File size: 61,409 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 |
{
"paper_id": "I08-1016",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:40:17.712906Z"
},
"title": "Entity-driven Rewrite for Multi-document Summarization",
"authors": [
{
"first": "Ani",
"middle": [],
"last": "Nenkova",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Pennsylvania",
"location": {}
},
"email": "nenkova@seas.upenn.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In this paper we explore the benefits from and shortcomings of entity-driven noun phrase rewriting for multi-document summarization of news. The approach leads to 20% to 50% different content in the summary in comparison to an extractive summary produced using the same underlying approach, showing the promise the technique has to offer. In addition, summaries produced using entity-driven rewrite have higher linguistic quality than a comparison non-extractive system. Some improvement is also seen in content selection over extractive summarization as measured by pyramid method evaluation.",
"pdf_parse": {
"paper_id": "I08-1016",
"_pdf_hash": "",
"abstract": [
{
"text": "In this paper we explore the benefits from and shortcomings of entity-driven noun phrase rewriting for multi-document summarization of news. The approach leads to 20% to 50% different content in the summary in comparison to an extractive summary produced using the same underlying approach, showing the promise the technique has to offer. In addition, summaries produced using entity-driven rewrite have higher linguistic quality than a comparison non-extractive system. Some improvement is also seen in content selection over extractive summarization as measured by pyramid method evaluation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Two of the key components of effective summarizations are the ability to identify important points in the text and to adequately reword the original text in order to convey these points. Automatic text summarization approaches have offered reasonably well-performing approximations for identifiying important sentences (Lin and Hovy, 2002; Schiffman et al., 2002; Erkan and Radev, 2004; Mihalcea and Tarau, 2004; Daum\u00e9 III and Marcu, 2006) but, not surprisingly, text (re)generation has been a major challange despite some work on sub-sentential modification (Jing and McKeown, 2000; Knight and Marcu, 2000 ; Barzilay and McKeown, 2005 ). An additional drawback of extractive approaches is that estimates for the importance of larger text units such as sentences depend on the length of the sentence (Nenkova et al., 2006) .",
"cite_spans": [
{
"start": 319,
"end": 339,
"text": "(Lin and Hovy, 2002;",
"ref_id": "BIBREF8"
},
{
"start": 340,
"end": 363,
"text": "Schiffman et al., 2002;",
"ref_id": "BIBREF13"
},
{
"start": 364,
"end": 386,
"text": "Erkan and Radev, 2004;",
"ref_id": "BIBREF4"
},
{
"start": 387,
"end": 412,
"text": "Mihalcea and Tarau, 2004;",
"ref_id": "BIBREF11"
},
{
"start": 413,
"end": 439,
"text": "Daum\u00e9 III and Marcu, 2006)",
"ref_id": "BIBREF3"
},
{
"start": 559,
"end": 583,
"text": "(Jing and McKeown, 2000;",
"ref_id": "BIBREF5"
},
{
"start": 584,
"end": 606,
"text": "Knight and Marcu, 2000",
"ref_id": "BIBREF6"
},
{
"start": 609,
"end": 635,
"text": "Barzilay and McKeown, 2005",
"ref_id": "BIBREF0"
},
{
"start": 800,
"end": 822,
"text": "(Nenkova et al., 2006)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Sentence simplification or compaction algorithms are driven mainly by grammaticality considerations. Whether approaches for estimating importance can be applied to units smaller than sentences and used in text rewrite in the summary production is a question that remains unanswered. The option to operate on smaller units, which can be mixed and matched from the input to give novel combinations in the summary, offers several possible advantages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Improve content Sometimes sentences in the input can contain both information that is very appropriate to include in a summary and information that should not appear in a summary. Being able to remove unnecessary parts can free up space for better content. Similarly, a sentence might be good overall, but could be further improved if more details about an entity or event are added in. Overall, a summarizer capable of operating on subsentential units would in principle be better at content selection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Improve readability Linguistic quality evaluation of automatic summaries in the Document Understanding Conference reveals that summarizers perform rather poorly on several readability aspects, including referential clarity. The gap between human and automatic performance is much larger for linguistic quality aspects than for content selection. In more than half of the automatic summaries there were entities for which it was not clear what/who they were and how they were related to the story. The ability to add in descriptions for entities in the summaries could improve the referential clarity of summaries and can be achieved through text rewrite of subsentential units.