File size: 83,885 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 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 | {
"paper_id": "P12-1011",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T09:27:16.243709Z"
},
"title": "Labeling Documents with Timestamps: Learning from their Time Expressions",
"authors": [
{
"first": "Nathanael",
"middle": [],
"last": "Chambers",
"suffix": "",
"affiliation": {},
"email": "nchamber@usna.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Temporal reasoners for document understanding typically assume that a document's creation date is known. Algorithms to ground relative time expressions and order events often rely on this timestamp to assist the learner. Unfortunately, the timestamp is not always known, particularly on the Web. This paper addresses the task of automatic document timestamping, presenting two new models that incorporate rich linguistic features about time. The first is a discriminative classifier with new features extracted from the text's time expressions (e.g., 'since 1999'). This model alone improves on previous generative models by 77%. The second model learns probabilistic constraints between time expressions and the unknown document time. Imposing these learned constraints on the discriminative model further improves its accuracy. Finally, we present a new experiment design that facilitates easier comparison by future work.",
"pdf_parse": {
"paper_id": "P12-1011",
"_pdf_hash": "",
"abstract": [
{
"text": "Temporal reasoners for document understanding typically assume that a document's creation date is known. Algorithms to ground relative time expressions and order events often rely on this timestamp to assist the learner. Unfortunately, the timestamp is not always known, particularly on the Web. This paper addresses the task of automatic document timestamping, presenting two new models that incorporate rich linguistic features about time. The first is a discriminative classifier with new features extracted from the text's time expressions (e.g., 'since 1999'). This model alone improves on previous generative models by 77%. The second model learns probabilistic constraints between time expressions and the unknown document time. Imposing these learned constraints on the discriminative model further improves its accuracy. Finally, we present a new experiment design that facilitates easier comparison by future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "This paper addresses a relatively new task in the NLP community: automatic document dating. Given a document with unknown origins, what characteristics of its text indicate the year in which the document was written? This paper proposes a learning approach that builds constraints from a document's use of time expressions, and combines them with a new discriminative classifier that greatly improves previous work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The temporal reasoning community has long depended on document timestamps to ground rela-tive time expressions and events (Mani and Wilson, 2000; Llid\u00f3 et al., 2001) . For instance, consider the following passage from the TimeBank corpus (Pustejovsky et al., 2003) :",
"cite_spans": [
{
"start": 122,
"end": 145,
"text": "(Mani and Wilson, 2000;",
"ref_id": "BIBREF12"
},
{
"start": 146,
"end": 165,
"text": "Llid\u00f3 et al., 2001)",
"ref_id": "BIBREF11"
},
{
"start": 238,
"end": 264,
"text": "(Pustejovsky et al., 2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "And while there was no profit this year from discontinued operations, last year they contributed 34 million, before tax.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Reconstructing the timeline of events from this document requires extensive temporal knowledge, most notably, the document's creation date to ground its relative expressions (e.g., this year = 2012). Not only did the latest TempEval competitions (Verhagen et al., 2007; Verhagen et al., 2009) include tasks to link events to the (known) document creation time, but state-of-the-art event-event ordering algorithms also rely on these timestamps (Chambers and Jurafsky, 2008; Yoshikawa et al., 2009) . This knowledge is assumed to be available, but unfortunately this is not often the case, particularly on the Web. Document timestamps are growing in importance to the information retrieval (IR) and management communities as well. Several IR applications depend on knowledge of when documents were posted, such as computing document relevance (Li and Croft, 2003; Dakka et al., 2008) and labeling search queries with temporal profiles (Diaz and Jones, 2004; Zhang et al., 2009) . Dating documents is similarly important to processing historical and heritage collections of text. Some of the early work that motivates this paper arose from the goal of automatically grounding documents in their historical contexts (de Jong et al., 2005; Kanhabua and Norvag, 2008; Kumar et al., 2011) . This paper builds on their work by incorporating more linguistic knowledge and explicit reasoning into the learner.",
"cite_spans": [
{
"start": 246,
"end": 269,
"text": "(Verhagen et al., 2007;",
"ref_id": "BIBREF15"
},
{
"start": 270,
"end": 292,
"text": "Verhagen et al., 2009)",
"ref_id": "BIBREF16"
},
{
"start": 444,
"end": 473,
"text": "(Chambers and Jurafsky, 2008;",
"ref_id": "BIBREF0"
},
{
"start": 474,
"end": 497,
"text": "Yoshikawa et al., 2009)",
"ref_id": "BIBREF17"
},
{
"start": 842,
"end": 862,
"text": "(Li and Croft, 2003;",
"ref_id": "BIBREF10"
},
{
"start": 863,
"end": 882,
"text": "Dakka et al., 2008)",
"ref_id": "BIBREF1"
},
{
"start": 934,
"end": 956,
"text": "(Diaz and Jones, 2004;",
"ref_id": "BIBREF4"
},
{
"start": 957,
"end": 976,
"text": "Zhang et al., 2009)",
"ref_id": "BIBREF18"
},
{
"start": 1213,
"end": 1235,
"text": "(de Jong et al., 2005;",
"ref_id": "BIBREF3"
},
{
"start": 1236,
"end": 1262,
"text": "Kanhabua and Norvag, 2008;",
"ref_id": "BIBREF6"
},
{
"start": 1263,
"end": 1282,
