File size: 83,624 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 | {
"paper_id": "D08-1015",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T16:30:39.966755Z"
},
"title": "Ranking Reader Emotions Using Pairwise Loss Minimization and Emotional Distribution Regression",
"authors": [
{
"first": "Kevin",
"middle": [],
"last": "Hsin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Taiwan University No",
"location": {
"addrLine": "1 Roosevelt Rd. Sec. 4",
"settlement": "Taipei",
"country": "Taiwan"
}
},
"email": ""
},
{
"first": "Yih",
"middle": [],
"last": "Lin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Taiwan University No",
"location": {
"addrLine": "1 Roosevelt Rd. Sec. 4",
"settlement": "Taipei",
"country": "Taiwan"
}
},
"email": ""
},
{
"first": "Hsin-Hsi",
"middle": [],
"last": "Chen",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Taiwan University No",
"location": {
"addrLine": "1 Roosevelt Rd. Sec. 4",
"settlement": "Taipei",
"country": "Taiwan"
}
},
"email": "hhchen@csie.ntu.edu.tw"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper presents two approaches to ranking reader emotions of documents. Past studies assign a document to a single emotion category, so their methods cannot be applied directly to the emotion ranking problem. Furthermore, whereas previous research analyzes emotions from the writer's perspective, this work examines readers' emotional states. The first approach proposed in this paper minimizes pairwise ranking errors. In the second approach, regression is used to model emotional distributions. Experiment results show that the regression method is more effective at identifying the most popular emotion, but the pairwise loss minimization method produces ranked lists of emotions that have better correlations with the correct lists.",
"pdf_parse": {
"paper_id": "D08-1015",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper presents two approaches to ranking reader emotions of documents. Past studies assign a document to a single emotion category, so their methods cannot be applied directly to the emotion ranking problem. Furthermore, whereas previous research analyzes emotions from the writer's perspective, this work examines readers' emotional states. The first approach proposed in this paper minimizes pairwise ranking errors. In the second approach, regression is used to model emotional distributions. Experiment results show that the regression method is more effective at identifying the most popular emotion, but the pairwise loss minimization method produces ranked lists of emotions that have better correlations with the correct lists.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Emotion analysis is an increasingly popular research topic due to the emergence of large-scale emotion data on the web. Previous work primarily studies emotional contents of texts from the writer's perspective, where it is typically assumed that a writer expresses only a single emotion in a document. Unfortunately, this premise does not hold when analyzing a document from the reader's perspective, because readers rarely agree unanimously on the emotion that a document instills. Figure 1 . Emotional responses of 626 people after reading a Yahoo! News article about an Iranian refugee mother and her two children who finally reunited with their family in the March of 2007 after been stranded in a Moscow airport for 10 months due to false passports.",
"cite_spans": [],
"ref_spans": [
{
"start": 483,
"end": 491,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "readers' responses are distributed among different emotion categories. In fact, none of the emotions in Figure 1 has a majority (i.e., more than 50%) of the votes. Intuitively, it is better to provide a ranking of emotions according to their popularity rather than associating a single reader emotion with a document. As a result, current writer-emotion analysis techniques for classifying a document into a single emotion category are not suitable for analyzing reader emotions. New methods capable of ranking emotions are required. Reader-emotion analysis has potential applications that differ from those of writer-emotion analysis. For example, by integrating emotion ranking into information retrieval, users will be able to retrieve documents that contain relevant contents and at the same time produce desired feelings. In addition, reader-emotion analysis can assist writers in foreseeing how their work will influence readers emotionally.",
"cite_spans": [],
"ref_spans": [
{
"start": 104,
"end": 112,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we present two approaches to ranking reader emotions. The first approach is inspired by the success of the pairwise loss minimization framework used in information retrieval to rank documents. Along a similar line, we devise a novel scheme to minimize the number of incorrectly-ordered emotion pairs in a document. In the second approach, regression is used to model reader-emotion distributions directly. Experiment results show that the regression method is more effective at identifying the most popular emotion, but the pairwise loss minimization method produces ordered lists of emotions that have better correlations with the correct lists.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The rest of this paper is organized as follows. Section 2 describes related work. In Section 3, details about the two proposed approaches are provided. Section 4 introduces the corpus and Section 5 presents how features are extracted from the corpus. Section 6 shows the experiment procedures and results. Section 7 concludes the paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Only a few studies in the past deal with the reader aspect of emotion analysis. For example, Lin et al. (2007; 2008) classify documents into readeremotion categories. Most previous work focuses on the writer's perspective. Pang et al. (2002) design an algorithm to determine whether a document's author expresses a positive or negative sentiment. They discover that using Support Vector Machines (SVM) with word unigram features results in the best performance. Since then, more work has been done to find features better than unigrams. In (Hu et al., 2005) , word sentiment information is exploited to achieve better classification accuracy.",
"cite_spans": [
{
"start": 93,
"end": 110,
"text": "Lin et al. (2007;",
"ref_id": "BIBREF7"
},
{
"start": 111,
"end": 116,
"text": "2008)",
"ref_id": "BIBREF8"
},
{
"start": 223,
"end": 241,
