File size: 78,992 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 | {
"paper_id": "D08-1013",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T16:31:07.412320Z"
},
"title": "Adding Redundant Features for CRFs-based Sentence Sentiment Classification",
"authors": [
{
"first": "Jun",
"middle": [],
"last": "Zhao",
"suffix": "",
"affiliation": {
"laboratory": "National Laboratory of Pattern Recognition",
"institution": "Chinese Academy of Sciences",
"location": {
"postCode": "100190",
"settlement": "Beijing",
"country": "China"
}
},
"email": "jzhao@nlpr.ia.ac.cn"
},
{
"first": "Kang",
"middle": [],
"last": "Liu",
"suffix": "",
"affiliation": {
"laboratory": "National Laboratory of Pattern Recognition",
"institution": "Chinese Academy of Sciences",
"location": {
"postCode": "100190",
"settlement": "Beijing",
"country": "China"
}
},
"email": "kliu@nlpr.ia.ac.cn"
},
{
"first": "Gen",
"middle": [],
"last": "Wang",
"suffix": "",
"affiliation": {
"laboratory": "National Laboratory of Pattern Recognition",
"institution": "Chinese Academy of Sciences",
"location": {
"postCode": "100190",
"settlement": "Beijing",
"country": "China"
}
},
"email": "gwang@nlpr.ia.ac.cn"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In this paper, we present a novel method based on CRFs in response to the two special characteristics of \"contextual dependency\" and \"label redundancy\" in sentence sentiment classification. We try to capture the contextual constraints on sentence sentiment using CRFs. Through introducing redundant labels into the original sentimental label set and organizing all labels into a hierarchy, our method can add redundant features into training for capturing the label redundancy. The experimental results prove that our method outperforms the traditional methods like NB, SVM, MaxEnt and standard chain CRFs. In comparison with the cascaded model, our method can effectively alleviate the error propagation among different layers and obtain better performance in each layer.",
"pdf_parse": {
"paper_id": "D08-1013",
"_pdf_hash": "",
"abstract": [
{
"text": "In this paper, we present a novel method based on CRFs in response to the two special characteristics of \"contextual dependency\" and \"label redundancy\" in sentence sentiment classification. We try to capture the contextual constraints on sentence sentiment using CRFs. Through introducing redundant labels into the original sentimental label set and organizing all labels into a hierarchy, our method can add redundant features into training for capturing the label redundancy. The experimental results prove that our method outperforms the traditional methods like NB, SVM, MaxEnt and standard chain CRFs. In comparison with the cascaded model, our method can effectively alleviate the error propagation among different layers and obtain better performance in each layer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "There are a lot of subjective texts in the web, such as product reviews, movie reviews, news, editorials and blogs, etc. Extracting these subjective texts and analyzing their orientations play significant roles in many applications such as electronic commercial, etc. One of the most important tasks in this field is sentiment * Contact: Jun ZHAO, jzhao@nlpr.ia.ac.cn classification, which can be performed in several levels: word level, sentence level, passage level, etc. This paper focuses on sentence level sentiment classification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction *",
"sec_num": "1"
},
{
"text": "Commonly, sentiment classification contains three layers of sub-tasks. From upper to lower, (1) Subjective/Objective classification: the subjective texts are extracted from the corpus teeming with both subjective and objective texts. (2) Polarity classification: a subjective text is classified into \"positive\" or \"negative\" according to the sentimental expressions in the text. (3) Sentimental strength rating: a subjective text is classified into several grades which reflect the polarity degree of \"positive\" or \"negative\". It is a special multi-class classification problem, where the classes are ordered. In machine learning, this kind of problem is also regarded as an ordinal regression problem (Wei Wu et al. 2005) . In this paper, we mainly focus on this problem in sentiment classification.",
"cite_spans": [
{
"start": 702,
"end": 722,
"text": "(Wei Wu et al. 2005)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction *",
"sec_num": "1"
},
{
"text": "Sentiment classification in sentence level has its special characteristics compared with traditional text classification tasks. Firstly, the sentiment of each sentence in a discourse is not independent to each other. In other words, the sentiment of each sentence is related to those of other adjacent sentences in the same discourse. The sentiment of a sentence may vary in different contexts. If we detach a sentence from the context, its sentiment may not be inferred correctly. Secondly, there is redundancy among the sentiment classes, especially in sentimental strength classes. For example:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction *",
"sec_num": "1"
},
{
"text": "\"I love the scenario of \"No country for old man\" very much!!\" \"This movie sounds good.\" The first sentence is labeled as \"highly praised\" class and the second one is labeled as \"something good\" class. Both the sentences express positive sentiment for the movie, but the former expresses stronger emotion than the latter. We can see that both \"highly praised\" and \"something good\" belong to an implicit class \"positive\", which can be regarded as the relation between them. If we add these implicit classes in the label set, the sentiment classes will form a hierarchical structure. For example, \"positive\" can be regarded as the parent class of \"highly praised\" and \"something good\", \"subjective\" can be regarded as the parent class of \"positive\" and \"negative\". This implicit hierarchical structure among labels should not be neglected because it may be beneficial for improving the accuracy of sentiment classification. In the paper, we call this characteristic of sentiment classification as \"label redundancy\". Unfortunately, in our knowledge most of the current research treats sentiment classification as a traditional multi-classification task or an ordinal regression task, which regard the sentimental classes being independent to each other and each sentence is also independent to the adjacent sentences in the context. In other words, they neglect the contextual information and the redundancy among sentiment classes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction *",