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "IP issues Another very practical reason to be interested in altering the original wording of sentences in summaries in a news browsing system involves intellectual property issues. Newspapers are not willing to allow verbatim usage of long passages of their articles on commercial websites. Being able to change the original wording can thus allow companies to include longer than one sentence summaries, which would increase user satisfaction .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "These considerations serve as direct motivation for exploring how a simple but effective summarizer framework can accommodate noun phrase rewrite in multi-document summarization of news. The idea is for each sentence in a summary to automatically examine the noun phrases in it and decide if a different noun phrase is more informative and should be included in the sentence in place of the original. Consider the following example:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Sentence 1 The arrest caused an international controversy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The arrest in London of former Chilean dictator Augusto Pinochet caused an international controversy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence 2",
"sec_num": null
},
{
"text": "Now, consider the situation where we need to express in a summary that the arrest was controversial and this is the first sentence in the summary, and sentence 1 is available in the input (\"The arrest caused an international controversy\"), as well as an unrelated sentence such as \"The arrest in London of former Chilean dictator Augusto Pinochet was widely discussed in the British press\". NP rewrite can allow us to form the rewritten sentence 2, which would be a much more informative first sentence for the summary: \"The arrest in London of former Chilean dictator Augusto Pinochet caused an international controversy\". Similarly, if sentence 2 is available in the input and it is selected in the summary after a sentence that expresses the fact that the arrest took place, it will be more appropriate to rewrite sentence 2 into sentence 1 for inclusion in the summary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence 2",
"sec_num": null
},
{
"text": "This example shows the potential power of noun phrase rewrite. It also suggests that context will play a role in the rewrite process, since different noun phrase realizations will be most appropriate depending on what has been said in the summary up to the point at which rewrite takes place.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence 2",
"sec_num": null
},
{
"text": "Frequency and frequency-related measures of importance have been traditionally used in text summarization as indicators of importance (Luhn, 1958; Lin and Hovy, 2000; Conroy et al., 2006) . Notably, a greedy frequency-driven approach leads to very good results in content selection (Nenkova et al., 2006) . In this approach sentence importance is measured as a function of the frequency in the input of the content words in that sentence. The most important sentence is selected, the weight of words in it are adjusted, and sentence weights are recomputed for the new weights beofre selecting the next sentence. This conceptually simple summarization approach can readily be extended to include NP rewrite and allow us to examine the effect of rewrite capabilities on overall content selection and readability. The specific algorithm for frequency-driven summarization and rewrite is as follows:",
"cite_spans": [
{
"start": 134,
"end": 146,
"text": "(Luhn, 1958;",
"ref_id": "BIBREF9"
},
{
"start": 147,
"end": 166,
"text": "Lin and Hovy, 2000;",
"ref_id": "BIBREF7"
},
{
"start": 167,
"end": 187,
"text": "Conroy et al., 2006)",
"ref_id": "BIBREF2"
},
{
"start": 282,
"end": 304,
"text": "(Nenkova et al., 2006)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Step 1 Estimate the importance of each content word w i based on its frequency in the input",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "n i , p(w i ) = n i N .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Step 2 For each sentence S j in the input, estimate its importance based on the words in the sentence w i \u2208 S j : the weight of the sentence is equal to the average weight of content words appearing in it.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "W eight(S j ) = w i \u2208S j p(w i ) |w i \u2208S j |",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Step 3 Select the sentence with the highest weight.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Step 4 For each maximum noun phrase N P k in the selected sentence 4.1 For each coreferring noun phrase N P i , such that N P i \u2261 N P k from all input documents, compute a weight W eight(N P i ) = F RW (w r \u2208 N P i ). 