File size: 64,280 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 | {
"paper_id": "U07-1012",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T03:08:53.550977Z"
},
"title": "Charting Democracy Across Parsers",
"authors": [
{
"first": "Scott",
"middle": [],
"last": "Nowson",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Macquarie University Sydney",
"location": {
"country": "Australia"
}
},
"email": ""
},
{
"first": "Robert",
"middle": [],
"last": "Dale",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Macquarie University Sydney",
"location": {
"country": "Australia"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Different parsers trained on the same corpus deliver different results, both in terms of overall performance and in terms of the individual analyses they provide. In particular, for any given sentence, one parser may provide a correct analysis, while another will produce an incorrect analysis; but when faced with a different sentence, the first parser may be in error while the second is correct. In this paper, we leverage this observation by exploring how the results of a number of different parsers may be combined to provide a better performance than any single parser. The method involves constructing a chart that contains edges contributed by a collection of parsers, with a simple voting mechanism to choose the most preferred constituents; this provides a significant improvement in performance over any individual parser. More sophisticated voting mechanisms are also discussed.",
"pdf_parse": {
"paper_id": "U07-1012",
"_pdf_hash": "",
"abstract": [
{
"text": "Different parsers trained on the same corpus deliver different results, both in terms of overall performance and in terms of the individual analyses they provide. In particular, for any given sentence, one parser may provide a correct analysis, while another will produce an incorrect analysis; but when faced with a different sentence, the first parser may be in error while the second is correct. In this paper, we leverage this observation by exploring how the results of a number of different parsers may be combined to provide a better performance than any single parser. The method involves constructing a chart that contains edges contributed by a collection of parsers, with a simple voting mechanism to choose the most preferred constituents; this provides a significant improvement in performance over any individual parser. More sophisticated voting mechanisms are also discussed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Parsers make mistakes. This is perhaps most apparent when a parser trained on a given corpus is applied to data from a domain or genre different to that of the training corpus. One can, of course, retrain the parser on new data that is more representative of the texts to be handled; but annotation is an expensive process, and the literature does not provide a great deal of guidance as to how much annotation is 1 Scott Nowson is now at Appen Pty Ltd. required in order to obtain an acceptable result (but see Reichart and Rappoport (2007a) for some recent interesting results in this area).",
"cite_spans": [
{
"start": 512,
"end": 542,
"text": "Reichart and Rappoport (2007a)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Unfortunately, parsers make mistakes even on the corpora on which they are trained. Before we begin to consider how we might adapt a parser to a new domain, we are therefore interested in how we might improve the performance of existing parsers on the corpora used to derive their models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We make the observation that different parsers have different 'error profiles', by which we mean that different parsers do not necessarily make the same mistakes. Consider the following verb phrase taken from our test corpus: Figure 1 shows the analyses provided for this verb phrase by three different parsers, as an illustration of the kinds of disagreements that are common. In the first analysis, in is misclassified as a preposition, while in the second and third analyses it is correctly analysed as a particle. However, the second parse contains a misparse of the embedded VP buying futures, while this is correctly analysed in the first and third parses.",
"cite_spans": [],
"ref_spans": [
{
"start": 226,
"end": 234,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This leads us to the hypothesis that, if we were able to select for each parser those parts of individual parses that are more likely to be correct, then the overall result would be an improvement upon the analysis of any individual parser. We explore this hypothesis in this paper, by providing a framework within which the analyses of different parsers can be combined, and the overall best parse selected. The idea of combining the results of different parsers is not in itself new, so in Section 2 we briefly survey related work in this area. In Section 3 we describe our approach, which takes advantage of the central ideas in chart parsing to provide a way of combining parse results, and we describe the parsers used in our experiments. Section 4 describes the results achieved by our method, demonstrating a significant improvement upon the performance achieved by any individual parser. Section 5 discusses how the simple voting mechanism presented here can be made more elaborate, with the prospect of even better improvements in performance, and Section 6 concludes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": ". . . lock in profits by buying futures when futures prices fall",