"sec_num": "1"
},
{
"text": "In order to consider the contextual information in the process of the sentence sentiment classification, some research defines contextual features and some uses special graph-based formulation, like (Bo Pang, et al. 2005) . In order to consider the label redundancy, one potential solution is to use a cascaded framework which can combine subjective/objective classification, polarity classification and sentimental strength classification together, where the classification results of the preceding step will be the input of the subsequent one. However, the subsequent classification cannot provide constraint and correction to the results of the preceding step, which will lead to the accumulation and propagation of the classification errors. As a result, the performance of sentiment analysis of sentences is often not satisfactory. This paper focuses on the above two special characteristics of the sentiment classification problem in the sentence level. To the first characteristic, we regard the sentiment classification as a sequence labeling problem and use conditional random field (CRFs) model to capture the relation between two adjacent sentences in the context. To the second characteristic, we propose a novel method based on a CRF model, in which the original task is mapped to a classification on a hierarchical structure, which is formed by the original label set and some additional implicit labels. In the hierarchical classification framework, the relations between the labels can be represented as the additional features in classification. Because these features are related to the original labels but unobserved, we name them as \"redundant features\" in this paper. They can be used to capture the redundant and hierarchical relation between different sentiment classes. In this way, not only the performance of sentimental strength rating is improved, the accuracies of subjective/objective classification and polarity classification are also improved compared with the traditional sentiment classification method. And in comparison with the cascaded method, the proposed approach can effectively alleviate error propagation. The experimental results on movie reviews prove the validity of our method.",
"cite_spans": [
{
"start": 199,
"end": 221,
"text": "(Bo Pang, et al. 2005)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction *",
"sec_num": "1"
},
{
"text": "For capturing the influence of the contexts to the sentiment of a sentence, we treat original sentiment classification as a sequence labeling problem. We regard the sentiments of all the sentences throughout a paragraph as a sequential flow of sentiments, and we model it using a conditional model. In this paper, we choose Conditional Random Fields (CRFs) (Lafferty et al, 2001) because it has better performance than other sequence labeling tools in most NLP applications. CRFs are undirected graphical models used to calculate the conditional probability of a set of labels given a set of input variables. We cite the definitions of CRFs in (Lafferty et al, 2001) . It defines the conditional probability proportional to the product of potential functions on cliques of the graph, exp ( , )",
"cite_spans": [
{
"start": 357,
"end": 379,
"text": "(Lafferty et al, 2001)",
"ref_id": "BIBREF10"
},
{
"start": 644,
"end": 666,
"text": "(Lafferty et al, 2001)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Contextual Influence for Sentiment Classification",
"sec_num": "2"
},
{
"text": "( | ) ( ) F Y X P Y X Z X \u03bb \u03bb \u22c5 = (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Contextual Influence for Sentiment Classification",
"sec_num": "2"
},
{
"text": "where X is a set of input random variables and Y is a set of random labels. ( , ) F Y X is an arbitrary feature function over its arguments, \u03bb is a learned weight for each feature function and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Contextual Influence for Sentiment Classification",
"sec_num": "2"
},
{
"text": "( ) exp( ( , )) y X Z F Y X \u03bb = \u22c5 \u2211 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Contextual Influence for Sentiment Classification",
"sec_num": "2"
},
{
"text": "The training of CRFs is based on Maximum Likelihood Principle (Fei Sha et al. 2003) . The log likelihood function is",
"cite_spans": [
{
"start": 62,
"end": 83,
"text": "(Fei Sha et al. 2003)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Contextual Influence for Sentiment Classification",
"sec_num": "2"
},
{
"text": "[ ] ( ) ( , ) log ( ) k k k k L F Y X Z X \u03bb \u03bb \u03bb = \u22c5 \u2212 \u2211",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Contextual Influence for Sentiment Classification",
"sec_num": "2"
},
{
"text": "Therefore, Limited-memory BFGS (L-BFGS) algorithm is used to find this nonlinear optimization parameters.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Contextual Influence for Sentiment Classification",
"sec_num": "2"
},
{
"text": "In this section, we explain the \"label redundancy\" in sentiment classification mentioned in the first section. We will analyze the effect of the label redundancy on the performance of sentiment classification from the experimental view. We conduct the experiments of polarity classification and sentimental strength rating on the corpus which will be introduced in section 5 later. The class set is also illustrated in that section.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Label Redundancy in Sentiment Classification",
"sec_num": "3"
},
{