4.2 Select the noun phrase with the highest weight and insert it in the sentence in place of the original NP. In case of ties, select the shorter noun phrase.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Step 5 For each content word in the rewritten sentence, update its weight by setting it to 0.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Step 6 If the desired summary length has not been reached, go to step 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Step 4 is the NP rewriting step. The function F RW is the rewrite composition function that assigns weights to noun phrases based on the importance of words that appear in the noun phrase. The two options that we explore here are F RW \u2261 Avr and F RW \u2261 Sum; the weight of an NP equals the average weight or sum of weights of content words in the NP respectively. The two selections lead to different behavior in rewrite. F RW \u2261 Avr will generally prefer the shorter noun phrases, typically consisting of just the noun phrase head and it will overall tend to reduce the selected sentence. F RW \u2261 Sum will behave quite differently: it will insert relevant information that has not been conveyed by the summary so far (add a longer noun phrase) and will reduce the NP if the words in it already appear in the summary. This means that F RW \u2261 Sum will have the behavior close to what we expect for entity-centric rewrite: inluding more descriptive information at the first mention of the entity, and using shorter references at subsequent mentions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Maximum noun phrases are the unit on which NP rewrite operates. They are defined in a dependency parse tree as the subtree that has as a root a noun such that there is no other noun on the path between it and the root of the tree. For example , there are two maximum NPs, with heads \"police\" and \"Augusto Pinochet\" in the sentence \"British police arrested former Chilean dictator Augusto Pinochet\". The noun phrase \"former chilean dictator\" is not a maximum NP, since there is a noun (augusto pinochet) on the path in the dependency tree between the noun \"dictator\" and the root of the tree. By definition a maximum NP includes all nominal and adjectival premodifiers of the head, as well as postmodifiers such as prepositional phrases, appositions, and relative clauses. This means that maximum NPs can be rather complex, covering a wide range of production rules in a context-free grammar.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "The dependency tree definition of maximum noun phrase makes it easy to see why these are a good unit for subsentential rewrite: the subtree that has the head of the NP as a root contains only modifiers of the head, and by rewriting the noun phrase, the amount of information expressed about the head entity can be varied.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "In our implementation, a context free grammar probabilistic parser (Charniak, 2000) was used to parse the input. The maximum noun phrases were identified by finding sequences of <np>...</np> tags in the parse such that the number of opening and closing tags is equal. Each NP identified by such tag spans was considered as a candidate for rewrite.",
"cite_spans": [
{
"start": 67,
"end": 83,
"text": "(Charniak, 2000)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "Coreference classes A coreference class CR m is the class of all maximum noun phrases in the input that refer to the same entity E m . The general problem of coreference resolution is hard, and is even more complicated for the multi-document summarization case, in which cross-document resolution needs to be performed. Here we make a simplifying assumption, stating that all noun phrases that have the same noun as a head belong to the same coreference class. While we expected that this assumption would lead to some wrong decisions, we also suspected that in most common summarization scenarios, even if there are more than one entities expressed with the same noun, only one of them would be the main focus for the news story and will appear more often across input sentences. References to such main entities will be likely to be picked in a sentence for inclusion in the summary by chance more often than other competeing entities. We thus used the head noun equivalance to form the classes. A post-evaluation inspection of the summaries confirmed that our assumption was correct and there were only a small number of errors in the rewritten summaries that were due to coreference errors, which were greatly outnumbered by parsing errors for example. In a future evaluation, we will evaluate the rewrite module assuming perfect coreference and parsing, in order to see the impact of the core NP-rewrite approach itself.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-rewrite enhanced frequency summarizer",
"sec_num": "2"
},
{