"sec_num": null
},
{
"text": "The combination of the results of several different components that carry out the same tasksometimes referred to as the ensemble-based approach-has been employed and shown to be successful in a number of fields such as part-of-speech tagging (Halteren et al., 1998) , word sense disambiguation (Pederson, 2000) and question answering (Chu-Carroll et al., 2003) .",
"cite_spans": [
{
"start": 242,
"end": 265,
"text": "(Halteren et al., 1998)",
"ref_id": "BIBREF6"
},
{
"start": 294,
"end": 310,
"text": "(Pederson, 2000)",
"ref_id": null
},
{
"start": 334,
"end": 360,
"text": "(Chu-Carroll et al., 2003)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background: Combining Parsers",
"sec_num": "2"
},
{
"text": "There are a number of approaches that have been employed for parser combination. Henderson and Brill (1999) describe experiments that fall within two general approaches they label parse hybridization and parse switching. The most basic form of hybridization is constituent voting, whereby constituents in a parse are included if they can be found in the majority of contributing parses. A second approach is to use a na\u00efve Bayes classifier in order to learn how much each parser should be trusted.",
"cite_spans": [
{
"start": 81,
"end": 107,
"text": "Henderson and Brill (1999)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background: Combining Parsers",
"sec_num": "2"
},
{
"text": "The alternative to this approach is to deal only with complete parses. Henderson and Brill again experimented with two approaches: similarity switching, whereby the parse chosen is the one which scores highest when judged for similarity to the remaining parses in the set; and a second na\u00efve Bayes approach to selecting the parse with the highest probability of being the best. All four of Henderson and Brill's approaches produced better results than any of the contributing parsers achieved: their best result was a 30% reduction of precision error rate, a 6% reduction of recall error rate and an absolute Fscore increase of 1.58%. These ideas have been ex-tended to take into account more context in adaptation to dependency based parsers with similarly successful results (Zeman and\u017dabokrtsk\u00fd, 2005) .",
"cite_spans": [
{
"start": 777,
"end": 804,
"text": "(Zeman and\u017dabokrtsk\u00fd, 2005)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background: Combining Parsers",
"sec_num": "2"
},
{
"text": "Henderson and Brill (2000) followed their earlier combination approach with one based on creating an ensemble of complementary parsers. Each parser was based upon the same underlying algorithm, but trained on different data. By using bagging and boosting approaches, their ensemble outperformed all single parsers, with a 0.6% absolute improvement in F-score. In a similar vein, Reichart and Rappoport (2007b) generated a number of parsing models by training one parser on slightly different training corpora. The resulting outputs were compared in order to judge the parse quality: the greater the number of models in agreement, the higher the quality.",
"cite_spans": [
{
"start": 379,
"end": 409,
"text": "Reichart and Rappoport (2007b)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background: Combining Parsers",
"sec_num": "2"
},
{
"text": "Clegg and Shepherd (2005) explored a number of alternative approaches to ensemble parsing when deploying trained parsers in a new domain. Using basic constituent voting based on Brill and Henderson's (1999) method, they report similar improvements, mostly to precision but also to recall. They achieved equally promising results from their variants of parse switching. The first of these was fallback cascades in which parsers are stacked in order of decreasing levels of sophistication. When the more complex model fails, the next parser attempts to parse. The bottom parser may be less accurate, but will be the least likely to fail. Their second whole-parse approach they simply termed parse selection, though it is similar to Henderson and Brill's similarity switching. Clegg and Shepherd varied this by trying different similarity metrics, such as constituent overlap or lineage similarity. Sagae and Lavie (2006) apply a notion of reparsing to a two stage parser combination chartbased approach. Once all single parses are complete, the first stage is to store all possible constituents in a chart with a label, start and end positions, and a weighting. Identical constituents from different parses are merged by adding their weights. The second stage of the process is to run a bottomup parsing algorithm, but rather than use a weighted grammar, the parser is guided by the weighted set of constituents. They experimented with different approaches to setting the initial weights of each nonterminal label. By combining five parsers they were able to achieve a error reduction of 44% for preci-sion and 14% for recall, and an absolute F-score increase of 1.1% (though it is worth noting each of these are best improvements, made across a different run made with different settings).",