"text": "Pang et al. (2002)",
"ref_id": "BIBREF10"
},
{
"start": 540,
"end": 557,
"text": "(Hu et al., 2005)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Experiments have been done to extract emotional information from texts at granularities finer than documents. Wiebe (2000) investigates the subjectivity of words, whereas Aman and Szpakowicz (2007) manually label phrases with emotional categories. In 2007, the SemEval-2007 workshop organized a task on the unsupervised annotation of news headlines with emotions (Strapparava and Mihalcea, 2007) .",
"cite_spans": [
{
"start": 110,
"end": 122,
"text": "Wiebe (2000)",
"ref_id": "BIBREF14"
},
{
"start": 363,
"end": 395,
"text": "(Strapparava and Mihalcea, 2007)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "As for the task of ranking, many machinelearning algorithms have been proposed in information retrieval. These techniques generate ranking functions which predict the relevance of a document. One class of algorithms minimizes the errors resulting from ordering document pairs incorrectly. Examples include (Joachims, 2002) , (Freund et al., 2003) and (Qin et al., 2007) . In particular, the training phase of the Joachims' Ranking SVM (Joachims, 2002) is formulated as the following SVM optimization problem:",
"cite_spans": [
{
"start": 306,
"end": 322,
"text": "(Joachims, 2002)",
"ref_id": "BIBREF5"
},
{
"start": 325,
"end": 346,
"text": "(Freund et al., 2003)",
"ref_id": "BIBREF2"
},
{
"start": 351,
"end": 369,
"text": "(Qin et al., 2007)",
"ref_id": "BIBREF11"
},
{
"start": 435,
"end": 451,
"text": "(Joachims, 2002)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "min \u2211 + k j i C k j i , , T 2 1 , , \u03be \u03be w w w, subject to: 0 : 1 )) , ( ) , ( ( : | ) , ( ), , ( , , , , T , , \u2265 \u2200 \u2200 \u2200 \u2212 \u2265 \u03a6 \u2212 \u03a6 > \u2208 \u2200 k j i k j i j k i k j k i k j k i k k j i d q d q s s V d q d q \u03be \u03be w (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "where V is the training corpus, \u03a6(q k , d i ) is the feature vector of document d i with respect to query q k , s k,i is the relevance score of d i with respect to q k , w is a weight vector, C is the SVM cost parameter, and \u03be i,j,k are slack variables. The set of constraints at (1) means that document pairwise orders should be preserved. Unfortunately, the above scheme for exploiting pairwise order information cannot be applied directly to the emotion ranking task, because the task requires us to rank emotions within a document rather than provide a ranking of documents. In particular, the definitions of \u03a6",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "(q k ,d i ), \u03a6(q k ,d j )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": ", s k,i and s k,j do not apply to emotion ranking. In the next section, we will show how the pairwise loss minimization concept is adapted for emotion ranking.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "In this section, we provide the formal description of the reader-emotion ranking problem. Then we describe the pairwise loss minimization (PLM) approach and the emotional distribution regression (EDR) approach to ranking emotions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Ranking Reader Emotions",
"sec_num": "3"
},
{
"text": "The reader emotion ranking problem is defined as follows. Let D = {d 1 , d 2 , \u2026, d N } be the document space, and E = {e 1 , e 2 , \u2026, e M } be the emotion space. Let f i : E \u2192 \u211c be the emotional probability function of d i \u2208D. That is, f i (e j ) outputs the fraction of readers who experience emotion e j after reading document d i . Our goal is to find a function r : D \u2192 E M such that r(d i ) = (e \u03c0(1) , e \u03c0(2) , \u2026, e \u03c0(M) ) where \u03c0 is Input: Set of emotion ordered pairs P 1. G \u2190 a graph with emotions as vertices and no edge 2. while (P \u2260 \u2205) 3. remove (e j ,e k ) with the highest confidence from P 4. if adding edge (e j ,e k ) to G produces a loop 5. then add (e k ,e j ) to G 6. else add (e j ,e k ) to G 7. return topological sort of G a permutation on {1, 2, \u2026, M}, and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Specification",
"sec_num": "3.1"
},
{
"text": "f i (e \u03c0(1) ) \u2265 f i (e \u03c0(2) ) \u2265 \u2026 \u2265 f i (e \u03c0(M) ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Specification",
"sec_num": "3.1"
},
{
"text": "As explained in Section 2, the information retrieval framework for exploiting pairwise order information cannot be applied directly to the emotion ranking problem. Hence, we introduce a novel formulation of the emotion ranking problem into an SVM optimization problem with constraints based on pairwise loss minimization. Algorithm 1. Merge Pairwise Orders.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "We now describe how we rank the emotions of a previously unseen document using the M(M -1)/2 pairwise ranking functions g jk created during the training phase. First, all of the pairwise ranking functions are applied to the unseen document, which generates the relative orders of every pair of emotions. These pairwise orders need to be combined together to produce a ranked list of all the emotions. Algorithm 1 does exactly this.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "Whereas Ranking SVM generates only a single ranking function, our method creates a pairwise ranking function g jk : D \u2192 \u211c for each pair of emotions e j and e k , aiming at satisfying the maximum number of the inequalities:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "In Algorithm 1, the confidence of an emotion ordered pair at Line 3 is the probability value returned by a LIBSVM classifier for predicting the order. LIBSVM's method for generating this probability is described in (Wu et al., 2003) . Lines 4 and 5 resolve the problem of conflicting emotion ordered pairs forming a loop in the ordering of emotions. The ordered list of emotions returned by Algorithm 1 at Line 7 is the final output of the PLM method.",
"cite_spans": [
{
"start": 215,
"end": 232,
