280, 506, 338 ] }, { "type": "title", "bbox": [ 108, 349, 221, 361 ], "lines": [ { "bbox": [ 106, 350, 222, 362 ], "spans": [ { "bbox": [ 106, 350, 222, 362 ], "score": 1.0, "content": "A.2 HYPERPARAMETERS", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 370, 505, 414 ], "lines": [ { "bbox": [ 105, 370, 505, 383 ], "spans": [ { "bbox": [ 105, 370, 431, 383 ], "score": 1.0, "content": "For VG-HuBERT, we run MBR selection on the combination of insertion gap", "type": "text" }, { "bbox": [ 431, 370, 484, 382 ], "score": 0.85, "content": "\\{ 0 . 1 , 0 . 2 , 0 . 3 \\}", "type": "inline_equation" }, { "bbox": [ 484, 370, 505, 383 ], "score": 1.0, "content": "sec-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 381, 505, 393 ], "spans": [ { "bbox": [ 105, 381, 211, 393 ], "score": 1.0, "content": "onds, segmentation layer", "type": "text" }, { "bbox": [ 211, 381, 252, 393 ], "score": 0.7, "content": "\\{ 9 , 1 0 , 1 1 \\}", "type": "inline_equation" }, { "bbox": [ 252, 381, 407, 393 ], "score": 1.0, "content": ", attention magnitude threshold at top", "type": "text" }, { "bbox": [ 407, 381, 478, 393 ], "score": 0.9, "content": "\\{ 3 0 \\% , 2 0 \\% , 1 0 \\% \\}", "type": "inline_equation" }, { "bbox": [ 478, 381, 505, 393 ], "score": 1.0, "content": ", three", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 392, 505, 406 ], "spans": [ { "bbox": [ 105, 392, 505, 406 ], "score": 1.0, "content": "training random seeds, and model snapshots at training step 20k, 30k, 40k, 50k, 60k. This gives 405", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 404, 196, 414 ], "spans": [ { "bbox": [ 106, 404, 196, 414 ], "score": 1.0, "content": "combinations in total.", "type": "text" } ], "index": 24 } ], "index": 22.5, "bbox_fs": [ 105, 370, 505, 414 ] }, { "type": "title", "bbox": [ 108, 428, 227, 439 ], "lines": [ { "bbox": [ 107, 428, 228, 440 ], "spans": [ { "bbox": [ 107, 428, 228, 440 ], "score": 1.0, "content": "A.3 FULL RESULTS TABLE", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "title", "bbox": [ 108, 448, 252, 460 ], "lines": [ { "bbox": [ 106, 447, 252, 461 ], "spans": [ { "bbox": [ 106, 447, 252, 461 ], "score": 1.0, "content": "A.4 WORD SEGMENTATION VIZ", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 109, 469, 505, 502 ], "lines": [ { "bbox": [ 106, 468, 505, 482 ], "spans": [ { "bbox": [ 106, 468, 505, 482 ], "score": 1.0, "content": "We show more examples of word segmentation generated by our improved VG-HuBERT in Figure 4.", "type": "text" } ], "index": 27 }, { "bbox": [ 107, 480, 505, 492 ], "spans": [ { "bbox": [ 107, 480, 201, 492 ], "score": 1.0, "content": "Segments marked with", "type": "text" }, { "bbox": [ 201, 480, 218, 491 ], "score": 0.73, "content": "\" + \"", "type": "inline_equation" }, { "bbox": [ 218, 480, 505, 492 ], "score": 1.0, "content": "are inserted segments, and vertical blue dotted lines are inferred word", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 490, 156, 504 ], "spans": [ { "bbox": [ 106, 490, 156, 504 ], "score": 1.0, "content": "boundaries.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 106, 468, 505, 504 ] }, { "type": "title", "bbox": [ 108, 516, 289, 527 ], "lines": [ { "bbox": [ 106, 516, 290, 529 ], "spans": [ { "bbox": [ 106, 516, 290, 529 ], "score": 1.0, "content": "A.5 VISUALIZATION OF INDUCED TREES", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 108, 536, 276, 548 ], "lines": [ { "bbox": [ 106, 535, 279, 550 ], "spans": [ { "bbox": [ 106, 535, 279, 550 ], "score": 1.0, "content": "We visualize the induced trees in Figure 5.", "type": "text" } ], "index": 31 } ], "index": 31, "bbox_fs": [ 106, 535, 279, 550 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 124, 139, 486, 400 ], "blocks": [ { "type": "table_body", "bbox": [ 124, 139, 486, 400 ], "group_id": 0, "lines": [ { "bbox": [ 124, 139, 486, 400 ], "spans": [ { "bbox": [ 124, 139, 486, 400 ], "score": 0.985, "html": "
ModelOutput SelectionSAIoU
