File size: 101,136 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 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 |
{
"paper_id": "I13-1004",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:15:23.799079Z"
},
"title": "Learning a Replacement Model for Query Segmentation With Consistency in Search Logs",
"authors": [
{
"first": "Wei",
"middle": [],
"last": "Zhang",
"suffix": "",
"affiliation": {},
"email": "zhangw3@i2r.a-star.edu.sg"
},
{
"first": "Yunbo",
"middle": [],
"last": "Cao",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Microsoft Research Asia \u2020 National University of Singapore",
"location": {}
},
"email": "yunbo.cao@microsoft.com"
},
{
"first": "Chin-Yew",
"middle": [],
"last": "Lin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Microsoft Research Asia \u2020 National University of Singapore",
"location": {}
},
"email": ""
},
{
"first": "Jian",
"middle": [],
"last": "Su",
"suffix": "",
"affiliation": {},
"email": "sujian@i2r.a-star.edu.sg"
},
{
"first": "Chew-Lim",
"middle": [],
"last": "Tan",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Query segmentation is to split a query into a sequence of non-overlapping segments that completely cover all tokens in the query. The majority of methods are unsupervised, however, they are usually not as accurate as supervised methods due to the lack of guidance from labeled data. In this paper, we propose a new paradigm of learning a replacement model with consistency (LRMC), to enable unsupervised training with guidance from search log data. In LRMC, we first assume the existence of a base segmenter (an implementation of any existing approach). Then, we utilize a key observation that queries with a similar intent tend to have consistent segmentations, to automatically collect a set of labeled data from the outputs of the base segmenter by leveraging search log data. Finally, we employ the auto-collected data to train a replacement model for selecting the correct segmentation of a new query from the outputs of the base segmenter. The results show LRMC can improve state-of-the-art methods by an F-Score of around 7%. * Wei Zhang did this work when he was an intern at Microsoft Research Asia.",
"pdf_parse": {
"paper_id": "I13-1004",
"_pdf_hash": "",
"abstract": [
{
"text": "Query segmentation is to split a query into a sequence of non-overlapping segments that completely cover all tokens in the query. The majority of methods are unsupervised, however, they are usually not as accurate as supervised methods due to the lack of guidance from labeled data. In this paper, we propose a new paradigm of learning a replacement model with consistency (LRMC), to enable unsupervised training with guidance from search log data. In LRMC, we first assume the existence of a base segmenter (an implementation of any existing approach). Then, we utilize a key observation that queries with a similar intent tend to have consistent segmentations, to automatically collect a set of labeled data from the outputs of the base segmenter by leveraging search log data. Finally, we employ the auto-collected data to train a replacement model for selecting the correct segmentation of a new query from the outputs of the base segmenter. The results show LRMC can improve state-of-the-art methods by an F-Score of around 7%. * Wei Zhang did this work when he was an intern at Microsoft Research Asia.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Nowadays keyword queries have been adopted as the de-facto query interface by most search engines. Query tokens are not independent or unordered symbols but rather ordered and structured words and phrases with syntactic relationships. Understanding the structure of a query is crucial for achieving better search performance. Such an understanding will also ease other search related applications such as query suggestion and rewriting, where one is able to work on semantic concepts instead of individual tokens. Query segmentation (QS), a process of splitting a query into a sequence of non-overlapping segments that completely cover all tokens, aims to address these challenges. It requires that every segment rendered is a phrase or a semantic unit. For example, given a query \"download adobe writer\", four different ways of segmentation are possible. The challenge is to determine which one is correct.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The majority of QS methods are unsupervised, however, they are not as accurate as supervised methods due to lack of guidance from labeled data. On the other hand, supervised models suffer from the problems: (1) new phrases/words are introduced on the web daily, which quickly invalidate static supervised models trained on a certain manually labeled set; (2) it is not feasible to develop a set of labeled data covering all domains on the web. In this paper, we propose a paradigm of learning a replacement model with consistency (LRMC), to enable unsupervised training and it improves various unsupervised QS systems.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "LRMC first assumes the existence of a base segmentation system (hereafter referred to as 'base segmenter') which can output top-n segmentations for any query. Then it tries to learn a replacement model capable of selecting the correct segmentation of a new query (if one exists) from the output of the base segmenter. Our study on three state-of-the-art systems (Section 5.2) shows that for more than 35% of queries the correct segmentations are not ranked as top-1 but included in the top-5 results of the base segmenter, which implies the potential of LRMC. The keys to our proposal include: (a) how to automatically acquire labeled data (i.e., for a query in the labeled data, what its correct segmentation is) and then (b) how to use the labeled data to learn the replacement model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our method for the automatic acquisition of the labeled data is motivated by the observation: Queries with a similar intent tend to have consistent segmentation results. In this paper, we say that a set of queries have similar intents if and only if they lead to the same set of web documents (i.e., clicks). For example, when issuing to a web search engine any of the three queries in Table 1 , we search for the same set of web pages which can provide 'free download of Adobe writer'. We denote such a set of queries as 'query intent set'. For the queries in the same query intent set, naturally we wish to explain them in the same way and thus require that their segmentations be consistent with each other. We say that q 1 and q 2 are inconsistent in segmentation if there exist more than one common subsequence of tokens having different segment boundaries. In Table 1 , we also include the top-2 segmentation results that can possibly be generated by any base segmenter. If we check only the 'rank-1' results, we observe that the segmentation 'download adobe | writer' disagrees with the other two. This means that we interpret the same sequence of tokens differently for different queries with a same intent, which is not what we expect to have. Instead, we expect to have the bolded segmentations in which none of the individual segments for one query disagrees with the segments for another query. In this paper, we propose two methods for selecting such correct segmentations from top-n segmentation results that are about the same query intent sets. With these methods, we can automatically build up a training data set, which allows us to train a reliable model.",