"text": "(Wu et al., 2003)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "\u2200d i \u2208D | f i (e j ) > f i (e k ) : g jk (d i ) > 0 \u2200d i \u2208D | f i (e j ) < f i (e k ) : g jk (d i ) < 0",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "In other words, we want to minimize the number of incorrectly-ordered emotion pairs. We further require g jk (d i ) to have the linear form",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "w T \u2126(d i ) + b,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "where w is a weight vector, b is a constant, and \u2126(d i ) is the feature vector of d i . Details of feature extraction will be presented in Section 5.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "As Joachims (2002) points out, the above type of problem is NP-Hard. However, an approximate solution to finding g ik can be obtained by solving the following SVM optimization problem:",
"cite_spans": [
{
"start": 3,
"end": 18,
"text": "Joachims (2002)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Loss Minimization",
"sec_num": "3.2"
},
{
"text": "In the second approach to ranking emotions, we use regression to model f i directly. A regression function h j : D \u2192 \u211c is generated for each e j \u2208E by learning from the examples (\u2126",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "(d i ), f i (e j )) for all documents d i in the training corpus. min \u2211 + i b C i \u03be \u03be , w w w T 2 1 , subject to: 0 : 1 ) ) ( ( : ) ( ) ( | 1 ) ( : ) ( ) ( | T T \u2265 \u2200 \u2212 \u2265 + \u2126 \u2212 < \u2208 \u2200 \u2212 \u2265 + \u2126 > \u2208 \u2200 i i i k i j i i i i k i j i i i b d e f e f Q d b d e f e f Q d \u03be \u03be \u03be w w",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "The regression framework we adopt is Support Vector Regression (SVR), which is a regression analysis technique based on SVM (Sch\u00f6lkopf et al., 2000) . We require h j to have the form",
"cite_spans": [
{
"start": 124,
"end": 148,
"text": "(Sch\u00f6lkopf et al., 2000)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "w T \u2126(d i ) + b.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "Finding h j is equivalent to solving the following optimization problem:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "where C is the SVM cost parameter, \u03be i are slack variables, and Q is the training corpus. We assume each document d i \u2208Q is labeled with f i (e j ) for every emotion e j \u2208E.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "min ) ( 2 , 1 , T 2 1 , , 2 , 1 , i i b C i i \u03be \u03be \u03be \u03be , + + \u2211 w w w subject to:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "When formulated as an SVM optimization problem, finding g jk is equivalent to training an SVM classifier for classifying a document into the e j or e k category. Hence, we use LIBSVM, which is an SVM implementation, to obtain the solution. 1 0 , : Figure 2 . News articles in the entire corpus grouped by the percentage of votes received by the most popular emotion.",
"cite_spans": [],
"ref_spans": [
{
"start": 248,
"end": 256,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": ") ( ) ) ( ( ) ) ( ( ) ( : 2 , 1 , 2 , T 1 , T \u2265 \u2200 \u2212 \u2265 \u2212 + \u2126 \u2212 \u2265 + \u2126 \u2212 \u2208 \u2200 i i i j i i i i j i i i e f b d b d e f Q d \u03be \u03be \u03be \u03b5 \u03be \u03b5 w w",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "where C is the cost parameter, \u03b5 is the maximum difference between the predicted and actual values we wish to maintain, \u03be i,1 and \u03be i,2 are slack variables, and Q is the training corpus. To solve the above optimization problem, we use SVM light 's SVR implementation. 2 When ranking the emotions of a previously unseen document d k , we sort the emotions e j \u2208E in descending order of h j (d k ).",
"cite_spans": [
{
"start": 268,
"end": 269,
"text": "2",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Emotional Distribution Regression",
"sec_num": "3.3"
},
{
"text": "The training and test corpora used in this study comprise Chinese news articles from Yahoo! Kimo News 3 , which allows a user to cast a vote for one of eight emotions to express how a news article makes her feel. Each Yahoo! news article contains a list of eight emotions at the bottom of the webpage. A reader may select one of the emotions and click on a submit button to submit the emotion. As with many websites which collect user responses, such as the Internet Movie Database, users are not forced to submit their responses. After submitting a response, the user can view a distribution of emotions indicating how other readers feel about the same article. Figure 1 shows the voting results of a Yahoo! news article.",
"cite_spans": [],
"ref_spans": [
{
"start": 663,
"end": 671,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Constructing the Corpus",
"sec_num": "4"
},
{
"text": "The eight available emotions are happy, sad, angry, surprising, boring, heartwarming, awesome, and useful. Useful is not a true emotion. Rather, it means that a news article contains practical information. The value f i (e j ) is derived by normalizing the number of votes for emotion e j in document d i by the total number votes in d i .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constructing the Corpus",
"sec_num": "4"
},
{
"text": "The entire corpus consists of 37,416 news articles dating from January 24, 2007 to August 7, 2007. News articles prior to June 1, 2007 form the training corpus (25,975 articles), and the remaining ones form the test corpus (11,441 articles). We collect articles a week after their publication dates to ensure that the vote counts have stabilized.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constructing the Corpus",
"sec_num": "4"
},
{
"text": "As mentioned earlier, readers rarely agree unanimously on the emotion of a document. Figure 2 illustrates this. In 41% of all the news articles in the entire corpus, the most popular emotion receives less than 60% of the votes.",
"cite_spans": [],
"ref_spans": [
{
"start": 85,
"end": 94,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Constructing the Corpus",
"sec_num": "4"
},
{
"text": "After obtaining news articles, the next step is to determine how to convert them into feature vectors for SVM and SVR. That is, we want to instantiate \u2126. For this purpose, three types of features are extracted.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extracting Features",
"sec_num": "5"
},
{
"text": "The first feature type consists of Chinese character bigrams, which are taken from the headline and content of each news article. The presence of a bigram is indicated by a binary feature value.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extracting Features",
"sec_num": "5"
},
{