Syntax InductionSegmentationSeg.Representation (continuous/discrete)
Right-BranchingVG-HuBERT+MBR100.546
Right-BranchingDPDP0.478
AV-NSLVG-HuBERT+MBR10VG-HuBERT1o (continuous)MBR0.516
AV-NSLVG-HuBERT+MBR10VG-HuBERT11 (continuous)MBR0.498
AV-NSLVG-HuBERT+MBR10VG-HuBERT12 (continuous)MBR0.492
AV-NSLVG-HuBERT+MBR10VG-HuBERT10,11,12 (continuous)MBR0.521
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+1k km (discrete)last ckpt.0.454
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+2k km (discrete)last ckpt.0.444
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+4k km (discrete)last ckpt.0.499
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+8k km (discrete)last ckpt.0.481
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+12k km (discrete)last ckpt.0.473
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+16k km (discrete)last ckpt.0.471
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+20k km (discrete)last ckpt.0.454
DPDP-cPCFGDPDPHuBERT2+1k km (discrete)last ckpt.0.434
DPDP-cPCFGDPDPHuBERT2+2k km (discrete)last ckpt.0.465
DPDP-cPCFGDPDPHuBERT2+4k km (discrete)last ckpt.0.444
DPDP-cPCFGDPDPHuBERT2+8k km (discrete)last ckpt.0.387
DPDP-cPCFGDPDPHuBERT2+12k km (discrete)last ckpt.0.447
DPDP-cPCFGDPDPHuBERT2+16k km (discrete)last ckpt.0.360
DPDP-cPCFGDPDPHuBERT1o+1k km (discrete)last ckpt.0.403
DPDP-cPCFGDPDPHuBERT1o+2k km (discrete)last ckpt.0.426
DPDP-cPCFGDPDPHuBERT1o+4k km (discrete)last ckpt.0.415
DPDP-cPCFGDPDPHuBERT1o+8k km (discrete)last ckpt.0.367
DPDP-cPCFGDPDPHuBERT1o+12k km (discrete)last ckpt.0.415
DPDP-cPCFGDPDPHuBERT1o+16k km (discrete)last ckpt.0.414
", "type": "table", "image_path": "c8bfe2ce8e91555d64875d1c67f70ec397468babd399bfd308e4ef7703972653.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 124, 139, 486, 226.0 ], "spans": [], "index": 0 }, { "bbox": [ 124, 226.0, 486, 313.0 ], "spans": [], "index": 1 }, { "bbox": [ 124, 313.0, 486, 400.0 ], "spans": [], "index": 2 } ] }, { "type": "table_caption", "bbox": [ 132, 414, 478, 425 ], "group_id": 0, "lines": [ { "bbox": [ 132, 413, 479, 426 ], "spans": [ { "bbox": [ 132, 413, 479, 426 ], "score": 1.0, "content": "Table 7: Fully-unsupervised phrase structure induction results evaluated with SAIOU.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "table", "bbox": [ 117, 549, 494, 646 ], "blocks": [ { "type": "table_body", "bbox": [ 117, 549, 494, 646 ], "group_id": 1, "lines": [ { "bbox": [ 117, 549, 494, 646 ], "spans": [ { "bbox": [ 117, 549, 494, 646 ], "score": 0.982, "html": "
ModelSegmentationSeg.Representationtree targetOutput SelectionSAIoU
trainvaltest
s-BeneparVG-HuBERT+MBR10HuBERT2AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT4AV-NSLAV-NSLoraclelast ckpt.0.536
s-BeneparVG-HuBERT+MBR10HuBERT6AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT8AV-NSLAV-NSLoraclelast ckpt.0.532
s-BeneparVG-HuBERT+MBR10HuBERT10AV-NSLAV-NSLoraclelast ckpt.0.537
s-BeneparVG-HuBERT+MBR10HuBERT12AV-NSLAV-NSLoraclelast ckpt.0.536
s-BeneparVG-HuBERT+MBR10HuBERT2,4,6,8,10,12AV-NSLAV-NSLoracleMBR0.536
", "type": "table", "image_path": "2c1f7953a8ea2567e5f04d2844f40f85e041afa5746eec08ef5cabdc3b2b3e3a.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 117, 549, 494, 581.3333333333334 ], "spans": [], "index": 4 }, { "bbox": [ 117, 581.3333333333334, 494, 613.6666666666667 ], "spans": [], "index": 5 }, { "bbox": [ 117, 613.6666666666667, 494, 646.0000000000001 ], "spans": [], "index": 6 } ] }, { "type": "table_caption", "bbox": [ 199, 659, 412, 671 ], "group_id": 1, "lines": [ { "bbox": [ 199, 659, 412, 672 ], "spans": [ { "bbox": [ 199, 659, 412, 672 ], "score": 1.0, "content": "Table 8: Self-training results evaluated with SAIOU.", "type": "text" } ], "index": 7 } ], "index": 7 } ], "index": 6.0 } ], "page_idx": 15, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "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": "16", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 124, 139, 486, 400 ], "blocks": [ { "type": "table_body", "bbox": [ 124, 139, 486, 400 ], "group_id": 0, "lines": [ { "bbox": [ 124, 139, 486, 400 ], "spans": [ { "bbox": [ 124, 139, 486, 400 ], "score": 0.985, "html": "
