File size: 67,945 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 |
{
"paper_id": "I08-1040",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:41:55.438634Z"
},
"title": "Unsupervised Classification of Sentiment and Objectivity in Chinese Text",
"authors": [
{
"first": "Taras",
"middle": [],
"last": "Zagibalov",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Sussex",
"location": {
"postCode": "BN1 9QH",
"settlement": "Brighton",
"country": "UK"
}
},
"email": "t.zagibalov@sussex.ac.uk"
},
{
"first": "John",
"middle": [],
"last": "Carroll",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Sussex",
"location": {
"postCode": "BN1 9QH",
"settlement": "Brighton",
"country": "UK"
}
},
"email": "j.a.carroll@sussex.ac.uk"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We address the problem of sentiment and objectivity classification of product reviews in Chinese. Our approach is distinctive in that it treats both positive / negative sentiment and subjectivity / objectivity not as distinct classes but rather as a continuum; we argue that this is desirable from the perspective of would-be customers who read the reviews. We use novel unsupervised techniques, including a one-word 'seed' vocabulary and iterative retraining for sentiment processing, and a criterion of 'sentiment density' for determining the extent to which a document is opinionated. The classifier achieves up to 87% F-measure for sentiment polarity detection.",
"pdf_parse": {
"paper_id": "I08-1040",
"_pdf_hash": "",
"abstract": [
{
"text": "We address the problem of sentiment and objectivity classification of product reviews in Chinese. Our approach is distinctive in that it treats both positive / negative sentiment and subjectivity / objectivity not as distinct classes but rather as a continuum; we argue that this is desirable from the perspective of would-be customers who read the reviews. We use novel unsupervised techniques, including a one-word 'seed' vocabulary and iterative retraining for sentiment processing, and a criterion of 'sentiment density' for determining the extent to which a document is opinionated. The classifier achieves up to 87% F-measure for sentiment polarity detection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Automatic classification of sentiment has been a focus of a number of recent research efforts (e.g. (Turney, 2002; Pang et al., 2002; Dave at al., 2003) . An important potential application of such work is in business intelligence: brands and company image are valuable property, so organizations want to know how they are viewed by the media (what the 'spin' is on news stories, and editorials), business analysts (as expressed in stock market reports), customers (for example on product review sites) and their own employees. Another important application is to help people find out others' views about products they have purchased (e.g. consumer electronics), services and entertainment (e.g. movies), stocks and shares (from investor bulletin boards), and so on. In the work reported in this paper we focus on product reviews, with the intended users of the processing being would-be customers.",
"cite_spans": [
{
"start": 100,
"end": 114,
"text": "(Turney, 2002;",
"ref_id": "BIBREF14"
},
{
"start": 115,
"end": 133,
"text": "Pang et al., 2002;",
"ref_id": "BIBREF8"
},
{
"start": 134,
"end": 152,
"text": "Dave at al., 2003)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our approach is based on the insight that positive and negative sentiments are extreme points in a continuum of sentiment, and that intermediate points in this continuum are of potential interest. For instance, in one scenario, someone might want to get an idea of the types of things people are saying about a particular product through reading a sample of reviews covering the spectrum from highly positive, through balanced, to highly negative. (We call a review balanced if it is an opinionated text with an undecided or weak sentiment direction). In another scenario, a would-be customer might only be interested in reading balanced reviews, since they often present more reasoned arguments with fewer unsupported claims. Such a person might therefore want to avoid reviews such as Example (1) -written by a Chinese purchaser of a mobile phone (our English gloss).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The software is bad, some sent SMS are never received by the addressee; compatibility is also bad, on some mobile phones the received messages are in a scrambled encoding! And sometimes the phone 'dies'! Photos are horrible! It doesn't have a cyclic or pro-grammable alarm-clock, you have to set it every time, how cumbersome! The back cover does not fit! The original software has many holes!",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In a third scenario, someone might decide they would like only to read opinionated, weakly negative reviews such as Example (2), since these often contain good argumentation while still identifying the most salient bad aspects of a product.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The response time of this mobile is very long, MMS should be less than 30kb only to be downloaded, also it doesn't support MP3 ring tones, (while) the built-in tunes are not good, and from time to time it 'dies', but when I was buying it I really liked it: very original, very nicely matching red and white colours, it has its individuality, also it's not expensive, but when used it always causes trouble, makes one's head ache",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The review contains both positive and negative sentiment covering different aspects of the product, and the fact that it contains a balance of views means that it is likely to be useful for a would-be customer. Moving beyond review classification, more advanced tasks such as automatic summarization of reviews (e.g. Feiguina & LaPalme, 2007) might also benefit from techniques which could distinguish more shades of sentiment than just a binary positive / negative distinction.",