"text": "Kumar et al., 2011)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The first part of this paper describes a novel learning approach to document dating, presenting a discriminative model and rich linguistic features that have not been applied to document dating. Further, we introduce new features specific to absolute time expressions. Our model outperforms the generative models of previous work by 77%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The second half of this paper describes a novel learning algorithm that orders time expressions against the unknown timestamp. For instance, the phrase the second quarter of 1999 might be labeled as being before the timestamp. These labels impose constraints on the possible timestamp and narrow down its range of valid dates. We combine these constraints with our discriminative learner and see another relative improvement in accuracy by 9%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Most work on dating documents has come from the IR and knowledge management communities interested in dating documents with unknown origins. de Jong et al. (2005) was among the first to automatically label documents with dates. They learned unigram language models (LMs) for specific time periods and scored articles with log-likelihood ratio scores. Kanhabua and Norvag (2008; extended this approach with the same model, but expanded its unigrams with POS tags, collocations, and tf-idf scores. They also integrated search engine results as features, but did not see an improvement. Both works evaluated on the news genre.",
"cite_spans": [
{
"start": 144,
"end": 162,
"text": "Jong et al. (2005)",
"ref_id": "BIBREF3"
},
{
"start": 351,
"end": 377,
"text": "Kanhabua and Norvag (2008;",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work",
"sec_num": "2"
},
{
"text": "Recent work by Kumar et al. (2011) focused on dating Gutenberg short stories. As above, they learned unigram LMs, but instead measured the KLdivergence between a document and a time period's LM. Our proposed models differ from this work by applying rich linguistic features, discriminative models, and by focusing on how time expressions improve accuracy. We also study the news genre.",
"cite_spans": [
{
"start": 15,
"end": 34,
"text": "Kumar et al. (2011)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work",
"sec_num": "2"
},
{
"text": "The only work we are aware of within the NLP community is that of Dalli and Wilks (2006) . They computed probability distributions over different time periods (e.g., months and years) for each observed token. The work is similar to the above IR work in its bag of words approach to classification.",
"cite_spans": [
{
"start": 66,
"end": 88,
"text": "Dalli and Wilks (2006)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work",
"sec_num": "2"
},
{
"text": "They focused on finding words that show periodic spikes (defined by the word's standard deviation in its distribution over time), weighted with inverse document frequency scores. They evaluated on a subset of the Gigaword Corpus (Graff, 2002) .",
"cite_spans": [
{
"start": 229,
"end": 242,
"text": "(Graff, 2002)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work",
"sec_num": "2"
},
{
"text": "The experimental setup in the above work (except Kumar et al. who focus on fiction) all train on news articles from a particular time period, and test on articles in the same time period. This leads to possible overlap of training and testing data, particularly since news is often reprinted across agencies the same day. In fact, one of the systems in Kanhabua and Norvag (2008) simply searches for one training document that best matches a test document, and assigns its timestamp. We intentionally deviate from this experimental design and instead create temporally disjoint train/test sets (see Section 5).",
"cite_spans": [
{
"start": 353,
"end": 379,
"text": "Kanhabua and Norvag (2008)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work",
"sec_num": "2"
},
{
"text": "Finally, we extend this previous work by focusing on aspects of language not yet addressed for document dating: linguistic structure and absolute time expressions. The majority of articles in our dataset contain time expressions (e.g., the year 1998), yet these have not been incorporated into the models despite their obvious connection to the article's timestamp. This paper first describes how to include time expressions as traditional features, and then describes a more sophisticated temporal reasoning component that naturally fits into our classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work",
"sec_num": "2"
},
{
"text": "Labeling documents with timestamps is similar to topic classification, but instead of choosing from topics, we choose the most likely year (or other granularity) in which it was written. We thus begin with a bag-of-words approach, reproducing the generative model used by both de Jong (2005) and Kanhabua and Norvag (2008; . The subsequent sections then introduce our novel classifiers and temporal reasoners to compare against this model.",
"cite_spans": [
{
"start": 285,
"end": 291,
"text": "(2005)",
"ref_id": null
},
{
"start": 296,
"end": 322,
"text": "Kanhabua and Norvag (2008;",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Timestamp Classifiers",
"sec_num": "3"
},
{
"text": "The model of de Jong et al. (2005) uses the normalized log-likelihood ratio (NLLR) to score documents. It weights tokens by the ratio of their probability in a specific year to their probability over the entire corpus. The model thus requires an LM for each year and an LM for the entire corpus:",
"cite_spans": [
{
"start": 16,
"end": 34,
"text": "Jong et al. (2005)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Language Models",
"sec_num": "3.1"
},
{
"text": "N LLR(D, Y ) = w\u2208D P (w|D) * log( P (w|Y ) P (w|C) ) (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Language Models",
"sec_num": "3.1"
},
{
"text": "where D is the target document, Y is the time span (e.g., a year), and C is the distribution of words in the corpus across all years. A document is labeled with the year that satisfies argmax Y N LLR(D, Y ). They adapted this model from earlier work in the IR community (Kraaij, 2004) . We apply Dirichletsmoothing to the language models (as in de Jong et al.), although the exact choice of \u03b1 did not significantly alter the results, most likely due to the large size of our training corpus. Kanhabua and Norvag added an entropy factor to the summation, but we did not see an improvement in our experiments. The unigrams w are lowercased tokens. We will refer to this de Jong et al. model as the Unigram NLLR. Follow-up work by Kanhabua and Norvag (2008) applied two filtering techniques to the unigrams in the model: 1. Word Classes: include only nouns, verbs, and adjectives as labeled by a POS tagger 2. IDF Filter: include only the top-ranked terms by tf-idf score",
"cite_spans": [
{
"start": 270,
"end": 284,
"text": "(Kraaij, 2004)",
"ref_id": "BIBREF8"
},
{
"start": 728,
"end": 754,
"text": "Kanhabua and Norvag (2008)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Language Models",
"sec_num": "3.1"
},
{