ModelOutput SelectionSAIoU
Syntax InductionSegmentationSeg.Representation (continuous/discrete)
Right-BranchingVG-HuBERT+MBR100.546
Right-BranchingDPDP0.478
AV-NSLVG-HuBERT+MBR10VG-HuBERT1o (continuous)MBR0.516
AV-NSLVG-HuBERT+MBR10VG-HuBERT11 (continuous)MBR0.498
AV-NSLVG-HuBERT+MBR10VG-HuBERT12 (continuous)MBR0.492
AV-NSLVG-HuBERT+MBR10VG-HuBERT10,11,12 (continuous)MBR0.521
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+1k km (discrete)last ckpt.0.454
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+2k km (discrete)last ckpt.0.444
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+4k km (discrete)last ckpt.0.499
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+8k km (discrete)last ckpt.0.481
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+12k km (discrete)last ckpt.0.473
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+16k km (discrete)last ckpt.0.471
AV-cPCFGVG-HuBERT+MBR10VG-HuBERT1o+20k km (discrete)last ckpt.0.454
DPDP-cPCFGDPDPHuBERT2+1k km (discrete)last ckpt.0.434
DPDP-cPCFGDPDPHuBERT2+2k km (discrete)last ckpt.0.465
DPDP-cPCFGDPDPHuBERT2+4k km (discrete)last ckpt.0.444
DPDP-cPCFGDPDPHuBERT2+8k km (discrete)last ckpt.0.387
DPDP-cPCFGDPDPHuBERT2+12k km (discrete)last ckpt.0.447
DPDP-cPCFGDPDPHuBERT2+16k km (discrete)last ckpt.0.360
DPDP-cPCFGDPDPHuBERT1o+1k km (discrete)last ckpt.0.403
DPDP-cPCFGDPDPHuBERT1o+2k km (discrete)last ckpt.0.426
DPDP-cPCFGDPDPHuBERT1o+4k km (discrete)last ckpt.0.415
DPDP-cPCFGDPDPHuBERT1o+8k km (discrete)last ckpt.0.367
DPDP-cPCFGDPDPHuBERT1o+12k km (discrete)last ckpt.0.415
DPDP-cPCFGDPDPHuBERT1o+16k km (discrete)last ckpt.0.414
", "type": "table", "image_path": "c8bfe2ce8e91555d64875d1c67f70ec397468babd399bfd308e4ef7703972653.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 124, 139, 486, 226.0 ], "spans": [], "index": 0 }, { "bbox": [ 124, 226.0, 486, 313.0 ], "spans": [], "index": 1 }, { "bbox": [ 124, 313.0, 486, 400.0 ], "spans": [], "index": 2 } ] }, { "type": "table_caption", "bbox": [ 132, 414, 478, 425 ], "group_id": 0, "lines": [ { "bbox": [ 132, 413, 479, 426 ], "spans": [ { "bbox": [ 132, 413, 479, 426 ], "score": 1.0, "content": "Table 7: Fully-unsupervised phrase structure induction results evaluated with SAIOU.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "table", "bbox": [ 117, 549, 494, 646 ], "blocks": [ { "type": "table_body", "bbox": [ 117, 549, 494, 646 ], "group_id": 1, "lines": [ { "bbox": [ 117, 549, 494, 646 ], "spans": [ { "bbox": [ 117, 549, 494, 646 ], "score": 0.982, "html": "
ModelSegmentationSeg.Representationtree targetOutput SelectionSAIoU
trainvaltest
s-BeneparVG-HuBERT+MBR10HuBERT2AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT4AV-NSLAV-NSLoraclelast ckpt.0.536
s-BeneparVG-HuBERT+MBR10HuBERT6AV-NSLAV-NSLoraclelast ckpt.0.538
s-BeneparVG-HuBERT+MBR10HuBERT8AV-NSLAV-NSLoraclelast ckpt.0.532
s-BeneparVG-HuBERT+MBR10HuBERT10AV-NSLAV-NSLoraclelast ckpt.0.537
s-BeneparVG-HuBERT+MBR10HuBERT12AV-NSLAV-NSLoraclelast ckpt.0.536
s-BeneparVG-HuBERT+MBR10HuBERT2,4,6,8,10,12AV-NSLAV-NSLoracleMBR0.536
", "type": "table", "image_path": "2c1f7953a8ea2567e5f04d2844f40f85e041afa5746eec08ef5cabdc3b2b3e3a.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 117, 549, 494, 581.3333333333334 ], "spans": [], "index": 4 }, { "bbox": [ 117, 581.3333333333334, 494, 613.6666666666667 ], "spans": [], "index": 5 }, { "bbox": [ 117, 613.6666666666667, 494, 646.0000000000001 ], "spans": [], "index": 6 } ] }, { "type": "table_caption", "bbox": [ 199, 659, 412, 671 ], "group_id": 1, "lines": [ { "bbox": [ 199, 659, 412, 672 ], "spans": [ { "bbox": [ 199, 659, 412, 672 ], "score": 1.0, "content": "Table 8: Self-training results evaluated with SAIOU.", "type": "text" } ], "index": 7 } ], "index": 7 } ], "index": 6.0 } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 158, 94, 453, 334 ], "blocks": [ { "type": "table_body", "bbox": [ 158, 94, 453, 334 ], "group_id": 0, "lines": [ { "bbox": [ 158, 94, 453, 334 ], "spans": [ { "bbox": [ 158, 94, 453, 334 ], "score": 0.982, "html": "