"cite_spans": [
{
"start": 317,
"end": 342,
"text": "Feiguina & LaPalme, 2007)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A second dimension, orthogonal to positive / negative, is opinionated / unopinionated (or equivalently subjective / objective). When shopping for a product, one might be interested in the physical characteristics of the product or what features the product has, rather than opinions about how well these features work or about how well the product as a whole functions. Thus, if one is looking for a review that contains more factual information than opinion, one might be interested in reviews like Example (3).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(3) (My) overall feeling about this mobile is not bad, it features: 5 alarm-clocks that switch the phone on (off), phone book for 800 items (500 people), lunar and solar calendars, fast switching between time and date modes, WAP networking, organizer,notebook and so on.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This review is mostly neutral (unopinionated), but contains information that could be useful to a would-be customer which might not be in a product specification document, e.g. fast switching between different operating modes. Similarly, wouldbe customers might be interested in retrieving completely unopinionated documents such as technical descriptions and user manuals. Again, as with sentiment classification, we argue that opinionated and unopinionated texts are not easily distinguishable separate sets, but form a continuum. In this continuum, intermediate points are of interest as well as the extremes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A major obstacle for automatic classification of sentiment and objectivity is lack of training data, which limits the applicability of approaches based on supervised machine learning. With the rapid growth in textual data and the emergence of new domains of knowledge it is virtually impossible to maintain corpora of tagged data that cover all -or even most -areas of interest. The cost of manual tagging also adds to the problem. Reusing the same corpus for training classifiers for new domains is also not effective: several studies report decreased accuracy in cross-domain classification (Engstr\u00f6m, 2004; Aue & Gamon, 2005) a similar problem has also been observed in classification of documents created over different time periods (Read, 2005) .",
"cite_spans": [
{
"start": 593,
"end": 609,
"text": "(Engstr\u00f6m, 2004;",
"ref_id": "BIBREF3"
},
{
"start": 610,
"end": 628,
"text": "Aue & Gamon, 2005)",
"ref_id": "BIBREF1"
},
{
"start": 737,
"end": 749,
"text": "(Read, 2005)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper we describe an unsupervised classification technique which is able to build its own sentiment vocabulary starting from a very small seed vocabulary, using iterative retraining to enlarge the vocabulary. In order to avoid problems of domain dependence, the vocabulary is built using text from the same source as the text which is to be classified. In this paper we work with Chinese, but using a very small seed vocabulary may mean that this approach would in principle need very little linguistic adjustment to be applied to a different language. Written Chinese has some specific features, one of which is the absence of explicitly marked word boundaries, which makes word-based processing problematic. In keeping with our unsupervised, knowledge-poor approach, we do not use any preliminary word segmentation tools or higher level grammatical analysis.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The paper is structured as follows. Section 2 reviews related work in sentiment classification and more generally in unsupervised training of classifiers. Section 3 describes our datasets, and Section 4 the techniques we use for unsupervised classification and iterative retraining. Sections 5 and 6 describe a number of experiments into how well the approaches work, and Section 7 concludes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Most previous work on the problem of categorizing opinionated texts has focused on the binary classification of positive and negative sentiment (Turney, 2002; Pang et al., 2002; Dave at al., 2003) . However, Pang & Lee (2005) describe an approach closer to ours in which they determine an author's evaluation with respect to a multi-point scale, similar to the 'five-star' sentiment scale widely used on review sites. However, authors of reviews are inconsistent in assigning fine-grained ratings and quite often star systems are not consistent between critics. This makes their approach very author-dependent. The main differences are that Pang and Lee use discrete classes (although more than two), not a continuum as in our approach, and use supervised machine learning rather than unsupervised techniques. A similar approach was adopted by Hagedorn et al. (2007) , applied to news stories: they defined five classes encoding sentiment intensity and trained their classifier on a manually tagged training corpus. They note that world knowledge is necessary for accurate classification in such open-ended domains.",