"text": "Polarity classification is a three-class classification process, and sentimental strength rating is a fiveclass classification process. We use first 200 reviews as the training set which contains 6,079 sentences, and other 49 reviews, totally 1,531 sentences, are used as the testing set. Both the three-class classification and the five-class classification use standard CRFs model with the same feature set. The results are shown in Table 1 , 2 and 3, where \"Answer\" denotes the results given by human, \"Results\" denotes the results given by \uff0c CRFs model \"Correct\" denotes the number of correct samples which is labeled by CRFs model. We use precision, recall and F1 value as the evaluation metrics. Table 1 gives the result of sentimental strength rating. Table 2 shows the polarity classification results extracted from the results of sentimental strength rating in Table 1 . The extraction process is as follows. In the sentimental strength rating results, we combine the sentences with \"PP\" class and the sentences with \"P\" class into \"Pos\" class, and the sentences with \"NN\" class and the sentences with \"N\" class into \"Neg\" class. So the results of five-class classification are transformed into the results of three-class classification. Table 3 is the results of performing polarity classification in the data set by CRFs directly. Table 3 .",
"cite_spans": [],
"ref_spans": [
{
"start": 435,
"end": 442,
"text": "Table 1",
"ref_id": null
},
{
"start": 702,
"end": 709,
"text": "Table 1",
"ref_id": null
},
{
"start": 759,
"end": 766,
"text": "Table 2",
"ref_id": "TABREF0"
},
{
"start": 870,
"end": 877,
"text": "Table 1",
"ref_id": null
},
{
"start": 1247,
"end": 1254,
"text": "Table 3",
"ref_id": null
},
{
"start": 1342,
"end": 1349,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Label Redundancy in Sentiment Classification",
"sec_num": "3"
},
{
"text": "From the results we can find the following phenomena.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Result of Polarity Classification",
"sec_num": null
},
{
"text": "(1) The corpus is severely unbalanced, the objective sentences take the absolute majority in the corpus, which leads to the poor accuracy for classifying subjective sentences. The experiment in Table 1 puts polarity classification and sentimental strength rating under a unique CRFs model, without considering the redundancy and hierarchical structure between different classes. As a result, the features for polarity classification will usually cover the features for sentimental strength rating. These reasons can explain why there is only one sample labeled as \"NN\" correctly and only 5 samples labeled as \"PP\" correctly.",
"cite_spans": [],
"ref_spans": [
{
"start": 194,
"end": 201,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Result of Polarity Classification",
"sec_num": null
},
{
"text": "(2) Comparing Table 2 with 3, we can find that, the F1 value of the polarity classification results extracted from sentimental strength rating results is lower than that of directly conducting polarity classification. That is because the redundancy between sentimental strength labels makes the classifier confused to determine the polarity of the sentence. Therefore, we should deal with the sentiment analysis in a hierarchical frame which can consider the redundancy between the different classes and make full use of the subjective and polarity information implicitly contained in sentimental strength classes.",
"cite_spans": [],
"ref_spans": [
{
"start": 14,
"end": 21,
"text": "Table 2",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Result of Polarity Classification",
"sec_num": null
},
{
"text": "As mentioned above, it's important for a classifier to consider the redundancy between different labels. However, from the standard CRFs described in formula (1), we can see that the training of CRFs only maximizes the probabilities of the observed labels Y in the training corpus. Actually, the redundant relation between sentiment labels is unobserved. The standard CRFs still treats each class as an isolated item so that its performance is not satisfied. In this section, we propose a novel method for sentiment classification, which can capture the redundant relation between sentiment labels through adding redundant features. In the following, we firstly show how to add these redundant features, then illustrate the characteristics of this method. After that, for the sentiment analysis task, the process of feature generation will be presented.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Capturing Label Redundancy for CRFs via Adding Redundant Features",
"sec_num": "4"
},
{
"text": "Adding redundant features has two steps. Firstly, an implicit redundant label set is designed, which can form a multi-layer hierarchical structure together with the original labels. Secondly, in the hierarchical classification framework, the implicit labels, which reflect the relations between the original labels, can be used as redundant features in the training process. We will use the following example to illustrate the first step for sentimental strength rating task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "For the task of sentimental strength rating, the original label set is {\"PP (highly praised)\", \"P (something good)\", \"Neu (objective description)\", \"N (something that needs improvement)\" and \"NN (strong aversion)\"}. In order to introduce redundant labels, the 5-class classification task is decomposed into the following three layers shown in Figure 1 . The label set in the first layer is {\"subjective\", \"objective\"}, The label set in the second layer is for polarity classification {\"positive\", \"objective\", \"negative\"}, and the label set in the third layer is the original set. Actually, the labels in the first and second layers are unobserved redundant labels, which will not be reflected in the final classification result obviously. sentimental labels In the second step, with these redundant labels, some implicit features can be generated for CRFs. So the standard CRFs can be rewritten as follows. T Y Y Y Y = , which contains implicit labels in the hierarchical structure shown in Figure 1 . The difference between our method and the standard chain CRFs is that we make some implicit redundant features to be active when training. The original feature function",
"cite_spans": [],
"ref_spans": [
{
"start": 343,
"end": 351,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 992,
"end": 1001,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "exp( ( , ) ) ( | ) ( ) exp( ( , ) ) exp( ( , ) ) T m j j j j m j j j T j F X T P T X Z X F X Y F X Y \u03bb \u03bb \u03bb = = \u22c5 = \u22c5 = \u22c5 \u2211 \u2211 \u2211 (2)",
"eq_num": "where 1"
}