"cite_spans": [
{
"start": 896,
"end": 918,
"text": "Sagae and Lavie (2006)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background: Combining Parsers",
"sec_num": "2"
},
{
"text": "3 Our Approach",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background: Combining Parsers",
"sec_num": "2"
},
{
"text": "Our approach is based on the central idea in chart parsing (Earley 1970; Kay 1980) : for any ambiguous string, the constituents derived from multiple parses can be maintained in one data structure, so that subsequent parses can reuse previously derived partial analyses. The insight leveraged here is that the same idea can be applied to multiple parsers: just as the chart can contain multiple analyses for a string as delivered by one parser, it can just as easily contain multiple analyses delivered by several parsers, thus providing a single unified view of all the different analyses, and allowing us to easily determine where parsers agree and where they disagree.",
"cite_spans": [
{
"start": 59,
"end": 72,
"text": "(Earley 1970;",
"ref_id": null
},
{
"start": 73,
"end": 82,
"text": "Kay 1980)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Basic Idea",
"sec_num": "3.1"
},
{
"text": "This is a very simple idea, but one which enables the development of a variety of approaches to choosing which edges should be used in building a preferred parse; and, as we demonstrate below, even the simplest methods provide good results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Basic Idea",
"sec_num": "3.1"
},
{
"text": "Our approach to combination is built upon a basic voting strategy, methodologically similar to Henderson and Brill (1999) and Sagae and Lavie (2006) , with implementational similarity to the latter. In its purest form, voting is purely democratic: all nominated constituents are considered equally suitable candidates to fill a position in the parse, and the candidate with the most votes-i.e., the candidate proposed by a majority of the parsers-is the winner. The algorithm is simple:",
"cite_spans": [
{
"start": 126,
"end": 148,
"text": "Sagae and Lavie (2006)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Basic Idea",
"sec_num": "3.1"
},
{
"text": "1. Each sentence is parsed by multiple parsers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Basic Idea",
"sec_num": "3.1"
},
{
"text": "converted into a chart representation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The resulting Penn Treebank parse strings are",
"sec_num": "2."
},
{
"text": "3. Starting with the root node, voting takes place as to what the children of that node should be.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The resulting Penn Treebank parse strings are",
"sec_num": "2."
},
{
"text": "Step 3 is then repeated for each successful child node.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "5. When the tree is fully populated by terminal nodes, the final chart is returned as a Penn Treebank parse representation for evaluation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "We now describe these stages in more detail.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "The first stage in our process is to parse each sentence with the individual contributing parsers. In the experiments reported here, we use three parsers: the Stanford lexicalised parser (Klein and Manning, 2003) ; Collins generative parsing model number 2 (Collins, 1999) as re-implemented by Bikel 2004; and the OpenNLP parser (Baldridge et al., 2003) . These were chosen for two reasons:",
"cite_spans": [
{
"start": 187,
"end": 212,
"text": "(Klein and Manning, 2003)",
"ref_id": "BIBREF10"
},
{
"start": 257,
"end": 272,
"text": "(Collins, 1999)",
"ref_id": "BIBREF4"
},
{
"start": 329,
"end": 353,
"text": "(Baldridge et al., 2003)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing",
"sec_num": "3.2"
},
{
"text": "\u2022 all three parsers output parses in the standard Penn Treebank notation, making conversion to our chart representation the same process for all; and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing",
"sec_num": "3.2"
},
{
"text": "\u2022 all three are provided with Java API functionality making incorporation into one system more straightforward.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing",
"sec_num": "3.2"
},
{
"text": "This latter advantage also reduces computation times by enabling just a single parser initialisation step before parsing all sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Parsing",
"sec_num": "3.2"
},
{