"text": "The NP rewrite summarization algorithm was applied to the 50 test sets for generic multi-document summarization from the 2004 Document Understanding Conference. Two examples of its operation with F RW \u2261 Avr are shown below. Original.1 While the British government defended the arrest, it took no stand on extradition of Pinochet to Spain. NP-Rewite.1 While the British government defended the arrest in London of former Chilean dictator Augusto Pinochet, it took no stand on extradition of Pinochet to Spain. Original.2 Duisenberg has said growth in the euro area countries next year will be about 2.5 percent, lower than the 3 percent predicted earlier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP rewrite evaluation",
"sec_num": "3"
},
{
"text": "European Central Bank, has said growth in the euro area will be about 2.5 percent, lower than just 1 percent in the euro-zone unemployment predicted earlier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-Rewrite.2 Wim Duisenberg, the head of the new",
"sec_num": null
},
{
"text": "We can see that in both cases, the NP rewrite pasted into the sentence important additional information. But in the second example we also see an error that was caused by the simplifying assumption for the creation of the coreference classes according to which the percentage of unemployment and growth have been put in the same class.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-Rewrite.2 Wim Duisenberg, the head of the new",
"sec_num": null
},
{
"text": "In order to estimate how much the summary is changed because of the use of the NP rewrite, we computed the unigram overlap between the original extractive summary and the NP-rewrite summary. As expected, F F W \u2261 Sum leads to bigger changes and on average the rewritten summaries contained only 54% of the unigrams from the extractive summaries; for F RW \u2261 Avr, there was a smaller change between the extractive and the rewritten summary, with 79% of the unigrams being the same between the two summaries.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NP-Rewrite.2 Wim Duisenberg, the head of the new",
"sec_num": null
},
{
"text": "Noun phrase rewrite has the potential to improve the referential clarity of summaries, by inserting in the sentences more information about entities when such is available. It is of interest to see how the rewrite version of the summarizer would compare to the extractive version, as well as how its linguistic quality compares to that of other summarizers that participated in DUC. Four summarizers were evaluated: peer 117, which was a system that used generation techniques to produce the summary and was the only real non-extractive summarizer participant at DUC 2004 (Vanderwende et al., 2004) ; the extractive frequency summarizer, and the two versions of the rewrite algorithm (Sum and Avr). The evaluated rewritten summaries had potential errors coming from different sources, such as coreference resolution, parsing errors, sentence splitting errors, as well as errors coming directly from rewrite, in which an unsuitable NP is chosen to be included in the summary. Improvements in parsing for example could lead to better overall rewrite results, but we evaluated the output as is, in order to see what is the performance that can be expected in a realistic setting for fully automatic rewrite. The evaluation was done by five native English speakers, using the five DUC linguistic quality questions on grammaticality (Q 1 ), repetition (Q 2 ), referential clarity (Q 3 ), focus (Q 4 ) and coherence (Q 5 ). Five evaluators were used so that possible idiosyncratic preference of a single evaluator could be avoided. Each evaluator evaluated all five summaries for each test set, presented in a random order. The results are shown in table 3.1. Each summary was evaluated for each of the properties on a scale from 1 to 5, with 5 being very good with respect to the quality and 1, very bad.",
"cite_spans": [
{
"start": 572,
"end": 598,
"text": "(Vanderwende et al., 2004)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic quality evaluation",
"sec_num": "3.1"
},
{
"text": "SYSTEM Q 1 Q 2 Q 3 Q 4 Q 5 SUM",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic quality evaluation",
"sec_num": "3.1"
},
{
"text": "Comparing NP rewrite to extraction Here we would be interested in comparing the extractive frequency summarizer (SUM Id ), and the two version of systems that rewrite noun phrases: SUM Avr (which changes about 20% of the text) and SUM Sum (which changes about 50% of the text). The general trend that we see for all five dimensions of linguistic quality is that the more the text is automatically altered, the worse the linguistic quality of the summary gets. In particular, the grammaticality of the summaries drops significantly for the rewrite systems. The increase of repetition is also significant between SUM Id and SUM Sum . Error analysis showed that sometimes increased repetition occurred in the process of rewrite for the following reason: the context weight update for words is done only after each noun phrase in the sentence has been rewritten. Occasionally, this led to a situation in which a noun phrase was augmented with