"text": "Chinese words form the second type of features. Unlike English words, consecutive Chinese words in a sentence are not separated by spaces. To deal with this problem, we utilize Stanford NLP Group's Chinese word segmenter to split a sentence into words. 4 As in the case of bigrams, binary feature values are used.",
"cite_spans": [
{
"start": 253,
"end": 254,
"text": "4",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Extracting Features",
"sec_num": "5"
},
{
"text": "We use character bigram features in addition to word features to increase the coverage of Chinese words. A Chinese word is formed by one or more contiguous Chinese characters. As mentioned earlier, Chinese words in a sentence are not separated by any boundary symbol (e.g., a space), so a Chinese word segmentation tool is always required to extract words from a sentence. However, a word segmenter may identify word boundaries erroneously, resulting in the loss of correct Chinese words. This problem is particularly severe if there are a lot of out-of-vocabulary words in a dataset. In Chinese, around 70% of all Chinese words are Chinese character bigrams (Chen et al., 1997) . Thus, using Chinese character bigrams as features will allow us to identify a lot of Chinese words, which when combined with the words extracted by the word segmenter, will give us a wider coverage of Chinese words.",
"cite_spans": [
{
"start": 659,
"end": 678,
"text": "(Chen et al., 1997)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Extracting Features",
"sec_num": "5"
},
{
"text": "The third feature type is extracted from news metadata. A news article's metadata are its news NDCG@k is used because ACC@k has the disadvantage of not taking emotional distributions into account. Take Figure 1 as an example. In the figure, heartwarming and happy have 31.3% and 30.7% of the votes, respectively. Since the two percentages are very close, it is reasonable to say that predicting happy as the first item in a ranked list may also be acceptable. However, doing so would be completely incorrect according to ACC@k. In contrast, NDCG@k would consider it to be partially correct, and the extent of correctness depends on how much heartwarming and happy's percentages of votes differ. To be exact, if happy is predicted as the first item, then the corresponding NDCG@1 would be 30.7% / 31.3% = 0.98. category, agency, hour of publication, reporter, and event location. Examples of news categories include sports and political. Again, we use binary feature values. News metadata are used because they may contain implicit emotional information.",
"cite_spans": [],
"ref_spans": [
{
"start": 202,
"end": 210,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Extracting Features",
"sec_num": "5"
},
{
"text": "The experiments are designed to achieve the following four goals: (i) to compare the ranking performance of different methods, (ii) to analyze the pairwise ranking quality of PLM, (iii) to analyze the distribution estimation quality of EDR, and (iv) to compare the ranking performance of different feature sets. The Yahoo! News training and test corpora presented in Section 4 are used in all experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "6"
},
{
"text": "The third metric is SACC@k, or set accuracy at k. It is a variant of ACC@k. According to SACC@k, a predicted ranked list is correct if the set of its first k items is the same as the true ranked list's set of first k items. In effect, SACC@k evaluates a ranking method's ability to place the top k most important items in the first k positions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "6"
},
{
"text": "We employ three metrics as indicators of ranking quality: ACC@k, NDCG@k and SACC@k.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Metrics for Ranking",
"sec_num": "6.1"
},
{
"text": "ACC@k stands for accuracy at position k. According to ACC@k, a predicted ranked list is correct if the list's first k items are identical (i.e., same items in the same order) to the true ranked list's first k items. If two emotions in a list have the same number of votes, then their positions are interchangeable. ACC@k is computed by dividing the number of correctly-predicted instances by the total number of instances.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation Metrics for Ranking",
"sec_num": "6.1"
},
{
"text": "SVM and SVR are employed in PLM and EDR, respectively. Both SVM and SVR have the adjustable C cost parameter, and SVR has an additional \u03b5 parameter. To estimate the optimal C value for a combination of SVM and features, we perform 4fold cross-validation on the Yahoo! News training corpus, and select the C value which results in the highest binary classification accuracy during crossvalidation. The same procedure is used to estimate the best C and \u03b5 values for a combination of SVR and features. The C-\u03b5 pair which results in the lowest mean squared error during cross-validation is chosen. The candidate C values for both SVM and SVR are 2 -10 , 2 -9 , \u2026, 2 -6 . The candidate \u03b5 values for SVR are 10 -2 and 10 -1 . All cross-validations are performed solely on the training data. The test data are not used to tune the parameters. Also, SVM and SVR allow users to specify the type of kernel to use. Linear kernel is selected for both SVM and SVR. NDCG@k, or normalized discounted cumulative gain at position k (J\u00e4rvelin and Kek\u00e4l\u00e4inen, 2002) , is a metric frequently used in information retrieval to judge the quality of a ranked list when multiple levels of relevance are considered. This metric is defined as",
"cite_spans": [
{
"start": 1015,
"end": 1046,
"text": "(J\u00e4rvelin and Kek\u00e4l\u00e4inen, 2002)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Tuning SVM and SVR Parameters",
"sec_num": "6.2"
},
{
"text": "\u2211 = + = k i i k i rel z k 1 2 ) 1 ( log @ NDCG",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tuning SVM and SVR Parameters",
"sec_num": "6.2"
},
{
"text": "where rel i is the relevance score of the predicted item at position i, and z k is a normalizing factor which ensures that a correct ranked list has an NDCG@k value of 1. In the emotion ranking problem, rel i is the percentage of reader votes received by the emotion at position i. Note that the log 2 (i+1) value in the denominator is a discount factor which decreases the weights of items ranked later in a list. NDCG@k has the range [0, 1], where 1 is the best. In the experiment results, NDCG@k values are averaged over all instances in the test corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tuning SVM and SVR Parameters",