"text": "We also tested with these filters, choosing a cutoff for the top-ranked terms that optimized performance on our development data. We also stemmed the words as Kanhabua and Norvag suggest. This model is the Filtered NLLR.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Language Models",
"sec_num": "3.1"
},
{
"text": "Kanhabua and Norvag also explored what they termed collocation features, but lacking details on how collocations were included (or learned), we could not reproduce this for comparison. However, we instead propose using NER labels to extract what may have counted as collocations in their data. Named entities are important to document dating due to the nature of people and places coming in and out of the news at precise moments in time. We compare the NER features against the Unigram and Filtered NLLR models in our final experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Language Models",
"sec_num": "3.1"
},
{
"text": "In addition to reproducing the models from previous work, we also trained a new discriminative version with the same features. We used a MaxEnt model and evaluated with the same filtering methods based on POS tags and tf-idf scores. The model performed best on the development data without any filtering or stemming. The final results (Section 6) only use the lowercased unigrams. Ultimately, this MaxEnt model vastly outperforms these NLLR models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Models",
"sec_num": "3.2"
},
{
"text": "The above language modeling and MaxEnt approaches are token-based classifiers that one could apply to any topic classification domain. Barring other knowledge, the learners solely rely on the observed frequencies of unigrams in order to decide which class is most likely. However, document dating is not just a simple topic classification application, but rather relates to temporal phenomena that is often explicitly described in the text itself. Language contains words and phrases that discuss the very time periods we aim to recover. These expressions should be better incorporated into the learner.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Models with Time Expressions",
"sec_num": "3.3"
},
{
"text": "Let the following snippet serve as a text example with an ambiguous creation time:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Motivation",
"sec_num": "3.3.1"
},
{
"text": "Then there's the fund-raiser at the American Museum of Natural History, which plans to welcome about 1,500 guests paying $1,000 to $5,000. Their tickets will entitle them to a preview of...the new Hayden Planetarium.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Motivation",
"sec_num": "3.3.1"
},
{
"text": "Without extremely detailed knowledge about the American Museum of Natural History, the events discussed here are difficult to place in time, let alone when the author reported it. However, time expressions are sometimes included, and the last sentence in the original text contains a helpful relative clause: This one clause is more valuable than the rest of the document, allowing us to infer that the document's timestamp is before February, 2000. An educated guess might surmise the article appeared in the year prior, 1999, which is the correct year. At the very least, this clause should eliminate all years after 2000 from consideration. Previous work on document dating does not integrate this information except to include the unigram '2000' in the model. This paper discusses two complementary ways to learn and reason about this information. The first is to simply add richer time-based features into the model. The second is to build separate learners that can assign probabilities to entire ranges of dates, such as all years following 2000 in the example above. We begin with the feature-based model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Motivation",
"sec_num": "3.3.1"
},
{
"text": "To our knowledge, the following time features have not been used in a document dating setting. We use the freely available Stanford Parser and NER system 1 to generate the syntactic interpretation for these features. We then train a MaxEnt classifier and compare against previous work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Features",
"sec_num": "3.3.2"
},
{
"text": "Typed Dependency: The most basic time feature is including governors of year mentions and the relation between them. This covers important contexts that determine the semantics of the time frame, like prepositions. For example, consider the following context for the mention 1997:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Features",
"sec_num": "3.3.2"
},
{
"text": "Torre, who watched the Kansas City Royals beat the Yankees, 13-6, on Friday for the first time since 1997.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Features",
"sec_num": "3.3.2"
},
{
"text": "The resulting feature is 'since pobj 1997'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Features",
"sec_num": "3.3.2"
},
{
"text": "Typed Dependency POS: Similar to Typed Dependency, this feature uses POS tags of the dependency relation's governor. The feature from the previous example is now 'PP pobj 1997'. This generalizes the features to capture time expressions with prepositions, as noun modifiers, or other constructs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Features",
"sec_num": "3.3.2"
},
{
"text": "Verb Tense: An important syntactic feature for temporal positioning is the tense of the verb that dominates the time expression. A past tense verb situates the phrase in 2003 differently than one in the future. We traverse the sentence's parse tree until a governor with a VB* tag is found, and determine its tense through hand constructed rules based on the structure of the parent VP. The verb tense feature takes a value of past, present, future, or undetermined.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Features",
"sec_num": "3.3.2"
},
{
"text": "Verb Path: The verb path feature is the dependency path from the nearest verb to the year expression. The following snippet will include the feature, 'expected prep in pobj 2002'. Named Entities: Although not directly related to time expressions, we also include n-grams of tokens that are labeled by an NER system using Person, Organization, or Location. People and places are often discussed during specific time periods, particularly in the news genre. Collecting named entity mentions will differentiate between an article discussing a bill and one discussing the US President, Bill Clinton. We extract NER features as sequences of uninterrupted tokens labeled with the same NER tag, ignoring unigrams (since unigrams are already included in the base model). Using the Verb Path example above, the bigram feature Hu Jintao is included.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Time Features",