ModelOutput SelectionF1
Syntax InductionSegmentationSeg.Representation
Randomoracle32.77
Left-Branchingoracle24.56
Right-Branchingoracle57.39
VG-NSLword embeddingsSupervised53.11
AV-NSLoraclelog-Mel spectrogramSupervised42.01
AV-NSLoracleHuBERT2Supervised55.51
AV-NSLoracleHuBERT-L2454.63
Supervised
AV-NSLoracleHuBERT2MBR54.99
AV-NSLoracleHuBERT4MBR53.25
AV-NSLoracleHuBERT6MBR53.46
AV-NSLoracleHuBERT8MBR53.14
AV-NSLoracleHuBERT10MBR36.67
AV-NSLoracleHuBERT12MBR48.51
AV-NSLoracleHuBERT-L24MBR54.39
AV-NSLoracleHuBERT2,4,6,8,10,12MBR55.56
AV-NSLoracleHuBERT2,4,6,8,10,12,24MBR55.96
AV-NSL →s-BeneparoracleHuBERT2MBR57.24
AV-NSL→s-BeneparoracleHuBERT4MBR57.08
AV-NSL→s-BeneparoracleHuBERT6MBR56.81
AV-NSL →s-BeneparoracleHuBERT8MBR56.94
AV-NSL→s-BeneparoracleHuBERT10MBR57.16
AV-NSL →s-BeneparoracleHuBERT12MBR57.33
", "type": "table", "image_path": "f4745ca7f36246708634794e135ac26df1beaea22c11c210c8b4487beeb85f22.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 158, 94, 453, 106.63157894736842 ], "spans": [], "index": 0 }, { "bbox": [ 158, 106.63157894736842, 453, 119.26315789473685 ], "spans": [], "index": 1 }, { "bbox": [ 158, 119.26315789473685, 453, 131.89473684210526 ], "spans": [], "index": 2 }, { "bbox": [ 158, 131.89473684210526, 453, 144.52631578947367 ], "spans": [], "index": 3 }, { "bbox": [ 158, 144.52631578947367, 453, 157.15789473684208 ], "spans": [], "index": 4 }, { "bbox": [ 158, 157.15789473684208, 453, 169.7894736842105 ], "spans": [], "index": 5 }, { "bbox": [ 158, 169.7894736842105, 453, 182.4210526315789 ], "spans": [], "index": 6 }, { "bbox": [ 158, 182.4210526315789, 453, 195.0526315789473 ], "spans": [], "index": 7 }, { "bbox": [ 158, 195.0526315789473, 453, 207.68421052631572 ], "spans": [], "index": 8 }, { "bbox": [ 158, 207.68421052631572, 453, 220.31578947368413 ], "spans": [], "index": 9 }, { "bbox": [ 158, 220.31578947368413, 453, 232.94736842105254 ], "spans": [], "index": 10 }, { "bbox": [ 158, 232.94736842105254, 453, 245.57894736842096 ], "spans": [], "index": 11 }, { "bbox": [ 158, 245.57894736842096, 453, 258.21052631578937 ], "spans": [], "index": 12 }, { "bbox": [ 158, 258.21052631578937, 453, 270.8421052631578 ], "spans": [], "index": 13 }, { "bbox": [ 158, 270.8421052631578, 453, 283.47368421052624 ], "spans": [], "index": 14 }, { "bbox": [ 158, 283.47368421052624, 453, 296.1052631578947 ], "spans": [], "index": 15 }, { "bbox": [ 158, 296.1052631578947, 453, 308.7368421052631 ], "spans": [], "index": 16 }, { "bbox": [ 158, 308.7368421052631, 453, 321.36842105263156 ], "spans": [], "index": 17 }, { "bbox": [ 158, 321.36842105263156, 453, 334.0 ], "spans": [], "index": 18 } ] }, { "type": "table_caption", "bbox": [ 115, 347, 492, 358 ], "group_id": 0, "lines": [ { "bbox": [ 118, 345, 492, 360 ], "spans": [ { "bbox": [ 118, 345, 476, 360 ], "score": 1.0, "content": "Table 9: Phrase structure induction with oracle segmentation given results evaluated with", "type": "text" }, { "bbox": [ 477, 347, 488, 358 ], "score": 0.87, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 489, 345, 492, 360 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 19 } ], "index": 19 } ], "index": 14.0 }, { "type": "table", "bbox": [ 181, 393, 429, 508 ], "blocks": [ { "type": "table_body", "bbox": [ 181, 393, 429, 508 ], "group_id": 1, "lines": [ { "bbox": [ 181, 393, 429, 508 ], "spans": [ { "bbox": [ 181, 393, 429, 508 ], "score": 0.98, "html": "
ModelF1Constituent Recall
NPVPPPADJP
VG-NSL (Shi et al.,2019)50.479.626.242.022.0
VG-NSL + HI53.374.632.566.521.7
VG-NSL +HI+FastText54.478.824.465.622.0
AV-NSL (oracle seg.+ HuBERT2)55.655.568.166.622.1
AV-NSL (oracle seg.+HuBERT4)53.757.456.861.321.3
