File size: 47,342 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T02:10:58.689711Z"
},
"title": "An Empirical Study of Compound PCFGs",
"authors": [
{
"first": "Yanpeng",
"middle": [],
"last": "Zhao",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Ivan",
"middle": [],
"last": "Titov Eae",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Compound probabilistic context-free grammars (C-PCFGs) have recently established a new state of the art for phrase-structure grammar induction. However, due to the high timecomplexity of chart-based representation and inference, it is difficult to investigate them comprehensively. In this work, we rely on a fast implementation of C-PCFGs to conduct evaluation complementary to that of Kim et al. (2019). We highlight three key findings: (1) C-PCFGs are data-efficient, (2) C-PCFGs make the best use of global sentence-level information in preterminal rule probabilities, and (3) the best configurations of C-PCFGs on English do not always generalize to morphology-rich languages.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Compound probabilistic context-free grammars (C-PCFGs) have recently established a new state of the art for phrase-structure grammar induction. However, due to the high timecomplexity of chart-based representation and inference, it is difficult to investigate them comprehensively. In this work, we rely on a fast implementation of C-PCFGs to conduct evaluation complementary to that of Kim et al. (2019). We highlight three key findings: (1) C-PCFGs are data-efficient, (2) C-PCFGs make the best use of global sentence-level information in preterminal rule probabilities, and (3) the best configurations of C-PCFGs on English do not always generalize to morphology-rich languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Probabilistic context-free grammars (PCFGs) have been used for unsupervised constituency grammar learning since decades ago (Lari and Young, 1990) , while learning PCFGs with the Expectation Maximization algorithm (Dempster et al., 1977) has been difficult as being involving non-convex optimization. Recently, Kim et al. (2019) propose compound PCFGs, an over-parameterized neural model that extends corpus-level PCFGs by defining a mixture of PCFGs per sentence. C-PCFGs have achieved the state-of-the-art performance on English and Chinese treebanks. They are also shown to be effective in a visually-grounded learning setting (Zhao and Titov, 2020) . However, because of the high time-complexity of chart-based representation and inference, it is hard to inspect C-PCFGs comprehensively.",
"cite_spans": [
{
"start": 124,
"end": 146,
"text": "(Lari and Young, 1990)",
"ref_id": "BIBREF3"
},
{
"start": 214,
"end": 237,
"text": "(Dempster et al., 1977)",
"ref_id": "BIBREF1"
},
{
"start": 311,
"end": 328,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
},
{
"start": 630,
"end": 652,
"text": "(Zhao and Titov, 2020)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work, we rely on a fast implementation 1 of C-PCFGs to conduct a set of experiments complementary to those of Kim et al. (2019) . Our 1 https://github.com/zhaoyanpeng/cpcfg. first experiment concerns data efficiency and length generalization of C-PCFGs. We empirically find that though trained only on short sentences, e.g., shorter than 30 tokens, C-PCFGs can generalize to longer sentences while maintaining high performance (54.8% F1) at test time. We further investigate which factors contribute to the good performance of C-PCFGs. Since a major difference between C-PCFGs and vanilla PCFGs is that C-PCFGs define sentence-dependent rule probabilities by using global sentence-level information, we ablate C-PCFGs by removing it from start, preterminal, and nonterminal rules, 2 individually. Our experimental results show that sentence-level information is most effective for preterminal rules. Despite the impressive performance of C-PCFGs on English, it is still unclear whether they can generalize to other languages. We thus conduct multilingual evaluation of C-PCFGs on the SPMRL dataset (Seddah et al., 2014) . The experimental results suggest that the best configurations of C-PCFGs on English do not necessarily generalize to morphology-rich languages.",