"text": "Once each sentence has been parsed, the resulting Penn Treebank parse strings are converted into charts. In the standard approach, a chart is a collection of vertices that span sequences of one or more words of the input, with pairs of vertices that are connected by grammatically labeled edges; where a sequence of words is amenable to more than one analysis, each analysis is represented by a separate edge. Subsequent decisions, or some other choice mechanism, then determine which of the multiple analyses should be chosen.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Representation",
"sec_num": "3.3"
},
{
"text": "In the implementation used here, vertices are defined in terms of character positions, while edges are defined by the start and end positions and given a grammatical label. As a step towards efficient implementation, each edge contributed by a given parser also indicates the constituent edges-the grammatical children-contained within the span of that edge. By example, consider the sentence the cat sat, which is analysed as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Representation",
"sec_num": "3.3"
},
{
"text": "(S (NP (DT the) (NN cat)) (VP (VBD sat))) and is spanned thus:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Representation",
"sec_num": "3.3"
},
{
"text": "t h e c a t s a t 0 1 2 3 4 5 6 7 8 9 10 11 This analysis would be represented as a chart consisting of edges:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Representation",
"sec_num": "3.3"
},
{
"text": "(0, 12, S, {(0,7,NP), (8,11,VP)}} (0, 7, NP, {(0,3,DT), (4,7,NN)}) (0, 3, DT, {(0,3,the)}) (4, 7, NN, {(4,7,cat)}) ...",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Representation",
"sec_num": "3.3"
},
{
"text": "The fundamental difference between our approach and those of Henderson and Brill (1999) and Sagae and Lavie (2006) described earlier is in the strategy used when selecting constituents. Previous approaches have considered constituents in isolation: Sagae and Lavie's charts contain all possible constituents, each assigned a weight based on their presence across individual parsers, and these are merely used to inform a second stage, bottom-up reparsing. By comparison, our system could be described as a single-stage, top-down process which operates across the prior parses. Similar to Sagae and Lavie, we employ simple voting to determine the choice of constituents, but we consider only the nominated children of each already-decided constituent. Since each such set of children corresponds to a valid parse, we can ensure there will not be any crossing brackets, and that the resulting parse will be grammatically sound.",
"cite_spans": [
{
"start": 61,
"end": 87,
"text": "Henderson and Brill (1999)",
"ref_id": "BIBREF7"
},
{
"start": 92,
"end": 114,
"text": "Sagae and Lavie (2006)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Voting",
"sec_num": "3.4"
},
{
"text": "Each grammatical constituent is defined by an edge within the chart. For each edge that covers the same span of words 1 across the individual parser outputs, the set of potential analyses can be retrieved. Each such solution provides a potential constituent analysis with which to continue the parse, and for whom votes can be tallied. So, in a democratic manner, any child nominated by all contributing parsers is unanimously voted into the parse. Similarly, any constituents that obtain a majority vote also succeed. In the case of a tie, we resort to arbitrarily choosing between the potential solutions. The only restriction is that children are chosen so that the entire span is accounted for and a complete tree is created.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Voting",
"sec_num": "3.4"
},
{
"text": "Of course pure democracy, at least in the case of parser combination, is quite na\u00efve. It treats all candidates as equal and does not take past performance of parsers into consideration; nor does it take into account the possibility that some parsers may perform better in specific situations. Clearly a sensible step forward here is to move towards a more meritocratic approach, as discussed in Section 5 below.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Chart Voting",
"sec_num": "3.4"
},
{
"text": "As introduced earlier, Figure 1 illustrates how three different parsers can construct parses that differ or are similar in different ways. In this section, we walk through the combination of these parses to provide an example of how our approach works. To save space and to aid clarity, we represent the span of any node simply by listing the words contained in that span.",
"cite_spans": [],
"ref_spans": [
{
"start": 23,
"end": 31,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "An Example",
"sec_num": "3.5"
},
{
"text": "In our example, we begin part-way through the parse, where the the current node of interest is the VP which spans from lock to futures. The analyses of this node are retrieved from the charts delivered by the three parsers, and votes are calculated across the children: VBP ('lock') 3 votes PRT ('in') 2 votes NP ('profits') 2 votes PP ('by . . . futures') 2 votes PP ('in . . . futures') 1 vote Three votes represents a unanimous decision, while two is a majority; so, the decomposition of the VP node that is common to the second and third analyses is chosen.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Example",
"sec_num": "3.5"
},
{
