File size: 66,382 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 |
{
"paper_id": "O08-1009",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:02:37.623038Z"
},
"title": "A Realistic and Robust Model for Chinese Word Segmentation",
"authors": [
{
"first": "Huang",
"middle": [],
"last": "Chu-Ren",
"suffix": "",
"affiliation": {},
"email": "churenhuang@gmail.com"
},
{
"first": "Yo",
"middle": [],
"last": "Ting-Shuo",
"suffix": "",
"affiliation": {},
"email": "tingshuo.yo@gmail.com"
},
{
"first": "Petr",
"middle": [],
"last": "\u0160imon",
"suffix": "",
"affiliation": {},
"email": "petr.simon@gmail.com"
},
{
"first": "Hsieh",
"middle": [],
"last": "Shu-Kai",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Normal University",
"location": {
"country": "Taiwan"
}
},
"email": "shukai@gmail.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "A realistic Chinese word segmentation tool must adapt to textual variations with minimal training input and yet robust enough to yield reliable segmentation result for all variants. Various lexicon-driven approaches to Chinese segmentation, e.g. [1,16], achieve high f-scores yet require massive training for any variation. Text-driven approach, e.g. [12], can be easily adapted for domain and genre changes yet has difficulty matching the high f-scores of the lexicon-driven approaches. In this paper, we refine and implement an innovative text-driven word boundary decision (WBD) segmentation model proposed in [15]. The WBD model treats word segmentation simply and efficiently as a binary decision on whether to realize the natural textual break between two adjacent characters as a word boundary. The WBD model allows simple and quick training data preparation converting characters as contextual vectors for learning the word boundary decision. Machine learning experiments with four different classifiers show that training with 1,000 vectors and 1 million vectors achieve comparable and reliable results. In addition, when applied to SigHAN Bakeoff 3 competition data, the WBD model produces OOV recall rates that are higher than all published results. Unlike all previous work, our OOV recall rate is comparable to our own F-score. Both experiments support the claim that the WBD model is a realistic model for Chinese word segmentation as it can be easily adapted for new variants with robust result. In conclusion, we will discuss linguistic ramifications as well as future implications for the WBD approach.",
"pdf_parse": {
"paper_id": "O08-1009",
"_pdf_hash": "",
"abstract": [
{
"text": "A realistic Chinese word segmentation tool must adapt to textual variations with minimal training input and yet robust enough to yield reliable segmentation result for all variants. Various lexicon-driven approaches to Chinese segmentation, e.g. [1,16], achieve high f-scores yet require massive training for any variation. Text-driven approach, e.g. [12], can be easily adapted for domain and genre changes yet has difficulty matching the high f-scores of the lexicon-driven approaches. In this paper, we refine and implement an innovative text-driven word boundary decision (WBD) segmentation model proposed in [15]. The WBD model treats word segmentation simply and efficiently as a binary decision on whether to realize the natural textual break between two adjacent characters as a word boundary. The WBD model allows simple and quick training data preparation converting characters as contextual vectors for learning the word boundary decision. Machine learning experiments with four different classifiers show that training with 1,000 vectors and 1 million vectors achieve comparable and reliable results. In addition, when applied to SigHAN Bakeoff 3 competition data, the WBD model produces OOV recall rates that are higher than all published results. Unlike all previous work, our OOV recall rate is comparable to our own F-score. Both experiments support the claim that the WBD model is a realistic model for Chinese word segmentation as it can be easily adapted for new variants with robust result. In conclusion, we will discuss linguistic ramifications as well as future implications for the WBD approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The paper deals with the fundamental issue why Chinese word segmentation remains a research topic and not a language technology application after more than twenty years of intensive study. Chinese text is typically presented as a continuous string of characters without conventionalized demarcation of word boundaries. Hence tokenization of words, commonly called word segmentation in literature, is a pre-requisite first step for Chinese language processing. Recent advances in Chinese word segmentation (CWS) include popular standardized competitions run by ACL SigHAN and typically high F-scores around 0.95 from leading teams [8] . However, these results are achieved at the cost of high computational demands, including massive resources and long machine learning time. In fact, all leading systems are expected to under-perform substantially without prior substantial training. It is also important to note that SigHAN competitions are conducted under the assumption that a segmentation program must be tuned separately for different source texts and will perform differently. This is a bow to the fact that different communities may conventionalize the concept of word differently; but also an implicit concession that it is hard for existing segmentation programs to deal with textual variations robustly. [15] proposed an innovative model for Chinese word segmentation which formulates it as simple two class classification task without having to refer to massive lexical knowledge base.",
"cite_spans": [
{
"start": 630,
"end": 633,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 1314,
"end": 1318,
"text": "[15]",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background and Motivation",
"sec_num": "1."