information that was expressed later in the original sentence. The referential clarity of rewritten summaries also drops significantly, which is a rather disappointing result, since one of the motivations for doing noun phrase rewrite was the desire to improve referential clarity by adding information where such is necessary. One of the problems here is that it is almost impossible for human evaluators to ignore grammatical errors when judging referential clarity. Grammatical errors decrease the overall readability and a summary that is given a lower grammaticality score tends to also receive lower referential clarity score. This fact of quality perception is a real challenge for summarizeration systems that move towards abstraction and alter the original wording of sentences since certainly automatic approaches are likely to introduce ingrammaticalities.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic quality evaluation",
"sec_num": "3.1"
},
{
"text": "Comparing SUM Sum and peer 117 We now turn to the comparison of between SUM Sum and the generation based system 117. This system is unique among the DUC 2004 systems, and the only one that year that experimented with generation techniques for summarization. System 117 is verbdriven: it analizes the input in terms of predicateargument triples and identifies the most important triples. These are then verbalized by a generation system originally developed as a realization component in a machine translation engine. As a result, peer 117 possibly made even more changes to the original text then the NP-rewrite system. The results of the comparison are consistent with the observation that the more changes are made to the original sentences, the more the readability of summaries decreases. SUM Sum is significantly better than peer 117 on all five readability aspects, with notable difference in the grammaticality and referential quality, for which SUM Sum outperforms peer 117 by a full point. This indicates that NPs are a good candidate granularity for sentence changes and it can lead to substantial altering of the text while preserving significantly better overall readability.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic quality evaluation",
"sec_num": "3.1"
},
{
"text": "We now examine the question of how the content in the summaries changed due to the NP-rewrite, since improving content selection was the other motivation for exploring rewrite. In particular, we are interested in the change in content selection between SUM Sum and SUM Id (the extractive version of the summarizer). We use SUM Sum for the comparison because it led to bigger changes in the summary text compared to the purely extractive version. We used the pyramid evaluation method: four human summaries for each input were manually analyzed to identify shared content units. The weight of each content unit is equal to the number of model summaries that express it. The pyramid score of an automatic summary is equal to the weight of the content units expressed in the summary divided by the weight of an ideally informative summary of the same length (the content unit identification is again done manually by an annotator).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Content selection evaluation",
"sec_num": "3.2"
},
{
"text": "Of the 50 test sets, there were 22 sets in which the NP-rewritten version had lower pyramid scores than the extractive version of the summary, 23 sets in which the rewritten summaries had better scores, and 5 sets in which the rewritten and extractive summaries had exactly the same scores. So we see that in half of the cases the NP-rewrite actually improved the content of the summary. The summarizer version that uses NP-rewrite has overall better content selection performance than the purely extractive system. The original pyramid score increased from 0.4039 to 0.4169 for the version with rewrite. This improvement is not significant, but shows a trend in the expected direction of improvement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Content selection evaluation",
"sec_num": "3.2"
},
{
"text": "The lack of significance in the improvement is due to large variation in performance: when np rewrite worked as expected, content selection improved. But on occasions when errors occurred, both readability and content selection were noticeably compromised. Here is an example of summaries for the same input in which the NP-rewritten version had better content. After each summary, we list the content units from the pyramid content analysis that were expressed in the summary. The weight of each content unit is given in brackets before the label of the unit and content units that differ between the extractive and rewritten version are displayed in italic. The rewritten version conveys high weight content units that do not appear in the extractive version, with weights 4 (maximum weight here) and 3 respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Content selection evaluation",
"sec_num": "3.2"
},
{