"cite_spans": [
{
"start": 144,
"end": 158,
"text": "(Turney, 2002;",
"ref_id": "BIBREF14"
},
{
"start": 159,
"end": 177,
"text": "Pang et al., 2002;",
"ref_id": "BIBREF8"
},
{
"start": 178,
"end": 196,
"text": "Dave at al., 2003)",
"ref_id": "BIBREF2"
},
{
"start": 208,
"end": 225,
"text": "Pang & Lee (2005)",
"ref_id": "BIBREF11"
},
{
"start": 844,
"end": 866,
"text": "Hagedorn et al. (2007)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "2.1"
},
{
"text": "There has also been previous work on determining whether a given text is factual or expresses opinion (Yu & Hatzivassiloglu, 2003; Pang & Lee, 2004) ; again this work uses a binary distinction, and supervised rather than unsupervised approaches.",
"cite_spans": [
{
"start": 102,
"end": 130,
"text": "(Yu & Hatzivassiloglu, 2003;",
"ref_id": null
},
{
"start": 131,
"end": 148,
"text": "Pang & Lee, 2004)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "2.1"
},
{
"text": "Recent work on classification of terms with respect to opinion (Esuli & Sebastiani, 2006 ) uses a three-category system to characterize the opinionrelated properties of word meanings, assigning numerical scores to Positive, Negative and Objective categories. The visualization of these scores somewhat resembles our graphs in Section 5, although we use two orthogonal scales rather than three categories; we are also concerned with classification of documents rather than terms.",
"cite_spans": [
{
"start": 63,
"end": 88,
"text": "(Esuli & Sebastiani, 2006",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "2.1"
},
{
"text": "Abney 2002compares two major kinds of unsupervised approach to classification (co-training and the Yarowsky algorithm). As we do not use multiple classifiers our approach is quite far from cotraining. But it is close to the paradigm described by Yarowsky (1995) and Turney (2002) as it also employs self-training based on a relatively small seed data set which is incrementally enlarged with unlabelled samples. But our approach does not use point-wise mutual information. Instead we use relative frequencies of newly found features in a training subcorpus produced by the previous iteration of the classifier. We also use the smallest possible seed vocabulary, containing just a single word; however there are no restrictions regarding the maximum number of items in the seed vocabulary.",
"cite_spans": [
{
"start": 246,
"end": 261,
"text": "Yarowsky (1995)",
"ref_id": "BIBREF15"
},
{
"start": 266,
"end": 279,
"text": "Turney (2002)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Unsupervised Classification",
"sec_num": "2.2"
},
{
"text": "Our approach starts out with a seed vocabulary consisting of a single word, (good). This word is tagged as a positive vocabulary item; initially there are no negative items. The choice of word was arbitrary, and other words with strongly positive or negative meaning would also be plausible seeds. Indeed, might not be the best possible seed, as it is relatively ambiguous: in some contexts it means to like or acts as the adverbial very, and is often used as part of other words (although usually contributing a positive meaning). But since it is one of the most frequent units in the Chinese language, it is likely to occur in a relatively large number of reviews, which is important for the rapid growth of the vocabulary list.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Seed Vocabulary",
"sec_num": "3.1"
},
{
"text": "Our test corpus is derived from product reviews harvested from the website IT168 1 . All the reviews were tagged by their authors as either positive or negative overall. Most reviews consist of two or three distinct parts: positive opinions, negative opinions, and comments ('other') -although some reviews have only one part. We removed duplicate reviews automatically using approximate matching, giving a corpus of 29531 reviews of which 23122 are positive (78%) and 6409 are negative (22%). The total number of different products in the corpus is 10631, the number of product categories is 255, and most of the reviewed products are either software products or consumer electronics. Unfortunately, it appears that some users misused the sentiment tagging facility on the website so quite a lot of reviews have incorrect tags. However, the parts of the reviews are much more reliably identified as being positive or negative so we used these as the items of the test corpus. In the experiments described in this paper we used 2317 reviews of mobile phones of which 1158 are negative and 1159 are positive. Thus random choice would have approximately 50% accuracy if all items were tagged either as negative or positive 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Test Corpus",
"sec_num": "3.2"
},
{
"text": "As discussed in Section 1, we do not carry out any word segmentation or grammatical processing of input documents. We use a very broad notion of words (or phrases) in the Chinese language. The basic units of processing are 'lexical items', each of which is a sequence of one or more Chinese characters excluding punctuation marks (which may actually form part of a word, a whole word or a sequence of words), and`zones', each of which is a sequence of characters delimited by punctuation marks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "4.1"