"cite_spans": [],
"ref_spans": [
{
"start": 386,
"end": 393,
"text": "Table 1",
"ref_id": null
},
{
"start": 866,
"end": 873,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The replacement model concerns about whether or not a 'rank-1' segmentation S a generated by a base segmenter should be replaced by a 'rank-k' (k > 1) segmentation S b . The decision of the replacement can be made by collectively considering one or multiple local transformations in the form of '",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "w i w i+1 \u2192 w i |w i+1 ' or 'w i |w i+1 \u2192 w i w i+1 '. 'w i w i+1 \u2192 w i |w i+1 '",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "means that S a does not include a segment boundary between tokens w i and w i+1 and S b does; Similarly, 'w i |w i+1 \u2192 w i w i+1 ' means the reverse. For example, for the first query in Table 1 , we can have the local transformations 'download adobe \u2192 download | adobe' and 'adobe | writer \u2192 adobe writer'. The proposed model estimates the score of every local transformation using a binary classifier and then aggregates the individual scores to reach its final decision.",
"cite_spans": [],
"ref_spans": [
{
"start": 186,
"end": 193,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We conduct extensive experiments using two public data sets. The results show that (a) our method for automatically constructing a set of labeled data with a base segmenter and a set of query intent sets as inputs is effective, capable of discovering correct segmentations missed by the evaluated base segmenters for more than 20% of queries (See M2 in terms of Acc qry in Table 3) ; and (b) our replacement model benefits existing QS approaches and boosts their performance significantly (e.g. the improvement of > 7% F-Score on the data WQ10-Majority in Table 4) .",
"cite_spans": [],
"ref_spans": [
{
"start": 373,
"end": 381,
"text": "Table 3)",
"ref_id": "TABREF4"
},
{
"start": 556,
"end": 564,
"text": "Table 4)",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We summarize our contributions as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(1) on the basis of the observation that queries with a similar intent tend to have consistent segmentations, we propose a method for automatically collecting from search log data a set of labeled data for QS. The method first groups queries in search log data into what we call a 'query intent set' and then select correct segmentations by examining the consistency among segmentations for the queries in the same 'query intent sets'. 2With the automatically-collected data, we develop a 'replacement model' for the purpose of checking whether or not a 'rank-1' segmentation generated by a base segmenter should be replaced by a 'rank-k' (k > 1) segmentation. (3) We conduct extensive experiments with two publicly available data sets and show that our proposal can effectively boost the performance of state-of-the-art systems (Hagen et al., 2010; Hagen et al., 2011) . Bergsma and Wang (2007) considered the decision to segment or not between each pair of adjacent words as a binary classification problem. Guo et al. (2008) , Yu and Shi (2009) , and Kiseleva et al. (2010) used methods based on CRF. As the cost of obtaining labeled data is high, they are usually not feasible to develop a set of labeled data covering all the domains on the web and then train a scalable QS model for web search.",
"cite_spans": [
{
"start": 829,
"end": 849,
"text": "(Hagen et al., 2010;",
"ref_id": "BIBREF7"
},
{
"start": 850,
"end": 869,
"text": "Hagen et al., 2011)",
"ref_id": "BIBREF8"
},
{
"start": 872,
"end": 895,
"text": "Bergsma and Wang (2007)",
"ref_id": "BIBREF2"
},
{
"start": 1010,
"end": 1027,
"text": "Guo et al. (2008)",
"ref_id": "BIBREF6"
},
{
"start": 1030,
"end": 1047,
"text": "Yu and Shi (2009)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The work for web-scale QS are usually unsupervised and utilized various statistics such as mutual information (MI) and frequency count collected from various sources such as web data, query logs, and etc (Risvik et al., 2003; Jones et al., 2006; Huang et al., 2010; Zhang et al., 2009) . also used the language model estimated from click-through documents to backoff the generating process of QS. Tan and Peng (2008) used n-gram frequencies from a large web corpus as well as Wikipedia. Hagen et al. (2010) showed that the raw n-gram could be exploited with an appropriate normalization scheme and achieved surprisingly good accuracy. Later, they enriched the work by including the use of Wikipedia (Hagen et al., 2011) . In our evaluation, we compare our proposal with the last two work which represent state-of-the-art.",
"cite_spans": [
{
"start": 204,
"end": 225,
"text": "(Risvik et al., 2003;",
"ref_id": "BIBREF15"
},
{
"start": 226,
"end": 245,
"text": "Jones et al., 2006;",
"ref_id": "BIBREF11"
},
{
"start": 246,
"end": 265,
"text": "Huang et al., 2010;",
"ref_id": "BIBREF10"
},
{
"start": 266,
"end": 285,
"text": "Zhang et al., 2009)",
"ref_id": "BIBREF21"
},
{
"start": 397,
"end": 416,
"text": "Tan and Peng (2008)",
"ref_id": "BIBREF18"
},
{
"start": 487,
"end": 506,
"text": "Hagen et al. (2010)",
"ref_id": "BIBREF7"
},
{
"start": 699,
"end": 719,
"text": "(Hagen et al., 2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Our proposal is orthogonal to all the above approaches. LRMC assumes the existence of a base segmenter (an implementation of any above approaches) and it focuses on how to leverage search log data to learn a replacement model for improving the output of base segmenters.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "QS. Let q = [w 1 , w 2 , \u2022 \u2022 \u2022 , w n ] denote a query consisting of n keywords. A segment s = [w i , \u2022 \u2022 \u2022 , w j ](1 \u2264 i \u2264 j \u2264 n) is a subsequence of the query. A segmentation S = [s 1 |s 2 | \u2022 \u2022 \u2022 |s K ]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Settings",
"sec_num": "3"
},
{
"text": "for query q is then defined as a sequence of nonoverlapping segments. '|' denotes a segmentation boundary. If we assume there is no order dependency of s, we can then treat S as a set {s k } K k=1 . Query Intent. There exist many definitions on query intent. In this paper we introduce an operational definition on query intent.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Settings",
"sec_num": "3"
},
{
"text": "The query intent(s) of a query q is defined as the set of URLs (U rls(q)) which are clicked for q by users of a web search engine.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definition 1",
"sec_num": null
},
{
"text": "Because most queries are ambiguous or multifaceted (Clarke et al., 2009) , we manage to restrict the number of intents into one or a few by grouping more queries together, which leads to the definition of 'query intent set'.",
"cite_spans": [
{
"start": 51,
"end": 72,
"text": "(Clarke et al., 2009)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Definition 1",
"sec_num": null