"text": "Extractive summary Italy's Communist Refounding Party rejected Prime Minister Prodi's proposed 1999 budget. By one vote, Premier Romano Prodi's center-left coalition lost a confidence vote in the Chamber of Deputies Friday, and he went to the presidential palace to rsign. Three days after the collapse of Premier Romano Prodi's center-left government, Italy's president began calling in political leaders Monday to try to reach a consensus on a new government. Prodi has said he would call a confidence vote if he lost the Communists' support.\" I have always acted with coherence,\" Prodi said before a morning meeting with President Oscar Luigi.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Content selection evaluation",
"sec_num": "3.2"
},
{
"text": "(4) Prodi lost a confidence vote (4) The Refounding Party is Italy's Communist Party (4) The Refounding Party rejected the government's budget",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Content selection evaluation",
"sec_num": "3.2"
},
{
"text": "(3) The dispute is over the 1999 budget (2) Prodi's coalition was center-left coalition (2) The confidence vote was lost by only 1 vote (1) Prodi is the Italian Prime Minister (1) Prodi wants a confidence vote from Parliament NP-rewrite version Communist Refounding, a fringe group of hard-line leftists who broke with the minstream Communists after they overhauled the party following the collapse of Communism in Eastern Europe rejected Prime Minister Prodi's proposed 1999 budget. By only one vote, the center-left prime minister of Italy, Romano Prodi, lost The vote in the lower chamber of Parliament 313 against the confidence motion brought by the government to 312 in favor in Parliament Friday and was toppled from power. President Oscar Luigi Scalfaro, who asked him to stay on as caretaker premier while the head of state decides whether to call elections. Below is another example, showing the worse deterioration of the rewritten summary compared to the extractive one, both in terms of grammaticality and content. Here, the problem with repetition during rewrite arises: the same person is mentioned twice in the sentence and at both places the same overly long description is selected during rewrie, rendering the sentence practically unreadable.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Content selection evaluation",
"sec_num": "3.2"
},
{
"text": "McKinney lured Shepard from the bar by saying they too were gay and one of their girlfriends said Shepard had embarrassed one of the men by making a pass at him. 1,000 people mourned Matthew Shepherd, the gay University of Wyoming student who was severely beaten and left to die tied to a fence. With passersby spontaneously joining the protest group, two women held another sign that read,\" No Hate Crimes in Wyoming.\" Two candlelight vigils were held Sunday night. Russell Anderson, 21, and Aaron McKinney, 21, were charged with attempted murder.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "(2) There were candlelight vigils in support for the victim",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "(1) Russell Henderson and Aaron McKinney are the names of the people responsible for the death NP-rewrite version Police said Henderson and McKinney lured the The slight, soft-spoken 21year-old Shepard, a freshman at the University of Wyoming, who became an overnight symbol of antigay violence after he was found dangling from the fence by a passerby from a bar by saying they too were gay and one of their girlfriends said the The slight, soft-spoken 21-year-old Shepard, a freshman at the University of Wyoming, who became an overnight symbol of anti-gay violence after he was found dangling from the fence by a passerby had embarrassed one of the new ads in that supposedly hate-free crusade.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "(4) The victim was a student at the University of Wyoming",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "(3)The nearly lifeless body was tied to a fence (1) A passerby found the victim",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "(1) Russell Henderson and Aaron McKinney are the names of the people responsible for the death",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "(1) The victim was 22-year old",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "Even from this unsuccessful attempt for rewrite we can see how changes of the original text can be desirable, since some of the newly introduced information is in fact suitable for the summary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extractive summary Police said Henderson and",
"sec_num": null
},
{
"text": "We have demonstrated that an entity-driven approach to rewrite in multi-document summarization can lead to considerably different summary, in terms of content, compared to the extractive version of the same system. Indeed, the difference leads to some improvement measurable in terms of pyramid method evaluation. The approach also significantly outperforms in linguistic quality a non-extractive event-centric system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "4"