},
{
"text": "We refine and implement this Word Boundary Decision (WBD) model and show that it is indeed realistic and robust. With drastically smaller demand on computational resources, we achieved comparable F-score with leading Bakeoff3 teams and outperform all on OOV recall, the most reliable criterion to show that our system deals with new events effectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and Motivation",
"sec_num": "1."
},
{
"text": "In what follows, we will discuss modeling issues and survey previous work in the first section. The WBD model will be introduced in the second section. This is followed by a description of the machine learning model is trained in Section 4. Results of applying this implementation to SigHAN Bakeoff3 data is presented in Section 5. We conclude with discussion of theoretical ramifications and implications in Section 6.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background and Motivation",
"sec_num": "1."
},
{
"text": "The performance of CWS systems is directly influenced by their design criteria and how Chinese word segmentation task is modeled. These modeling issues did not receive in-depth discussion in previous literature:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "How to model Chinese word segmentation",
"sec_num": "2."
},
{
"text": "Modeling Segmentation. The input to Chinese word segmentation is a string of characters.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "How to model Chinese word segmentation",
"sec_num": "2."
},
{
"text": "However, the task of segmentation can be modeled differently. All previous work share the assumption that the task of segmentation is to find our all segments of the string that are words. This can be done intuitively by dictionary lookup, or be looking at strength of collocation within a string, e.g. [12] . Recent studies, e.g. [14, 16, 5, 17] , reduce the complexity of this model and avoided the thorny issue of the elusive concept of word at the same time by modeling segmentation as learning the likelihood of characters being the edges of these word strings. These studies showed that, with sufficient features, machine learning algorithms can learn from training corpus and use their inherent model to tokenize Chinese text satisfactorily. The antagonistic null hypothesis of treating segmentation as simply identifying inherent textual breaks between two adjacent characters was never pursued.",
"cite_spans": [
{
"start": 303,
"end": 307,
"text": "[12]",
"ref_id": "BIBREF11"
},
{
"start": 331,
"end": 335,
"text": "[14,",
"ref_id": "BIBREF13"
},
{
"start": 336,
"end": 339,
"text": "16,",
"ref_id": "BIBREF15"
},
{
"start": 340,
"end": 342,
"text": "5,",
"ref_id": "BIBREF4"
},
{
"start": 343,
"end": 346,
"text": "17]",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "How to model Chinese word segmentation",
"sec_num": "2."
},
{
"text": "Out-of-Vocabulary Words. Identification of Out-of Vocabulary words (OOV, sometimes conveniently referred to as new words) has been a challenge to all systems due to data sparseness problem, as well as for dealing with true neologisms which cannot be learned from training data per se. This requirement means that CWS system design must incorporate explicit or implicit morphology knowledge to assure appropriate sensitivity to context in which potential words occur as previously unseen character sequences. developed different textual conventions as well as lexical items. This is compounded by the usual text type, domain, and genre contrasts. A robust CWS system must be able to adapt to these variations without requiring massive retraining. A production environment with it's time restrictions possesses great demands on the segmentation system to be able to quickly accommodate even to mixture of text types, since such a mixture would introduce confusing contexts and confuse system that would rely too heavily on text type, i.e. particular lexicon choice and specific morphology, and too large a context. Space and time demands. Current CWS systems cannot avoid long training times and large memory demands. This is a consequence of the segmentation model employed. This is acceptable when CWS systems are used for offline tasks such as corpora preprocessing, where time and space can be easily provided and when needed. However, for any typically web-based practical language engineering applications, such high demand on computing time is not acceptable.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "How to model Chinese word segmentation",
"sec_num": "2."