"cite_spans": [
{
"start": 118,
"end": 135,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
},
{
"start": 1106,
"end": 1127,
"text": "(Seddah et al., 2014)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Compound PCFGs provide a novel parameterization of PCFGs. Unlike PCFGs, which assign each grammar rule r a non-negative scalar \u03c0 r such that r:A \u03b3 \u03c0 r = 1 for each given left-hand-side symbol A, C-PCFGs relax the strong context-free assumption of PCFGs by assuming that rule probabilities follow a compound distribution:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "\u03c0 r = g r (z; \u03b8), z \u223c p(z) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "where p(z) is a prior distribution and allows for capturing interdependencies between the rules; g r (z; \u03b8) takes a latent z as input and incorporates the interdependencies into rule probabilities. Typically, g r (z; \u03b8) is parameterized by flexible neural networks and is amenable to gradient-based optimization techniques (we refer interested readers to Kim et al. (2019) for the detailed parameterization).",
"cite_spans": [
{
"start": 355,
"end": 372,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "Learning C-PCFGs is formalized as maximizing the log likelihood of each observed sentence w = w 1 w 2 . . . w n :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "log p \u03b8 (w) = log z t\u2208T G (w) p \u03b8 (t|z)p(z) dz ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "where T G (w) consists of all parses of a sentence w under a PCFG G. As standard in learning latent variable models, C-PCFGs resort to variational inference for tractable learning and instead maximize the evidence lower bound (ELBO):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "log p \u03b8 (w) \u2265 ELBO(w; \u03c6, \u03b8) = E q \u03c6 (z|w) [log p \u03b8 (w|z)] \u2212 KL[q \u03c6 (z|w)||p(z)] ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "where the first term computes the expected log likelihood under a variational posterior q \u03c6 (z|w); the KL term can be estimated analytically when p(z) and q \u03c6 (z|w) are normally distributed. q \u03c6 (z|w) is parameterized by a neural network and defines a distribution over the latent z.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "C-PCFGs satisfy the context-free assumption conditioned on z and thus admit tractable inference for each given z. Inference with C-PCFGs seeks the most probable parse t * of w:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "t * = argmax z p \u03b8 (t|w, z)p \u03b8 (z|w) dz .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "Though given z, the maximum a posterior (MAP) inference over p \u03b8 (t|w, z) can be exactly solved by using the CYK algorithm, the integral over z renders inference intractable. The MAP inference is instead approximated by:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "t * \u2248 argmax z p \u03b8 (t|w, z)\u03b4(z \u2212 \u00b5 \u03c6 (w)) dz ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "where \u03b4(\u2022) is the Dirac delta function and \u00b5 \u03c6 (w) is the mean vector of the variational posterior.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "Similarly to C-PCFGs, neural PCFGs (N-PCFGs) also use neural networks to parameterize PCFGs, but their parameterization does not rely on the sentence-dependent z. In the following discussion, we will refer to z as 'sentence embedding'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Compound PCFGs",
"sec_num": "2"
},
{
"text": "Datasets: We investigate the parsing performance of C-PCFGs across ten languages. Specifically, we conduct experiments on the Wall Street Journal (WSJ) corpus of the Penn Treebank (Marcus et al., 1994) for English, the Penn Chinese Treebank 5.1 (CTB) (Xue et al., 2005) for Chinese, and eight additional treebanks from the SPMRL 2014 shared task (Seddah et al., 2014) for the other eight languages (Basque, German, French, Hebrew, Hungarian, Korean, Polish, Swedish). We use the standard data splits for each treebank. Following Kim et al. (2019) , punctuation is removed from all data; the top 10000 frequent tokens in the training data of each treebank are kept as the vocabulary. 3 Unless otherwise specified, we train C-PCFGs on sentences no longer than 40 tokens.",