},
{
"text": "Definition 2 A query intent set Q IN T is a set of queries satisfying the following conditions:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definition 1",
"sec_num": null
},
{
"text": "a) q\u2208Q INT U rls(q) = \u2205; b) |Q IN T | > c.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definition 1",
"sec_num": null
},
{
"text": "where |Q IN T | denotes the number of elements in Q IN T , and c is a parameter to control how specific a query intent is; a larger value for c usually means that the query intent is more specific and thus less ambiguous.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definition 1",
"sec_num": null
},
{
"text": "Query intent sets used in our experiments will be detailed in Section 5.1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Definition 1",
"sec_num": null
},
{
"text": "First, the paradigm LRMC assumes the existence of a base segmenter that is able to output top-n segmentations for any query. Then it tries to learn a replacement model capable of replacing the rank-1 segmentation generated by this base segmenter with one rank-k (k > 1) segmentation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "LRMC can be illustrated by the following flowchart. First, a query q is fed into a base segmenter. As a result, a set of segmentations",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "{S i } n i=1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "regarding q are generated. Subscript i denotes the rank of the corresponding segmentation. Next,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "{S i } n i=1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "are fed into a replacement model. The replacement model tries every possible replacement S i (i > 1) for the rank-1 segmentation S 1 (as indicated by the curved arrows). The trial ends with two possible results: (a) None of the replacements is valid (S 1 cannot be replaced); and (b) one segmentation S * i (i * > 1) is the most likely replacement and thus chosen as the final segmentation for q (e.g., the replacement of the solid curve).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "q base segmenter \u2212 \u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212 \u2192 S1 S2 S3 \u2022 \u2022 \u2022 Sn replacement model \u2212 \u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212\u2212 \u2192 S1 S2 S3 \u2022 \u2022 \u2022 Sn",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "LRMC is motivated by the following observation: for most cases, the correct segmentation for a query is included in its top-n segmentation results already. Usually, there are not that many likely segmentations for a query and thus correct segmentations cannot be ranked too low by a base segmenter. For example, for any base segmenter in our experiment, more than 93% of queries can have a segmentation that is agreed upon by at least one of the annotators in its top-5 results. Given this observation, what we have to do is not to generate or propose a new segmentation, but to tell which segmentation is correct in the top-n results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "Next, we detail how the replacement model is learned. Specifically, we first introduce how we automatically extract from search log data a set of labeled data with 'consistency' as a guidance and then explain how a 'replacement model' can be learned from this data set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Proposed Paradigm",
"sec_num": "4.1"
},
{
"text": "Assume that we have a query intent set",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Supervision",
"sec_num": "4.2"
},
{
"text": "Q IN T = {q i } m i=1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Supervision",
"sec_num": "4.2"
},
{
"text": ". With a base segmenter, we generate the top-n segmentation results {S ij } (1 \u2264 j \u2264 n) for each query q i , which forms the following matrix:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Supervision",
"sec_num": "4.2"
},
{
"text": "S Q INT = S11 S12 S13 \u2022 \u2022 \u2022 S1n S21 S22 S23 \u2022 \u2022 \u2022 S2n \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 Sm1 Sm2 Sm3 \u2022 \u2022 \u2022 Smn \u2190 q1 \u2190 q2 \u2022 \u2022 \u2022 \u2190 qm",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Supervision",
"sec_num": "4.2"
},
{
"text": "What we manage to achieve is to collect a set of 'labeled' data from {S ij }. In the 'labeled' data, each query q i has only one segmentation S ij * (S ij * \u2208 {S ij } n j=1 ), which we consider 'correct'. We make use of two types of strategies to choose the 'correct' segmentations from S Q INT (e.g., those underlined ones in",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Supervision",
"sec_num": "4.2"
},
{
"text": "S Q INT , namely S 12 , S 22 , \u2022 \u2022 \u2022 , S m1 ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Supervision",
"sec_num": "4.2"
},
{
"text": "Before explaining the two strategies, let us first introduce how we measure the consistency between two segmentations. The consistency cst(S, S \u2032 ) between segmentations S and S \u2032 is defined as the number of segments they share, i.e.,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Supervision",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u2032 ) = |S \u2229 S \u2032 |",
"eq_num": "(1)"
}
],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "The first strategy (M1) that we use as 'supervision' for the acquisition of the labeled data is as follows: The correct segmentations for the m queries in the same query intent set should be very consistent with or similar to each other although the segmentations cannot be exactly the same. Thus, the correct segmentations can be chosen with the objective function:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "(j * 1 , \u2022 \u2022 \u2022 , j * m ) = arg max 1\u2264j 1 ,\u2022\u2022\u2022 ,jm\u2264n 1\u2264i<i \u2032 \u2264m cst(Sij i , S i \u2032 j i \u2032 )",
"eq_num": "(2)"
}
],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "where j * i denote the index (or rank) of the correct segmentation for query q i .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "The other strategy (M2) is on the basis of the observation: Although at most only one top-n segmentation can be correct for a query, most segmentations are not totally incorrect, i.e., they include some correct segments while having some incorrect segments as well. Thus, those incorrect segmentations also provide some clues about what can be correct. In addition, as the choices for 'incorrect segment' are usually more than those for correct segment, it is relatively hard for incorrect segments to converge to a few. As a result, a correct segment should be more popular than any one single incorrect segment. Given this discussion, we can have the second objective function:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "(j * 1 , \u2022 \u2022 \u2022 , j * m ) = arg max 1\u2264j 1 ,\u2022\u2022\u2022 ,jm\u2264n ( 1\u2264i,i \u2032 \u2264m 1\u2264j \u2032 \u2264n cst(Sij i , S i \u2032 j \u2032 ) \u2212 1\u2264i\u2264m cst(Sij i , Sij i ))",
"eq_num": "(3)"
}