],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "( , ) F Y X is replaced by 1 ( , ) m j j j F X Y = \u2211",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": ". We use an example to illustrate the process of feature generation. When a sentence including the word \"good\" is labeled as \"PP\", our model not only generate the state feature (good, \"PP\"), but also two implicit redundant state feature (good, \"positive\") and (good, \"subjective\"). Through adding larger-granularity labels \"positive\" and \"negative\" into the model, our method can increase the probability of \"positive\" and decrease the probability of \"negative\". Furthermore, \"P\" and \"PP\" will share the probability gain of \"positive\", therefore the probability of \"P\" will be larger than that of \"N\". For the transition feature, the same strategy is used. Therefore the complexity of its",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "training procedure is ( ) m j j O M N F l \u00d7 \u00d7 \u00d7 \u2211 where M",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "is the number of the training samples, N is the average sentence length, j F is the average number of activated features in the j th layer, l is the number of the original labels and m is the number of the layers. For the complexity of the decoding procedure, our method has (",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": ") m j j O N F l \u00d7 \u00d7 \u2211 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "It's worth noting that, (1) transition features are extracted in each layer separately rather than across different layers. For example, feature (good, \"subjective\", \"positive\") will never be extracted because \"subjective\" and \"positive\" are from different layers;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "(2) if one sentence is labeled as \"Neu\", no implicit redundant features will be generated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adding Redundant Features for CRFs",
"sec_num": "4.1"
},
{
"text": "Our method allows that the label sets are dependent and redundant. As a result, it can improve the performance of not only the classifier for the original sentimental strength rating task, but also the classifiers for other tasks in the hierarchical frame, i.e. polarity classification and subjective/objective classification. This kind of dependency and redundancy can lead to two characteristics of the proposed method for sentiment classification compared with traditional methods, such as the cascaded method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Characteristics of Our Method",
"sec_num": "4.2"
},
{
"text": "(1) Error-correction: Two dependent tasks in the neighboring layers can correct the errors of each other relying on the inconsistent redundant information. For example, if in the first layer, the features activated by \"objective\" get larger scores than the features activated by \"subjective\", and in the second layer the features activated by \"positive\" get larger scores than the features activated by \"objective\", then inconsistency emerges. At this time, our method can globally select the label with maximum probability. This characteristic can make up the deficiency of the cascaded method which may induce error propagation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Characteristics of Our Method",
"sec_num": "4.2"
},
{
"text": "(2) Differentiating the ordinal relation among sentiment labels: Our method organizes the ordinal sentiment labels into a hierarchy through introducing redundant labels into standard chain CRFs, in this way the degree of classification errors can be controlled. In the different layers of sentiment analysis task, the granularities of classification are different. Therefore, when an observation cannot be correctly labeled on a smaller-granularity label set, our method will use the larger-granularity labels in the upper layer to control the final classification labels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Characteristics of Our Method",
"sec_num": "4.2"
},
{
"text": "For feature selection, our method selects different features for each layer in the hierarchical frame.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Selection in Different Layers",
"sec_num": "4.3"
},
{
"text": "In the top layer of the frame shown in Figure 1 , for subjective/objective classification task, we use not only adjectives and the verbs which contain subjective information (e.g., \"believe\", \"think\") as the features, but also the topic words. The topic words are defined as the nouns or noun phases which frequently appear in the corpus. We believe that some topic words contain subjective information.",
"cite_spans": [],
"ref_spans": [
{
"start": 39,
"end": 47,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Feature Selection in Different Layers",
"sec_num": "4.3"
},
{
"text": "In the middle and bottom layers, we not only use the features in the first layer, but also some special features as follows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Selection in Different Layers",
"sec_num": "4.3"
},
{
"text": "(1) The prior orientation scores of the sentiment words: Firstly, a sentiment lexicon is generated by extending the synonymies and antonyms in WordNet 2 from a positive and negative seed list. Then, the positive score and the negative score of a sentiment word are individually accumulated and weighted according to the polarity of its synonymies and antonyms. At last we scale the normalized distance of the two scores into 5 levels, which will be the prior orientation of the word. When there is a negative word, like {not, no, can't, merely, never, \u2026}, occurring nearby the feature word in the range of 3 words size window, the orientation of this word will be reversed and \"NO\" will be added in front of the original feature word for creating a new feature word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Selection in Different Layers",
"sec_num": "4.3"
},
{
"text": "(2) Sentence transition features: We consider two types of sentence transition features. The first type is the conjunctions and the adverbs occurring in the beginning of this sentence. These conjunctions and adverbs are included in a word list which is manually selected, like {and, or, but, though, however, generally, contrarily, \u2026}. The second type of the sentence transition feature is the position of the sentence in one review. The reason lies in that: the reviewers often follow some writing patterns, for example some reviewers prefer to concede an opposite factor before expressing his/her real sentiment. Therefore, we divide a review into five parts, and assign each sentence with the serial number of the part which the sentence belongs to.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Selection in Different Layers",