"cite_spans": [
{
"start": 180,
"end": 201,
"text": "(Marcus et al., 1994)",
"ref_id": "BIBREF4"
},
{
"start": 251,
"end": 269,
"text": "(Xue et al., 2005)",
"ref_id": "BIBREF7"
},
{
"start": 346,
"end": 367,
"text": "(Seddah et al., 2014)",
"ref_id": "BIBREF6"
},
{
"start": 529,
"end": 546,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "3"
},
{
"text": "We re-implement C-PCFGs relying on Torch-Struct (Rush, 2020) and adopt the same hyperparameter settings as in Kim et al. (2019) . We train C-PCFGs for each language separately. On each treebank we run C-PCFGs four times with different random seeds and for 30 epochs. The best model in each run is selected according to the perplexity on the validation data. At test time, trivial spans, such as single-word and sentence-level spans, are ignored. We report average corpus-and sentence-level F1 numbers as well as the unbiased standard deviations.",
"cite_spans": [
{
"start": 48,
"end": 60,
"text": "(Rush, 2020)",
"ref_id": "BIBREF5"
},
{
"start": 110,
"end": 127,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model hyperparameters and evaluation:",
"sec_num": null
},
{
"text": "We compare C-PCFGs against three trivial baselines (left-/ right-branching model and random trees) and a neural PCFG model. In short, C-PCFGs beats all baselines in terms of corpus-and sentence-level F1 (see the second row of Table 1 ). Our re-implementation of C-PCFGs reaches the highest sentence-level F1, slightly outperforming the model of Kim et al. (2019) by 0.5% F1. To give an in-depth analysis of the model gains, we present recall numbers on six most frequent constituent labels in the test data (NP, VP, PP, SBAR, ADJP, ADVP). Unsurprisingly, C-PCFGs achieve the best recall for most labels (4 out of 6 constituent labels). However, on verb phrases (VPs) they fall far behind the right-branching baseline (-30 Table 1 : Recall on six frequent constituent labels (NP, VP, PP, SBAR, ADJP, ADVP) in the WSJ test data, corpuslevel F1 (C-F1), and sentence-level F1 (S-F1) results. The best mean number in each column is in bold. \u2020 denotes results reported by Kim et al. (2019) . L# indicates that the models are trained on sentences no longer than # tokens. recall), presumably because VPs are longer and involve more complex linguistic structures. We further plot distributions of the six labels across constituent lengths (see Figure 1) . We can see that VPs are nearly uniformly distributed over different constituent lengths. In contrast, noun phrases (NPs) account for 61% of short constituents that have less than 6 tokens and cover 51% of total constituents. It suggests that C-PCFGs can recognize local and short constituents with a high accuracy but struggles with long constituents; there is clearly a room for improvement on VPs. 0.03 0.04 0.02 0.01 0.01 0.01 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.21 0.14 0.09 0.07 0.05 0.04 0.04 0.03 0.02 0.02 0.02 ",
"cite_spans": [
{
"start": 345,
"end": 362,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
},
{
"start": 717,
"end": 721,
"text": "(-30",
"ref_id": null
},
{
"start": 966,
"end": 983,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [
{
"start": 226,
"end": 233,
"text": "Table 1",
"ref_id": null
},
{
"start": 722,
"end": 729,
"text": "Table 1",
"ref_id": null
},
{
"start": 1236,
"end": 1245,
"text": "Figure 1)",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Main results",
"sec_num": "4.1"
},
{
"text": "A crucial aspect of human languages is their compositionality. Humans can derive grammar rules from a few sentences and combine the rules to generate new sentences compositionally. As C-PCFGs are backed by context-free grammar, we hypothesize that C-PCFGs are data-efficient and have a good generalizability to unseen sentences and constituents.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data efficiency and length generalization",
"sec_num": "4.2"
},
{