},
{
"text": "Results also show that in terms of linguistic quality, extractive systems will be curently superior to systems that alter the original wording from the input. Sadly, extractive and abstractive systems are evaluated together and compared against each other, putting pressure on system developers and preventing them from fully exploring the strengths of generation techniques. It seems that if researchers in the field are to explore non-extractive methods, they would need to compare their systems separately from extractive systems, at least in the beginning exploration stages. The development of nonextractive approaches in absolutely necessary if automatic summarization were to achieve levels of performance close to human, given the highly abstractive form of summaries written by people.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "4"
},
{
"text": "Results also indicate that both extractive and nonextractive systems perform rather poorly in terms of the focus and coherence of the summaries that they produce, identifying macro content planning as an important area for summarization.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "4"
},
{
"text": "(4) The victim was a student at the University of Wyoming (4) The victim was brutally beaten (4) The victim was openly gay (3) The crime was widely denounced (3) The nearly lifeless body was tied to a fence (3) The victim died (3) The victim was left to die (2) The men were arrested on charges of kidnapping and attempted first degree murder",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Sentence fusion for multidocument news summarization",
"authors": [
{
"first": "Regina",
"middle": [],
"last": "Barzilay",
"suffix": ""
},
{
"first": "Kathleen",
"middle": [],
"last": "Mckeown",
"suffix": ""
}
],
"year": 2005,
"venue": "Computational Linguistics",
"volume": "",
"issue": "3",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Regina Barzilay and Kathleen McKeown. 2005. Sen- tence fusion for multidocument news summarization. Computational Linguistics, 31(3).",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A maximum-entropy-inspired parser",
"authors": [
{
"first": "Eugene",
"middle": [],
"last": "Charniak",
"suffix": ""
}
],
"year": 2000,
"venue": "NAACL-2000",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eugene Charniak. 2000. A maximum-entropy-inspired parser. In NAACL-2000.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Topic-focused multi-document summarization using an approximate oracle score",
"authors": [
{
"first": "John",
"middle": [],
"last": "Conroy",
"suffix": ""
},
{
"first": "Judith",
"middle": [],
"last": "Schlesinger",
"suffix": ""
},
{
"first": "Dianne O'",
"middle": [],
"last": "Leary",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of ACL, companion volume",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John Conroy, Judith Schlesinger, and Dianne O'Leary. 2006. Topic-focused multi-document summarization using an approximate oracle score. In Proceedings of ACL, companion volume.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Bayesian queryfocused summarization",
"authors": [
{
"first": "Hal",
"middle": [],
"last": "Daum\u00e9",
"suffix": ""
},
{
"first": "Iii",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Conference of the Association for Computational Linguistics (ACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hal Daum\u00e9 III and Daniel Marcu. 2006. Bayesian query- focused summarization. In Proceedings of the Confer- ence of the Association for Computational Linguistics (ACL), Sydney, Australia.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Lexrank: Graph-based centrality as salience in text summarization",
"authors": [
{
"first": "Gunes",
"middle": [],
"last": "Erkan",
"suffix": ""
},
{
"first": "Dragomir",
"middle": [],
"last": "Radev",
"suffix": ""
}
],
"year": 2004,
"venue": "Journal of Artificial Intelligence Research",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gunes Erkan and Dragomir Radev. 2004. Lexrank: Graph-based centrality as salience in text summa- rization. Journal of Artificial Intelligence Research (JAIR).",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Cut and paste based text summarization",
"authors": [
{
"first": "Hongyan",
"middle": [],
"last": "Jing",
"suffix": ""
},
{
"first": "Kathleen",
"middle": [],
"last": "Mckeown",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 1st Conference of the North American Chapter of the Association for Computational Linguistics (NAACL'00)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hongyan Jing and Kathleen McKeown. 2000. Cut and paste based text summarization. In Proceedings of the 1st Conference of the North American Chap- ter of the Association for Computational Linguistics (NAACL'00).",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Statistics-based summarization -step one: Sentence compression",
"authors": [