},
{
"text": "Each zone is classified as either positive or negative based whether positive or negative vocabulary items predominate. In more detail, a simple maximum match algorithm is used to find all lexical items (character sequences) in the zone that are in the vocabulary list. As there are two parts of the vocabulary (positive and negative), we correspondingly calculate two scores using Equation 1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "4.1"
},
{
"text": "3 , S i = L d L phrase S d N d (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "4.1"
},
{
"text": "where Ld is the length in characters of a matching lexical item, Lphrase is the length of the current zone in characters, Sd is the current sentiment score of the matching lexical item (initially 1.0), and Nd is a negation check coefficient. The negation check is a regular expression which determines if the lexical item is preceded by a negation within its enclosing zone. If a negation is found then Nd is set to -1. The check looks for six frequently occurring negations:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "4.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "(bu), (buhui), (meiyou), (baituo),",
"eq_num": "(mianqu)"
}
],
"section": "Sentiment Classification",
"sec_num": "4.1"
},
{
"text": ", and (bimian). The sentiment score of a zone is the sum of sentiment scores of all the items found in it. In fact there are two competing sentiment scores for every zone: one positive (the sum of all scores of items found in the positive part of the vocabulary list) and one negative (the sum of the scores for the items in the negative part). The sentiment direction of a zone is determined from the maximum of the absolute values of the two competing scores for the zone.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "4.1"
},
{
"text": "This procedure is applied to all zones in a document, classifying each zone as positive, negative, or neither (in cases where there are no positive or negative vocabulary items in the zone). To determine the sentiment direction of the whole document, the classifier computes the difference between the number of positive and negative zones. If the result is greater than zero the document is classified as positive, and vice versa. If the result is zero the document is balanced or neutral for sentiment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentiment Classification",
"sec_num": "4.1"
},
{
"text": "The task of iterative retraining is to enlarge the initial seed vocabulary (consisting of a single word as discussed in Section 3.1) into a comprehensive vocabulary list of sentiment-bearing lexical items. In each iteration, the current version of the classifier is run on the product review corpus to classify each document, resulting in a training subcorpus of positive and a negative documents. The subcorpus is used to adjust the scores of existing positive and negative vocabulary items and to find new items to be included in the vocabulary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iterative Retraining",
"sec_num": "4.2"
},
{
"text": "Each lexical item that occurs at least twice in the corpus is a candidate for inclusion in the vocabulary list. After candidate items are found, the system calculates their relative frequencies in both the positive and negative parts of the current training subcorpus. The system also checks for negation while counting occurrences: if a lexical item is preceded by a negation, its count is reduced by one. This results in negative counts (and thus negative relative frequencies and scores) for those items that are usually used with negation; for example, (the quality is far too bad) is in the positive part of the vocabulary with a score of -1.70. This means that the item was found in reviews classified by the system as positive but it was preceded by a negation. If during classification this item is found in a document it will reduce the positive score for that document (as it is in the positive part of the vocabulary), unless the item is preceded by a negation. In this situation the score will be reversed (multiplied by -1), and the positive score will be increased -see Equation 1above.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iterative Retraining",
"sec_num": "4.2"
},
{
"text": "For all candidate items we compare their relative frequencies in the positive and negative documents in the subcorpus using Equation 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iterative Retraining",
"sec_num": "4.2"
},
{
"text": "difference= | F p \u2212 F n | \ue09e F p \ue083 F n \ue09f/2 (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iterative Retraining",
"sec_num": "4.2"
},
{
"text": "If difference < 1, then the frequencies are similar and the item does not have enough distinguishing power, so it is not included in the vocabulary. Otherwise the the sentiment score of the item is (re-) calculated -according to Equation (3) for positive items, and analogously for negative items.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iterative Retraining",
"sec_num": "4.2"
},
{
"text": "F p F p \ue083F n (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iterative Retraining",
"sec_num": "4.2"
},
{
"text": "Finally, the adjusted vocabulary list with the new scores is ready for the next iteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Iterative Retraining",