"text": "We design a length-generalization test to verify our hypothesis. Specifically, we train C-PCFGs using training sentences of length equal to or below a chosen sentence length. We choose five sentence lengths, 10, 20, 30, 40 and 50, indicated by L10, L20, L30, L40 and L50, respectively (see the third row of Table 1)). Figure 3 illustrates sentence-level F1 numbers on the test data of WSJ. Overall, training C-PCFGs on more / longer sentences results in higher F1 numbers. But using training sentences longer than 40 tokens only trivially improves the performance (+0.2% F1). Given that 97% test sentences are shorter than 40 tokens, we conjecture that training sentences shorter than 40 tokens can adequately cover lexical / structural characteristics in the test data. On the other hand, longer sentences have a larger tree space and probably make it harder for the model to learn. Notably, discarding training sentences longer than 30 tokens only decreases the model performance by 1.2% F1, suggesting that C-PCFGs are data-efficient.",
"cite_spans": [],
"ref_spans": [
{
"start": 318,
"end": 326,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data efficiency and length generalization",
"sec_num": "4.2"
},
{
"text": "We also conduct a constituent-length generalization test to study the generalizability of C-PCFGs on unseen long constituents. Since the test data of WSJ is too small to provide reliable statistics across constituent lengths, we test C-PCFGs on training sentences and report F1 numbers across constituent lengths (see Figure 2) . In general, F1 numbers become lower as constituent length increases. This is reasonable because large constituents merge from small constituents; errors in small constituents accumulate when composing larger constituents.",
"cite_spans": [],
"ref_spans": [
{
"start": 318,
"end": 327,
"text": "Figure 2)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data efficiency and length generalization",
"sec_num": "4.2"
},
{
"text": "We further investigate the influence of training data on length generalization. We use sentence lengths 30 and 40 as an illustration. Compare with L40C-PCFG, when tested on constituents longer than 40 tokens, L30C-PCFG shows a slightly better generalizability (see Figure 2b) . It consistently outperforms L40C-PCFG on sentences of length ranging from 30 to 40, though L40C-PCFG can access all sentences shorter than 40 tokens during training. This implies that C-PCFGs generalize well from short sentences; using additional long training sentences may hurt the generalizability. We also plots the proportions of constituent lengths. For example, there are about 6400 constituents of length from 30 to 40, which account for about 1.1% of total constituents, suggesting that the conclusion about the better generalizability of L30C-PCFGs is reliable. Figure 2b visualizes the performance of an L40N-PCFG. Surprisingly, L40N-PCFG shows the best generalizability on long constituents. Where does the F1 improvement of C-PCFGs over N-PCFGs come from? Look at the F1 numbers on shorter constituents in Figure 2a , clearly C-PCFGs are better on constituents that are shorter than 11 tokens, while L40N-PCFGs consistently outperform C-PCFGs on constituents of length above 11. L30C-PCFGs fall in between L40C-PCFGs and L40N-PCFGs, once again showing that restricting training to short sentences can endow C-PCFGs good parsing performance as well as lead to improved generalization.",
"cite_spans": [],
"ref_spans": [
{
"start": 265,
"end": 275,
"text": "Figure 2b)",
"ref_id": null
},
{
"start": 850,
"end": 859,
"text": "Figure 2b",
"ref_id": null
},
{
"start": 1097,
"end": 1106,
"text": "Figure 2a",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data efficiency and length generalization",
"sec_num": "4.2"
},
{
"text": "C-PCFGs demonstrate a significant improvement over N-PCFGs. Compare with N-PCFGs, C-PCFGs use an additional sentence embedding (i.e., the latent variable z, see Section 2) to parameterize sentence-specific PCFGs. Concretely, the sentence embedding is used to parameterize three types of rules: preterminal rules (P), nonterminal rules (N), and start rules (R). We would like to know which type of rules makes the best use of the sentence embedding? To this end, we let a C-PCFG use corpus-level parameters for each of the three types of rules, individually, i.e., parameters for a rule type are shared among sentences. Interestingly, C-PCFGs degenerate into N-PCFGs when using corpus-level parameters for preterminal rules (see Figure 4 ). It implies that the sentence embedding is most crucial for the parameterization of preterminal rules, presumably because the sentence embedding helps preterminal rules derive the knowledge of Kim et al. (2019) , we observe that C-PCFGs suffer a huge variance, e.g., on the Chinese and Swedish treebanks. part-of-speech tags, which is beneficial for parsing. ",