"text": "Note that the PRT daughter of the VP node receives two votes in total, and subsequently so in turn does it's daughter, the RP. However, though the NP node also received just two votes, the NNS node at the next level of analysis receives three votes. This is because in the case of one of these analyses the NP node is buried deeper in the tree. Similarly, though the PP in buying futures was voted twice for its position in the tree, it can be found in all three parses at some level.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Example",
"sec_num": "3.5"
},
{
"text": "However, we have a disagreement as to the decomposition of the PP: IN ('by') 3 votes S ('buying futures') 2 votes NP ('buying futures') 1 vote Table 2 : Precision, Recall and F-score for individual parsers and their combination; all sentences (n = 2416).",
"cite_spans": [],
"ref_spans": [
{
"start": 143,
"end": 150,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "An Example",
"sec_num": "3.5"
},
{
"text": "Consequently, the chosen analysis of the PP is that proposed in the first and third trees.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Example",
"sec_num": "3.5"
},
{
"text": "We evaluate the results of our approach using the PARSEVAL standard Evalb (Sekine et al., 2006) . The input to the system is Section 23 of the Wall Street Journal (WSJ). All sentences are pretokenised to ensure standard input, though each parser executes its own part-of-speech tagging. The system outputs four sets of parse strings: one for each of the three constituent parsers, and one for the final combined result. The sets of parses are compared against the gold standard.",
"cite_spans": [
{
"start": 74,
"end": 95,
"text": "(Sekine et al., 2006)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "3.6"
},
{
"text": "We report the bracketing precision, recall and Fscore for sentences of length less than 40 words in Table 1 , and for all sentences in Table 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 100,
"end": 107,
"text": "Table 1",
"ref_id": "TABREF0"
},
{
"start": 135,
"end": 142,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "4"
},
{
"text": "It is clear that the combined system performs the best. Considering all sentences, we have achieved an error reduction of 22% for precision and 5% for recall, along with an absolute F-score increase of 2.3% over the best single contributor. In order to compare our results with those of previous studies, we reproduce the results of Henderson and Brill (1999) and Sagae and Lavie (2006) alongside our own in Ta-ble 3. Our results are comparable directly with those of Henderson and Brill; Sagae and Lavie's scores are a compilation of their best scores across three separate systems tuned to maximise each dimension, hence the high increase in precision and recall.",
"cite_spans": [
{
"start": 333,
"end": 359,
"text": "Henderson and Brill (1999)",
"ref_id": "BIBREF7"
},
{
"start": 364,
"end": 386,
"text": "Sagae and Lavie (2006)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "4"
},
{
"text": "As a further investigation, we employed a simple measure of confidence in a parse as a function of the number of parsers in the system, the total number of edges in the final chart and the total number of votes cast over just those successful edges: confidence = votes edges \u00d7 parsers Confidence will be highest if all the parsers agreed on each edge (had the same parse throughout) and will be lower the less they agree. Average confidence across our output is 0.88, which suggests that overall there was a high degree of agreement across parsers. The confidence measure also shows a significant correlation (p < .001) with the precision and recall scores across all sentences. This suggests that the system is most likely to be wrong when it is least confident in its output, and so the confidence metric is a good one.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "4"
},
{
"text": "The performance values reported in Tables 1 and 2 show that the combined system produces more accurate results than the original individual parsers, as we had hoped. By simply taking a majority vote on constituents, our system results in more correct constituent analyses than those proposed by the individual parsers. However, the combined result is not a huge improvement over the highest performing of its contributing proposals.",
"cite_spans": [],
"ref_spans": [
{
"start": 35,
"end": 49,