"sec_num": "4.2"
},
{
"text": "Given a sentiment classification for each zone in a document, we compute sentiment density as the proportion of opinionated zones with respect to the total number of zones in the document. Sentiment density measures the proportion of opinionated text in a document, and thus the degree to which the document as a whole is opinionated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Objectivity Classification",
"sec_num": "4.3"
},
{
"text": "It should be noted that neither sentiment score nor sentiment density are absolute values, but are relative and only valid for comparing one document with other. Thus, a sentiment density of 0.5 does not mean that the review is half opinionated, half not. It means that the review is less opinionated than a review with density 0.9.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Objectivity Classification",
"sec_num": "4.3"
},
{
"text": "We ran the system on the product review corpus (Section 3.2) for 20 iterations. The results for bina-ry sentiment classification are shown in Table 1 . We see increasing F-measure up to iteration 18, after which both precision and recall start to descrease; we therefore use the version of the classifier as it stood after iteration 18 4 . These figures are only indicative of the classification accuracy of the system. Accuracy might be lower for unseen text, although since our approach is unsupervised we could in principle perform further retraining iterations on any sample of new text to tune the vocabulary list to it.",
"cite_spans": [],
"ref_spans": [
{
"start": 142,
"end": 149,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "We also computed a (strong) baseline, using as the vocabulary list the NTU Sentiment Dictionary (Ku et al., 2006) 5 which is intended to contain only sentiment-related words and phrases. We assigned each positive and negative vocabulary item a score of 1 or -1 respectively. This setup achieved 87.77 precision and 77.09 recall on the product review corpus.",
"cite_spans": [
{
"start": 96,
"end": 113,
"text": "(Ku et al., 2006)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "In Section 1 we argued that sentiment and objectivity should both be considered as continuums, not binary distinctions. Section 4.1 describes how our approach compares the number of positive and negative zones for a document and treats the difference as a measure of the 'positivity' or 'negativity' of a review. The document in Example (2), with 12 zones, is assigned a score of -1 (the least negative score possible): the review contains some positive sentiment but the overall sentiment direction of the review is negative. In contrast, Example (1) is identified as a highly negative review, as would be expected, with a score of -8, from 11 zones. Similarly, with regard to objectivity, the sentiment density of the text in Example (3) is 0.53, which reflects its more factual character compared to Example (1), which has a score of 0.91. We can represent sentiment and objectivity on the following scales:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "Balanced Positive",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Negative",
"sec_num": null
},
{
"text": "The scales are orthogonal, so we can combine them into a single coordinate system:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unopinionated Neutral Opinionated",
"sec_num": null
},
{
"text": "We would expect most product reviews to be placed towards the top of the the coordinate system (i.e. opinionated), and stretch from left to right. Figure 1 plots the results of sentiment and objectivity classification of the test corpus in this two dimensional coordinate system, where X represents sentiment (with scores scaled with respect to the number of zones so that -100 is the most negative possible and +100 the most positive), and Y represents sentiment density (0 being unopinionated and 1 being highly opinionated).",
"cite_spans": [],
"ref_spans": [
{
"start": 147,
"end": 155,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Opinionated Negative Positive",
"sec_num": null
},
{
"text": "Most of the reviews are located in the upper part of the coordinate system, indicating that they have been classified as opinionated, with either positive or negative sentiment direction. Looking at the overall shape of the plot, more opinionated documents tend to have more explicit sentiment direction, while less opinionated texts stay closer to the balanced / neutral region (around X = 0). ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Opinionated Negative Positive",
"sec_num": null
},
{
"text": "As can be seen in Figure 1 , the classifier managed to map the reviews onto the coordinate system. However, there are very few points in the neutral region, that is, on the same X = 0 line as balanced but with low sentiment density. By inspection, we know that there are neutral reviews in our data set. We therefore conducted a further experiment to investigate what the problem might be. We took Wikipedia 6 articles written in Chinese on mobile telephony and related issues, as well as several articles about the technology, the market and the history of mobile telecommunications, and split them into small parts (about a paragraph long, to make their size close to the size of the reviews) resulting in a corpus of 115 documents, which we assume to be mostly unopinionated. We processed these documents with the trained classifier and found that they were mapped