"cite_spans": [
{
"start": 932,
"end": 949,
"text": "Kim et al. (2019)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [
{
"start": 728,
"end": 736,
"text": "Figure 4",
"ref_id": "FIGREF5"
}
],
"eq_spans": [],
"section": "Model ablation",
"sec_num": "4.3"
},
{
"text": "Despite the surprisingly good performance on English, it is still unclear whether C-PCFGs can generalize to languages beyond English. We thus conduct multilingual evaluation of C-PCFGs on nine additional languages (see Tablel 2). When training C-PCFGs on the nine languages, we use the hyperparameters of the best-performing C-PCFG on English, i.e., we tune C-PCFGs only on WSJ and use the best configurations on the other treebanks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multilingual evaluation",
"sec_num": "4.4"
},
{
"text": "We can see that C-PCFGs achieve the highest overall mean F1 (average F1 number over all treebanks), though they have two fewer winning treebanks than N-PCFGs. Notably, both C-PCFGs and N-PCFGs outperform the trivial baselines by a large margin, suggesting their nice generalizability on languages beyond English. However, they are worse than the right-branching baseline on the Polish and Swedish treebanks. As these languages have rich morphologies, we anticipate an improvement from encoding the knowledge of morphologies into the sentence embedding.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multilingual evaluation",
"sec_num": "4.4"
},
{
"text": "We have presented an in-depth analysis of C-PCFGs from a quantitative perspective. The analysis concerns three aspects of C-PCFGs: data efficiency and length generalization, the role of the latent sentence embedding, and multilingual performance. Our experimental results show that C-PCFGs can learn well only from short sentences and maintain good performance at test time. The latent sentence embedding is crucial for the good performance of C-PCFGs. Among the three rule types, preterminal rules make the most of it. However, the configurations of the best-performing C-PCFGs on English do not always generalize to morphologyrich languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
},
{
"text": "Start rules generate a nonterminal symbol from the start symbol S (e.g., S A), preterminal rules generate a word from a nonterminal symbol (e.g., A w), and nonterminal rules are binary rules of the form A BC, which involve only nonterminal symbols.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "A unified data preprocessing pipeline is available at https://github.com/zhaoyanpeng/xcfg.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We would like to thank anonymous reviewers for their suggestions and comments. The project was supported by the European Research Council (ERC Starting Grant BroadSem 678254) and the Dutch National Science Foundation (NWO VIDI 639.022.518).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Model Chinese Basque German French Hebrew Hungarian Korean Polish Swedish Mean",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Model Chinese Basque German French Hebrew Hungarian Korean Polish Swedish Mean",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Maximum likelihood from incomplete data via the em algorithm",
"authors": [
{
"first": "A",
"middle": [
"P"
],
"last": "References",
"suffix": ""
},
{
"first": "N",
"middle": [
"M"
],
"last": "Dempster",
"suffix": ""
},
{
"first": "D",
"middle": [
"B"
],
"last": "Laird",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Rubin",
"suffix": ""
}
],
"year": 1977,
"venue": "Journal of the Royal Statistical Society: Series B (Methodological)",
"volume": "39",
"issue": "1",
"pages": "1--22",
"other_ids": {
"DOI": [
"10.1111/j.2517-6161.1977.tb01600.x"