],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "Note that cst(S ij i , S ij i ) = |S ij i |. Given one selected segmentation S ij i , the objective is to sum up the consistencies between itself and any of the rest in matrix S Q INT . Thus, by this objective, we choose the segmentations whose segments are agreed with by most top-n segmentations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "Both strategies assume that correct segments are more popular than incorrect segments in the topn output of one reasonably-performing base segmenter. Both strategies will fail if the assumption is not true. Our experiments in Section 5.2, in which both strategies are able to find more correct segmentations than the base segmenters, can be seen as a support for the assumption.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "cst(S, S",
"sec_num": null
},
{
"text": "The replacement model is to tell whether or not a segmentation ranked as top-1 by a base segmenter should be replaced by another segmentation with a rank of j (1 < j \u2264 n). For example, we have a query q whose top-n segmentations are {S qj } n j=1 . Then, the input of the replacement model will be a possible replacement S q1 \u2192 S qj (j > 1) and the output will be a label '1' or '0'. Label '1' means S q1 should be replaced by S qj and '0' means 'not'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "With that in mind, we can then make use of 'consistency' to create a labeled data set. For example, if query q belongs to a query intent set Q IN T and its correct segmentation chosen by the objective (2) or (3) is S qj * , we can generate the labeled instance(s) as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "D q = {(S q1 \u2192 S qj , 0)} j =1 if j * = 1 {(S q1 \u2192 S qj * , 1)} otherwise",
"eq_num": "(4)"
}
],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "By combining all such data sets together, we then have the final labeled data set D = \u222a q D q . Note that query q can come from multiple query intent sets (not just one single set).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "Next, we explain how to use the above training data to learn a replacement model. The decision of whether or not to do the replacement of S q1 \u2192 S qj can be made by collectively considering one or multiple local transformations in the form of '",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "w i w i+1 \u2192 w i |w i+1 ' or 'w i |w i+1 \u2192 w i w i+1 '. 'w i w i+1 \u2192 w i |w i+1 '",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "means that S q1 does not include a segment boundary between tokens w i and w i+1 and S qj does; 'w i |w i+1 \u2192 w i w i+1 ' means the reverse.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "Let T (S q1 \u2192 S qj ) denote the set of all possible local transformations from S q1 to S qj and x denote one element from the set (i.e., one local transformation). If we know the likelihood f (x) of every individual transformation x being valid, the score of replacing S q1 by S qj can then be estimated as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "x\u2208T (S q1 \u2192S qj ) f (x).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "The likelihood of a local transformation x being valid can be estimated with a binary classifier. We employ SVM as the classifier. Given an instance x, SVM assigns a score to it based on f (x) = w T x + b, where w denotes a weight vector and b denotes an intercept. Given a replacement (S q1 \u2192 S qj , y) where y \u2208 {0, 1}, a set of labeled data for the binary classifier is prepared as: {x, y} x\u2208T (S q1 \u2192S qj ) . By considering all the replacements in D, we will have a final training data set",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "{(x i , y i )} N i=1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "for SVM. On the basis of that, we can do the replacement as follows: If for certain j (j > 1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "x\u2208T (S q1 \u2192S qj ) f (x) > 0, we will use the segmentation with arg max 1<j\u2264n",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "x\u2208T (S q1 \u2192S qj ) f (x) as its index to replace the top-1 segmentation; Otherwise, we will keep using the top-1 segmentation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Replacement Model",
"sec_num": "4.3"
},
{
"text": "In this section, we describe the features for representing a local transformation x, which is in the form of either '",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning Features",
"sec_num": "4.4"
},
{
"text": "w i 0 w i 0 +1 \u2192 w i 0 |w i 0 +1 ' or 'w i 0 |w i 0 +1 \u2192 w i 0 w i 0 +1 '.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Learning Features",
"sec_num": "4.4"
},
{
"text": "We utilize four categories of features which are possible indicators of a transformation, representing a variety of information such as lexical, syntactic, semantic and etc. Contextual Features: Lexical. The left and right tokens around the decision position, w i 0 and w i 0 +1 , are a good signal of the transformation. In the example of \"google desktop | download\", the token 'download' is separated from its left neighbor. Such common query tokens in the training data with the property of usually being separated from or being connected to its left/right neighbor can help predict new transformations (e.g. \"adobe writer download \u2192 adobe writer | download\"). On the basis of this observation, we adopt the left token w i 0 and the right token w i 0 +1 as the features for representing a transformation x. Furthermore, sometimes one word alone can not perfectly characterize a transformation. For example, to reject the transformation \"diet plan \u2192 diet | plan\", we have to use the token bigram <diet plan>. Thus, we include all the token bigrams in the form of < w i 0 w i 0 +1 > as features as well. As we all know, lexical features usually suffer from a data sparse-ness issue when used in various tasks (Bagga and Baldwin, 1998; Sriram et al., 2010) . Fortunately, the web-scale training data we collect from query intent sets (Section 4.2) enables us to have a good coverage of lexical features.",
"cite_spans": [
{
"start": 1210,
"end": 1235,
"text": "(Bagga and Baldwin, 1998;",
"ref_id": "BIBREF0"
},
{
"start": 1236,
"end": 1256,
"text": "Sriram et al., 2010)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Learning Features",
"sec_num": "4.4"
},
{
"text": "Contextual Features: POS Tag. Bergsma et al. (2007) show that part-of-speech (POS) tags are useful in their segmentation classification. We also exploit the POS tag pair of w i 0 and w i 0 +1 as features. For example, intuitively, \"NN NN \u2192 NN | NN\" is more likely to occur than \"JJ NN \u2192 JJ | NN\". The POS tags that we consider include all types of POS tags. Note that this is different from Bergsma et al. (2007) . As their segmentation model only takes care of noun phrase queries, their POS tags are restricted to determiners, adjectives, and nouns. The POS tagger by (Roth and Zelenko, 1998) is used in this paper.",
"cite_spans": [
{
"start": 30,
"end": 51,
"text": "Bergsma et al. (2007)",
"ref_id": "BIBREF2"
},
{
"start": 391,
"end": 412,
"text": "Bergsma et al. (2007)",
"ref_id": "BIBREF2"
},
{
"start": 570,
"end": 594,
"text": "(Roth and Zelenko, 1998)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Learning Features",
"sec_num": "4.4"
},
{