almost exactly where balanced documents should be (see Figure 2 ). Most of these documents have weak sentiment direction (X = -5 to +10), but are classified as relatively opinionated (Y > 0.5). The former is to be expected, whereas the latter is not. When investigating the possible reasons for this behavior we noticed that the classifier found not only feature descriptions (like nice touch) or expressions which describe attitude ( (one) like(s)), but also product features (for example, MMS or TV) to be opinionated. This is because the presence of some advanced features such as MMS in mobile phones is often regarded as a positive by 6 www.wikipedia.org authors of reviews. In addition, the classifier found words that were used in reviews to describe situations connected with a product and its features: for example, (service) was often used in descriptions of quite unpleasant situations when a user had to turn to a manufacturer's post-sales service for repair or replacement of a malfunctioning phone, and (user) was often used to describe what one can do with some advanced features. Thus the classifier was able to capture some product-specific as well as market-specific sentiment markers, however, it was not able to distinguish the context these generally objective words were used in. This resulted in relatively high sentiment density of neutral texts which contained these words but used in other types of context.",
"cite_spans": [],
"ref_spans": [
{
"start": 18,
"end": 26,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 923,
"end": 931,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "To verify this hypothesis we applied the same processing to our corpus derived from Wikipedia articles, but using as the vocabulary list the NTU Sentiment Dictionary. The results (Figure 3) show that most of the neutral texts are now mapped to the lower part of the opinionation scale (Y < 0.5), as expected. Therefore, to successfully distinguish between balanced reviews and neutral documents a classifier should be able to detect when product features are used as sentiment markers and when they are not.",
"cite_spans": [],
"ref_spans": [
{
"start": 179,
"end": 189,
"text": "(Figure 3)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "We have described an approach to classification of documents with respect to sentiment polarity and objectivity, representing both as a continuum, and mapping classified documents onto a coordinate system that also represents the difference between balanced and neutral text. We have presented a novel, unsupervised, iterative retraining procedure for deriving the classifier, starting from the most minimal size seed vocabulary, in conjunction with a simple negation check. We have verified that the approach produces reasonable results. The approach is extremely minimal in terms of language processing technology, giving it good possibilities for porting to different genres, domains and languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "We also found that the accuracy of the method depends a lot on the seed word chosen. If the word has a relatively low frequency or does not have a definite sentiment-related meaning, the results may be very poor. For example, an antonymous word to (good) in Chinese is (bad), but the latter is not a frequent word: the Chinese prefer to say (not good). When this word was used as the seed word, accuracy was little more than 15%. Although the first iteration produced high precision (82%), the size of the extracted subcorpus was only 24 items, resulting in the system being unable to produce a good classifier for the following iterations. Every new iteration produced an even poorer result as each new extracted corpus was of lower accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "On the other hand, it seems that a seed list consisting of several low-frequency one-character words can compensate each other and produce better results by capturing a larger part of the corpus (thus increasing recall). Nevertheless a single word may also produce results even better than those for multiword seed lists. For example, the two-character word (comfortable) as seed reached 91% accuracy with 90% recall. We can conclude that our method relies on the quality of the seed word. We therefore need to investigate ways of choosing 'lucky' seeds and avoiding 'unlucky' ones.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "Future work should also focus on improving classification accuracy: adding a little languagespecific knowledge to be able to detect some word boundaries should help; we also plan to experiment with more sophisticated methods of sentiment score calculation. In addition, the notion of 'zone' needs refining and language-specific adjustments (for example, a 'reversed comma' should not be considered to be a zone boundary marker, since this punctuation mark is generally used for the enumeration of related objects).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "More experiments are also necessary to determine how the approach works across domains, and further investigation into methods for distinguishing between balanced and neutral text.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "Finally, we need to produce a new corpus that would enable us to evaluate the performance of a pre-trained version of the classifier that did not have any prior access to the documents it was classifying: we need the reviews to be tagged not in a binary way as they are now, but in a way that reflects the two continuums we use (sentiment and objectivity).