AV-NSL (oracle seg.+HuBERT6)53.959.455.459.321.2
AV-NSL (oracle seg.+HuBERT8)53.956.058.064.922.5
AV-NSL (oracle seg.+HuBERT10)50.655.848.157.020.5
AV-NSL (oracle seg. + HuBERT12)49.062.534.445.017.4
", "type": "table", "image_path": "050364f2d38b65ea288aae27fb645c76ef0557b032099c9586e1766bc706389a.jpg" } ] } ], "index": 21, "virtual_lines": [ { "bbox": [ 181, 393, 429, 431.3333333333333 ], "spans": [], "index": 20 }, { "bbox": [ 181, 431.3333333333333, 429, 469.66666666666663 ], "spans": [], "index": 21 }, { "bbox": [ 181, 469.66666666666663, 429, 507.99999999999994 ], "spans": [], "index": 22 } ] }, { "type": "table_caption", "bbox": [ 107, 522, 505, 555 ], "group_id": 1, "lines": [ { "bbox": [ 105, 520, 505, 534 ], "spans": [ { "bbox": [ 105, 520, 333, 534 ], "score": 1.0, "content": "Table 10: Recall of specific typed phrases, and overall", "type": "text" }, { "bbox": [ 334, 522, 346, 533 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 346, 520, 505, 534 ], "score": 1.0, "content": "score, evaluated on the SpokenCOCO", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 533, 505, 545 ], "spans": [ { "bbox": [ 106, 533, 505, 545 ], "score": 1.0, "content": "test split. VG-NSL numbers are taken directly from (Shi et al., 2019). AV-NSL here are trained on", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 544, 390, 556 ], "spans": [ { "bbox": [ 106, 544, 390, 556 ], "score": 1.0, "content": "oracle segmentation with vanilla HuBERT as the layer representations.", "type": "text" } ], "index": 25 } ], "index": 24 } ], "index": 22.5 }, { "type": "table", "bbox": [ 171, 590, 441, 680 ], "blocks": [ { "type": "table_body", "bbox": [ 171, 590, 441, 680 ], "group_id": 2, "lines": [ { "bbox": [ 171, 590, 441, 680 ], "spans": [ { "bbox": [ 171, 590, 441, 680 ], "score": 0.98, "html": "
ModelVisual EmbeddingF1
Syntax InductionSegmentationSeg.Representation
AV-NSLoracleHuBERT2ResNet10155.51
AV-NSLuniformHuBERT2ResNet10148.97
AV-NSLoracleHuBERT10ResNet10150.50
AV-NSLuniformHuBERT10ResNet10136.62
AV-NSLoracleHuBERT2DINO55.71
AV-NSLoracleHuBERT2random31.23
", "type": "table", "image_path": "2ddcfd0083b47d3e9a76616a61125f968e94c7fb52ddb6a784f94a6f20fcd18e.jpg" } ] } ], "index": 27, "virtual_lines": [ { "bbox": [ 171, 590, 441, 620.0 ], "spans": [], "index": 26 }, { "bbox": [ 171, 620.0, 441, 650.0 ], "spans": [], "index": 27 }, { "bbox": [ 171, 650.0, 441, 680.0 ], "spans": [], "index": 28 } ] }, { "type": "table_caption", "bbox": [ 107, 692, 504, 715 ], "group_id": 2, "lines": [ { "bbox": [ 106, 692, 505, 705 ], "spans": [ { "bbox": [ 106, 692, 505, 705 ], "score": 1.0, "content": "Table 11: Top rows: Impact of segmentation quality for AV-NSL with number of words segments", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 704, 406, 715 ], "spans": [ { "bbox": [ 106, 704, 406, 715 ], "score": 1.0, "content": "known in advance. Bottom rows: Impact of visual embedding for AV-NSL", "type": "text" } ], "index": 30 } ], "index": 29.5 } ], "index": 28.25 } ], "page_idx": 16, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "17", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 158, 94, 453, 334 ], "blocks": [ { "type": "table_body", "bbox": [ 158, 94, 453, 334 ], "group_id": 0, "lines": [ { "bbox": [ 158, 94, 453, 334 ], "spans": [ { "bbox": [ 158, 94, 453, 334 ], "score": 0.982, "html": "
ModelOutput SelectionF1
Syntax InductionSegmentationSeg.Representation
Randomoracle32.77
Left-Branchingoracle24.56
Right-Branchingoracle57.39
VG-NSLword embeddingsSupervised53.11
AV-NSLoraclelog-Mel spectrogramSupervised42.01
AV-NSLoracleHuBERT2Supervised55.51
AV-NSLoracleHuBERT-L2454.63
Supervised
AV-NSLoracleHuBERT2MBR54.99
AV-NSLoracleHuBERT4MBR53.25
AV-NSLoracleHuBERT6MBR53.46
AV-NSLoracleHuBERT8MBR53.14
AV-NSLoracleHuBERT10MBR36.67
AV-NSLoracleHuBERT12MBR48.51
AV-NSLoracleHuBERT-L24MBR54.39
AV-NSLoracleHuBERT2,4,6,8,10,12MBR55.56
AV-NSLoracleHuBERT2,4,6,8,10,12,24MBR55.96
AV-NSL →s-BeneparoracleHuBERT2MBR57.24