"sec_num": "4.3"
},
{
"text": "In order to evaluate the performance of our method, we conducted experiments on a sentence level annotation corpus obtained from Purdue University, which is also used in (Mao and Lebanon 07) . This corpus contains 249 movie reviews and 7,610 sentences totally, which is randomly selected from the Cornell sentence polarity dataset v1.0. Each sentence was hand-labeled with one of five classes: PP (highly praised), P (something good), Neu (objective description), N (something that needs improvement) and NN (strong aversion), which contained the orientation polarity of each sentence. Based on the 5-class manually labeled results mentioned above, we also assigned each sentence with one of three classes: Pos (positive polarity), Neu (objective description), Neg (negative polarity). Data statistics for the corpus are given in Table 4 . Table 4 . Data Statistics for Movies Reviews Corpus There is a problem in the dataset that more than 70% of the sentences are labeled as \"Neu\" and labels are seriously unbalanced. As a result, the \"Neu\" label is over-emphasized. For this problem, Mao and Lebanon (2007) made a balanced data set (equal number sentences for different labels) which is sampled in the original corpus. Since randomly sampling sentences from the original corpus will break the intrinsic relationship between two adjacent sentences in the context, we don't create balanced label data set.",
"cite_spans": [
{
"start": 170,
"end": 190,
"text": "(Mao and Lebanon 07)",
"ref_id": null
},
{
"start": 1087,
"end": 1109,
"text": "Mao and Lebanon (2007)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 830,
"end": 837,
"text": "Table 4",
"ref_id": null
},
{
"start": 840,
"end": 847,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data and Baselines",
"sec_num": "5.1"
},
{
"text": "For the evaluation of our method, we choose accuracy as the evaluation metrics and some classical methods as the baselines. They are Na\u00efve Bayes (NB), Support Vector Machine (SVM), Maximum Entropy (MaxEnt) (Kamal Nigam et al. 1999) and standard chain CRFs (Fei et al. 2003) . We also regard cascaded-CRFs as our baseline for comparing our method with the cascaded-based method. For NB, we use Laplace smoothing method. For SVM, we use the LibSVM 3 with a linear kernel function 4 . For MaxEnt, we use the implementation in the toolkit Mallet 5 . For CRFs, Table 7 . The accuracy of Subjective/Objective Classification we use the implementation in Flex-CRFs 6 . We set the iteration number to 120 in the training process of the method based on CRFs. In the cascaded model we set 3 layers for sentimental strength rating, where the first layer is subjective/objective classification, the second layer is polarity classification and the last layer is sentimental strength classification. The upper layer passes the results as the input to the next layer.",
"cite_spans": [
{
"start": 206,
"end": 231,
"text": "(Kamal Nigam et al. 1999)",
"ref_id": "BIBREF9"
},
{
"start": 251,
"end": 273,
"text": "CRFs (Fei et al. 2003)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 556,
"end": 563,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data and Baselines",
"sec_num": "5.1"
},
{
"text": "In the first experiment, we evaluate the performance of our method for sentimental strength rating. Experimental results for each method are given in Table 5 . We not only give the overall accuracy of each method, but also the performance for each sentimental strength label. All baselines use the same feature space mentioned in section 4.3, which combine all the features in the three layers together, except cascaded CRFs and our method. In cascaded-CRFs and our method, we use different features in different layers mentioned in section 4.3. These results were gathered using 5-fold cross validation with one fold for testing and the other 4 folds for training. From the results, we can obtain the following conclusions. (1) The three versions of CRFs perform consistently better than Na\u00efve Bayes, SVM and MaxEnt methods. We think that is because CRFs model considers the contextual influence of each sentence. (2) Comparing the performance of cascaded CRFs with that of standard sequence CRFs, we can see that not only the overall accuracy but also the accuracy for each sentimental strength label are improved, where the overall accuracy is increased by 3%. It proves that taking the hierarchical relationship between labels into account is very essential for sentiment classification. The reason is that: the cascaded model performs sentimental strength rating in three hierarchical layers, while standard chain CRFs model treats each label as an independent individual. So the performance of the cascaded model is superior to the standard chain CRFs. 3The experimental results also show that our method performs better than the Cascaded CRFs. The classification accuracy is improved from 71.53% to 75.21%. We think that is because our method adds the label redundancy among the sentimental strength labels into consideration through adding redundant features into the feature sets, and the three subtasks in the cascaded model are merged into a unified model. So the output result is a global optimal result. In this way, the problem of error propagation in the cascaded frame can be alleviated.",
"cite_spans": [],
"ref_spans": [
{
"start": 150,
"end": 157,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Sentimental Strength Rating",
"sec_num": "5.2"
},
{
"text": "In the second experiment, we evaluate the performance of our method for sentiment polarity classification. Our method is based on a hierarchical frame, which can perform different tasks in different layers at the same time. For example, it can determine the polarity of sentences when sentimental strength rating is performed. Here, the polarity classification results of our method are extracted from the results of the sentimental strength rating mentioned above. In the sentimental strength rating results, we combine the sentences with PP label and the sentences with P label into one set, and the sentences with NN label and the sentences with N label into one set. So the results of 5-class classification are transformed into the results of 3-class classification. Other methods like NB, SVM, MaxEnt, standard chain CRFs perform 3-class classification directly, and their label sets in the training corpus is {Pos, Neu, Neg}. The parameter setting is the same as sentimental strength rating. For the cascaded-CRFs method, we firstly perform subjective/objective classification, and then determine the polarity of the sentences based on the subjective sentences. The experimental results are given in Table 6 .",