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "http://product.it168.com",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "This corpus is publicly available at http://www.informatics. sussex.ac.uk/users/tz21/it168test.zip 3 In the first iteration, when we have only one item in the vocabulary, negative zones are found by means of the negation check (so not + good = negative item).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The size of the sentiment vocabulary after iteration 18 was 22530 (13462 positive and 9068 negative).5 Ku et al. automatically generated the dictionary by enlarging",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The first author is supported by the Ford Foundation International Fellowships Program.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Bootstrapping",
"authors": [
{
"first": "Steven",
"middle": [],
"last": "Abney",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "360--367",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Abney, Steven (2002) Bootstrapping. In Proceedings of the 40th Annual Meeting of the Association for Com- putational Linguistics, Philadelphia, PA. 360-367.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Customizing sentiment classifiers to new domains: a case study",
"authors": [
{
"first": "Anthony & Michael",
"middle": [],
"last": "Aue",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Gamon",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of RANLP-2005",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aue, Anthony & Michael Gamon (2005) Customizing sentiment classifiers to new domains: a case study. In Proceedings of RANLP-2005.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Mining the peanut gallery: opinion extraction and semantic classification of product reviews",
"authors": [
{
"first": "Kushal",
"middle": [],
"last": "Dave",
"suffix": ""
},
{
"first": "Steve",
"middle": [],
"last": "Lawrence",
"suffix": ""
},
{
"first": "&",
"middle": [],
"last": "David",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Pennock",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the Twelfth International World Wide Web Conference",
"volume": "",
"issue": "",
"pages": "519--528",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dave, Kushal, Steve Lawrence & David M. Pennock (2003) Mining the peanut gallery: opinion extraction and semantic classification of product reviews. In Proceedings of the Twelfth International World Wide Web Conference. 519-528.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Topic dependence in sentiment classification. Unpublished MPhil dissertation",
"authors": [
{
"first": "Charlotte",
"middle": [],
"last": "Engstr\u00f6m",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Engstr\u00f6m, Charlotte (2004) Topic dependence in senti- ment classification. Unpublished MPhil dissertation, University of Cambridge.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "SENTI-WORDNET: a publicly available lexical resource for opinion mining",
"authors": [
{
"first": "Andrea & Fabrizio",
"middle": [],
"last": "Esuli",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Sebastiani",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of LREC-06, the 5th Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Esuli, Andrea & Fabrizio Sebastiani (2006) SENTI- WORDNET: a publicly available lexical resource for opinion mining. In Proceedings of LREC-06, the 5th Conference on Language Resources and Evaluation, Genoa, Italy.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Jordi Atserias (2007) World knowledge in broad-coverage information filtering",
"authors": [
{
"first": "Bennett",
"middle": [],
"last": "Hagedorn",
"suffix": ""
},
{
"first": "Massimiliano",
"middle": [],
"last": "Ciaramita",
"suffix": ""
}
],
"year": null,
"venue": "Proceedings of the 30th ACM SIGIR Conference on Research and Development in Information Retrieval",
"volume": "",
"issue": "",
"pages": "801--802",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hagedorn, Bennett, Massimiliano Ciaramita & Jordi At- serias (2007) World knowledge in broad-coverage in- formation filtering. In Proceedings of the 30th ACM SIGIR Conference on Research and Development in Information Retrieval. 801-802.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Opinion extraction, summarization and tracking in news and blog corpora",
"authors": [
{
"first": "Lun-Wei",
"middle": [],
"last": "Ku",
"suffix": ""
},
{
"first": "Yu-Ting",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "& Hsin-Hsi",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the AAAI-2006 Spring Symposium on Computational Approaches to Analyzing Weblogs, AAAI Technical Report",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ku, Lun-Wei, Yu-Ting Liang & Hsin-Hsi Chen (2006) Opinion extraction, summarization and tracking in news and blog corpora. In Proceedings of the AAAI- 2006 Spring Symposium on Computational Ap- proaches to Analyzing Weblogs, AAAI Technical Re- port.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Query-based summarization of customer reviews",
"authors": [
{
"first": "Olga & Guy",
"middle": [],