"sec_num": "6.2"
},
{
"text": "The nearest neighbor (NN) method is used as the baseline. The ranked emotion list of a news article in the test corpus is predicted as follows. First, the test news article is compared to every training news article using cosine similarity, which is defined as In Figure 3 , EDR's ACC@1 (0.751) is higher than those of PLM and NN, and the differences are statistically significant with p-value < 0.01. So, EDR is the best method at predicting the most popular emotion. However, PLM has the best ACC@k for k \u2265 2, and the differences from the other two methods are all significant with p-value < 0.01. This means that PLM's predicted ranked lists better resemble the true ranked lists. Figure 3 displays a sharp decrease in ACC@k values as k increases. This trend indicates the hardness of predicting a ranked list correctly. Looking from a different angle, the ranking task under the ACC@k metric is equivalent to the classification of news articles into one of 8!/(8 -k)! classes, where we regard each unique emotion sequence of length k as a class. In fact, computing ACC@8 for a ranking method is the same as evaluating the method's ability to classify a news article into one of 8! = 40,320 classes. So, producing a completelycorrect ranked list is a difficult task.",
"cite_spans": [],
"ref_spans": [
{
"start": 264,
"end": 272,
"text": "Figure 3",
"ref_id": null
},
{
"start": 684,
"end": 692,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Nearest Neighbor Baseline",
"sec_num": "6.3"
},
{
"text": "| | | | | | ) , ( cos i i j i j i D D D D d d \u00d7 \u2229 =",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nearest Neighbor Baseline",
"sec_num": "6.3"
},
{
"text": "In Figure 4 , all of PLM and EDR's NDCG@k improvements over NN are statistically significant with p-value < 0.01. For some values of k, the difference in NDCG@k between PLM and EDR is not significant. The high NDCG@k values (i.e., greater than 0.8) of PLM and EDR imply that although it is difficult for PLM and EDR to generate completely-correct ranked lists, these two methods are effective at placing highly popular emotions to the beginning of ranked lists.",
"cite_spans": [],
"ref_spans": [
{
"start": 3,
"end": 11,
"text": "Figure 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Nearest Neighbor Baseline",
"sec_num": "6.3"
},
{
"text": "In Figure 5 , PLM outperforms the other two methods for 2 \u2264 k \u2264 7, and the differences are all statistically significant with p-value < 0.01. For small values of k (e.g., 2 \u2264 k \u2264 3), PLM's higher SACC@k values mean that PLM is better at placing the highly popular emotions in the top positions of a ranked list.",
"cite_spans": [],
"ref_spans": [
{
"start": 3,
"end": 11,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Nearest Neighbor Baseline",
"sec_num": "6.3"
},
{
"text": "To further compare PLM and EDR, we examine their performance on individual test instances. Figure 6 shows the percentage of test instances where both PLM and EDR give incorrect lists, only PLM gives correct lists, only EDR gives ranked lists, and both methods give correct lists. The \"Only PLM Correct\" and \"Only EDR Correct\" categories are nonzero, so neither PLM nor EDR is always better than the other.",
"cite_spans": [],
"ref_spans": [
{
"start": 91,
"end": 99,
"text": "Figure 6",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Nearest Neighbor Baseline",
"sec_num": "6.3"
},
{
"text": "In summary, EDR is the best at predicting the most popular emotion according to ACC@1, NDCG@1 and SACC@1. However, PLM generates ranked lists that better resemble the correct ranked lists according to ACC@k and SACC@k for k \u2265 2. Further analysis shows that neither method is always better than the other.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Nearest Neighbor Baseline",
"sec_num": "6.3"
},
{
"text": "In this subsection, we evaluate the performance of PLM in predicting pairwise orders. We first examine the quality of ranked lists generated by PLM in terms of pairwise orders. To do this, we use Kendall's \u03c4 b correlation coefficient, which is a statistical measure for determining the correlation between two ranked lists when there may be ties between two items in a list (Liebetrau, 1983) . The value of \u03c4 b is determined based on the number of concordant pairwise orders and the number of discordant pairwise orders between two ranked lists. Therefore, this measure is appropriate for evaluating the effectiveness of PLM at predicting pairwise orders correctly. \u03c4 b has the range [-1, 1] , where 1 means a perfect positive correlation, and -1 means two lists are the reverse of each other.",
"cite_spans": [
{
"start": 374,
"end": 391,
"text": "(Liebetrau, 1983)",
"ref_id": "BIBREF6"
},
{
"start": 684,
"end": 691,
"text": "[-1, 1]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": "When computing \u03c4 b of two ranked lists, we also calculate a p-value to indicate whether the correlation is statistically significant.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": "We compute \u03c4 b statistics between a predicted ranked list and the corresponding true ranked list. Table 1 shows the results. In Table 1 , numbers in the \"Average \u03c4 b \" and \"Average p-value\" columns are averaged over all test instances. The statistics for EDR and NN are also included for comparison. From the table, we see that PLM has the highest average \u03c4 b value and the lowest average p-value, so PLM is better at preserving pairwise orders than EDR and NN methods. This observation verifies that PLM's minimization of pairwise loss leads to better prediction of pairwise orders.",
"cite_spans": [],
"ref_spans": [
{
"start": 98,
"end": 105,
"text": "Table 1",
"ref_id": "TABREF0"
},
{
"start": 128,
"end": 135,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": "We now look at the individual performance of the 28 pairwise emotion rankers g jk . As mentioned in Section 3.2, each pairwise emotion ranker g jk is equivalent to a binary classifier for classifying a document into the e j or e k category. So, we look at their classification accuracies in Table 2 . In the table, accuracy ranges from 0.75 for the awesomesurprising pair to 0.91 for the useful-boring pair.",
"cite_spans": [],
"ref_spans": [
{
"start": 291,