"text": ". Following previous work (Section 2), we also adopt MI between w i 0 and w i 0 +1 as our feature. The work (Bergsma and Wang, 2007) also considered the case of a noun phrase with multiple modifiers (e.g. \"female bus driver\"). To make the segmentation decision between 'female' and 'bus', M I('female', 'driver') is more suitable to represent the information of not separating them than M I('female', 'bus'). Thus, we also incorporate M I(w i 0 \u22121 , w i 0 +1 ) and M I(w i 0 , w i 0 +2 ) into our feature set.",
"cite_spans": [
{
"start": 108,
"end": 132,
"text": "(Bergsma and Wang, 2007)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Mutual Information (MI)",
"sec_num": null
},
{
"text": "Most previous work on QS only can use wordbased MI as introduced above. However, in some cases, the MI between tokens can not provide sufficient information for a segmentation decision. For instance, assume that we have the following two queries with their correct segmentations: (1) \"download | call of duty | free\"; (2) \"duty free | shops | sfo\". Only using the token-based mutual information M I('duty', 'free') can not discriminate the two queries from each other and thus can not give different segmentations for 'duty free' in the two queries. In our work, as the query has been segmented by a base segmenter, we propose to also use the segment-based MI. In the 'duty free' example, M I('call of duty', 'free') will be incorporated for the transformation decision related to \"download | call of duty free \u2192 download | call of duty | free\", where the token-based M I('duty', 'free') does not work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mutual Information (MI)",
"sec_num": null
},
{
"text": "We define the semantic features on the basis of segments. For a transformation 'w i 0 w i 0 +1 \u2192 w i 0 |w i 0 +1 ' , let us denote the segment including w i 0 w i 0 +1 (in the first segmentation) as s 1 and denote the segments including w i 0 and w i 0 +1 separately (in the second segmentation) as s 2 and s 3 , respectively. To obtain semantic labels for the above three segments, we make use of a web-scale knowledge base of entities, namely Freebase (Bollacker et al., 2008) . First, we map the three segments to the Freebase entities by string matching, and then use the names or aliases of the associated categories of the mapped entities as their semantic labels. Finally, the semantic labels for s 1 , s 2 and s 3 are used as features. Due to the ambiguity, a phrase in a query may be mistakenly linked to a certain entity in the knowledge base. Thus, the semantic features include some noises. However, even with noises such features can still contribute to QS. To illustrate how the semantic features work, consider the query with the assigned semantic label as follows, As pointed out by (Tan et al., 2008) , QS approaches which are only based on statistical information (e.g. M I and frequencies of n-grams) collected from the Web, cannot guarantee that the resulting segments are meaningful ones. For the query 'history of the search engine', a possible segmentation is 'history of the | search engine', as both 'history of the' and 'search engine' occur on the web frequently. In contrast, semantic information can distinguish 'search engine' from 'history of the', as 'history of the' is labeled as NULL and 'search engine' is labeled as 'computer software genre'. Moreover, the learning algorithm can also learn some implicit relations between transformations and semantic labels, e.g. some particular combination of labels for s 1 , s 2 and s 3 may often trigger or prevent a transformation. Table 2 shows the values of these features. The rank feature is designed to distinguish among the different segmentation rankings of a base segmenter. For example, this feature can capture the intuition that for a good base segmenter, top ranked segmentations should have more of a chance to be selected. The direction feature is used to distinguish the two kinds of transformations: 'w i w i+1 \u2192 w i |w i+1 ' and 'w i |w i+1 \u2192 w i w i+1 '. The position feature considers decision positions, as transformations in different positions may have different chances.",
"cite_spans": [
{
"start": 454,
"end": 478,
"text": "(Bollacker et al., 2008)",
"ref_id": "BIBREF3"
},
{
"start": 1099,
"end": 1117,
"text": "(Tan et al., 2008)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [
{
"start": 1909,
"end": 1916,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Semantic Features.",
"sec_num": null
},
{
"text": "j, the rank of the segmentation to which we transform the top-1 segmentation. Direction 1, if \"wiwi+1 \u2192 wi|wi+1\"; 0, reverse. Position lef t",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Rank",
"sec_num": null
},
{
"text": "Number of words from the decision position to the beginning/end of query. Break Level. The decision of break is whether or not to insert a segment boundary between two tokens in the query. The break-level accuracy (Acc brk ) is defined as the proportion of the correctly-made decisions out of all such decisions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "Segment Level. Let Q eval denote the set of queries. S sys q is the segmentation generated by a system and S eval q is given by a human. Then,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P sg = q\u2208Q eval |S sys q \u2229 S eval q | |S sys q | R sg = q\u2208Q eval |S sys q \u2229 S eval q | |S eval q | F sg = 2 \u2022 P sg \u2022 R sg P seg + R sg",
"eq_num": "(6)"
}
],
"section": "Position right",
"sec_num": null
},
{
"text": "We use two data sets as introduced in (Bergsma and Wang, 2007) and (Hagen et al., 2010) , denoted as 'Bergsma-Wang-07' (BW07) and 'Webis-QSeC-10' (WQ10). BW07 includes 500 test queries which all were noun phrase queries. Each query was segmented manually by three annotators (denoted as annotator A, B, and C) respectively. For 44% of the queries, all three annotators agree on the segmentations. Such an agreement between annotators cannot be considered as 'strong', which to some extent implies that human annotations may not be so reliable when used for training a segmentation model capable of consistently working over different queries. WQ10 includes 4,850 queries. Each query can be any type of query, not necessarily a noun phrase query. Each query was annotated by ten annotators.",
"cite_spans": [
{
"start": 38,
"end": 62,
"text": "(Bergsma and Wang, 2007)",
"ref_id": "BIBREF2"
},
{
"start": 67,
"end": 87,
"text": "(Hagen et al., 2010)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "We made use of the mining method in (Hu et al., 2011) for collecting the query intent sets. With the search log data and clicks (Apr 1, 2009-Mar 31, 2010) as input, we finally obtained 9,412,308 query intent sets, which totally include 30,902,284 unique queries. The similar queries in each set share more than 10 clicks. Each set includes 2\u223c11 queries. We denote this data set as QSet. Note that this data set does not have any annotations. Thus, we also tried to construct another data set (denoted as QSet ann ) by intersecting QSet with WQ10. QSet ann includes 1,554 queries. Every query in QSet ann is then associated with the human annotations from WQ10 and linked to at least one query intent set in QSet.",
"cite_spans": [