"text": "Tables 1 and 2",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "It is of course possible that for the most part, all parsers get the same things wrong -the rare and infrequent syntactic constructions. This would present a simple voting system with no way to select the correct analysis. However, it is likely that systems that get the same things wrong do so in the same way. Such agreement on incorrectness still represents an agreement, which would provide a high level of confidence in the incorrect choice. However, looking at our confidence scores, this incorrect agreement does not appear to be the case: errors appear to follow a lack of confidence -where there is most disagreement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "The biggest weakness in our approach lies in the arbitrary decision-making procedure used in breaking tied situations. In such tied situations, if the wrong result is chosen, then all the constituent analyses below that point have a high likelihood of being incorrect. This is a particular weakness of our top-down approach, in contrast to Sagae and Lavie's bottom-up method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "There are a variety of ways in which the basic model developed here could be extended. Of course, one could extend the mechanism beyond the three parsers that we use to incorporate a larger number of parsers. However, a more interesting direction is to improve the voting mechanism. The greatest number of errors appears to stem from situations of low agreement, when voting is tied. 2 One approach to resolving deadlocked situations such as these might be to employ a lookahead approach. As illustrated in our example in section 3.5 upon voting across the top level VP, the NP receives two votes. However, this is only because it was directly under the VP in two cases; the NP was in fact still present in the third analysis, but buried further down in the tree. In a tied situation, this fact would have argued for one analysis over the other.",
"cite_spans": [
{
"start": 384,
"end": 385,
"text": "2",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Another approach is to observe that, while democracy is fair other things being equal, parsers are more akin to experts to be consulted. For example, we might think of each parser as having particular areas of expertise, in the sense that its performance on some kinds of constituents might be better than others. If a given parser has a track record of performing well in the analysis of particular kinds of constituents or substructures, then that parser's vote should carry more weight.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "There are a number of approaches we might take to developing a more meritocratic decision procedure. Table 3 : Precision, Recall and F-score for parsers; The best individual parser from each study, plus the best combined results, and the differences between them.",
"cite_spans": [],
"ref_spans": [
{
"start": 101,
"end": 108,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "the system agree with a parser, the more popular it becomes, and the heavier its weighting. The downside, however, is that should one system perform well early on, its weighting may be so much that a local maxima may be reached.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Previous Track Record: This is also a general measure of performance, but is static and relies on external information. Weightings are set based on the prior performance of a parser: those that have previously produced most accurate results will be trusted more and weighted higher. One source for this data would be previously published, preferably comparable, results. However, as we noted at the start of the paper, good performance in one domain or genre does not guarantee similar results in another.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Two other measures, as suggested by Henderson and Brill (1999) , take context into account:",
"cite_spans": [
{
"start": 36,
"end": 62,
"text": "Henderson and Brill (1999)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Constituent-Level Track Record: The previous approach gives higher weighting to the parsers that have previously performed best overall, but this does not mean they were the best at everything. In this approach, we narrow the focus to performance over individual constituent types: higher weighting is given to a parser's vote, if upon prior evaluation it has proven successful at selecting the specific nominated constituent. The prerequisite to this is that performance analysis must have been carried out at the level of individual constituents. Alternatives might include using machine learning techniques to automatically determine which parsers do best in which situations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "Structural-Level Track Record: The approach above could be further extended to take account of a larger amount of syntactic context; for example, it might be the case that some parers are better at subject NPs but less good at object NPs. Here we would need to compute weights based on past performance on correct annotation of subtrees in an analysis; clearly this could be done at varying levels of granularity, modulo the problem of sparse data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "This paper reports work concerned with combining parsers using a chart based representation and voting scheme. It has introduced the methodology we will employ in our future parsing work: the outputs from multiple parsers are transformed into a chart representation; by voting over children these charts are combined into a single chart combining those constituents for which there is the strongest evidence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "The combination process pursued here is based on the simplest interpretation of evidence, where we pursue a purely democratic approach. This approach is most obviously deficient when we have to deal with ties. Nonetheless, the resulting parses prove more accurate than the single nominees that contributed to their creation, and performance compares well to previous studies that employ more complex and sophisticated methods. This suggests our approach has considerable scope for subsequent improvement, some possible directions for which we have outlined in the latter part of this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "Matching edges are defined by the tuple <start pos, end pos, label>.