"cite_spans": [],
"ref_spans": [
{
"start": 1207,
"end": 1214,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Sentiment Polarity Classification",
"sec_num": "5.3"
},
{
"text": "From the experimental results, we can obtain the following conclusion for sentiment polarity classification, which is similar to the conclusion for sentimental strength rating mentioned in section 5.2. That is both our model and the cascaded model can get better performance than other traditional methods, such as NB, SVM, MaxEnt, etc. But the performance of the cascaded CRFs (76.94%) is lower than that of our method (78.55%). This indicates that because our method exploits the label redundancy in the different layers, it can increase the accuracies of both polarity classification and sentimental strength rating at the same time compared with other methods.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Polarity Classification",
"sec_num": "5.3"
},
{
"text": "In the last experiment, we test our method for subjective/objective classification. The subjective/objective label of the data is extracted from its original label like section 5.3. As the same as the experiment for polarity classification, all baselines perform subjective/objective classification directly. It's no need to perform the cascaded-based method because it's a 2-class task. The results of our method are extracted from the results of the sentimental strength rating too. The results are shown in Table 7 . From it, we can obtain the similar conclusion, i.e. our method outperforms other methods and has the 80.18% classification accuracy. Our method, which introduces redundant features into training, can increase the accuracies of all tasks in the different layers at the same time compared with other baselines. It proves that considering label redundancy are effective for promoting the performance of a sentimental classifier.",
"cite_spans": [],
"ref_spans": [
{
"start": 510,
"end": 517,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Subjective/Objective Classification",
"sec_num": "5.4"
},
{
"text": "Recently, many researchers have devoted into the problem of the sentiment classification. Most of researchers focus on how to extract useful textual features (lexical, syntactic, punctuation, etc.) for determining the semantic orientation of the sentences using machine learning algorithm (Bo et al. 2002; Kim and Hovy, 2004; Bo et al. 2005 , Hu et al. 2004 Alina et al 2008; Alistair et al 2006) . But fewer researchers deal with this problem using CRFs model.",
"cite_spans": [
{
"start": 289,
"end": 305,
"text": "(Bo et al. 2002;",
"ref_id": "BIBREF2"
},
{
"start": 306,
"end": 325,
"text": "Kim and Hovy, 2004;",
"ref_id": "BIBREF8"
},
{
"start": 326,
"end": 340,
"text": "Bo et al. 2005",
"ref_id": "BIBREF4"
},
{
"start": 341,
"end": 357,
"text": ", Hu et al. 2004",
"ref_id": null
},
{
"start": 358,
"end": 375,
"text": "Alina et al 2008;",
"ref_id": "BIBREF0"
},
{
"start": 376,
"end": 396,
"text": "Alistair et al 2006)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "6"
},
{
"text": "For identifying the subjective sentences, there are several research, like (Wiebe et al, 2005) . For polarity classification on sentence level, (Kim and Hovy, 2004) judged the sentiment by classifying a pseudo document composed of synonyms of indicators in one sentence. (Pang and Lee, 04) proposed a semi-supervised machine learning method based on subjectivity detection and minimum-cut in graph.",
"cite_spans": [
{
"start": 75,
"end": 94,
"text": "(Wiebe et al, 2005)",
"ref_id": null
},
{
"start": 144,
"end": 164,
"text": "(Kim and Hovy, 2004)",
"ref_id": "BIBREF8"
},
{
"start": 271,
"end": 285,
"text": "(Pang and Lee,",
"ref_id": null
},
{
"start": 286,
"end": 289,
"text": "04)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "6"
},
{
"text": "Cascaded models for sentiment classification were studied by (Pang and Lee, 2005) . Their work mainly used the cascaded frame for determining the orientation of a document and the sentences. In that work, an initial model is used to determine the orientation of each sentence firstly, then the top subjective sentences are input into a documentlevel model to determine the document's orientation.",
"cite_spans": [
{
"start": 61,
"end": 81,
"text": "(Pang and Lee, 2005)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "6"
},
{
"text": "The CRFs has previously been used for sentiment classification. Those methods based on CRFs are related to our work. (Mao et al, 2007) used a sequential CRFs regression model to measure the polarity of a sentence in order to determine the sentiment flow of the authors in reviews. However, this method must manually select a word set for constraints, where each selected word achieved the highest correlation with the sentiment. The performance of isotonic CRFs is strongly related to the selected word set. (McDonald et al 2007; Ivan et al 2008) proposed a structured model based on CRFs for jointly classifying the sentiment of text at varying levels of granularity. They put the sentence level and document level sentiment analysis in an integrated model and employ the orientation of the document to influence the decision of sentence's orientation. Both the above two methods didn't consider the redundant and hierarchical relation between sentimental strength labels. So their methods cannot get better results for the problem mentioned in this paper.",
"cite_spans": [
{
"start": 117,
"end": 134,
"text": "(Mao et al, 2007)",
"ref_id": "BIBREF15"
},
{
"start": 508,
"end": 529,
"text": "(McDonald et al 2007;",
"ref_id": null
},
{
"start": 530,
"end": 546,
"text": "Ivan et al 2008)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "6"
},
{
"text": "Another solution to this problem is to use a joint multi-layer model, such as dynamic CRFs, multilayer CRFs, etc. Such kind of models can treat the three sub-tasks in sentiment classification as a multi-task problem and can use a multi-layer or hierarchical undirected graphic to model the sentiment of sentences. The main difference between our method and theirs is that we consider the problem from the feature representation view. Our method expands the feature set according to the number of layers in the hierarchical frame. So the complexity of its decoding procedure is lower than theirs, for example the complexity of the multi-layer CRFs is ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "6"