AV-NSL→s-BeneparoracleHuBERT4MBR57.08
AV-NSL→s-BeneparoracleHuBERT6MBR56.81
AV-NSL →s-BeneparoracleHuBERT8MBR56.94
AV-NSL→s-BeneparoracleHuBERT10MBR57.16
AV-NSL →s-BeneparoracleHuBERT12MBR57.33
", "type": "table", "image_path": "f4745ca7f36246708634794e135ac26df1beaea22c11c210c8b4487beeb85f22.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 158, 94, 453, 106.63157894736842 ], "spans": [], "index": 0 }, { "bbox": [ 158, 106.63157894736842, 453, 119.26315789473685 ], "spans": [], "index": 1 }, { "bbox": [ 158, 119.26315789473685, 453, 131.89473684210526 ], "spans": [], "index": 2 }, { "bbox": [ 158, 131.89473684210526, 453, 144.52631578947367 ], "spans": [], "index": 3 }, { "bbox": [ 158, 144.52631578947367, 453, 157.15789473684208 ], "spans": [], "index": 4 }, { "bbox": [ 158, 157.15789473684208, 453, 169.7894736842105 ], "spans": [], "index": 5 }, { "bbox": [ 158, 169.7894736842105, 453, 182.4210526315789 ], "spans": [], "index": 6 }, { "bbox": [ 158, 182.4210526315789, 453, 195.0526315789473 ], "spans": [], "index": 7 }, { "bbox": [ 158, 195.0526315789473, 453, 207.68421052631572 ], "spans": [], "index": 8 }, { "bbox": [ 158, 207.68421052631572, 453, 220.31578947368413 ], "spans": [], "index": 9 }, { "bbox": [ 158, 220.31578947368413, 453, 232.94736842105254 ], "spans": [], "index": 10 }, { "bbox": [ 158, 232.94736842105254, 453, 245.57894736842096 ], "spans": [], "index": 11 }, { "bbox": [ 158, 245.57894736842096, 453, 258.21052631578937 ], "spans": [], "index": 12 }, { "bbox": [ 158, 258.21052631578937, 453, 270.8421052631578 ], "spans": [], "index": 13 }, { "bbox": [ 158, 270.8421052631578, 453, 283.47368421052624 ], "spans": [], "index": 14 }, { "bbox": [ 158, 283.47368421052624, 453, 296.1052631578947 ], "spans": [], "index": 15 }, { "bbox": [ 158, 296.1052631578947, 453, 308.7368421052631 ], "spans": [], "index": 16 }, { "bbox": [ 158, 308.7368421052631, 453, 321.36842105263156 ], "spans": [], "index": 17 }, { "bbox": [ 158, 321.36842105263156, 453, 334.0 ], "spans": [], "index": 18 } ] }, { "type": "table_caption", "bbox": [ 115, 347, 492, 358 ], "group_id": 0, "lines": [ { "bbox": [ 118, 345, 492, 360 ], "spans": [ { "bbox": [ 118, 345, 476, 360 ], "score": 1.0, "content": "Table 9: Phrase structure induction with oracle segmentation given results evaluated with", "type": "text" }, { "bbox": [ 477, 347, 488, 358 ], "score": 0.87, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 489, 345, 492, 360 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 19 } ], "index": 19 } ], "index": 14.0 }, { "type": "table", "bbox": [ 181, 393, 429, 508 ], "blocks": [ { "type": "table_body", "bbox": [ 181, 393, 429, 508 ], "group_id": 1, "lines": [ { "bbox": [ 181, 393, 429, 508 ], "spans": [ { "bbox": [ 181, 393, 429, 508 ], "score": 0.98, "html": "
ModelF1Constituent Recall
NPVPPPADJP
VG-NSL (Shi et al.,2019)50.479.626.242.022.0
VG-NSL + HI53.374.632.566.521.7
VG-NSL +HI+FastText54.478.824.465.622.0
AV-NSL (oracle seg.+ HuBERT2)55.655.568.166.622.1
AV-NSL (oracle seg.+HuBERT4)53.757.456.861.321.3
AV-NSL (oracle seg.+HuBERT6)53.959.455.459.321.2
AV-NSL (oracle seg.+HuBERT8)53.956.058.064.922.5
AV-NSL (oracle seg.+HuBERT10)50.655.848.157.020.5
AV-NSL (oracle seg. + HuBERT12)49.062.534.445.017.4
", "type": "table", "image_path": "050364f2d38b65ea288aae27fb645c76ef0557b032099c9586e1766bc706389a.jpg" } ] } ], "index": 21, "virtual_lines": [ { "bbox": [ 181, 393, 429, 431.3333333333333 ], "spans": [], "index": 20 }, { "bbox": [ 181, 431.3333333333333, 429, 469.66666666666663 ], "spans": [], "index": 21 }, { "bbox": [ 181, 469.66666666666663, 429, 507.99999999999994 ], "spans": [], "index": 22 } ] }, { "type": "table_caption", "bbox": [ 107, 522, 505, 555 ], "group_id": 1, "lines": [ { "bbox": [ 105, 520, 505, 534 ], "spans": [ { "bbox": [ 105, 520, 333, 534 ], "score": 1.0, "content": "Table 10: Recall of specific typed phrases, and overall", "type": "text" }, { "bbox": [ 334, 