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Note that these situations are even more likely to occur if the system were to employ an even number of parsers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The authors acknowledge the financial support of the Capital Markets Cooperative Research Centre.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF1": {
"ref_id": "b1",
"title": "Distributional Analysis of a Lexicalized Statistical Parsing Model. Procedings of EMNLP",
"authors": [
{
"first": "M",
"middle": [],
"last": "Daniel",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Bikel",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel M. Bikel. 2004. Distributional Analysis of a Lexicalized Statistical Parsing Model. Procedings of EMNLP 2004, NJ.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Question Answering, Two Heads Are Better Than One. Proceedings of HLT-NAACL 2003",
"authors": [
{
"first": "Jennifer",
"middle": [],
"last": "Chu-Carroll",
"suffix": ""
},
{
"first": "Krzysztof",
"middle": [],
"last": "Czuba",
"suffix": ""
},
{
"first": "John",
"middle": [],
"last": "Prager",
"suffix": ""
},
{
"first": "Abraham",
"middle": [],
"last": "Ittycheriah",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "24--31",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jennifer Chu-Carroll, Krzysztof Czuba, John Prager and Abraham Ittycheriah 2003. In Question Answering, Two Heads Are Better Than One. Proceedings of HLT- NAACL 2003, 24-31.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Evaluating and integrating treebank parsers on a biomedical corpus",
"authors": [
{
"first": "B",
"middle": [],
"last": "Andrew",
"suffix": ""
},
{
"first": "Adrian",
"middle": [
"J"
],
"last": "Clegg",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Shepherd",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the 43rd Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrew B. Clegg and Adrian J. Shepherd. 2005. Evalu- ating and integrating treebank parsers on a biomedical corpus. Proceedings of the 43rd Meeting of the Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Head-Driven Statistical Models for Natural Language Parsing",
"authors": [
{
"first": "Michael",
"middle": [],
"last": "Collins",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Michael Collins. 1999. Head-Driven Statistical Mod- els for Natural Language Parsing. PhD Dissertation, University of Pennsylvania.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "An efficient context-free parsing algorithm",
"authors": [
{
"first": "J",
"middle": [],
"last": "Early",
"suffix": ""
}
],
"year": 1986,
"venue": "Readings in natural language processing",
"volume": "",
"issue": "",
"pages": "25--33",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Early. 1986. An efficient context-free parsing algo- rithm. In B. J. Grosz, K. Sparck-Jones & B. L. Web- ber (Eds), Readings in natural language processing, 25-33. Morgan Kaufmann, San Francisco, CA.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Improving data driven wordclass tagging by system combination",
"authors": [
{
"first": "Jakub",
"middle": [],
"last": "Hans Van Halteren",
"suffix": ""
},
{
"first": "Walter",
"middle": [],
"last": "Zavrel",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Daelemans",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of the 17th International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "491--497",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hans van Halteren, Jakub Zavrel and Walter Daelemans. 1998 Improving data driven wordclass tagging by sys- tem combination. Proceedings of the 17th Interna- tional Conference on Computational Linguistics, 491- 497.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Exploiting Diversity in Natural Language Processing: Combining Parsers",
"authors": [
{
"first": "C",
"middle": [],
"last": "John",
"suffix": ""
},
{
"first": "Eric",
"middle": [],
"last": "Henderson",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Brill",
"suffix": ""
}
],
"year": 1999,
"venue": "Proceedings of EMNLP 1999",
"volume": "",
"issue": "",
"pages": "187--194",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John C. Henderson and Eric Brill. 1999. Exploiting Di- versity in Natural Language Processing: Combining Parsers. Proceedings of EMNLP 1999, 187-194.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Bagging and Boosting a Treebank Parser",
"authors": [
{
"first": "C",
"middle": [],
"last": "John",
"suffix": ""
},
{
"first": "Eric",
"middle": [],