{
"start": 36,
"end": 53,
"text": "(Hu et al., 2011)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "As each query has more than one segmentation due to different annotators, we select segmentation as our reference under two schemes: 'Majority' where the segmentations agreed upon by a majority of the annotators are chosen as the reference, and 'Best' where the annotated segmentations that maximize the break accuracy Acc brk of the evaluated segmenter are chosen as the reference.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "We mainly utilized three unsupervised systems as base segmenters. They are described in (Hagen et al., 2010) , (Hagen et al., 2011) and (Risvik et al., 2003) , denoted as Base H-1 , Base H-2 and Base CN respectively. They can represent the stateof-the-art QS performance. For example, Base H-2 on on data BW07(A) achieves 69.2%F sg which slightly outperforms the recent unsupervised system ",
"cite_spans": [
{
"start": 88,
"end": 108,
"text": "(Hagen et al., 2010)",
"ref_id": "BIBREF7"
},
{
"start": 111,
"end": 131,
"text": "(Hagen et al., 2011)",
"ref_id": "BIBREF8"
},
{
"start": 136,
"end": 157,
"text": "(Risvik et al., 2003)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "(69.0%F sg ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "As we focus on web QS, we did not compare LRMC with supervised methods which are only designed for one particular domain. For example, Yu et al. (2009) 's method is for queries of relational databases. The work (Bergsma and Wang, 2007) and the supervised stage of (Bendersky et al., 2009) are only for noun-phrases.",
"cite_spans": [
{
"start": 135,
"end": 151,
"text": "Yu et al. (2009)",
"ref_id": "BIBREF20"
},
{
"start": 211,
"end": 235,
"text": "(Bergsma and Wang, 2007)",
"ref_id": "BIBREF2"
},
{
"start": 264,
"end": 288,
"text": "(Bendersky et al., 2009)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Position right",
"sec_num": null
},
{
"text": "LRMC relies on a training data which is automatically collected with the help of query intent sets. Thus, in this section, we evaluate the training set collected by M1 and M2 (Section 4.2).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Weak Supervision",
"sec_num": "5.2"
},
{
"text": "In the experiments, we first applied a base segmenter to the queries in QSet and then managed to choose one segmentation as correct from the output for each query with either M1 or M2. Last, we evaluated the new output by checking only the segmentations for the queries in subset QSet ann . Some queries in QSet ann may belong to different intent-sets and in each intent-set may have different segmentation labels as 'correct'. In our evaluation, we randomly selected one of them as the final label. Besides, we also included an ideal method Oracle by which the correct segmentation can always be identified and used as the new output if the segmentation exists in the top-n re-sults of the base segmenter. Note that Oracle is an upper-bound result obtained by directly matching with human's annotation and cannot be applied to query intent sets QSet for collecting labeled data. Table 3 provides the results, where top-k (1 \u2264 k \u2264 5) means that the input to M1/M2 is the top-k results of the corresponding base segmenter. Note that the top-1 results are the performance of the base segmenters.",
"cite_spans": [],
"ref_spans": [
{
"start": 880,
"end": 887,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Consistency as Weak Supervision",
"sec_num": "5.2"
},
{
"text": "By checking the results of Oracle, we can find that for every base segmenter, more than 35% of the correct segmentations in the top-5 results are not covered by the top-1 results (in terms of Acc qry ). In addition, around 90% correct segmentation boundaries (Acc brk ) are included in the top-5 results of the base segmenters. These findings indicate the feasibility of our replacement model, which tries to replace a rank-1 segmentation by a rank-k (k > 1) segmentation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Weak Supervision",
"sec_num": "5.2"
},
{
"text": "From the table, we also see that both M1 and M2 are able to significantly perform better than the base segmenters do (p < 0.05, t-test). This can be observed through all the measures. For example, using Acc qry as the evaluation metric, the percentage of the correct segmentations that M2 discovers more than the base segmenters do ranges from 20.2% to 24.6%. These improvements prove the underlying assumption that queries with a similar intent tend to have consistent segmentation. Besides, we can see that M2 can reach a satisfied performance to collect the labeled data. For example, break-level accuracy Acc brk can reach 80%. Table 3 also shows that M1 and M2 perform best by using top 3 or 4 results from base segmenter. This finding indicates that our framework should work with a reasonable base segmenter.",
"cite_spans": [],
"ref_spans": [
{
"start": 632,
"end": 639,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Consistency as Weak Supervision",
"sec_num": "5.2"
},
{
"text": "By comparing the results generated by M1 and M2 with all three measures, we see that M2 consistently performs better than M1, and the difference is statistical significant (p < 0.05, t-test). This tells us that consistency should be calculated with all the top-n segmentations rather than with only the selected segmentations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Consistency as Weak Supervision",
"sec_num": "5.2"
},
{
"text": "In this section, we investigate the effectiveness of our LRMC which is a combination of data collecting method and replacement model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Query Segmentation",
"sec_num": "5.3"
},
{
"text": "In the experiments, we first made use of M2 to automatically collect the training data from the query intent sets QSet. During the process of col- lecting the data, we took only the top-3 segmentations as input. The training data set collected by M2 contains around 45 million instances (pairs of segmentations). Then, all this labeled data is used to train the replacement model as introduced in Section 4.3. We made use of LIBSVM (Chang and Lin, 2011) and a linear kernel in our experiment. Finally, we applied the learned replacement models to the evaluation data sets BW07 and WQ10. Table 4 reports the QS results 1 . Following previous work (Bergsma and Wang, 2007; Hagen et al., 2011) , we report four groups of results with the data BW07. In each of the first three groups, only the reference segmentations from annotator A, B or C are used. The fourth group is 'Best' of BW07. We also report two groups of results ('Majority' and 'Best') with the data WQ10. Comparing each pair of 'Base' and 'LRMC', we can see that LRMC proposed in this paper can be successfully spliced onto different base segmenters and significantly improves them over different data sets under the three evaluation metrics Acc qry , Acc brk and F sg . (p < 0.05, t-test). Especially, the state-of-the-art systems Base H-1 and Base H-2 have been significantly improved by LRMC. The improvements prove that the automatically-collected labeled data can guide QS and our replacement model can take advantage of the data.",
"cite_spans": [
{
"start": 432,
"end": 453,