522, 346, 533 ], "score": 0.88, "content": "F _ { 1 }", "type": "inline_equation" }, { "bbox": [ 346, 520, 505, 534 ], "score": 1.0, "content": "score, evaluated on the SpokenCOCO", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 533, 505, 545 ], "spans": [ { "bbox": [ 106, 533, 505, 545 ], "score": 1.0, "content": "test split. VG-NSL numbers are taken directly from (Shi et al., 2019). AV-NSL here are trained on", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 544, 390, 556 ], "spans": [ { "bbox": [ 106, 544, 390, 556 ], "score": 1.0, "content": "oracle segmentation with vanilla HuBERT as the layer representations.", "type": "text" } ], "index": 25 } ], "index": 24 } ], "index": 22.5 }, { "type": "table", "bbox": [ 171, 590, 441, 680 ], "blocks": [ { "type": "table_body", "bbox": [ 171, 590, 441, 680 ], "group_id": 2, "lines": [ { "bbox": [ 171, 590, 441, 680 ], "spans": [ { "bbox": [ 171, 590, 441, 680 ], "score": 0.98, "html": "
ModelVisual EmbeddingF1
Syntax InductionSegmentationSeg.Representation
AV-NSLoracleHuBERT2ResNet10155.51
AV-NSLuniformHuBERT2ResNet10148.97
AV-NSLoracleHuBERT10ResNet10150.50
AV-NSLuniformHuBERT10ResNet10136.62
AV-NSLoracleHuBERT2DINO55.71
AV-NSLoracleHuBERT2random31.23
", "type": "table", "image_path": "2ddcfd0083b47d3e9a76616a61125f968e94c7fb52ddb6a784f94a6f20fcd18e.jpg" } ] } ], "index": 27, "virtual_lines": [ { "bbox": [ 171, 590, 441, 620.0 ], "spans": [], "index": 26 }, { "bbox": [ 171, 620.0, 441, 650.0 ], "spans": [], "index": 27 }, { "bbox": [ 171, 650.0, 441, 680.0 ], "spans": [], "index": 28 } ] }, { "type": "table_caption", "bbox": [ 107, 692, 504, 715 ], "group_id": 2, "lines": [ { "bbox": [ 106, 692, 505, 705 ], "spans": [ { "bbox": [ 106, 692, 505, 705 ], "score": 1.0, "content": "Table 11: Top rows: Impact of segmentation quality for AV-NSL with number of words segments", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 704, 406, 715 ], "spans": [ { "bbox": [ 106, 704, 406, 715 ], "score": 1.0, "content": "known in advance. Bottom rows: Impact of visual embedding for AV-NSL", "type": "text" } ], "index": 30 } ], "index": 29.5 } ], "index": 28.25 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 145, 505, 616 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 145, 505, 616 ], "group_id": 0, "lines": [ { "bbox": [ 106, 145, 505, 616 ], "spans": [ { "bbox": [ 106, 145, 505, 616 ], "score": 0.952, "type": "image", "image_path": "6eafec723763f1955cfa3db3f2fc4b4dc5b5264e53948ae6cd47f58566a414d3.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 145, 505, 302.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 302.0, 505, 459.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 459.0, 505, 616.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 622, 505, 646 ], "group_id": 0, "lines": [ { "bbox": [ 106, 622, 505, 635 ], "spans": [ { "bbox": [ 106, 622, 505, 635 ], "score": 1.0, "content": "Figure 4: Examples of attention segments generated by VG-HuBERT. Inserted segments are marked", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 633, 369, 646 ], "spans": [ { "bbox": [ 105, 633, 126, 646 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 126, 634, 142, 644 ], "score": 0.74, "content": "\" + \"", "type": "inline_equation" }, { "bbox": [ 142, 633, 369, 646 ], "score": 1.0, "content": ". Vertical blue dotted lines are inferred word boundaries.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 } ], "page_idx": 17, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 763 ], "spans": [ { "bbox": [ 299, 750, 313, 763 ], "score": 1.0, "content": "18", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 145, 505, 616 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 145, 505, 616 ], "group_id": 0, "lines": [ { "bbox": [ 106, 145, 505, 616 ], "spans": [ { "bbox": [ 106, 145, 505, 616 ], "score": 0.952, "type": "image", "image_path": "6eafec723763f1955cfa3db3f2fc4b4dc5b5264e53948ae6cd47f58566a414d3.