"end": 298,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": "From the psychological perspective, the relatively low accuracy of the awesome-surprising pair is expected, because awesome is surprising in a positive sense. So, readers should have a hard time distinguishing between these two emotions. And the SVM classifier, which models reader responses, should also find it difficult to discern these two emotions. Based on this observation, we suspect that the pairwise classification performance actually reflects the underlying emotional ambiguity experienced by readers. To verify this, we quantify the degree of ambiguity between two emotions, and compare the result to pairwise classification accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": "To quantify emotional ambiguity, we introduce the concept of discrimination value between two emotions e j and e k in a document d i , which is defined as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": ") ( ) ( ) ( ) ( k i j i k i j i e f e f e f e f + \u2212",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": "where f i is the emotional probability function defined in Section 3.1. Intuitively, the larger the discrimination value is, the smaller the degree of ambiguity between two emotions is. Figure 7 shows the relationship between pairwise classification accuracy and the average discrimination value of the corresponding emotion pair. The general pattern is that as accuracy increases, the discrimination value also increases. To provide concrete evidence, we use Pearson's product-moment correlation coefficient, which has the range of [-1, 1] , where 1 means a perfect positive correlation (Moore, 2006) . The coefficient for the data in Figure 7 is 0.726 with p-value < 0.01. Thus, pairwise emotion classification accuracy reflects the emotional ambiguity experienced by readers. In summary, PLM's pairwise loss minimization leads to better pairwise order predictions than EDR and NN. Also, the pairwise classification results reveal the inherent ambiguity between emotions.",
"cite_spans": [
{
"start": 533,
"end": 540,
"text": "[-1, 1]",
"ref_id": null
},
{
"start": 588,
"end": 601,
"text": "(Moore, 2006)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 186,
"end": 194,
"text": "Figure 7",
"ref_id": "FIGREF3"
},
{
"start": 636,
"end": 644,
"text": "Figure 7",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Pairwise Ranking Quality of PLM",
"sec_num": "6.5"
},
{
"text": "In this subsection, we evaluate EDR's performance in estimating the emotional probability function f i .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Distribution Estimation Quality of EDR",
"sec_num": "6.6"
},
{
"text": "With the prior knowledge that a news article's f i values sum to 1 over all emotions, and f i is between 0 and 1, we adjust EDR's f i predictions to produce proper distributions. It is done as follows. A predicted f i value greater than 1 or less than 0 is set to 1 and 0, respectively. Then the predicted f i values are normalized to sum to 1 over all emotions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Distribution Estimation Quality of EDR",
"sec_num": "6.6"
},
{
"text": "NN's distribution estimation performance is included for comparison. For NN, the predicted f i values of a test article are taken from the emotional distribution of the most similar training article. Figure 8 shows the mean squared error of EDR and NN for predicting f i . In the figure, the error generated by EDR is less than those by NN, and all the differences are statistically significant with pvalue < 0.01. Thus, EDR's use of regression leads to better estimation of f i than the NN. Figure 9 shows each of the three feature type's ACC@k for predicting test instances' ranked lists when PLM is used. The feature comparison graph for EDR is not shown, because it exhibits a very similar trend as PLM. For both PLM and EDR, bigrams are better than words, which are in turn better than news metadata. In Figure 9 , the combination of all three feature sets achieves the best performance. For both PLM and EDR, the improvements in ACC@k of using all features over words and metadata are all significant with p-value < 0.01, and the improvements over bigrams are significant for k \u2264 2. Hence, in general, it is better to use all three feature types together.",
"cite_spans": [],
"ref_spans": [
{
"start": 200,
"end": 208,
"text": "Figure 8",
"ref_id": null
},
{
"start": 492,
"end": 500,
"text": "Figure 9",
"ref_id": null
},
{
"start": 809,
"end": 817,
"text": "Figure 9",
"ref_id": null
}
],
"eq_spans": [],
"section": "Distribution Estimation Quality of EDR",
"sec_num": "6.6"
},
{
"text": "This paper presents two methods to ranking reader emotions. The PLM method minimizes pairwise loss, and the EDR method estimates emotional distribution through regression. Experiments with significant tests show that EDR is better at predicting the most popular emotion, but PLM produces ranked lists that have higher correlation with the correct lists. We further verify that PLM has better pairwise ranking performance than the other two methods, and EDR has better distribution estimation performance than NN.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "As for future work, there are several directions we can pursue. An observation is that PLM exploits pairwise order information, whereas EDR exploits emotional distribution information. We plan to combine these two methods together. Another research direction is to improve EDR by finding better features. We would also like to integrate emotion ranking into information retrieval.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "http://www.csie.ntu.edu.tw/~cjlin/libsvm/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://svmlight.joachims.org/ 3 http://tw.news.yahoo.com",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://nlp.stanford.edu/software/segmenter.shtml",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We are grateful to the Computer and Information Networking Center, National Taiwan University, for the support of high-performance computing facilities. The research in this paper was partially supported by National Science Council, Taiwan, under the contract NSC 96-2628-E-002-240-MY3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Percentage of Votes Received by Most Popular Emotion Number of News Articles References Saima Aman and Stan Szpakowicz",
"authors": [],
"year": 2007,
"venue": "Proceedings of 10th International Conference on Text, Speech and Dialogue",
"volume": "4629",
"issue": "",