"text": "(Chang and Lin, 2011)",
"ref_id": "BIBREF4"
},
{
"start": 646,
"end": 670,
"text": "(Bergsma and Wang, 2007;",
"ref_id": "BIBREF2"
},
{
"start": 671,
"end": 690,
"text": "Hagen et al., 2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [
{
"start": 587,
"end": 594,
"text": "Table 4",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Query Segmentation",
"sec_num": "5.3"
},
{
"text": "We have proposed a paradigm LRMC for QS. LRMC assumes the existence of a base segmenter 1 Note that the results for the base segmenters Base H-1 and Base H-2 are not exactly same as those reported in (Hagen et al., 2011) although they are very close. For example, Base H-1 and Base H-2 on WQ10 achieve 73.4% F sg and 74.2% F sg in the original paper. Ours are 71.2% and 72.1%. The reasons are as follows: For BW07, they used a cleaned version of the data set; for WQ10, they released just a subset of the data used in their experiments. and then learns how to select correct segmentations from the output of the base segmenter. The replacement model is trained by a labeled data set which can be automatically collected from query intent sets, instead of relying on any human annotation. There exist two interesting directions for future work: (1) we observe that there is still a big gap in performance between the proposed methods and Oracle. According to our analysis, most of the gap is caused by that the incorrect segmentations for some similar queries also happen to have a high consistency when measured by either proposed strategy. Thus, it is worth studying other methods that can address such performance gap.",
"cite_spans": [
{
"start": 200,
"end": 220,
"text": "(Hagen et al., 2011)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
},
{
"text": "(2) we would like to further explore the concept of query intent sets. In this paper, we assume that similar intent queries tend to have similar segmentations. A reasonable next step is to explore the idea that similar intent queries tend to have similar labels, which can be useful for the task of tagging query segments with semantic labels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "6"
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Entity-based crossdocument coreferencing using the vector space model",
"authors": [
{
"first": "A",
"middle": [],
"last": "Bagga",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Baldwin",
"suffix": ""
}
],
"year": 1998,
"venue": "COLING",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Bagga and B. Baldwin. Entity-based cross- document coreferencing using the vector space model. In COLING, 1998.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Twostage query segmentation for information retrieval",
"authors": [
{
"first": "M",
"middle": [],
"last": "Bendersky",
"suffix": ""
},
{
"first": "W",
"middle": [
"B"
],
"last": "Croft",
"suffix": ""
},
{
"first": "D",
"middle": [
"A"
],
"last": "Smith",
"suffix": ""
}
],
"year": 2009,
"venue": "SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Bendersky, W. B. Croft, and D. A. Smith. Two- stage query segmentation for information retrieval. In SIGIR, 2009.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Learning noun phrase query segmentation",
"authors": [
{
"first": "S",
"middle": [],
"last": "Bergsma",
"suffix": ""
},
{
"first": "Q",
"middle": [
"I"
],
"last": "Wang",
"suffix": ""
}
],
"year": 2007,
"venue": "EMNLP-CoNLL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Bergsma and Q. I. Wang. Learning noun phrase query segmentation. In EMNLP-CoNLL, 2007.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Freebase: a collaboratively created graph database for structuring human knowledge",
"authors": [
{
"first": "K",
"middle": [],
"last": "Bollacker",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Evans",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Paritosh",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Sturge",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Taylor",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Bollacker, C. Evans, P. Paritosh, T. Sturge, and J. Taylor. Freebase: a collaboratively created graph database for structuring human knowledge. In SIG- MOD, 2008.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "LIBSVM: A library for support vector machines. Intelligent Systems and Technology",
"authors": [
{
"first": "C",
"middle": [
"C"
],
"last": "Chang",
"suffix": ""
},
{
"first": "C",
"middle": [
"J"
],
"last": "Lin",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. C. Chang and C. J. Lin. LIBSVM: A library for sup- port vector machines. Intelligent Systems and Tech- nology, 2011.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Overview of the trec 2009 web track",
"authors": [
{
"first": "C",
"middle": [
"L A"
],
"last": "Clarke",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Craswell",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Soboroff",
"suffix": ""
}
],
"year": 2009,
"venue": "TREC",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. L. A. Clarke, N. Craswell, and I. Soboroff. Overview of the trec 2009 web track. In TREC, 2009.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "A unified and discriminative model for query refinement",
"authors": [
{
"first": "J",
"middle": [],
"last": "Guo",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Cheng",
"suffix": ""
}
],
"year": 2008,
"venue": "SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Guo, G. Xu, H. Li, and X. Cheng. A unified and dis- criminative model for query refinement. In SIGIR, 2008.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The power of naive query segmentation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Hagen",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Potthast",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Stein",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Br\u00e4utigam",
"suffix": ""
}
],
"year": 2010,
"venue": "SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Hagen, M. Potthast, B. Stein, and C. Br\u00e4utigam. The power of naive query segmentation. In SIGIR, 2010.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Query segmentation revisited",
"authors": [
{
"first": "M",
"middle": [],
"last": "Hagen",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Potthast",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Stein",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Br\u00e4utigam",
"suffix": ""
}
],
"year": 2011,
"venue": "WWW",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Hagen, M. Potthast, B. Stein, and C. Br\u00e4utigam. Query segmentation revisited. In WWW, 2011.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Mining query subtopics from search log data",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Qian",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Jiang",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Pei",
"suffix": ""
},
{
"first": "Q",
"middle": [],
"last": "Zheng",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Hu, Y. Qian, H. Li, D. Jiang, J. Pei, and Q. Zheng. Mining query subtopics from search log data. In WWW, 2011.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Exploring web scale language models for search query processing",
"authors": [
{
"first": "J",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Miao",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Behr",
"suffix": ""
},
{
"first": "C",
"middle": [
"L"
],
"last": "Giles",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Huang, J. Gao, J. Miao, X. Li, K. Wang, F. Behr, and C. L. Giles. Exploring web scale language models for search query processing. In WWW, 2010.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Generating query substitutions",
"authors": [
{
"first": "R",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Rey",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Madani",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Greiner",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Jones, B. Rey, O. Madani, and W. Greiner. Generat- ing query substitutions. In WWW, 2006.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Unsupervised query segmentation using click data: preliminary results",
"authors": [
{
"first": "J",
"middle": [],
"last": "Kiseleva",
"suffix": ""
},
{
"first": "Q",
"middle": [],
"last": "Guo",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Agichtein",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Billsus",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Chai",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Kiseleva, Q. Guo, E. Agichtein, D. Billsus, and W. Chai. Unsupervised query segmentation using click data: preliminary results. In WWW, 2010.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Unsupervised query segmentation using clickthrough for information retrieval",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "B.-J",
"middle": [
"P"
],
"last": "Hsu",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Zhai",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2011,
"venue": "SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Li, B.-J. P. Hsu, C. Zhai, and K. Wang. Unsuper- vised query segmentation using clickthrough for in- formation retrieval. In SIGIR, 2011.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Unsupervised query segmentation using only query logs",
"authors": [
{
"first": "N",
"middle": [],
"last": "Mishra",
"suffix": ""
},
{
"first": "R",
"middle": [
"S"
],
"last": "Roy",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Ganguly",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Laxman",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Choudhury",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. Mishra, R. S. Roy, N. Ganguly, S. Laxman, and M. Choudhury. Unsupervised query segmentation using only query logs. In WWW, 2011.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Query segmentation for web search",
"authors": [
{
"first": "K",
"middle": [
"M"
],
"last": "Risvik",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Mikolajewski",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Boros",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. M. Risvik, T. Mikolajewski, and P. Boros. Query segmentation for web search. In WWW, 2003.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Part of speech tagging using a network of linear separators",
"authors": [
{
"first": "D",
"middle": [],
"last": "Roth",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Zelenko",
"suffix": ""
}
],
"year": 1998,
"venue": "The 17th International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Roth and D. Zelenko. Part of speech tagging using a network of linear separators. In Coling- Acl, The 17th International Conference on Compu- tational Linguistics, 1998.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Short text classification in twitter to improve information filtering",
"authors": [
{
"first": "B",
"middle": [],
"last": "Sriram",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Fuhry",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Demir",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ferhatosmanoglu",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Demirbas",
"suffix": ""
}
],
"year": 2010,
"venue": "SIGIR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B. Sriram, D. Fuhry, E. Demir, H. Ferhatosmanoglu, and M. Demirbas. Short text classification in twitter to improve information filtering. In SIGIR, 2010.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Unsupervised query segmentation using generative language models and wikipedia",
"authors": [
{
"first": "B",
"middle": [],
"last": "Tan",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Peng",
"suffix": ""
}
],
"year": 2008,
"venue": "WWW",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B. Tan and F. Peng. Unsupervised query segmentation using generative language models and wikipedia. In WWW, 2008.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Unsupervised query segmentation using monolingual word alignment method",
"authors": [
{
"first": "D",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2012,
"venue": "Comp. and Info. Science",
"volume": "5",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Wu, Y. Zhang, and T. Liu. Unsupervised query segmentation using monolingual word alignment method. Comp. and Info. Science, 5(1), 2012.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Query segmentation using conditional random fields",
"authors": [
{
"first": "X",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Shi",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "X. Yu and H. Shi. Query segmentation using condi- tional random fields. In KEYS, 2009.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Query segmentation based on eigenspace similarity",
"authors": [
{
"first": "C",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Sun",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "T.-S",
"middle": [],
"last": "Chua",
"suffix": ""
}
],
"year": 2009,
"venue": "ACL-IJCNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Zhang, N. Sun, X. Hu, T. Huang, and T.-S. Chua. Query segmentation based on eigenspace similarity. In ACL-IJCNLP, 2009.",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"text": "Segmentation Result download adobe writer download adobe | writer download | adobe writer free adobe writer download free | adobe writer | download free | adobe | writer | download free adobe writer free | adobe writer free adobe | writer Table 1: Segmentation results for queries with a similar intent (Results in bold are considered correct.)",
"content": "<table><tr><td>queries</td><td>Rank-1 Segmentation Result</td><td>Rank-2</td></tr></table>",
"num": null,
"html": null,
"type_str": "table"
},
"TABREF2": {
"text": "The 'rank', 'direction' and 'position' features",
"content": "<table><tr><td colspan=\"2\">5 Experiments</td><td/></tr><tr><td colspan=\"2\">5.1 Experimental Setup</td><td/></tr><tr><td colspan=\"3\">Following Hagen et al. (2011), we evaluate a QS</td></tr><tr><td colspan=\"2\">system at three levels: Query Level:</td><td/></tr><tr><td>Acc qry =</td><td>#correctly segmented queries #queries in the evaluation data set</td><td>(5)</td></tr></table>",
"num": null,
"html": null,
"type_str": "table"
},
"TABREF4": {
"text": "Consistency as weak supervision on QSet ann (Majority)",
"content": "<table/>",
"num": null,
"html": null,
"type_str": "table"
},
"TABREF6": {
"text": "Performance on query segmentation",
"content": "<table/>",
"num": null,
"html": null,
"type_str": "table"
}
}
}
} |