},
{
"text": ") j j F O N l \u00d7 \u00d7 \u2211 ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "6"
},
{
"text": "where N is the average sentence length, j F is the average number of activated features in the j th layer, l is the number of the original labels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Works",
"sec_num": "6"
},
{
"text": "In the paper, we propose a novel method for sentiment classification based on CRFs in response to the two special characteristics of \"contextual dependency\" and \"label redundancy\" in sentence sentiment classification. We try to capture the contextual constraints on the sentence sentiment using CRFs. For capturing the label redundancy among sentiment classes, we generate a hierarchical framework through introducing redundant labels, under which redundant features can be introduced. The experimental results prove that our method outperforms the traditional methods (like NB, SVM, ME and standard chain CRFs). In comparison with cascaded CRFs, our method can effectively alleviate error propagation among different layers and obtain better performance in each layer.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "7"
},
{
"text": "For our future work, we will explore other hierarchical models for sentimental strength rating because the experiments presented in this paper prove this hierarchical frame is effective for ordinal regression. We would expand the idea in this paper into other models, such as Semi-CRFs and Hierarchical-CRFs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "7"
},
{
"text": "http://wordnet.princeton.edu/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://www.csie.ntu.tw/~cjlin/libsvm 4 http://svmlight.joachims.org/ 5 http://mallet.cs.umass.edu/index.php/Main_Page",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://flexcrfs.sourceforge.net",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "When Specialists and Generalists Work Together: Overcoming Domain Dependence in Sentiment Tagging",
"authors": [
{
"first": "Alina",
"middle": [
"A"
],
"last": "Sabine",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alina A. and Sabine B. 2008. When Specialists and Generalists Work Together: Overcoming Domain Dependence in Sentiment Tagging. In Proc. of ACL- 08",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Sentiment Classification of Movie Reviews Using Contextual Valence Shifters",
"authors": [
{
"first": "Alistair",
"middle": [],
"last": "Kennedy",
"suffix": ""
},
{
"first": "Diana",
"middle": [],
"last": "Inkpen",
"suffix": ""
}
],
"year": 2006,
"venue": "Computational Intelligence",
"volume": "22",
"issue": "2",
"pages": "110--125",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alistair Kennedy and Diana Inkpen. 2006. Sentiment Classification of Movie Reviews Using Contextual Valence Shifters. Computational Intelligence, 22(2), pages 110-125",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"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 EMNLP 2002",
"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 EMNLP 2002, pp.79-86.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts",
"authors": [
{
"first": "Bo",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "Lillian",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of ACL 2004",
"volume": "",
"issue": "",
"pages": "271--278",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of ACL 2004, pp.271-278.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales",
"authors": [
{
"first": "Bo",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "Lillian",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of ACL 2005",
"volume": "",
"issue": "",
"pages": "115--124",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of ACL 2005, pp.115-124.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A Joint Model o f Text and Aspect Ratings of Sentiment Summarization",
"authors": [
{
"first": "Ivan",
"middle": [],
"last": "Titov",
"suffix": ""
},
{
"first": "Ryan",
"middle": [],
"last": "Mcdonald",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08",
"volume": "",
"issue": "",
"pages": "308--316",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ivan Titov and Ryan McDonald. 2008. A Joint Model o f Text and Aspect Ratings of Sentiment Summarization. In Proceedings of ACL-08, pages 308-316",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Annotating expressions of opinions and emotions in lauguage. Language Resources and Evaluation",
"authors": [
{
"first": "Janyce",
"middle": [],
"last": "Webie",
"suffix": ""
},
{
"first": "Theresa",
"middle": [],
"last": "Wilson",
"suffix": ""
},
{
"first": "Claire",
"middle": [],
"last": "Cardie",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Janyce Webie, Theresa Wilson and Claire Cardie. 2005. Annotating expressions of opinions and emotions in lauguage. Language Resources and Evaluation 2005",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Shallow Parsing with Conditional Random Fields",
"authors": [
{
"first": "Fei",
"middle": [],
"last": "Sha",
"suffix": ""
},
{
"first": "Fernando",
"middle": [],
"last": "Pereira",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings ofHLT-NAACL 2003",
"volume": "",
"issue": "",
"pages": "213--220",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fei Sha and Fernando Pereira, 2003 Shallow Parsing with Conditional Random Fields, In Proceedings ofHLT-NAACL 2003, Edmonton, Canada, pp. 213- 220.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Determining the Sentiment of Opinions",
"authors": [
{
"first": "Edward",
"middle": [
"H"
],
"last": "Kim",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of COLING-04",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kim, S and Edward H. Hovy. 2004. Determining the Sentiment of Opinions. In Proceedings of COLING- 04.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Using Maximum Entropy for Text Classification",
"authors": [
{
"first": "Kamal",
"middle": [],
"last": "Nigam",
"suffix": ""
},
{
"first": "John",
"middle": [],
"last": "Lafferty",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Mccallum",
"suffix": ""
}
],
"year": 1999,
"venue": "Proceedings of IJCAI Workshop on Machine Learning for Information Filtering",