]
},
"num": null,
"urls": [],
"raw_text": "References A. P. Dempster, N. M. Laird, and D. B. Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Soci- ety: Series B (Methodological), 39(1):1-22.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Compound probabilistic context-free grammars for grammar induction",
"authors": [
{
"first": "Yoon",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Dyer",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Rush",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "2369--2385",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1228"
]
},
"num": null,
"urls": [],
"raw_text": "Yoon Kim, Chris Dyer, and Alexander Rush. 2019. Compound probabilistic context-free grammars for grammar induction. In Proceedings of the 57th An- nual Meeting of the Association for Computational Linguistics, pages 2369-2385, Florence, Italy. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "The estimation of stochastic context-free grammars using the insideoutside algorithm",
"authors": [
{
"first": "K",
"middle": [],
"last": "Lari",
"suffix": ""
},
{
"first": "S",
"middle": [
"J"
],
"last": "Young",
"suffix": ""
}
],
"year": 1990,
"venue": "Computer Speech and Language",
"volume": "4",
"issue": "1",
"pages": "35--56",
"other_ids": {
"DOI": [
"10.1016/0885-2308(90)90022-X"
]
},
"num": null,
"urls": [],
"raw_text": "K. Lari and S.J. Young. 1990. The estimation of stochastic context-free grammars using the inside- outside algorithm. Computer Speech and Language, 4(1):35 -56.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "The Penn Treebank: Annotating predicate argument structure",
"authors": [
{
"first": "Mitchell",
"middle": [],
"last": "Marcus",
"suffix": ""
},
{
"first": "Grace",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Mary",
"middle": [
"Ann"
],
"last": "Marcinkiewicz",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Macintyre",
"suffix": ""
},
{
"first": "Ann",
"middle": [],
"last": "Bies",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Ferguson",
"suffix": ""
},
{
"first": "Karen",
"middle": [],
"last": "Katz",
"suffix": ""
},
{
"first": "Britta",
"middle": [],
"last": "Schasberger",
"suffix": ""
}
],
"year": 1994,
"venue": "Human Language Technology: Proceedings of a Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mitchell Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. 1994. The Penn Treebank: Annotating predicate ar- gument structure. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Torch-struct: Deep structured prediction library",
"authors": [
{
"first": "Alexander",
"middle": [],
"last": "Rush",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
"volume": "",
"issue": "",
"pages": "335--342",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-demos.38"
]
},
"num": null,
"urls": [],
"raw_text": "Alexander Rush. 2020. Torch-struct: Deep structured prediction library. In Proceedings of the 58th An- nual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 335- 342, Online. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Introducing the SPMRL 2014 shared task on parsing morphologically-rich languages",
"authors": [
{
"first": "Djam\u00e9",
"middle": [],
"last": "Seddah",
"suffix": ""
},
{
"first": "Sandra",
"middle": [],
"last": "K\u00fcbler",
"suffix": ""
},
{
"first": "Reut",
"middle": [],
"last": "Tsarfaty",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the First Joint Workshop on Statistical Parsing of Morphologically Rich Languages and Syntactic Analysis of Non-Canonical Languages",
"volume": "",
"issue": "",
"pages": "103--109",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Djam\u00e9 Seddah, Sandra K\u00fcbler, and Reut Tsarfaty. 2014. Introducing the SPMRL 2014 shared task on pars- ing morphologically-rich languages. In Proceedings of the First Joint Workshop on Statistical Parsing of Morphologically Rich Languages and Syntactic Analysis of Non-Canonical Languages, pages 103- 109, Dublin, Ireland. Dublin City University.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The penn chinese treebank: Phrase structure annotation of a large corpus",
"authors": [
{
"first": "Naiwen",
"middle": [],
"last": "Xue",
"suffix": ""
},
{
"first": "Fei",