"sec_num": "3.3.2"
},
{
"text": "This section departs from the above document classifiers and instead classifies individual emphyear mentions. The goal is to automatically learn temporal constraints on the document's timestamp.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning Time Constraints",
"sec_num": "4"
},
{
"text": "Instead of predicting a single year for a document, a temporal constraint predicts a range of years. Each time mention, such as 'not since 2009', is a constraint representing its relation to the document's timestamp. For example, the mentioned year '2009' must occur before the year of document creation. This section builds a classifier to label time mentions with their relations (e.g., before, after, or simultaneous with the document's timestamp), enabling these mentions to constrain the document classifiers described above. Figure 1 gives an example of time mentions and the desired labels we wish to learn.",
"cite_spans": [],
"ref_spans": [
{
"start": 531,
"end": 539,
"text": "Figure 1",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Learning Time Constraints",
"sec_num": "4"
},
{
"text": "To better motivate the need for constraints, let The clause until February 2000 in a present tense context may not definitively identify the document's timestamp (1999 is a good guess), but as discussed earlier, it should remove all future years beyond 2000 from consideration. We thus want to impose a constraint based on this phrase that says, loosely, 'this document was likely written before 2000'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning Time Constraints",
"sec_num": "4"
},
{
"text": "The document classifiers described in previous sections cannot capture such ordering information. Our new time features in Section 3.3.2 add richer time information (such as until pobj 2000 and open prep until pobj 2000), but they compete with many other features that can mislead the final classification. An independent constraint learner may push the document classifier in the right direction.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning Time Constraints",
"sec_num": "4"
},
{
"text": "We learn several types of constraints between each year mention and the document's timestamp. Year mentions are defined as tokens with exactly four digits, numerically between 1900 and 2100. Let T be the document timestamp's year, and M the year mention. We define three core relations:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constraint Types",
"sec_num": "4.1"
},
{
"text": "1. Before Timestamp: M < T 2. After Timestamp: M > T 3. Same as Timestamp: M == T",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constraint Types",
"sec_num": "4.1"
},
{
"text": "We also experiment with 7 fine-grained relations: Obviously the more fine-grained a relation, the better it can inform a classifier. We experiment with these two granularities to compare performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constraint Types",
"sec_num": "4.1"
},
{
"text": "The learning process is a typical training environment where year mentions are treated as labeled training examples. Labels for year mentions are automatically computed by comparing the actual timestamp of the training document (all documents in Gigaword have dates) with the integer value of the year token. For example, a document written in 1997 might contain the phrase, \"in the year 2000\". The year token (2000) is thus three+ years after the timestamp (1997). We use this relation for the year mention as a labeled training example.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constraint Types",
"sec_num": "4.1"
},
{
"text": "Ultimately, we want to use similar syntactic constructs in training so that \"in the year 2000\" and \"in the year 2003\" mutually inform each other. We thus compute the label for each time expression, and replace the integer year with the generic YEAR token to generalize mentions. The text for this example becomes \"in the year YEAR\" (labeled as three+ years after). We train a MaxEnt model on each year mention, to be described next. Table 2 gives the overall counts for the core relations in our training data. The vast majority of year mentions are references to the future (e.g. after the timestamp).",
"cite_spans": [],
"ref_spans": [
{
"start": 433,
"end": 440,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Constraint Types",
"sec_num": "4.1"
},
{
"text": "The features we use to classify year mentions are given in Table 1 . The same time features in the document classifier of Section 3.3.2 are included, as well as several others specific to this constraint task.",
"cite_spans": [],
"ref_spans": [
{
"start": 59,
"end": 66,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Constraint Learner",
"sec_num": "4.2"
},
{
"text": "We use a MaxEnt classifier trained on the individual year mentions. Documents often contain multiple (and different) year mentions; all are included in training and testing. This classifier labels mentions with relations, but in order to influence the document classifier, we need to map the relations to individual year predictions. Let T d be the set of mentions in document d. We represent a MaxEnt classifier by P Y (R|t) for a time mention t \u2208 T d and possible relations R. We map this distribution over relations to a distribution over years by defining P year (Y |d):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constraint Learner",
"sec_num": "4.2"
},
{
"text": "P year (y|d) = 1 Z(T d ) t\u2208T d P Y (rel(val(t) \u2212 y)|t) (2) rel(x) = \uf8f1 \uf8f2 \uf8f3 bef ore if x < 0 af ter if x > 0 simultaneous otherwise (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constraint Learner",
"sec_num": "4.2"
},
{
"text": "where val(t) is the integer year of the year mention and Z(T d ) is the partition function. The rel(val(t)\u2212 y) function simply determines if the year mention t (e.g., 2003) is before, after, or overlaps the year we are predicting for the document's unknown timestamp y. We use a similar function for the seven finegrained relations. Figure 3 visually illustrates how P year (y|d) is constructed from three year mentions.",
"cite_spans": [],
"ref_spans": [
{
"start": 333,
"end": 341,
"text": "Figure 3",
"ref_id": "FIGREF8"
}
],
"eq_spans": [],
"section": "Constraint Learner",
"sec_num": "4.2"