"pages": "196--205",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Percentage of Votes Received by Most Popular Emotion Number of News Articles References Saima Aman and Stan Szpakowicz. 2007. Identifying Expressions of Emotion in Text. In Proceedings of 10th International Conference on Text, Speech and Dialogue, Lecture Notes in Computer Science 4629, 196-205. Springer, Plze\u0148, CZ.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Chinese Text Retrieval wihtout using a Dictionary",
"authors": [
{
"first": "Aitao",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Jianzhang",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Liangjie",
"middle": [],
"last": "Xu",
"suffix": ""
},
{
"first": "Frederic",
"middle": [],
"last": "Gey",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Meggs",
"suffix": ""
}
],
"year": 1997,
"venue": "Proceedings of 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval",
"volume": "",
"issue": "",
"pages": "42--49",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aitao Chen, Jianzhang He, Liangjie Xu, Frederic Gey, and Jason Meggs. 1997. Chinese Text Retrieval wihtout using a Dictionary. In Proceedings of 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 42-49. Association for Computing Machinery, Phila- delphia, US.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "An Efficient Boosting Algorithm for Combining Preferences",
"authors": [
{
"first": "Yoav",
"middle": [],
"last": "Freund",
"suffix": ""
},
{
"first": "Raj",
"middle": [
"D"
],
"last": "Iyer",
"suffix": ""
},
{
"first": "Robert",
"middle": [
"E"
],
"last": "Schapire",
"suffix": ""
},
{
"first": "Yoram",
"middle": [],
"last": "Singer",
"suffix": ""
}
],
"year": 2003,
"venue": "Journal of Machine Learning Research",
"volume": "4",
"issue": "",
"pages": "933--969",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoav Freund, Raj D. Iyer, Robert E. Schapire, and Yoram Singer. 2003. An Efficient Boosting Algorithm for Combining Preferences. Journal of Machine Learning Research, 4, 933-969.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A New Method for Sentiment Classification in Text Retrieval",
"authors": [
{
"first": "Yi",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "Jianyong",
"middle": [],
"last": "Duan",
"suffix": ""
},
{
"first": "Xiaoming",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Bingzhen",
"middle": [],
"last": "Pei",
"suffix": ""
},
{
"first": "Ruzhan",
"middle": [],
"last": "Lu",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of 2nd International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1--9",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yi Hu, Jianyong Duan, Xiaoming Chen, Bingzhen Pei, and Ruzhan Lu. 2005. A New Method for Sentiment Classification in Text Retrieval. In Proceedings of 2nd International Joint Conference on Natural Lan- guage Processing, 1-9. Jeju Island, KR.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Cumulative Gain-based Evaluation of IR Techniques",
"authors": [
{
"first": "Kalervo",
"middle": [],
"last": "J\u00e4rvelin",
"suffix": ""
},
{
"first": "Jaana",
"middle": [],
"last": "Kek\u00e4l\u00e4inen",
"suffix": ""
}
],
"year": 2002,
"venue": "ACM Transactions on Information Systems",
"volume": "20",
"issue": "4",
"pages": "422--446",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kalervo J\u00e4rvelin and Jaana Kek\u00e4l\u00e4inen. Cumulative Gain-based Evaluation of IR Techniques. 2002. ACM Transactions on Information Systems, 20(4), 422-446.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Optimizing Search Engines using Clickthrough Data",
"authors": [
{
"first": "Thorsten",
"middle": [],
"last": "Joachims",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thorsten Joachims. 2002. Optimizing Search Engines using Clickthrough Data. In Proceedings of 8th ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining. Association for Computing Machinery, Edmonton, CA.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Measures of Association",
"authors": [
{
"first": "Albert",
"middle": [
"M"
],
"last": "Liebetrau",
"suffix": ""
}
],
"year": 1983,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Albert M. Liebetrau. 1983. Measures of Association. Sage Publications, Newbury Park, US.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "What Emotions do News Articles Trigger in their Readers?",
"authors": [
{
"first": "Kevin",
"middle": [
"H"
],
"last": "Lin",
"suffix": ""
},
{
"first": "Changhua",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Hsin-Hsi",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of 30th ACM SIGIR Conference",
"volume": "",
"issue": "",
"pages": "733--734",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kevin H. Lin, Changhua Yang, and Hsin-Hsi Chen. 2007. What Emotions do News Articles Trigger in their Readers? In Proceedings of 30th ACM SIGIR Conference, 733-734. Association for Computing Machinery, Amsterdam, NL.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Emotion Classification of Online News Articles from the Reader's Perspective",
"authors": [
{
"first": "Kevin",
"middle": [
"H"
],
"last": "Lin",
"suffix": ""
},
{
"first": "Changhua",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Hsin-Hsi",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of International Conference on Web Intelligence. Institute of Electrical and Electronics Engineers",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kevin H. Lin, Changhua Yang, and Hsin-Hsi Chen. 2008. Emotion Classification of Online News Articles from the Reader's Perspective. In Proceedings of In- ternational Conference on Web Intelligence. Institute of Electrical and Electronics Engineers, Sydney, AU.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The Basic Practice of Statistics",
"authors": [
{
"first": "David",
"middle": [],
"last": "Moore",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Moore. 2006. The Basic Practice of Statistics. W.H. Freeman and Company, New York, US.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Thumbs up? Sentiment Classification Using Machine Learning Techniques",
"authors": [
{
"first": "Bo",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "Lillian",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Shivakumar",
"middle": [],
"last": "Vaithyanathan",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of 2002 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "79--86",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? Sentiment Classification Using Machine Learning Techniques. In Proceedings of 2002 Conference on Empirical Methods in Natural Language Processing, 79-86. Association for Com- putational Linguistics, Philadelphia, US.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Ranking with Multiple Hyperplanes",