"last": "Feiguina",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lapalme",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 20th Canadian Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "452--463",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Feiguina, Olga & Guy Lapalme (2007) Query-based summarization of customer reviews. In Proceedings of the 20th Canadian Conference on Artificial Intelli- gence, Montreal, Canada. 452-463.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Thumbs up? Sentiment classification using machine learning techniques",
"authors": [
{
"first": "Bo",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "Lillian",
"middle": [],
"last": "Lee & Shivakumar",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Vaithyanathan",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "79--86",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pang, Bo, Lillian Lee & Shivakumar Vaithyanathan (2002) Thumbs up? Sentiment classification using machine learning techniques. In Proceedings of the Conference on Empirical Methods in Natural Lan- guage Processing, Philadelphia, PA. 79-86.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A sentimental education: sentiment analysis using subjectivity summarization based on minimum cuts",
"authors": [
{
"first": "Bo & Lillian",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 42nd",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pang, Bo & Lillian Lee (2004) A sentimental education: sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Annual Meeting of the Association for Computational Linguistics",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "271--278",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Annual Meeting of the Association for Computation- al Linguistics, Barcelona, Spain. 271-278.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Seeing stars: exploiting class relationships for sentiment categorization with respect to rating scales",
"authors": [
{
"first": "Bo & Lillian",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the 43rd",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pang, Bo & Lillian Lee (2005) Seeing stars: exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Annual Meeting of the Association for Computational Linguistics",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "115--124",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Annual Meeting of the Association for Computation- al Linguistics, Ann Arbor, MI. 115-124.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Using emoticons to reduce dependency in machine learning techniques for sentiment classification",
"authors": [
{
"first": "Jonathon",
"middle": [],
"last": "Read",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the Student Research Workshop at ACL-05",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Read, Jonathon (2005) Using emoticons to reduce de- pendency in machine learning techniques for senti- ment classification. In Proceedings of the Student Research Workshop at ACL-05, Ann Arbor, MI.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Thumbs up or thumbs down? Semantic orientation applied to unsupervised classification of reviews",
"authors": [
{
"first": "Peter",
"middle": [
"D"
],
"last": "Turney",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "417--424",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Turney, Peter D. (2002) Thumbs up or thumbs down? Semantic orientation applied to unsupervised classifi- cation of reviews. In Proceedings of the 40th Annual Meeting of the Association for Computational Lin- guistics, Philadelphia, PA. 417-424.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Unsupervised word sense disambiguation rivaling supervised methods",
"authors": [
{
"first": "David",
"middle": [],
"last": "Yarowsky",
"suffix": ""
}
],
"year": 1995,
"venue": "Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "189--196",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yarowsky, David (1995) Unsupervised word sense dis- ambiguation rivaling supervised methods. In Pro- ceedings of the 33rd Annual Meeting of the Associa- tion for Computational Linguistics, Cambridge, MA. 189-196.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Towards answering opinion questions: separating facts from opinions and identifying the polarity of opinion sentences",
"authors": [
{
"first": "Hong & Vasileios",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hatzivassiloglou",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "129--136",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yu, Hong & Vasileios Hatzivassiloglou (2003) Towards answering opinion questions: separating facts from opinions and identifying the polarity of opinion sen- tences. In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing, Sapporo, Japan. 129-136.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "Reviews classified according to sentiment (X axis) and degree of opinionation (Y axis).",
"num": null,
"type_str": "figure"
},
"FIGREF1": {
"uris": null,
"text": "Classification of a sample of articles from Wikipedia.",
"num": null,
"type_str": "figure"
},
"FIGREF2": {
"uris": null,
"text": "Classification of a sample of articles from Wikipedia, using the NTU Sentiment Dictionary as the vocabulary list.",
"num": null,
"type_str": "figure"
},
"TABREF0": {
"text": "Results for binary sentiment classification during iterative retraining.",
"content": "<table><tr><td>an initial manually created seed vocabulary by consulting two</td></tr><tr><td>thesauri, including tong2yi4ci2ci2lin2 and the Academia Sini-</td></tr><tr><td>ca Bilingual Ontological WordNet 3.</td></tr></table>",
"num": null,
"type_str": "table",
"html": null
}
}
}
} |