},
{
"text": "Finally, given the document classifiers of Section 3 and the constraint classifier just defined in Section 4, we create a joint model combining the two with the following linear interpolation:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Classifier",
"sec_num": "4.3"
},
{
"text": "P (y|d) = \u03bbP doc (y|d) + (1 \u2212 \u03bb)P year (y|d) (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Classifier",
"sec_num": "4.3"
},
{
"text": "where y is a year, and d is the document. \u03bb was set to 0.35 by maximizing accuracy on the dev set. See . This optimal \u03bb = .35 weights the constraint classifier higher than the document classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Classifier",
"sec_num": "4.3"
},
{
"text": "This paper uses the New York Times section of the Gigaword Corpus (Graff, 2002) for evaluation. Most previous work on document dating evaluates on the news genre, so we maintain the pattern for consistency. Unfortunately, we cannot compare to these previous experiments because of differing evaluation setups. Dalli and Wilks (2006) is most similar in their use of Gigaword, but they chose a random set of documents that cannot be reproduced. We instead define specific segments of the corpus for evaluation. The main goal for this experiment setup was to establish specific training, development, and test sets. One of the potential difficulties in testing with news articles is that the same story is often reprinted with very minimal (or no) changes. Over 10% of the documents in the New York Times section of the Gigaword Corpus are exact or approximate duplicates of another document in the corpus 2 . A training set for document dating must not include duplicates from the test set.",
"cite_spans": [
{
"start": 66,
"end": 79,
"text": "(Graff, 2002)",
"ref_id": null
},
{
"start": 310,
"end": 332,
"text": "Dalli and Wilks (2006)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "5"
},
{
"text": "We adopt the intuition behind the experimental setup used in other NLP domains, like parsing, where the entire test set is from a contiguous section of the corpus (as opposed to randomly selected examples across the corpus). As the parsing community trains on sections 2-21 of the Penn Treebank (Marcus et al., 1993) and tests on section 23, we create Gigaword sections by isolating specific months. ",
"cite_spans": [
{
"start": 295,
"end": 316,
"text": "(Marcus et al., 1993)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Datasets",
"sec_num": "5"
},
{
"text": "Jan-May and Sep-Dec Development July",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": null
},
{
"text": "June and August",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Testing",
"sec_num": null
},
{
"text": "In other words, the development set includes documents from July 1995 , July 1996 , July 1997 . We chose the dev/test sets to be in the middle of the year so that the training set includes documents on both temporal sides of the test articles. We include years 1995-2001 and 2004-2006, but skip 2002 and 2003 due to their abnormally small size compared to the other years. Finally, we experiment in a balanced data setting, training and testing on the same number of documents from each year. The test set includes 11,300 documents in each year (months June and August) for a total of 113,000 test documents.",
"cite_spans": [
{
"start": 60,
"end": 69,
"text": "July 1995",
"ref_id": null
},
{
"start": 70,
"end": 81,
"text": ", July 1996",
"ref_id": null
},
{
"start": 82,
"end": 93,
"text": ", July 1997",
"ref_id": null
},
{
"start": 261,
"end": 308,
"text": "1995-2001 and 2004-2006, but skip 2002 and 2003",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Testing",
"sec_num": null
},
{
"text": "The development set includes 7,300 from July of each year. Training includes approximately 75,000 documents in each year with some years slightly less than 75,000 due to their smaller size in the corpus. The total number of training documents for the 10 evaluated years is 725,468. The full list of documents is online at www.usna.edu/Users/cs/nchamber/data/timestamp.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Testing",
"sec_num": null
},
{
"text": "We experiment on the Gigaword corpus as described in Section 5. Documents are tokenized and parsed with the Stanford Parser. The year in the timestamp is retrieved from the document's Gigaword ID which contains the year and day the article was re-trieved. Year mentions are extracted from documents by matching all tokens with exactly four digits whose integer is in the range of 1900 and 2100.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "6"
},
{
"text": "The MaxEnt classifiers are also from the Stanford toolkit, and both the document and year mention classifiers use its default settings (quadratic prior). The \u03bb factor in the joint classifier is optimized on the development set as described in Section 4.3. We also found that dev results improved when training ignores the border months of Jan, Feb, and Dec. The features described in this paper were selected solely by studying performance on the development set. The final reported results come from running on the test set once at the end of this study. Table 3 shows the results on the Test set for all document classifiers. We measure accuracy to compare overall performance since the test set is a balanced set (each year has the same number of test documents). Unigram NLLR and Filtered NLLR are the language model implementations of previous work as described in Section 3.1. MaxEnt Unigram is our new discriminative model for this task. MaxEnt Time is the discriminative model with rich time features (but not NER) as described in Section 3.3.2 (Time+NER includes NER). Finally, the Joint model is the combined document and year mention classifiers as described in Section 4.3. .54 .56 .62 .44 .48 .48 .46 .57 F1 .55 .52 .57 .46 .48 .49 .48 .58 7%, and adding NER by another 6%. Total relative improvement in accuracy is thus almost 77% from the Time+NER model over Filtered NLLR. Further, the temporal constraint model increases this best classifier by another 3.9%. All improvements are statistically significant (p < 0.000001, McNemar's test, 2-tailed). Table 6 shows that performance increased most on the documents that contain at least one year mention (60% of the corpus).",
"cite_spans": [
{
"start": 1186,
"end": 1252,
"text": ".54 .56 .62 .44 .48 .48 .46 .57 F1 .55 .52 .57 .46 .48 .49 .48 .58",
"ref_id": null
}
],
"ref_spans": [
{
"start": 556,
"end": 563,
"text": "Table 3",
"ref_id": "TABREF3"
},
{