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceeding of The American Association for Artificial Intelligence Conference (AAAI-2000)",
"volume": "",
"issue": "",
"pages": "703--710",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kevin Knight and Daniel Marcu. 2000. Statistics-based summarization -step one: Sentence compression. In Proceeding of The American Association for Artificial Intelligence Conference (AAAI-2000), pages 703-710.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The automated acquisition of topic signatures for text summarization",
"authors": [
{
"first": "Chin-Yew",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Eduard",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 18th conference on Computational linguistics",
"volume": "",
"issue": "",
"pages": "495--501",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chin-Yew Lin and Eduard Hovy. 2000. The automated acquisition of topic signatures for text summarization. In Proceedings of the 18th conference on Computa- tional linguistics, pages 495-501.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Automated multi-document summarization in neats",
"authors": [
{
"first": "Chin-Yew",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Eduard",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the Human Language Technology Conference (HLT2002 )",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chin-Yew Lin and Eduard Hovy. 2002. Automated multi-document summarization in neats. In Proceed- ings of the Human Language Technology Conference (HLT2002 ).",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The automatic creation of literature abstracts",
"authors": [
{
"first": "H",
"middle": [
"P"
],
"last": "Luhn",
"suffix": ""
}
],
"year": 1958,
"venue": "IBM Journal of Research and Development",
"volume": "2",
"issue": "2",
"pages": "159--165",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. P. Luhn. 1958. The automatic creation of literature abstracts. IBM Journal of Research and Development, 2(2):159-165.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Do summaries help? a task-based evaluation of multi-document summarization",
"authors": [
{
"first": "K",
"middle": [],
"last": "Mckeown",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Passonneau",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Elson",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Nenkova",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Hirschberg",
"suffix": ""
}
],
"year": 2005,
"venue": "SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. McKeown, R. Passonneau, D. Elson, A. Nenkova, and J. Hirschberg. 2005. Do summaries help? a task-based evaluation of multi-document summariza- tion. In SIGIR.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Textrank: Bringing order into texts",
"authors": [
{
"first": "R",
"middle": [],
"last": "Mihalcea",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Tarau",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of EMNLP 2004",
"volume": "",
"issue": "",
"pages": "404--411",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Mihalcea and P. Tarau. 2004. Textrank: Bringing or- der into texts. In Proceedings of EMNLP 2004, pages 404-411.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "A compositional context sensitive multidocument summarizer: exploring the factors that influence summarization",
"authors": [
{
"first": "Ani",
"middle": [],
"last": "Nenkova",
"suffix": ""
},
{
"first": "Lucy",
"middle": [],
"last": "Vanderwende",
"suffix": ""
},
{
"first": "Kathleen",
"middle": [],
"last": "Mckeown",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ani Nenkova, Lucy Vanderwende, and Kathleen McKe- own. 2006. A compositional context sensitive multi- document summarizer: exploring the factors that influ- ence summarization. In Proceedings of SIGIR.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Experiments in multidocument summarization",
"authors": [
{
"first": "Barry",
"middle": [],
"last": "Schiffman",
"suffix": ""
},
{
"first": "Ani",
"middle": [],
"last": "Nenkova",
"suffix": ""
},
{
"first": "Kathleen",
"middle": [],
"last": "Mckeown",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the Human Language Technology Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Barry Schiffman, Ani Nenkova, and Kathleen McKeown. 2002. Experiments in multidocument summarization. In Proceedings of the Human Language Technology Conference.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Event-centric summary generation",
"authors": [
{
"first": "Lucy",
"middle": [],
"last": "Vanderwende",
"suffix": ""
},
{
"first": "Michele",
"middle": [],
"last": "Banko",
"suffix": ""
},
{
"first": "Arul",
"middle": [],
"last": "Menezes",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the Document Understanding Conference (DUC'04)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lucy Vanderwende, Michele Banko, and Arul Menezes. 2004. Event-centric summary generation. In Pro- ceedings of the Document Understanding Conference (DUC'04).",
"links": null
}
},
"ref_entries": {}
}
} |