"last": "Henderson",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Brill",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of NAACL 2000",
"volume": "",
"issue": "",
"pages": "34--41",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John C. Henderson and Eric Brill. 2000. Bagging and Boosting a Treebank Parser. Proceedings of NAACL 2000, 34-41.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Algorithm schemata and data structures in syntactic processing",
"authors": [
{
"first": "M",
"middle": [],
"last": "Kay",
"suffix": ""
}
],
"year": 1986,
"venue": "Readings in natural language processing",
"volume": "",
"issue": "",
"pages": "35--70",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Kay. 1986. Algorithm schemata and data structures in syntactic processing. In B. J. Grosz, K. Sparck- Jones & B. L. Webber (Eds), Readings in natural lan- guage processing, 35-70. Morgan Kaufmann, San Francisco, CA.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Accurate Unlexicalized Parsing",
"authors": [
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 41st Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dan Klein and Christopher D. Manning. 2003. Accurate Unlexicalized Parsing. Proceedings of the 41st Meet- ing of the Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A Simple Approach to Building Ensembles of Naive Bayesian Classifiers for Word Sense Disambigusation",
"authors": [
{
"first": "Ted",
"middle": [],
"last": "Pedersen",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of NAACL 2000",
"volume": "",
"issue": "",
"pages": "63--69",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ted Pedersen. 2000. A Simple Approach to Building En- sembles of Naive Bayesian Classifiers for Word Sense Disambigusation. Proceedings of NAACL 2000, 63- 69.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Self-Training for Enhancement and Domain Adaptation of Statistical Parsers Trained on Small Datasets",
"authors": [
{
"first": "Roi",
"middle": [],
"last": "Reichart",
"suffix": ""
},
{
"first": "Ari",
"middle": [],
"last": "Rappoport",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 45th Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "616--623",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roi Reichart and Ari Rappoport. 2007. Self-Training for Enhancement and Domain Adaptation of Statisti- cal Parsers Trained on Small Datasets. Proceedings of the 45th Meeting of the Association for Computational Linguistics, 616-623.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "An Ensemble Method for Selection of High Quality Parses",
"authors": [
{
"first": "Roi",
"middle": [],
"last": "Reichart",
"suffix": ""
},
{
"first": "Ari",
"middle": [],
"last": "Rappoport",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 45th Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "408--415",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roi Reichart and Ari Rappoport. 2007. An Ensemble Method for Selection of High Quality Parses. Pro- ceedings of the 45th Meeting of the Association for Computational Linguistics, 408-415.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Parser Combination by Reparsing",
"authors": [
{
"first": "Kenji",
"middle": [],
"last": "Sagae",
"suffix": ""
},
{
"first": "Alon",
"middle": [],
"last": "Lavie",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of HLT-NAACL 2006",
"volume": "",
"issue": "",
"pages": "129--133",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kenji Sagae and Alon Lavie. 2006. Parser Combina- tion by Reparsing. Proceedings of HLT-NAACL 2006, 129-133.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Improving Parsing Accuracy by Combining Diverse Dependency Parser",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Zeman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Zden\u011bk\u017eabokrtsk\u00fd",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Ninth International Workshop on Parsing Technology",
"volume": "",
"issue": "",
"pages": "171--178",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Zeman and Zden\u011bk\u017dabokrtsk\u00fd 2006. Improving Parsing Accuracy by Combining Diverse Dependency Parser. Proceedings of the Ninth International Work- shop on Parsing Technology, 171-178.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"num": null,
"text": "Three analyses by different parsers",
"uris": null
},
"TABREF0": {
"num": null,
"text": "Precision, Recall and F-score for individual parsers and their combination; sentence length <= 40 words (n = 2245).",
"html": null,
"type_str": "table",
"content": "<table><tr><td>Parser</td><td>P</td><td>R</td><td>F</td></tr><tr><td>Stanford</td><td colspan=\"3\">87.0 85.7 86.4</td></tr><tr><td colspan=\"4\">OpenNLP 88.1 87.7 87.9</td></tr><tr><td>Collins</td><td colspan=\"3\">72.9 88.9 80.1</td></tr><tr><td colspan=\"4\">Combined 90.7 89.5 90.1</td></tr><tr><td>Parser</td><td>P</td><td>R</td><td>F</td></tr><tr><td>Stanford</td><td colspan=\"3\">86.4 85.0 85.7</td></tr><tr><td colspan=\"4\">OpenNLP 87.4 87.0 87.2</td></tr><tr><td>Collins</td><td colspan=\"3\">72.7 88.3 79.7</td></tr><tr><td colspan=\"4\">Combined 90.2 88.9 89.5</td></tr></table>"
}
}
}
} |