"start": 1565,
"end": 1572,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "6"
},
{
"text": "Finally, Table 5 shows the results of the temporal constraint classifiers on year mentions. Not surprisingly, the fine-grained performance is quite a bit lower than the core relations. The full Joint results in Table 3 use the three core relations, but the seven fine-grained relations give approximately the same results. Its lower accuracy is mitigated by the finer granularity (i.e., the majority class basline is lower).",
"cite_spans": [],
"ref_spans": [
{
"start": 9,
"end": 16,
"text": "Table 5",
"ref_id": "TABREF6"
},
{
"start": 211,
"end": 218,
"text": "Table 3",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "6"
},
{
"text": "The main contribution of this paper is the discriminative model (54% improvement) and a new set of Table 6 : Accuracy on all documents and documents with at least one year mention (about 60% of the corpus).",
"cite_spans": [],
"ref_spans": [
{
"start": 99,
"end": 106,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7"
},
{
"text": "features for document dating (14% improvement). Such a large performance boost makes clear that the log likelihood and entropy approaches from previous work are not as effective as discriminative models on a large training corpus. Further, token-based features do not capture the implicit references to time in language. Our richer syntax-based features only apply to year mentions, but this small textual phenomena leads to a surprising 13% relative improvement in accuracy. Table 6 shows that a significant chunk of this improvement comes from documents containing year mentions, as expected.",
"cite_spans": [],
"ref_spans": [
{
"start": 476,
"end": 483,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7"
},
{
"text": "The year constraint learner also improved performance. Although most of its features are in the document classifier, by learning constraints it captures a different picture of time that a traditional document classifier does not address. Combining this picture with the document classifier leads to another 3.9% relative improvement. Although we focused on year mentions here, there are several avenues for future study, including explorations of how other types of time expressions might inform the task. These constraints might also have applications to the ordering tasks of recent TempEval competitions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7"
},
{
"text": "Finally, we presented a new evaluation setup for this task. Previous work depended on having training documents in the same week and day of the test documents. We argued that this may not be an appropriate assumption in some domains, and particularly problematic for the news genre. Our proposed evaluation setup instead separates training and testing data across months. The results show that loglikelihood ratio scores do not work as well in this environment. We hope our explicit train/test environment encourages future comparison and progress on document dating.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7"
},
{
"text": "Approximate duplicate is defined as an article whose first two sentences exactly match the first two of another article. Only the second matched document is counted as a duplicate.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Many thanks to Stephen Guo and Dan Jurafsky for early ideas and studies on this topic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Jointly combining implicit constraints improves temporal ordering",
"authors": [
{
"first": "Nathanael",
"middle": [],
"last": "Chambers",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the Conference on Empirical Methods on Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nathanael Chambers and Dan Jurafsky. 2008. Jointly combining implicit constraints improves temporal or- dering. In Proceedings of the Conference on Em- pirical Methods on Natural Language Processing (EMNLP), Hawaii, USA.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Answering general time sensitive queries",
"authors": [
{
"first": "L",
"middle": [],
"last": "Dakka",
"suffix": ""
},
{
"first": "P",
"middle": [
"G"
],
"last": "Gravano",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ipeirotis",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 17th International ACM Conference on Information and Knowledge Management",
"volume": "",
"issue": "",
"pages": "1437--1438",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dakka, L. Gravano, and P. G. Ipeirotis. 2008. An- swering general time sensitive queries. In Proceedings of the 17th International ACM Conference on Informa- tion and Knowledge Management, pages 1437-1438.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Automatic dating of documents and temporal text classification",
"authors": [
{
"first": "Angelo",
"middle": [],
"last": "Dalli",
"suffix": ""
},
{
"first": "Yorick",
"middle": [],
"last": "Wilks",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Workshop on Annotating and Reasoning about Time and Events",
"volume": "",
"issue": "",
"pages": "17--22",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Angelo Dalli and Yorick Wilks. 2006. Automatic dat- ing of documents and temporal text classification. In Proceedings of the Workshop on Annotating and Rea- soning about Time and Events, pages 17-22.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Temporal language models for the disclosure of historical text",
"authors": [
{
"first": "Jong",
"middle": [],
"last": "Franciska De",
"suffix": ""
},
{
"first": "Henning",
"middle": [],
"last": "Rode",
"suffix": ""
},
{
"first": "Djoerd",
"middle": [],
"last": "Hiemstra",
"suffix": ""
}
],
"year": 2005,
"venue": "Humanities, computers and cultural heritage: Proceedings of the XVIth International Conference of the Association for History and Computing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franciska de Jong, Henning Rode, and Djoerd Hiemstra. 2005. Temporal language models for the disclosure of historical text. In Humanities, computers and cultural heritage: Proceedings of the XVIth International Con- ference of the Association for History and Computing (AHC 2005).",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Using temporal profiles of queries for precision prediction",
"authors": [
{
"first": "Fernando",
"middle": [],
"last": "Diaz",
"suffix": ""
},
{
"first": "Rosie",
"middle": [],
"last": "Jones",
"suffix": ""
}
],
"year": 2004,