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 145, 505, 302.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 302.0, 505, 459.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 459.0, 505, 616.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 622, 505, 646 ], "group_id": 0, "lines": [ { "bbox": [ 106, 622, 505, 635 ], "spans": [ { "bbox": [ 106, 622, 505, 635 ], "score": 1.0, "content": "Figure 4: Examples of attention segments generated by VG-HuBERT. Inserted segments are marked", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 633, 369, 646 ], "spans": [ { "bbox": [ 105, 633, 126, 646 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 126, 634, 142, 644 ], "score": 0.74, "content": "\" + \"", "type": "inline_equation" }, { "bbox": [ 142, 633, 369, 646 ], "score": 1.0, "content": ". Vertical blue dotted lines are inferred word boundaries.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 113, 314, 505, 457 ], "blocks": [ { "type": "image_body", "bbox": [ 113, 314, 505, 457 ], "group_id": 0, "lines": [ { "bbox": [ 113, 314, 505, 457 ], "spans": [ { "bbox": [ 113, 314, 505, 457 ], "score": 0.96, "type": "image", "image_path": "b9ba5c7f7252344a82826fa03c2de12600f34bc8ecb7e8a39029ac2f4d17e2ae.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 113, 314, 505, 361.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 113, 361.6666666666667, 505, 409.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 113, 409.33333333333337, 505, 457.00000000000006 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 463, 506, 497 ], "group_id": 0, "lines": [ { "bbox": [ 106, 462, 506, 476 ], "spans": [ { "bbox": [ 106, 462, 506, 476 ], "score": 1.0, "content": "Figure 5: Visualization of an example produced by AV-NSL (best viewed in color). Top (red and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 474, 506, 487 ], "spans": [ { "bbox": [ 105, 474, 506, 487 ], "score": 1.0, "content": "green): the ground-truth parse tree; bottom (blue and yellow): the generated parse tree. In each tree,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 486, 354, 497 ], "spans": [ { "bbox": [ 105, 486, 354, 497 ], "score": 1.0, "content": "a parent segment adjacently covers its two children segments.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ], "page_idx": 18, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 38 ], "lines": [ { "bbox": [ 107, 25, 308, 38 ], "spans": [ { "bbox": [ 107, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2023", "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": "19", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 113, 314, 505, 457 ], "blocks": [ { "type": "image_body", "bbox": [ 113, 314, 505, 457 ], "group_id": 0, "lines": [ { "bbox": [ 113, 314, 505, 457 ], "spans": [ { "bbox": [ 113, 314, 505, 457 ], "score": 0.96, "type": "image", "image_path": "b9ba5c7f7252344a82826fa03c2de12600f34bc8ecb7e8a39029ac2f4d17e2ae.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 113, 314, 505, 361.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 113, 361.6666666666667, 505, 409.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 113, 409.33333333333337, 505, 457.00000000000006 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 463, 506, 497 ], "group_id": 0, "lines": [ { "bbox": [ 106, 462, 506, 476 ], "spans": [ { "bbox": [ 106, 462, 506, 476 ], "score": 1.0, "content": "Figure 5: Visualization of an example produced by AV-NSL (best viewed in color). Top (red and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 474, 506, 487 ], "spans": [ { "bbox": [ 105, 474, 506, 487 ], "score": 1.0, "content": "green): the ground-truth parse tree; bottom (blue and yellow): the generated parse tree. In each tree,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 486, 354, 497 ], "spans": [ { "bbox": [ 105, 486, 354, 497 ], "score": 1.0, "content": "a parent segment adjacently covers its two children segments.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }