File size: 92,478 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 |
{
"paper_id": "I11-1009",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:31:17.110840Z"
},
"title": "Automatic Topic Model Adaptation for Sentiment Analysis in Structured Domains",
"authors": [
{
"first": "Geoffrey",
"middle": [],
"last": "Levine",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Illinois at Urbana-Champaign Urbana",
"location": {
"region": "IL"
}
},
"email": "levine@illinois.edu"
},
{
"first": "Gerald",
"middle": [],
"last": "Dejong",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Illinois at Urbana-Champaign Urbana",
"location": {
"region": "IL"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We present a novel topic modeling approach to sentiment analysis for documents organized into hierarchical categories. In our approach, positive, negative, and subject matter topics are learned and used to infer document labels. A Markov chain Monte Carlo model procedure adapts the number and structure of topics based on a minimum description length objective function. We apply our approach to Yelp.com business reviews and Amazon.com book reviews and demonstrate that 1) the model adaptation procedure selects a high quality model from the space of alternatives, and 2) the resulting model performs well relative to state of the art regression and topic modeling approaches.",
"pdf_parse": {
"paper_id": "I11-1009",
"_pdf_hash": "",
"abstract": [
{
"text": "We present a novel topic modeling approach to sentiment analysis for documents organized into hierarchical categories. In our approach, positive, negative, and subject matter topics are learned and used to infer document labels. A Markov chain Monte Carlo model procedure adapts the number and structure of topics based on a minimum description length objective function. We apply our approach to Yelp.com business reviews and Amazon.com book reviews and demonstrate that 1) the model adaptation procedure selects a high quality model from the space of alternatives, and 2) the resulting model performs well relative to state of the art regression and topic modeling approaches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Selecting an appropriate model is an important part of any machine learning endeavor. The model must be chosen in a manner so as to balance two objectives: 1) Be sufficiently rich to capture the relevant patterns in the data, and 2) Be simple enough to avoid spurious patterns in the training data (overfitting). In natural language processing tasks, there are often many modeling choices to be made regarding what feature granularities and interactions to consider. It is important to make these decisions in a manner such that the resulting model strikes a balance between these two somewhat contradictory objectives.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In order to appropriately make these choices, we must consider not only the task involved but also the training data available. With copious data we can reliably calibrate complex models, but with limited data complex models risk overfitting. Many general model selection techniques exist in which each candidate model is fit to the training data and scored with respect to a particular criterion. While these approaches allow us to compare a small number of models in order to select the most appropriate, they require calibrating each model's parameters to the training data. However, when there are many modeling choices to be made and thus a large space of alternative models, fitting all of them to the training data is computationally prohibitive.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we present a novel topic modeling approach for structured sentiment analysis domains and an automatic model adaptation approach that takes advantage of categorical metadata. This model adaptation approach resolves the structure of the metadata with the significant patterns in the training data to determine the number and range of latent topics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We demonstrate our approach on Yelp.com business reviews as well as Amazon.com book reviews. We show that our model adaptation approach selects an appropriate model given a particular amount of training data, and the resulting model is high quality relative to alternative regression and topic modeling approaches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Sentiment analysis, in which the opinion of the author is estimated from a document, has recently grown in popularity. Many works have explored unigram models (Pang and Lee, 2005; Snyder and Barzilay, 2007) . Higher-order n-gram models are explored in (Pang and Lee, 2008; Baccianella et al., 2009) . In order to combat the high dimensional feature space that accompanies such models, models restricting features based on part of speech patterns (Baccianella et al., 2009) or opinion templates (root, modifiers, negation words) (Qu et al., 2010) have been introduced.",
"cite_spans": [
{
"start": 159,
"end": 179,
"text": "(Pang and Lee, 2005;",
"ref_id": "BIBREF10"
},
{
"start": 180,
"end": 206,
"text": "Snyder and Barzilay, 2007)",
"ref_id": "BIBREF17"
},
{
"start": 252,
"end": 272,
"text": "(Pang and Lee, 2008;",
"ref_id": "BIBREF11"
},
{
"start": 273,
"end": 298,
"text": "Baccianella et al., 2009)",
"ref_id": "BIBREF1"
},
{
"start": 446,
"end": 472,
"text": "(Baccianella et al., 2009)",
"ref_id": "BIBREF1"
},
{
"start": 528,
"end": 545,
"text": "(Qu et al., 2010)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Topic models are generative models in which the words in a document are assumed to be asso-ciated with one of a number of abstract \"topics.\" Latent Dirichlet allocation (Blei et al., 2003) is a popular topic model in which the topic distribution per document is assumed to have a Dirichlet prior. In supervised LDA (Blei and McAuliffe, 2007) , the distribution of document topics is used to produce a document label. (Zhao et al., 2010; Titov and McDonald, 2008b; Titov and McDonald, 2008a ) focus on topic modeling based approaches to aspect-based sentiment summarization, identifying product features and the opinion associated with each.",
"cite_spans": [
{
"start": 169,
"end": 188,
"text": "(Blei et al., 2003)",
"ref_id": "BIBREF3"
},
{
"start": 315,
"end": 341,
"text": "(Blei and McAuliffe, 2007)",
"ref_id": "BIBREF2"
},
{
"start": 417,
"end": 436,
"text": "(Zhao et al., 2010;",
"ref_id": "BIBREF24"
},
{
"start": 437,
"end": 463,
"text": "Titov and McDonald, 2008b;",
"ref_id": "BIBREF20"
},
{
"start": 464,
"end": 489,
"text": "Titov and McDonald, 2008a",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Model selection is the act of using data to choose a statistical model from a set of candidates. Often, this task is performed by fitting each candidate model to the training data and using a criterion to score the models and select one. Popular criteria include the Akaike information criterion (AIC) (Akaike, 1974) , the Bayesian information criterion (BIC) (Schwarz, 1978) , and the minimum description length principle (MDL) (Rissanen, 1978; Grunwald, 2007) . Structural Risk Minimization (Vapnik, 1995) defines a general framework in which a nested hierarchy of hypotheses can be defined based on prior knowledge of the domain, such that a hypothesis balancing goodness of fit with simplicity can be identified. The work presented in this paper is closely related to the model adaptation procedure presented in (Levine et al., 2010) , in which a hill-climbing approach is used to explore a large space of generative models.",
"cite_spans": [
{
"start": 302,
"end": 316,
"text": "(Akaike, 1974)",
"ref_id": "BIBREF0"
},
{
"start": 360,
"end": 375,
"text": "(Schwarz, 1978)",
"ref_id": "BIBREF15"
},
{
"start": 429,
"end": 445,
"text": "(Rissanen, 1978;",
"ref_id": "BIBREF14"
},
{
"start": 446,
"end": 461,
"text": "Grunwald, 2007)",
"ref_id": "BIBREF7"
},
{
"start": 493,
"end": 507,
"text": "(Vapnik, 1995)",
"ref_id": "BIBREF21"
},
{
"start": 816,
"end": 837,
"text": "(Levine et al., 2010)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Our approach takes advantage of hierarchical categorical metadata. Formally, this hierarchy forms a tree structure, which we refer to as C (See Figure 1 ). Individual nodes in the tree are called categories, for which we use notation c. A categorization, c, is a set of categories,",
"cite_spans": [],
"ref_spans": [
{
"start": 144,
"end": 153,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "c = {c d,1 , c d,2 , ...}.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "c can be thought of as metadata about a product/service being reviewed. For example, with regard to a book review, c could equal {\"Fiction\", \"Fiction\\Drama\", \"Fiction\\Drama\\Romance\", ... }. c must be well formed. That is, if a node c \u2208 C appears in categorization c, all ancestors of c (in the tree C) must also appear in c. c can contain multiple distantly related categories. For example, a particular book could belong to both \"Fic-tion\\Poetry\" and \"Children\\Humor.\" Documents, or examples, are denoted d = Figure 1 : A subtree of the category tree, C, corresponding to the Amazon Books domain.",
"cite_spans": [],
"ref_spans": [
{
"start": 510,
"end": 518,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "x d , c d , y d . x d = [w d,1 , w d,2 , ..., w d,|x d | ]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "is a vector of words. Each word is an element from the vocabulary, V = {w 1 , w 2 , ..., w |V | }. c d is the document's categorization. y d is a numeric rating from a discrete space ({1,2,3,4,5} for our domains). The rating is an overall score given by the document's author to the product or service being reviewed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "We are given a collection of documents, D, and our goal is to learn a function f ( x, c ) to predict rating\u0177 from an unlabeled document so as to minimize the expected loss over the unknown distribution of documents:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "E (loss(y, f ( x, c )))",
"eq_num": "(1)"
}
],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "We use the squared error loss function.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Modeling for Sentiment Analysis in Structured Domains",
"sec_num": "3"
},
{
"text": "We will start by presenting our generative document topic model. In this model, each review is composed of a mixture of topics, and each topic is associated with a distribution over words. We use t \u2208 T to denote a topic, and P t to denote t's word distribution. Within a document, each word is assumed to be generated from a particular topic, although which topic is unobservable. In many topic model approaches, such as latent Dirichlet allocation (Blei et al., 2003) , topics are learned in an unsupervised or weekly supervised fashion (as is the case with supervised LDA). In our model, we assume each document is generated according to a rigid topic distribution (more similar to labeled LDA (Ramage et al., 2011) ). Each document is a mixture of three topics: 1) a positive topic (+), in which the reviewer is speaking favorably about the product/service, 2) a negative topic (-), in which the reviewer is speaking unfavorably, and 3) a subject topic (s i ) corresponding to general text about the content/features of the product.",
"cite_spans": [
{
"start": 449,
"end": 468,
"text": "(Blei et al., 2003)",
"ref_id": "BIBREF3"
},
{
"start": 696,
"end": 717,
"text": "(Ramage et al., 2011)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "The proportion of positive words to negative words is a function of the rating score. Subject topics reflect the language used when discussing a particular product or group of products, and do not directly influence a document's rating. Still, learning these topics appropriately is crucial to the performance of the model. When a word is indicative of either the positive or negative topic, it is important to account for its probability in the subject topic. For example, the word \"good\" may be less indicative of a book review's rating if the review discusses a book about ethics. Furthermore, if subject topics are not learned appropriately, words related to the subject matter of products/services with a disproportionate number of positive training reviews would be attributed to the positive word topic. This will lead to poor performance on unseen data. On the other hand, if these words are correctly attributed to the subject topic, then the high ratings will appropriately be attributed to the unconditional positive words appearing in the reviews.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "What constitutes a subject worthy of having its own topic? For books, should we only make broad distinctions such as fiction vs. non-fiction? Should we learn a unique subject topic for each book? Should we use something in between these two extremes? In answering these questions, we need to balance goodness of fit to the training data with model simplicity. There is no optimal answer, it is a function both of the domain (in that we need to make the most \"significant\" distinctions), and the amount of training data available to calibrate our model (more training data allows us to reliably learn the additional parameters introduced by making additional distinctions).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "There exists a many-to-one relationship between documents and subject topics. The mapping from document to subject topic is a function of the document's categorization,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "s i = g(c d ), s i \u2208 T .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "We call the function g the topic mapping function. The range of g is the set of subject topics, {s 1 , s 2 , ..., s N } \u2282 T . In Sections 3.1.1 and 3.1.2 we assume that g is fixed. In Section 3.2, we consider exploring the space of alternative topic mapping functions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "We assume that in expectation, a fixed but unknown fraction, \u03b1 of each document is composed of the subject topic. The remainder of the review is composed of the positive and negative topics, and the positive/negative ratio is related to the docu-ment's rating. Let y min and y max represent the minimum and maximum scores in the rating scale. For document d with score y d the expected fractional breakdown into topics is as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "Positive: f + (y d ) =(1 \u2212 \u03b1) y d \u2212 y min y max \u2212 y min Negative: f \u2212 (y d ) =(1 \u2212 \u03b1) y max \u2212 y d y max \u2212 y min Subject: f s (y d ) =\u03b1 (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "In total, a model M is composed of the topic mapping function, the value \u03b1, and the word distributions associated with each topic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "M = (g, \u03b1, P + , P \u2212 , P s 1 , P s 2 , ..., P s N ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Structure",
"sec_num": "3.1"
},
{
"text": "Expectation maximization (Hastie et al., 2001 ) can be used to train our topic model. The procedure works by iteratively updating 1) the assignment of words in each document to latent topics (Expectation Step), and 2) the word distributions associated with each topic (Maximization Step). EM proceeds as follows:",
"cite_spans": [
{
"start": 25,
"end": 45,
"text": "(Hastie et al., 2001",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "Expectation",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "Step Each word is assigned an expected topic distribution. For word i in document d:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "q d,i (+) = f + (y d )P + (w d,i ) Z d,i q d,i (\u2212) = f \u2212 (y d )P \u2212 (w d,i ) Z d,i q d,i (g(c d )) = f s (y d )P g(c d ) (w d,i ) Z d,i Z d,i = t\u2208{+,\u2212,g(c d )} f t (y d )P t (w d,i ) (3) Maximization",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "Step Topic word distributions are updated so as to maximize the likelihood of the training data. For each topic t:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P t (w) = d\u2208D |x d | i=1 I w (w d,i )q d,i (t) d\u2208D |x d | i=1 q d,i (t)",
"eq_num": "(4)"
}
],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "I w (w d,i ) = 1 if w d,i = w 0 otherwise (5) 3.1.2 Inference",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "Given the trained topic models we use Bayes' theorem to compute the probability that an unlabeled document x d , c d is associated with a particular rating. Let T d = {+, \u2212, g(c d )}:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (y|x d , c d ) = P (y)P (x d , c d |y) P (x d , c d ) = P (y) |x d | i=1 t\u2208T d f t (y)P t (w d,i ) y P (y ) |x d | i=1 t\u2208T d f t (y)P t (w d,i )",
"eq_num": "(6)"
}
],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "For evaluation purposes, we output the expected value of y and compute the squared error to the true value.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "In this section we introduce a Markov chain Monte Carlo approach to selecting a topic mapping function g. Here, we stochastically explore the space of topic mapping functions, driven by the minimum description length principle and estimates of the effect of modifications to g. This approach resists local minima and efficiently finds a high quality topic mapping function.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Adaptation",
"sec_num": "3.2"
},
{
"text": "Our goal is to find a model that balances fit to the training data with simplicity, and concentrates its flexibility where most useful to capture relevant patterns in the domain. We accomplish this by utilizing a two part minimum description length objective function. The objective is the sum of 1) the description length (in bits) required to encode the model and 2) the description length of the data given the model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Minimum Description Length Objective",
"sec_num": "3.2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(M, D) = L(M ) + L(D|M )",
"eq_num": "(7)"
}
],
"section": "Minimum Description Length Objective",
"sec_num": "3.2.1"
},
{
"text": "where L(M ) is a function of the number of model parameters (\u2248 the product of the number of topics and the vocabulary size) and L(D|M ) is the negative log likelihood of the data given the model. Thus the goal is to jointly minimize the complexity of the model and maximize the likelihood of the data given the model, and the objective can be rewritten as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Minimum Description Length Objective",
"sec_num": "3.2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(M, D) = \u03b2(N + 2)|V | + \u2212log (l(D|M ))",
"eq_num": "(8)"
}
],
"section": "Minimum Description Length Objective",
"sec_num": "3.2.1"
},
{
"text": "Figure 2: Two possible partitioning trees for the Amazon.com Books category tree (Figure 1 ). Tree b) is formed by splitting s 2 in a) based on membership in the \"Computer\\Software\" category.",
"cite_spans": [],
"ref_spans": [
{
"start": 81,
"end": 90,
"text": "(Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Minimum Description Length Objective",
"sec_num": "3.2.1"
},
{
"text": "where \u03b2 is a complexity penalty constant, which is selected via cross-validation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Minimum Description Length Objective",
"sec_num": "3.2.1"
},
{
"text": "The topic mapping function g maps from categorization c to a subject topic s i \u2208 T . We select a particular g from the space of binary partitioning trees, G. In a binary partitioning tree, each internal node references a category c, and each leaf node references a subject topic s i . See Figure 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 289,
"end": 297,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Topic Mapping Functions",
"sec_num": "3.2.2"
},
{
"text": "Starting at the root, a categorization, c is recursively assigned by each internal node to 1) the left subtree if the referenced category c is in c, and 2) the right subtree otherwise, until a leaf (with associated subject topic) is reached. For example, within the book review domain, a node may reference the category \"Computers.\" In this case, computer books are recursively assigned a subject topic by the left subtree, and all others by the right subtree. We allow only well formed partitioning trees: Any node in g that references a category c with parent category parent(c) \u2208 C must have an ancestor that references parent(c). For example, we do not allow a node in g to reference \"Com-puters\\Software\", unless we have already conditioned on the \"Computers\" category. This constraint guarantees that we partition the space of categorizations into coherent regions (we would never assign \"Computer\\Hardware\" and \"Fiction\" books to the same subject topic while assigning \"Computer\\Software\" to a different topic).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Topic Mapping Functions",
"sec_num": "3.2.2"
},
{
"text": "In order to guide the search through G, we consider 2 types of modification operations: We can 1) Split a leaf based on category c \u2208 C, splitting one partition into two, adding an additional subject topic to the model, or 2) Merge two leaves with the same parent, combining two partitions into one, removing a subject topic from the model. Given a particular g, there are a finite number of possible merge and split operations to the leaves. Key to our search is the fact that we can estimate the change to the description length objective that each possible modification will cause, using the latent topic distributions assigned to each topic during expectation maximization. Consider merging two subject topics s i and s j :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u2206L = \u2212\u03b2|V | \u2212 t\u2208{s i ,s j } w\u2208V # t (w)log P m i,j (w) P t (w)",
"eq_num": "(9)"
}
],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "# t (w) = d\u2208D |x d | i=1 I w (w d,i )q d,i (t) P m i,j (w) = t\u2208{s i ,s j } # t (w) t\u2208{s i ,s j } d\u2208(D) |x d | i=1 q d,i (t)",
"eq_num": "(10)"
}
],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "Now consider splitting subject topic s i based on category c:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u2206L = \u03b2|V | \u2212 t\u2208{s i,c ,s i,!c } w\u2208V # t (w)log P t (w) P s i (w)",
"eq_num": "(11)"
}
],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "# s i,c (w) = d=(x,c,y)\u2208D s.t.c\u2208c |x d | i=1 I w (w d,i )q d,i (s i ) P s i,c (w) = # s i,c (w) d=(x,c,y)\u2208D s.t.c\u2208c |x d | i=1 q d,i (s i ) # s i,!c (w) = d=(x,c,y)\u2208D s.t.c / \u2208c |x d | i=1 I w (w d,i )q d,i (s i ) P s i,!c (w) = # s i,!c (w) d=(x,c,y)\u2208D s.t.c / \u2208c |x d | i=1 q d,i (s i )",
"eq_num": "(12)"
}
],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "These estimates are upper bounds on the change to the description length objective function. Incorporating these changes (and the associated word distributions) and then retraining the model with expectation maximization may further reduce the objective. These bounds serve as a guide to estimate the objective for models that have not been fit to the training data, which will drive our search through G for the optimal topic mapping function.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adjacent Model Estimation",
"sec_num": "3.2.3"
},
{
"text": "Markov chain Monte Carlo (Gilks et al., 1996) stochastically steps through the space of alternative topic mapping functions. At each iteration of MCMC, the topic model with the current topic mapping function is fit to the training data and the objective change associated with all possible merges and splits is estimated. We then construct a proposal distribution for alternative models that can be reached with these operations. Limiting the proposal distribution to these candidate models, as in (Titov and Klementiev, 2011) and (Singh et al., 2011) induces a decomposable, feasible computation. A model is sampled from this distribution and adopted if certain criteria on its fitness are met.",
"cite_spans": [
{
"start": 25,
"end": 45,
"text": "(Gilks et al., 1996)",
"ref_id": "BIBREF6"
},
{
"start": 498,
"end": 526,
"text": "(Titov and Klementiev, 2011)",
"ref_id": "BIBREF18"
},
{
"start": 531,
"end": 551,
"text": "(Singh et al., 2011)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "MCMC will converge to a probability distribution over models. By making better models (those with a lower objective) more probable, the MCMC chain will be driven towards higher quality models. We use an exponential distribution over models:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (M ) = e \u2212L(M,D) Z P",
"eq_num": "(13)"
}
],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "with normalization factor Z P . The proposal distribution, Q assigns some probability to all candidate models that can be reached by a single merge or split to each of the leaves in the current partitioning tree. In Q, splits and merges to leaves without a common parent are independent by construction. Now, consider a particular leaf, l, that has the following possible splits, S = {c 1 , c 2 , ..., c l }, and cannot be merged with another leaf. For example, in Figure 2a , the leaf corresponding to s 1 meets this condition as it cannot be merged to another leaf and has possible splits {\"Fiction\\Drama\", \"Fiction\\Poetry\", \"Nonfiction\", \"Computers\", \"Children\", ... }. Let M l represent the subset of models where l is not split, and M l,c i represent the subset of candidate models where l has been split with respect to category c i .",
"cite_spans": [],
"ref_spans": [
{
"start": 465,
"end": 474,
"text": "Figure 2a",
"ref_id": null
}
],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "Q(M l ) = e \u2212\u03c4 L(M,D) Z l Q(M l,c i ) = e \u2212\u03c4 L(M l,c i ,D) |S|Z l Z l = e \u2212\u03c4 L(M,D) + c\u2208S 1 |S| e \u2212 L(M l,c ,D) (14)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "0 < \u03c4 \u2264 1 controls a balance between having the proposal distribution completely influenced by the estimated objectives vs. a uniform proposal distribution. For all pairs of leaves that share a common parent, we entertain a merge operation. In Figure 2a the leaves corresponding to s 2 and s 3 meet this criteria. Suppose two leaves, l and l have possible splits S = {c 1 , c 2 , ..., c l } and S = {c 1 , c 2 , ..., c l } respectively. In addition to the one merged alternative, there are w = (|S| + 1)(|S | + 1) alternatives that involve only splits to the two leaves. Let M l,l represent the subset of candidate models where l and l are merged",
"cite_spans": [],
"ref_spans": [
{
"start": 244,
"end": 253,
"text": "Figure 2a",
"ref_id": null
}
],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Q(M l,l ) = w \u2212 1 2 e \u2212\u03c4 L(M l,l ,D) w \u2212 1 2 e \u2212\u03c4 L(M l,l ,D) + (Z l )(Z l )",
"eq_num": "(15)"
}
],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "The factor of w \u2212 1 2 accounts for the difference between the number of neighbors that the models with l and l merged vs. split have. If the two leaves are not merged, then the conditional probability for each of the (|S| + 1)(|S | + 1) remaining structural alternatives is computed in Equation 14.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "A new topic mapping function g is sampled from Q and fit to the training data via the expectation maximization presented in Section 3.1.1. If a random value sampled uniformly from U[0, 1) is less than",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (M g )Q(g|g ) P (M g )Q(g |g)",
"eq_num": "(16)"
}
],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "then g is accepted as the new topic mapping function g t+1 . This guarantees that the Markov chain will converge to the distribution P as t \u2192 \u221e. Because the ratio P (M g )/P (M g ) appears in Equation 16, the normalization factor Z P in Equation 13 cancels out and does not need to be computed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MCMC exploration",
"sec_num": "3.2.4"
},
{
"text": "We perform a set of experiments to demonstrate the following: 1) Given the topic model structure outlined in Section 3.1, the model adaptation procedure in section 3.2 selects a high performing topic mapping function while only evaluating a small fraction of the total number of funtions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Empirical Evaluation",
"sec_num": "4"
},
{
"text": "2) The topic model resulting from model adaptation is high quality relative to alternative stateof-the-art approaches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Empirical Evaluation",
"sec_num": "4"
},
{
"text": "We demonstrate our approach to two structured sentiment analysis datasets. First, we gathered a collection of approximately 8,000 Yelp.com business reviews from the greater New York area. For this data, businesses are assigned into categories and subcategories based on the Yelp.com business hierarchy. There are 22 primary categories {Arts and Entertainment, Education, Financial Services, Restaurants,...}, each with 6 to 100 subcategories (restaurants have the most subcategories, {Japanese, Barbeque, Cafe, Fast Food, Burgers, Ultra High Enc, Formal, Full Bar,...}). Businesses can be assigned to multiple categories and subcategories within the hierarchy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "4.1"
},
{
"text": "Second, we utilize 20,000 Amazon.com book reviews, extracted from the data set first presented in (Qu et al., 2010) . Categorical distinctions in these domains are related to the Amazon.com product hierarchy. A small portion of the product hierarchy appears in Figure 1 . Books can be assigned to multiple distantly related categories. For example, the book Six Wives of Henry VIII belongs to \"History\\Europe\\England\\Tudor and Stuart,\" \"Biographies and Memoirs\\Specific Groups\\Women\" and three other categories For each domain, we have at most one review corresponding to any particular business/book. This allows for a broad coverage of the space of categorizations.",
"cite_spans": [
{
"start": 98,
"end": 115,
"text": "(Qu et al., 2010)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [
{
"start": 261,
"end": 269,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data",
"sec_num": "4.1"
},
{
"text": "To compensate for extreme variations in the training data we apply two smoothing steps. First, we found that for longer reviews, the assumption that each word is drawn independently from the document's topics is too strong, and so for reviews with more than 35 words, we scale the term counts such that the total is 35. Second, because of the large size of the vocabulary, after training, some rare words have zero or near zero probability in some of the topics. When these words are observed during inference, they have a very strong effect on the document's expected rating. We found that smoothing the subject topics with the overall word distribution across topics stabilizes the predicted ratings and improves performance. The amount of smoothing could be optimized to maximize the likelihood of the test data, but we found that performance varied little for a wide range of values and so we choose a 1 to 1 smoothing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Discussion",
"sec_num": "4.2"
},
{
"text": "From each dataset, we sample a subset of size 1000 for cross validation parameter tuning and use the remaining examples for experimentation. The validation data is used to learn the values of \u03b1, the subject topic fraction, and \u03b2, the complexity penalty. We found that setting \u03c4 , the MCMC smoothing factor, equal to .1 worked well across our datasets. For each trial, then, disjoint training and test sets are sampled from the remaining data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Discussion",
"sec_num": "4.2"
},
{
"text": "First, we apply our Markov chain Monte Carlo model adaptation procedure along with greedy and random alternatives to demonstrate the necessity of a directed and stochastic approach. In the greedy approach, at each iteration we estimate the objective for all candidate models that can be reached with a single split/merge to each subject topic and adopt the model with the minimum estimate. For the random approach, at each iteration we start with the simplest topic mapping function (mapping all categorizations to one subject topic), and uniformly at random add distinctions until the model has the same number of subject topics as the optimal model found by the MCMC approach. We choose this instead of sampling at random from all possible topic mapping functions as the vast majority of such functions have nearly as many subject topics as training examples. For each approach, at iteration i, we chart the test mean squared error for the best (lowest objective) model observed during training in iterations 1 to i. Figure 3 charts the per iteration mean squared error on the Yelp test data for the three model adaptation approaches. The greedy approach initially makes the fastest progress, but it is susceptible to local minima, and it levels off before being overtaken by the MCMC approach. As the random approach does not leverage the data in determining what distinctions to make, it fails to make progress at the rate of the other approaches. Its poor performance is indicative of the importance of having an efficient directed model adaptation approach, as high performing models are few and far between, even if we limit our search to models of the appropriate complexity level (number of subject topics). Figure 4 shows a representative partitioning tree learned from the Yelp.com dataset.",
"cite_spans": [],
"ref_spans": [
{
"start": 1019,
"end": 1027,
"text": "Figure 3",
"ref_id": "FIGREF0"
},
{
"start": 1717,
"end": 1725,
"text": "Figure 4",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Results and Discussion",
"sec_num": "4.2"
},
{
"text": "Next we compare our approach to alternative regression and topic modeling approaches. In order to implement regression, we 1) Form a vector of unigram (and optionally bigram) occurrences normalized to length 1 (which we found to work better than unnormalized or frequency vectors), and 2) Form a vector corresponding to categorical membership with one element for each node in category tree C. For each example, we set each element in the vector to value \u03b3 if the example belongs to the corresponding category, and 0 otherwise. The feature vector is the concatenation of these two vectors. We tested three regression approaches: ridge regression, lasso, and -support vector regression with a quadratic kernel (Chang and Lin, 2001 ). In each case, the cross validation dataset is used to tune the value of \u03b3 and the regularization parameter (for ridge regression and lasso) or and the cost parameter (for SVR). We found that in all cases, lasso and SVR were ouperformed by ridge regression, and so omit their results.",
"cite_spans": [
{
"start": 709,
"end": 729,
"text": "(Chang and Lin, 2001",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Discussion",
"sec_num": "4.2"
},
{
"text": "For the supervised latent Dirichlet allocation approach, as the space of labels is numeric and discrete, we can treat the task either as a regression problem (Blei and McAuliffe, 2007) , or as a multiclass classification problem We used an open source implementation of each approach, (Chang, 2010) and (Wang, 2009) , and found that utilizing the multiclass approach and predicting the expected rating based on the posterior likelihood of each class outperformed the regression approach, so we present these results. The cross validation data is used to learn the number of latent topics and Dirichlet distribution parameter. For the Markov chain Monte Carlo approach, in order to hasten learning for this comparison, starting from the simplest topic mapping function, we perform a greedy model adaptation until reaching an estimated local minimum, and then apply 50 additional iterations of MCMC model adaptation. Figure 5 shows the average mean squared error for each approach for various amounts of training data. Our topic model with model adaptation has lower error than each of the alternatives. Paired ttests reveal that the differences are statistically significant in all cases (p < .01 for all Yelp.com and p < .05 for all Amazon.com tests). Using MCMC model adaptation also outperforms using either the simplest topic mapping function or the most complex mapping function (which maps each distinct training categorization to a different subject topic).",
"cite_spans": [
{
"start": 158,
"end": 184,
"text": "(Blei and McAuliffe, 2007)",
"ref_id": "BIBREF2"
},
{
"start": 285,
"end": 298,
"text": "(Chang, 2010)",
"ref_id": "BIBREF5"
},
{
"start": 303,
"end": 315,
"text": "(Wang, 2009)",
"ref_id": "BIBREF23"
}
],
"ref_spans": [
{
"start": 915,
"end": 923,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results and Discussion",
"sec_num": "4.2"
},
{
"text": "Ridge regression with unigrams uses the same word and categorical representations as our approach. However, it is unable to entertain the nonlinear relationships between document categorizations and words and is outperformed in all cases. Bigrams improve the performance of ridge regression, especially for larger amounts of training data. This suggests that accounting for word ordering could potentially improve the performance of our topic model as well. sLDA is unable to take advantage of the categorical information during topic construction, and with the limited training data available, its performance is marginally better than guessing the mean label (MSE: 1.675 for Yelp.com data and 1.660 for Amazon.com data).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Discussion",
"sec_num": "4.2"
},
{
"text": "We present an approach to sentiment analysis for structured domains. In our approach, positive, negative, and subject topics are learned and used to infer document labels. Partitioning tree based topic mapping functions define the number and structure of subject topics. A Markov chain Monte Carlo model adaptation procedure explores the space of topic mapping functions based on a minimum description length objective. We demonstrate the approach on two sentiment analysis domains and show that the model adaptation procedure efficiently finds a high performance model that leverages the categorical structure of the documents to outperform other regression and topic modeling approaches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
}
],
"back_matter": [
{
"text": "This material is based upon work supported by the Office of Naval Research under Award No. ONR Grant N00014-09-1-0693. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of the Office of Naval Research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgment",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A new look at the statistical model identification",
"authors": [
{
"first": "H",
"middle": [],
"last": "Akaike",
"suffix": ""
}
],
"year": 1974,
"venue": "IEEE Transactions on Automatic Control",
"volume": "19",
"issue": "6",
"pages": "716--723",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Akaike. 1974. A new look at the statistical model identification. IEEE Transactions on Auto- matic Control, 19(6):716-723.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Muli-facet rating of product reviews",
"authors": [
{
"first": "S",
"middle": [],
"last": "Baccianella",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Esuli",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Sebastiani",
"suffix": ""
}
],
"year": 2009,
"venue": "The 31st European Conference on Information Retrieval Research",
"volume": "",
"issue": "",
"pages": "461--472",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Baccianella, A. Esuli, and F. Sebastiani. 2009. Muli-facet rating of product reviews. In The 31st European Conference on Information Retrieval Re- search, pages 461-472.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Supervised topic models",
"authors": [
{
"first": "D",
"middle": [],
"last": "Blei",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Mcauliffe",
"suffix": ""
}
],
"year": 2007,
"venue": "Neural Information Processing Systems",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Blei and J. McAuliffe. 2007. Supervised topic mod- els. In Neural Information Processing Systems.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Latent dirichlet allocation",
"authors": [
{
"first": "D",
"middle": [],
"last": "Blei",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Ng",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Jordan",
"suffix": ""
}
],
"year": 2003,
"venue": "Journal of Machine Learning Research",
"volume": "3",
"issue": "",
"pages": "993--1022",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Blei, A. Ng, and M. Jordan. 2003. Latent dirichlet allocation. Journal of Machine Learning Research, 3:993-1022.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "LIB-SVM: A Library for Support Vector Machines",
"authors": [
{
"first": "Chih-Chung",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Chih-Jen",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chih-Chung Chang and Chih-Jen Lin, 2001. LIB- SVM: A Library for Support Vector Machines. Software available at http://www.csie.ntu. edu.tw/\u02dccjlin/libsvm.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Lda: Collapsed gibbs sampling methods for topic models",
"authors": [
{
"first": "J",
"middle": [],
"last": "Chang",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Chang. 2010. Lda: Collapsed gibbs sampling meth- ods for topic models. online.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Markov Chain Monte Carlo in Practice",
"authors": [
{
"first": "W",
"middle": [],
"last": "Gilks",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Richardson",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Spiegelhalter",
"suffix": ""
}
],
"year": 1996,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. Gilks, S. Richardson, and D. Spiegelhalter. 1996. Markov Chain Monte Carlo in Practice. Chapman and Hall.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The Minimum Description Length Principle",
"authors": [
{
"first": "P",
"middle": [],
"last": "Grunwald",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Grunwald. 2007. The Minimum Description Length Principle. The MIT Press, Cambridge, Mass.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "The Elements of Statistical Learning",
"authors": [
{
"first": "T",
"middle": [],
"last": "Hastie",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Tibshirani",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Friedman",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Hastie, R. Tibshirani, and J. Friedman. 2001. The Elements of Statistical Learning. Springer, New York, NY.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Automatic model adaptation for complex structured domains",
"authors": [
{
"first": "G",
"middle": [],
"last": "Levine",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Dejong",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Samdani",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Vembu",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Roth",
"suffix": ""
}
],
"year": 2010,
"venue": "The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases",
"volume": "",
"issue": "",
"pages": "243--258",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "G. Levine, G. DeJong, L. Wang, R. Samdani, S. Vembu, and D. Roth. 2010. Automatic model adaptation for complex structured domains. In The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, pages 243-258.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales",
"authors": [
{
"first": "B",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2005,
"venue": "The 43rd Annual Meeting of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B. Pang and L. Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In The 43rd Annual Meet- ing of the Association for Computational Linguis- tics: Human Language Technologies, page 124.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval",
"authors": [
{
"first": "B",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "2",
"issue": "",
"pages": "1--135",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B. Pang and L. Lee. 2008. Opinion mining and senti- ment analysis. Foundations and Trends in Informa- tion Retrieval, 2(1-2):1-135.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "The bag-ofopinions method for review rating prediction from sparse text patterms",
"authors": [
{
"first": "L",
"middle": [],
"last": "Qu",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Ifrim",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Weikum",
"suffix": ""
}
],
"year": 2010,
"venue": "The International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "913--921",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "L. Qu, G. Ifrim, and G. Weikum. 2010. The bag-of- opinions method for review rating prediction from sparse text patterms. In The International Confer- ence on Computational Linguistics, pages 913-921.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Labeled lda: A supervised topic model for credit attribution in multi-labeled corpora",
"authors": [
{
"first": "D",
"middle": [],
"last": "Ramage",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Hall",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Nallapati",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2011,
"venue": "The Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Ramage, D. Hall, R. Nallapati, and C. Manning. 2011. Labeled lda: A supervised topic model for credit attribution in multi-labeled corpora. In The Conference on Empirical Methods in Natural Lan- guage Processing.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Modeling by shortest data description",
"authors": [
{
"first": "J",
"middle": [],
"last": "Rissanen",
"suffix": ""
}
],
"year": 1978,
"venue": "Automatica",
"volume": "14",
"issue": "",
"pages": "445--471",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Rissanen. 1978. Modeling by shortest data descrip- tion. Automatica, 14:445-471.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Estimating the dimension of a model",
"authors": [
{
"first": "G",
"middle": [
"E"
],
"last": "Schwarz",
"suffix": ""
}
],
"year": 1978,
"venue": "Annals of Statistics",
"volume": "6",
"issue": "2",
"pages": "461--464",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "G. E. Schwarz. 1978. Estimating the dimension of a model. Annals of Statistics, 6(2):461-464.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Large-scale cross-document coreference using distributed inference and hierarchical models",
"authors": [
{
"first": "S",
"middle": [],
"last": "Singh",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Subramanya",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Pereira",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Mccallum",
"suffix": ""
}
],
"year": 2011,
"venue": "The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Singh, A. Subramanya, F. Pereira, and A. McCallum. 2011. Large-scale cross-document coreference us- ing distributed inference and hierarchical models. In The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Tech- nologies.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Multiple aspect ranking using the good grief algorithm",
"authors": [
{
"first": "B",
"middle": [],
"last": "Snyder",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Barzilay",
"suffix": ""
}
],
"year": 2007,
"venue": "The 8th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "300--307",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B. Snyder and R. Barzilay. 2007. Multiple aspect rank- ing using the good grief algorithm. In The 8th An- nual Conference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technologies, pages 300-307.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "A bayesian model for unsupervised semantic parsing",
"authors": [
{
"first": "I",
"middle": [],
"last": "Titov",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Klementiev",
"suffix": ""
}
],
"year": 2011,
"venue": "The 49th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "I. Titov and A. Klementiev. 2011. A bayesian model for unsupervised semantic parsing. In The 49th An- nual Meeting of the Association for Computational Linguistics.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A joint model of text and aspect ratings for sentiment summarization",
"authors": [
{
"first": "I",
"middle": [],
"last": "Titov",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Mcdonald",
"suffix": ""
}
],
"year": 2008,
"venue": "The 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "308--316",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "I. Titov and R. McDonald. 2008a. A joint model of text and aspect ratings for sentiment summarization. In The 46th Annual Meeting of the Association for Computational Linguistics: Human Language Tech- nologies, pages 308-316.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Modeling online reviews with multi-grain topic models",
"authors": [
{
"first": "I",
"middle": [],
"last": "Titov",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Mcdonald",
"suffix": ""
}
],
"year": 2008,
"venue": "The Seventeenth International Conference on World Wide Web",
"volume": "",
"issue": "",
"pages": "111--120",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "I. Titov and R. McDonald. 2008b. Modeling online re- views with multi-grain topic models. In The Seven- teenth International Conference on World Wide Web, pages 111-120.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "The Nature of Statistical Learning Theory",
"authors": [
{
"first": "V",
"middle": [],
"last": "Vapnik",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "V. Vapnik. 1995. The Nature of Statistical Learning Theory. Springer-Verlag.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Simultaneous image classification and annotation",
"authors": [
{
"first": "C",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Blei",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Fei-Fei",
"suffix": ""
}
],
"year": 2009,
"venue": "IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Wang, D. Blei, and L. Fei-Fei. 2009. Simultane- ous image classification and annotation. In IEEE Conference on Computer Vision and Pattern Recog- nition.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Supervised latent dirichlet allocation for classification",
"authors": [
{
"first": "C",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Wang. 2009. Supervised latent dirichlet allocation for classification. online.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Jointly modeling aspects and opinions with a maxent-lda hybrid",
"authors": [
{
"first": "W",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Jiang",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Yan",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2010,
"venue": "The Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "56--65",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. Zhao, J. Jiang, H. Yan, and X. Li. 2010. Jointly modeling aspects and opinions with a maxent-lda hybrid. In The Conference on Empirical Methods in Natural Language Processing, pages 56-65.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Learning curves for three model sampling approaches on Yelp.com test data with 500 training examples (averaged over 20 trials).",
"uris": null,
"type_str": "figure",
"num": null
},
"FIGREF1": {
"text": "A representative partitioning tree learned from 500 training examples on the Yelp.com data.",
"uris": null,
"type_str": "figure",
"num": null
}
}
}
} |