"venue": "In Proceedings of the 27th Annual International ACM Special Interest Group on Information Retrieval Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fernando Diaz and Rosie Jones. 2004. Using temporal profiles of queries for precision prediction. In Pro- ceedings of the 27th Annual International ACM Spe- cial Interest Group on Information Retrieval Confer- ence.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Improving temporal language models for determining time of non-timestamped documents",
"authors": [
{
"first": "Nattiya",
"middle": [],
"last": "Kanhabua",
"suffix": ""
},
{
"first": "Kjetil",
"middle": [],
"last": "Norvag",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 12th European conference on Research and Advanced Technology for Digital Libraries",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nattiya Kanhabua and Kjetil Norvag. 2008. Improv- ing temporal language models for determining time of non-timestamped documents. In Proceedings of the 12th European conference on Research and Advanced Technology for Digital Libraries.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Using temporal language models for document dating. Lecture Notes in Computer Science: machine learning and knowledge discovery in databases",
"authors": [
{
"first": "Nattiya",
"middle": [],
"last": "Kanhabua",
"suffix": ""
},
{
"first": "Kjetil",
"middle": [],
"last": "Norvag",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nattiya Kanhabua and Kjetil Norvag. 2009. Using tem- poral language models for document dating. Lecture Notes in Computer Science: machine learning and knowledge discovery in databases, 5782.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Variations on language modeling for information retrieval",
"authors": [
{
"first": "W",
"middle": [],
"last": "Kraaij",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. Kraaij. 2004. Variations on language modeling for information retrieval. Ph.D. thesis, University of Twente.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Supervised language modeling for temporal resolution of texts",
"authors": [
{
"first": "Abhimanu",
"middle": [],
"last": "Kumar",
"suffix": ""
},
{
"first": "Matthew",
"middle": [],
"last": "Lease",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Baldridge",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of CIKM",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Abhimanu Kumar, Matthew Lease, and Jason Baldridge. 2011. Supervised language modeling for temporal res- olution of texts. In Proceedings of CIKM.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Time-based language models",
"authors": [
{
"first": "Xiaoyan",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "W. Bruce",
"middle": [],
"last": "Croft",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the twelfth international conference on Information and knowledge management",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Xiaoyan Li and W. Bruce Croft. 2003. Time-based lan- guage models. In Proceedings of the twelfth interna- tional conference on Information and knowledge man- agement.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Extracting temporal references to assign document event-time periods",
"authors": [
{
"first": "Dolores",
"middle": [
"M"
],
"last": "Llid\u00f3",
"suffix": ""
},
{
"first": "Rafael",
"middle": [],
"last": "Llavori",
"suffix": ""
},
{
"first": "Mari\u00e1",
"middle": [
"J"
],
"last": "Aramburu",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of the 12th International Conference on Database and Expert Systems Applications",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dolores M. Llid\u00f3, Rafael Llavori, and Mari\u00e1 J. Aram- buru. 2001. Extracting temporal references to assign document event-time periods. In Proceedings of the 12th International Conference on Database and Ex- pert Systems Applications.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Robust temporal processing of news",
"authors": [
{
"first": "Inderjeet",
"middle": [],
"last": "Mani",
"suffix": ""
},
{
"first": "George",
"middle": [],
"last": "Wilson",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 38th Annual Meeting on Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Inderjeet Mani and George Wilson. 2000. Robust tempo- ral processing of news. In Proceedings of the 38th An- nual Meeting on Association for Computational Lin- guistics.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Building a large annotated corpus of english: The penn treebank",
"authors": [
{
"first": "Mitchell",
"middle": [
"P"
],
"last": "Marcus",
"suffix": ""
},
{
"first": "Beatrice",
"middle": [],
"last": "Santorini",
"suffix": ""
},
{
"first": "Mary",
"middle": [
"Ann"
],
"last": "Marcinkiewicz",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated cor- pus of english: The penn treebank. Computational Linguistics, 19.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Marcia Lazo, Andrea Setzer, and Beth Sundheim",
"authors": [
{
"first": "James",
"middle": [],
"last": "Pustejovsky",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Hanks",
"suffix": ""
},
{
"first": "Roser",
"middle": [],
"last": "Sauri",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "See",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Day",
"suffix": ""
},
{
"first": "Lisa",
"middle": [],
"last": "Ferro",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Gaizauskas",
"suffix": ""
}
],
"year": 2003,
"venue": "The timebank corpus. Corpus Linguistics",
"volume": "",
"issue": "",
"pages": "647--656",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "James Pustejovsky, Patrick Hanks, Roser Sauri, Andrew See, David Day, Lisa Ferro, Robert Gaizauskas, Mar- cia Lazo, Andrea Setzer, and Beth Sundheim. 2003. The timebank corpus. Corpus Linguistics, pages 647- 656.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Semeval-2007 task 15: Tempeval temporal relation identification",
"authors": [
{
"first": "Marc",
"middle": [],
"last": "Verhagen",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Gaizauskas",
"suffix": ""
},
{
"first": "Frank",
"middle": [],
"last": "Schilder",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Hepple",
"suffix": ""
},
{
"first": "Graham",
"middle": [],
"last": "Katz",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Pustejovsky",
"suffix": ""
}
],
"year": 2007,
"venue": "Workshop on Semantic Evaluations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marc Verhagen, Robert Gaizauskas, Frank Schilder, Mark Hepple, Graham Katz, and James Pustejovsky. 2007. Semeval-2007 task 15: Tempeval temporal re- lation identification. In Workshop on Semantic Evalu- ations.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "The tempeval challenge: identifying temporal relations in text",