"volume": "",
"issue": "",
"pages": "61--67",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kamal Nigam, John Lafferty and Andrew McCallum. 1999. Using Maximum Entropy for Text Classification. In Proceedings of IJCAI Workshop on Machine Learning for Information Filtering, pages 61-67.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
"authors": [
{
"first": "J",
"middle": [],
"last": "Lafferty",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Mccallum",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Pereira",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of ICML-01",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J Lafferty, A McCallum, F Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of ICML-01, pages 282.289.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Movie review mining and summarization",
"authors": [
{
"first": "L",
"middle": [],
"last": "Zhuang",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Jing",
"suffix": ""
},
{
"first": "X",
"middle": [
"Y"
],
"last": "Zhu",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 15 th ACM international conference on Information and knowledge management (CIKM)",
"volume": "",
"issue": "",
"pages": "43--50",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "L. Zhuang, F. Jing, and X.Y. Zhu. 2006. Movie review mining and summarization. In Proceedings of the 15 th ACM international conference on Information and knowledge management (CIKM), pages 43-50.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Mining and summarizing customer reviews",
"authors": [
{
"first": "M",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 2004 ACM SIGKDD international conference on Knowledge discovery and data mining",
"volume": "",
"issue": "",
"pages": "168--177",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Hu and B. Liu. 2004a. Mining and summarizing customer reviews. In Proceedings of the 2004 ACM SIGKDD international conference on Knowledge discovery and data mining, pages 168-177.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Structured Models for Fine-to-Coarse Sentiment Analysis",
"authors": [
{
"first": "Ryan",
"middle": [],
"last": "Mcdonald",
"suffix": ""
},
{
"first": "Kerry",
"middle": [],
"last": "Hannan",
"suffix": ""
},
{
"first": "Tyler",
"middle": [],
"last": "Neylon",
"suffix": ""
}
],
"year": null,
"venue": "Proceedings of ACL 2007",
"volume": "",
"issue": "",
"pages": "432--439",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ryan McDonald, Kerry Hannan and Tyler Neylon et al. Structured Models for Fine-to-Coarse Sentiment Analysis. In Proceedings of ACL 2007, pp. 432-439.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Gaussian Processes for Oridinal Regression",
"authors": [
{
"first": "Wei",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Zoubin",
"middle": [],
"last": "Ghahraman",
"suffix": ""
}
],
"year": 2005,
"venue": "The Journal of Machine learning Research",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wei Wu, Zoubin Ghahraman, 2005. Gaussian Processes for Oridinal Regression. The Journal of Machine learning Research, 2005",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Isotonic Conditional Random Fields and Local Sentiment Flow",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Mao",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Lebanon",
"suffix": ""
}
],
"year": 2007,
"venue": "Advances in Neural Information Processing Systems",
"volume": "19",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Mao and G. Lebanon, 2007. Isotonic Conditional Random Fields and Local Sentiment Flow. Advances in Neural Information Processing Systems 19, 2007",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "The hierarchical structure of \u065b",
"type_str": "figure",
"num": null
},
"FIGREF1": {
"uris": null,
"text": "when decoding and our method only has (",
"type_str": "figure",
"num": null
},
"TABREF0": {
"text": "Result of Polarity Classification Extracted fromTable 1.",
"html": null,
"content": "<table><tr><td colspan=\"6\">Label Answer Results Correct Precision Recall</td><td>F1</td></tr><tr><td>PP</td><td>51</td><td>67</td><td>5</td><td>0.0746</td><td colspan=\"2\">0.0980 0.0847</td></tr><tr><td>P</td><td>166</td><td>177</td><td>32</td><td>0.1808</td><td colspan=\"2\">0.1928 0.1866</td></tr><tr><td>Neu</td><td>1190</td><td>1118</td><td>968</td><td>0.8658</td><td colspan=\"2\">0.81.34 0.8388</td></tr><tr><td>N</td><td>105</td><td>140</td><td>25</td><td>0.1786</td><td colspan=\"2\">0.2381 0.2041</td></tr><tr><td>NN</td><td>19</td><td>29</td><td>1</td><td>0.0345</td><td colspan=\"2\">0.0526 0.0417</td></tr><tr><td>Total</td><td>1531</td><td>1531</td><td>1031</td><td>0.67.34</td><td colspan=\"2\">0.6734 0.6734</td></tr><tr><td/><td colspan=\"5\">Table 1. Result of Sentimental Strength Rating</td></tr><tr><td colspan=\"6\">Label Answer Results Correct Precision Recall</td><td>F1</td></tr><tr><td>Pos</td><td>217</td><td>244</td><td>79</td><td>0.3238</td><td colspan=\"2\">0.3641 0.3427</td></tr><tr><td>Neu</td><td>1190</td><td>1118</td><td>968</td><td>0.8658</td><td colspan=\"2\">0.8134 0.8388</td></tr><tr><td>Neg</td><td>124</td><td>169</td><td>41</td><td>0.2426</td><td colspan=\"2\">0.3306 0.2799</td></tr><tr><td>Total</td><td>1531</td><td>1531</td><td>1088</td><td>0.7106</td><td colspan=\"2\">0.7106 0.7106</td></tr><tr><td colspan=\"6\">Label Answer Results Correct Precision Recall</td><td>F1</td></tr><tr><td>Pos</td><td>217</td><td>300</td><td>108</td><td>0.3600</td><td colspan=\"2\">0.4977 0.4178</td></tr><tr><td>Neu</td><td>1190</td><td>1101</td><td>971</td><td>0.8819</td><td colspan=\"2\">0.8160 0.8477</td></tr><tr><td>Neg</td><td>124</td><td>130</td><td>40</td><td>0.3077</td><td colspan=\"2\">0.3226 0.3150</td></tr><tr><td>Total</td><td>1531</td><td>1531</td><td>1119</td><td>0.7309</td><td colspan=\"2\">0.7309 0.7309</td></tr></table>",
"type_str": "table",
"num": null
},
"TABREF5": {
"text": "",
"html": null,
"content": "<table><tr><td/><td/><td colspan=\"2\">Results of Polarity Classification</td></tr><tr><td>Label</td><td>NB</td><td colspan=\"3\">SVM MaxEnt Standard CRF Our Method</td></tr><tr><td colspan=\"3\">Subjective 0.4743 0.5847 0.4872</td><td>0.5594</td><td>0.6764</td></tr><tr><td colspan=\"3\">Objective 0.8170 0.8248 0.8212</td><td>0.8312</td><td>0.8269</td></tr><tr><td>Total</td><td colspan=\"2\">0.7238 0.7536 0.7518</td><td>0.7561</td><td>0.8018</td></tr></table>",
"type_str": "table",
"num": null
}
}
}
} |