"middle": [],
"last": "Xia",
"suffix": ""
},
{
"first": "Fu-Dong",
"middle": [],
"last": "Chiou",
"suffix": ""
},
{
"first": "Marta",
"middle": [],
"last": "Palmer",
"suffix": ""
}
],
"year": 2005,
"venue": "Natural Language Engineering",
"volume": "11",
"issue": "2",
"pages": "207--238",
"other_ids": {
"DOI": [
"10.1017/S135132490400364X"
]
},
"num": null,
"urls": [],
"raw_text": "Naiwen Xue, Fei Xia, Fu-dong Chiou, and Marta Palmer. 2005. The penn chinese treebank: Phrase structure annotation of a large corpus. Natural Lan- guage Engineering, 11(2):207-238.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Visually grounded compound PCFGs",
"authors": [
{
"first": "Yanpeng",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "Ivan",
"middle": [],
"last": "Titov",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "4369--4379",
"other_ids": {
"DOI": [
"10.18653/v1/2020.emnlp-main.354"
]
},
"num": null,
"urls": [],
"raw_text": "Yanpeng Zhao and Ivan Titov. 2020. Visually grounded compound PCFGs. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4369-4379, Online. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"num": null,
"type_str": "figure",
"text": "Label distribution over constituent lengths on the WSJ test data. All denotes frequencies of constituent lengths. Zero frequencies are due to the limited numerical precision.",
"uris": null
},
"FIGREF2": {
"num": null,
"type_str": "figure",
"text": "F1 w.r.t. constituent length above 30",
"uris": null
},
"FIGREF3": {
"num": null,
"type_str": "figure",
"text": "F1 numbers broken down by constituent lengths on the WSJ training data. During training, constituents (sentences) longer than 30 tokens (L30) are unseen to L30C-PCFG and are unseen to L40C-PCFG and L40N-PCFG when longer than 40 tokens (L40). F1 numbers on the WSJ test data with varying maximum lengths of training sentences.",
"uris": null
},
"FIGREF5": {
"num": null,
"type_str": "figure",
"text": "F1 numbers on the WSJ test data. Shared P / N / R indicates C-PCFGs that use corpus-level parameters for preterminal / nonterminal / start rules (see Section 4.3).",
"uris": null
},
"TABREF0": {
"content": "<table><tr><td>Model NP</td><td>VP</td><td>PP</td><td>SBAR</td><td>ADJP</td><td>ADVP</td><td>C-F1</td><td>S-F1</td></tr><tr><td>Left Branching 10.4</td><td>0.5</td><td>5.0</td><td>5.3</td><td>2.5</td><td>8.0</td><td>6.0</td><td>8.7</td></tr><tr><td>Right Branching 24.1</td><td>71.5</td><td>42.4</td><td>68.7</td><td>27.7</td><td>38.1</td><td>36.1</td><td>39.5</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td>\u00b10.1</td></tr><tr><td>\u2020 N-PCFG 71.2</td><td>33.8</td><td>58.8</td><td>52.5</td><td>32.5</td><td>45.5</td><td/><td>50.8</td></tr><tr><td colspan=\"8\">N-\u00b13.8</td></tr><tr><td>\u2020 C-PCFG 74.7</td><td>41.7</td><td>68.8</td><td>56.1</td><td>40.4</td><td>52.5</td><td/><td>55.2</td></tr><tr><td>C-PCFG 76.7</td><td/><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td/><td/><td/><td/><td>.8%</td></tr></table>",
"text": "Random Trees 22.5 \u00b10.3 12.3 \u00b10.3 19.0 \u00b10.5 9.3 \u00b10.6 24.3 \u00b11.7 26.9 \u00b11.3 15.3 \u00b10.1 18.1 PCFG 72.2 \u00b14.8 31.4 \u00b19.7 66.8 \u00b14.7 50.2 \u00b19.1 46.3 \u00b15.7 55.2 \u00b15.0 49.0 \u00b13.5 50.8 \u00b12.0 40.7 \u00b15.5 71.3 \u00b12.1 53.8 \u00b13.1 45.9 \u00b12.8 64.2 \u00b12.8 53.5 \u00b11.4 55.7 \u00b11.3 L50C-PCFG 76.9 \u00b13.6 40.7 \u00b13.7 72.3 \u00b10.6 60.1 \u00b15.5 46.9 \u00b15.8 63.2 \u00b15.0 53.8 \u00b12.1 55.9 \u00b11.9 L40C-PCFG 76.7 \u00b12.0 40.7 \u00b15.5 71.3 \u00b12.1 53.8 \u00b13.1 45.9 \u00b12.8 64.2 \u00b12.8 53.5 \u00b11.4 55.7 \u00b11.3 L30C-PCFG 74.5 \u00b12.8 38.4 \u00b11.7 71.1 \u00b11.2 59.7 \u00b14.8 44.2 \u00b14.1 64.3 \u00b13.1 52.5 \u00b11.5 54.8 \u00b11.4 L20C-PCFG 72.4 \u00b12.3 36.5 \u00b11.1 69.2 \u00b11.7 54.1 \u00b13.2 41.9 \u00b12.3 58.1 \u00b17.1 50.6 \u00b10.9 52.8 \u00b10.7 L10C-PCFG 67.1 \u00b13.8 31.0 \u00b19.8 61.3 \u00b12.2 45.9 \u00b18.2 36.7 \u00b12.3 41.3 \u00b16.0 45.5 \u00b12.4 48.2 \u00b12.3",
"html": null,
"num": null,
"type_str": "table"
},
"TABREF1": {
"content": "<table/>",
"text": "Sentence-level F1 numbers on multilingual treebanks. Similarly to",
"html": null,
"num": null,
"type_str": "table"
}
}
}
} |