"authors": [
{
"first": "Marc",
"middle": [],
"last": "Verhagen",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Gaizauskas",
"suffix": ""
},
{
"first": "Frank",
"middle": [],
"last": "Schilder",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Hepple",
"suffix": ""
},
{
"first": "Jessica",
"middle": [],
"last": "Moszkowicz",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Pustejovsky",
"suffix": ""
}
],
"year": 2009,
"venue": "Special Issue: Computational Semantic Analysis of Language: SemEval-2007 and Beyond",
"volume": "43",
"issue": "",
"pages": "161--179",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marc Verhagen, Robert Gaizauskas, Frank Schilder, Mark Hepple, Jessica Moszkowicz, and James Puste- jovsky. 2009. The tempeval challenge: identifying temporal relations in text. Special Issue: Computa- tional Semantic Analysis of Language: SemEval-2007 and Beyond, 43(2):161-179.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Jointly identifying temporal relations with markov logic",
"authors": [
{
"first": "Katsumasa",
"middle": [],
"last": "Yoshikawa",
"suffix": ""
},
{
"first": "Sebastian",
"middle": [],
"last": "Riedel",
"suffix": ""
},
{
"first": "Masayuki",
"middle": [],
"last": "Asahara",
"suffix": ""
},
{
"first": "Yuji",
"middle": [],
"last": "Matsumoto",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Association for Computational Linguistics (ACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Katsumasa Yoshikawa, Sebastian Riedel, Masayuki Asa- hara, and Yuji Matsumoto. 2009. Jointly identify- ing temporal relations with markov logic. In Proceed- ings of the Association for Computational Linguistics (ACL).",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Search result re-ranking by feedback control adjustment for timesensitive query",
"authors": [
{
"first": "Ruiqiang",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Yi",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Zhaohui",
"middle": [],
"last": "Zheng",
"suffix": ""
},
{
"first": "Donald",
"middle": [],
"last": "Metzler",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Yun Nie",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ruiqiang Zhang, Yi Chang, Zhaohui Zheng, Donald Metzler, and Jian yun Nie. 2009. Search result re-ranking by feedback control adjustment for time- sensitive query. In Proceedings of the 2009 Annual Conference of the North American Chapter of the As- sociation for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"uris": null,
"text": "Their tickets will entitle them to a preview of...the new Hayden Planetarium, which does not officially open until February 2000.",
"type_str": "figure"
},
"FIGREF1": {
"num": null,
"uris": null,
"text": "1 http://nlp.stanford.edu/software Three year mentions and their relation to the document creation year. Relations can be correctly identified for training using known document timestamps.Supervising them is Vice President Hu Jintao, who appears to be Jiang's favored successor if he retires from leadership as expected in 2002.",
"type_str": "figure"
},
"FIGREF2": {
"num": null,
"uris": null,
"text": "Distribution over years for a single document as output by a MaxEnt classifier.",
"type_str": "figure"
},
"FIGREF3": {
"num": null,
"uris": null,
"text": "illustrate a typical distribution output by a document classifier for a training document. Two of the years appear likely (1999 and 2001), however, the document contains a time expression that seems to impose a strict constraint that should eliminate 2001 from consideration: Their tickets will entitle them to a preview of...the new Hayden Planetarium, which does not officially open until February 2000.",
"type_str": "figure"
},
"FIGREF4": {
"num": null,
"uris": null,
"text": "One year Before Timestamp: M == T \u2212 1 2. Two years Before Timestamp: M == T \u2212 2 3. Three+ years Before Timestamp: M < T \u2212 2 4. One year After Timestamp: M == T + 1 5. Two years After Timestamp: M == T + 2 6. Three+ years After Timestamp: M > T + 2 7. Same Year and Timestamp: M == T",
"type_str": "figure"
},
"FIGREF6": {
"num": null,
"uris": null,
"text": "Development set accuracy and \u03bb values.",
"type_str": "figure"
},
"FIGREF7": {
"num": null,
"uris": null,
"text": "Figure 4. This optimal \u03bb = .35 weights the constraint classifier higher than the document classifier.",
"type_str": "figure"
},
"FIGREF8": {
"num": null,
"uris": null,
"text": "Three year mentions in a document and the distributions output by the learner. The document is from 2002. The dots indicate the before, same, and after relation probabilities. The combination of three constraints results in a final distribution that gives the years 2001 and 2002 the highest probability. This distribution can help a document classifier make a more informed final decision.",
"type_str": "figure"
},
"TABREF1": {
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "",
"html": null
},
"TABREF2": {
"content": "<table><tr><td>Model</td><td>Overall Accuracy</td></tr><tr><td>Random Guess</td><td>10.0%</td></tr><tr><td>Unigram NLLR</td><td>24.1%</td></tr><tr><td>Filtered NLLR</td><td>29.1%</td></tr><tr><td>MaxEnt Unigram</td><td>45.1%</td></tr><tr><td>MaxEnt Time</td><td>48.3%</td></tr><tr><td>MaxEnt Time+NER</td><td>51.4%</td></tr><tr><td>Joint</td><td>53.4%</td></tr></table>",
"type_str": "table",
"num": null,
"text": "shows the F1 scores of the Joint model by year.Our new MaxEnt model outperforms previous work by 55% relative accuracy. Incorporating time features further improves the relative accuracy by",
"html": null
},
"TABREF3": {
"content": "<table><tr><td/><td>95 96 97 98 99 00 01 02</td></tr><tr><td>P</td><td>.57 .49 .52 .48 .47 .51 .51 .59</td></tr><tr><td>R</td><td/></tr></table>",
"type_str": "table",
"num": null,
"text": "Performance as measured by accuracy. The predicted year must exactly match the actual year.",
"html": null
},
"TABREF4": {
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "Yearly results for the Joint model. 2005/06 are omitted due to space, with F1 .56 and .63, respectively.",
"html": null
},
"TABREF6": {
"content": "<table><tr><td/><td>All</td><td>With Year Mentions</td></tr><tr><td>MaxEnt Unigram</td><td>45.1%</td><td>46.1%</td></tr><tr><td colspan=\"2\">MaxEnt Time+NER 51.4%</td><td>54.3%</td></tr><tr><td>Joint</td><td>53.4%</td><td>57.7%</td></tr></table>",
"type_str": "table",
"num": null,
"text": "Precision, recall, and F1 for the core relations. Accuracy for both core and fine-grained.",
"html": null
}
}
}
} |