"authors": [
{
"first": "Tao",
"middle": [],
"last": "Qin",
"suffix": ""
},
{
"first": "Tie-Yan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Lai",
"suffix": ""
},
{
"first": "Xu-Dong",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "De-Sheng",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Hang",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of 30 th ACM SIGIR Conference",
"volume": "",
"issue": "",
"pages": "279--286",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tao Qin, Tie-Yan Liu, Wei Lai, Xu-Dong Zhang, De- Sheng Wang, and Hang Li. 2007. Ranking with Mul- tiple Hyperplanes. In Proceedings of 30 th ACM SIGIR Conference, 279-286. Association for Com- puting Machinery, Amsterdam, NL.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "New Support Vector Algorithms",
"authors": [
{
"first": "Bernhard",
"middle": [],
"last": "Sch\u00f6lkopf",
"suffix": ""
},
{
"first": "Alex",
"middle": [
"J"
],
"last": "Smola",
"suffix": ""
},
{
"first": "Robert",
"middle": [
"C"
],
"last": "Williamson",
"suffix": ""
},
{
"first": "Peter",
"middle": [
"L"
],
"last": "Barlett",
"suffix": ""
}
],
"year": 2000,
"venue": "Neural Computation",
"volume": "12",
"issue": "5",
"pages": "1207--1245",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bernhard Sch\u00f6lkopf, Alex J. Smola, Robert C. William- son, and Peter L. Barlett. 2000. New Support Vector Algorithms. Neural Computation, 12(5), 1207-1245.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "SemEval-2007 Task 14: Affective Text",
"authors": [
{
"first": "Carlo",
"middle": [],
"last": "Strapparava",
"suffix": ""
},
{
"first": "Rada",
"middle": [],
"last": "Mihalcea",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of 4th International Workshop on Semantic Evaluations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Carlo Strapparava and Rada Mihalcea. 2007. SemEval- 2007 Task 14: Affective Text. In Proceedings of 4th International Workshop on Semantic Evaluations. Prague, CZ.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Learning Subjective Adjectives from Corpora",
"authors": [
{
"first": "M",
"middle": [],
"last": "Janyce",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wiebe",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of 17th Conference of the American Association for Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "735--740",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Janyce M. Wiebe. 2000. Learning Subjective Adjectives from Corpora. In Proceedings of 17th Conference of the American Association for Artificial Intelligence, 735-740. AAAI Press, Austin, US.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Probability Estimates for Multi-class Classification by Pairwise Coupling",
"authors": [
{
"first": "",
"middle": [],
"last": "Ting-Fan",
"suffix": ""
},
{
"first": "Chih-Jen",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Ruby",
"middle": [
"C"
],
"last": "Lin",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Weng",
"suffix": ""
}
],
"year": 2004,
"venue": "Journal of Machine Learning Research",
"volume": "5",
"issue": "",
"pages": "975--1005",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ting-Fan Wu, Chih-Jen Lin, and Ruby C. Weng. Prob- ability Estimates for Multi-class Classification by Pairwise Coupling. 2004. Journal of Machine Learn- ing Research, 5, 975-1005.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "Figure 1illustrates this phenomenon. In the figure,",
"uris": null,
"type_str": "figure"
},
"FIGREF1": {
"num": null,
"text": "Performance of PLM and EDR.",
"uris": null,
"type_str": "figure"
},
"FIGREF2": {
"num": null,
"text": "where d i and d j are two news articles, and D i and D j are sets of Chinese character bigrams in d i and d j , respectively. The ranked emotion list of the training article having the highest cosine similarity with the test article is used as the predicted ranked list. Figures 3 to 5 show the performance of different ranking methods on the test corpus. For both PLM and EDR, all of the bigram, word, and news metadata features are used.",
"uris": null,
"type_str": "figure"
},
"FIGREF3": {
"num": null,
"text": "Accuracy of pairwise emotion classification and the corresponding average discrimination value.",
"uris": null,
"type_str": "figure"
},
"FIGREF4": {
"num": null,
"text": "Mean squared error of NN and EDR for estimating the emotional distributions of the test corpus. PLM performance using different features.",
"uris": null,
"type_str": "figure"
},
"TABREF0": {
"num": null,
"html": null,
"text": "Kendall's \u03c4 b statistics.",
"type_str": "table",
"content": "<table><tr><td>Method</td><td/><td/><td/><td>Average \u03c4 b</td><td/><td/><td colspan=\"2\">Average p-value</td></tr><tr><td>PLM</td><td/><td/><td/><td>0.584</td><td/><td/><td/><td>0.068</td></tr><tr><td>EDR</td><td/><td/><td/><td>0.474</td><td/><td/><td/><td>0.114</td></tr><tr><td>NN</td><td/><td/><td/><td>0.392</td><td/><td/><td/><td>0.155</td></tr><tr><td/><td/><td>He</td><td>Su</td><td>Sa</td><td>Us</td><td>Ha</td><td>Bo</td><td>An</td></tr><tr><td>Aw</td><td colspan=\"2\">0.80</td><td>0.75</td><td>0.78</td><td>0.77</td><td>0.82</td><td>0.76</td><td>0.79</td></tr><tr><td>He</td><td/><td/><td>0.79</td><td>0.81</td><td>0.78</td><td>0.81</td><td>0.89</td><td>0.81</td></tr><tr><td>Su</td><td/><td/><td/><td>0.82</td><td>0.78</td><td>0.80</td><td>0.82</td><td>0.82</td></tr><tr><td>Sa</td><td/><td/><td/><td/><td>0.78</td><td>0.80</td><td>0.84</td><td>0.82</td></tr><tr><td>Us</td><td/><td/><td/><td/><td/><td>0.82</td><td>0.91</td><td>0.82</td></tr><tr><td>Ha</td><td/><td/><td/><td/><td/><td/><td>0.83</td><td>0.79</td></tr><tr><td>Bo</td><td/><td/><td/><td/><td/><td/><td/><td>0.80</td></tr><tr><td colspan=\"9\">Table 2. Classification accuracies of SVM pairwise</td></tr><tr><td colspan=\"9\">emotion classifiers on the test corpus. He = heartwarm-</td></tr><tr><td colspan=\"9\">ing, Su = surprising, Sa = sad, Us = useful, Ha = happy,</td></tr><tr><td colspan=\"5\">Bo = boring, and An = angry.</td><td/><td/><td/></tr><tr><td>Average Discrimination</td><td>Value of Emotion Pair</td><td>0.58 0.63 0.68 0.73</td><td/><td/><td/><td/><td/></tr><tr><td/><td/><td>0.53</td><td/><td/><td/><td/><td/></tr><tr><td/><td/><td>0.75</td><td/><td>0.8</td><td>0.85</td><td/><td>0.9</td></tr><tr><td/><td/><td/><td colspan=\"5\">Accuracy of Pairwise Emotion Classification</td></tr></table>"
}
}
}
} |