},
{
"text": "Two contrasting approaches to Chinese word segmentation summarize the dilemma of segmentation system design. A priori, one can argue that segmentation is the essential tool for building a (mental) lexicon hence segmentation cannot presuppose lexical knowledge. On the other hand, as a practical language technology issue, one can also argue that segmentation is simply matching all possible words from a (hypothetical) universal lexicon and can be simplified as mapping to a large yet incomplete lexicon. Hence we can largely divide previous approaches to Chinese word segmentation as lexicon-driven or text-driven.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Previous works: a critical review",
"sec_num": "2.1"
},
{
"text": "Text-Driven. Text-driven approach to segmentation relies on contextual information to identify words and do not assume any prior lexical knowledge. Researches in this approach typically emphasize the need for an empirical approach to define the concept of a word in a language. [12] work based on mutual information (MI) is the best-known and most comprehensive in this approach. The advantage of this approach it can be applied to all different variations of language and yet be highly adaptive. However, the basic implementation of MI applies bi-syllabic words only. In addition, it cannot differentiate between highly collocative bigrams (such as \u5c31\u4e0d jiubu -\u2026then not\u2026\u2016) and words. Hence it typically has lower recall and precision rate than current methods. Even though text-driven approaches are no longer popular, they are still widely used to deal with OOV with a lexicon-driven approach.",
"cite_spans": [
{
"start": 278,
"end": 282,
"text": "[12]",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous works: a critical review",
"sec_num": "2.1"
},
{
"text": "Tokenization. The classical lexicon-driven segmentation model, described in [1] and is still adopted in many recent works. Segmentation is typically divided into two stages: dictionary look up and OOV word identification. This approach requires comparing and matching tens of thousands of dictionary entries in addition to guessing a good number of OOV words. In other words, it has a 10 4 x 10 4 scale mapping problem with unavoidable data sparseness. This model also has the unavoidable problem of overlapping ambiguity where e.g. a string [C i-1 , C i , C i+1 ] contains multiple sub-strings, such as [C i-1 , C i ] and [C i , C i+1 ], which are entries in the dictionary. The degree of such ambiguities is estimated to fall between 5% to 20% [2, 6] .",
"cite_spans": [
{
"start": 76,
"end": 79,
"text": "[1]",
"ref_id": "BIBREF0"
},
{
"start": 746,
"end": 749,
"text": "[2,",
"ref_id": "BIBREF1"
},
{
"start": 750,
"end": 752,
"text": "6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous works: a critical review",
"sec_num": "2.1"
},
{
"text": "Character classification. Character classification or tagging, first proposed in [14] , became a very popular approach recently since it is proved to be very effective in addressing problems of scalability and data sparseness [14, 4, 16, 17] . Since it tries to model the possible position of a character in a word as character-strings, it is still lexicon-driven. This approach has been also successfully applied by to name entity resolution, e.g. [17] . This approach is closely related to the adoption of the machine learning algorithm of conditional random field (CRF), [7] . CRF has been shown [11] to be optimal algorithm for sequence classification. The major disadvantages are big memory and computational time requirement.",
"cite_spans": [
{
"start": 81,
"end": 85,
"text": "[14]",
"ref_id": "BIBREF13"
},
{
"start": 226,
"end": 230,
"text": "[14,",
"ref_id": "BIBREF13"
},
{
"start": 231,
"end": 233,
"text": "4,",
"ref_id": "BIBREF3"
},
{
"start": 234,
"end": 237,
"text": "16,",
"ref_id": "BIBREF15"
},
{
"start": 238,
"end": 241,
"text": "17]",
"ref_id": "BIBREF16"
},
{
"start": 449,
"end": 453,
"text": "[17]",
"ref_id": "BIBREF16"
},
{
"start": 574,
"end": 577,
"text": "[7]",
"ref_id": "BIBREF6"
},
{
"start": 599,
"end": 603,
"text": "[11]",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous works: a critical review",
"sec_num": "2.1"
},
{
"text": "Our approach is based on a simplified idea of Chinese text, which we have introduced earlier in [15] . Chinese text can be formalized as a sequence of characters and intervals as illustrated in Figure 1 .",
"cite_spans": [
{
"start": 96,
"end": 100,
"text": "[15]",
"ref_id": "BIBREF14"
}
],
"ref_spans": [
{
"start": 194,
"end": 202,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "There is no indication of word boundaries in Chinese text, only string of characters c i .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "Characters in this string can be conceived as being separated by interval I i . To obtain a segmented text, i.e. a text where individual words are delimited by some graphical mark such as space, we need to identify which of these intervals are to be replaced by such word delimiter.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "We can introduce a utility notion of imaginary intervals between characters, which we formally classify into two types:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "Type 0: a character boundary (CB) is an imaginary boundary between two characters Type 1: a word boundary (WB), an interval separating two words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "With such a formulation, segmentation task can be easily defined as a classification task and machine learning algorithms can be employed to solve it. For conventional machine learning algorithms, classifications are made based on a set of features, which identify certain properties of the target to be classified.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "In a segmented text, all the intervals between characters are labeled as a word boundary or as Figure 1 : Chinese text formalization a character boundary, however, characters are not considered as being part of any particular word. Their sole function is to act as a contextual aid for identification of the most probable interval label. Since the intervals between characters (be it a word boundary or a character boundary) don't carry any information at all, we need to rely on the information provided by group of characters surrounding them. Now we can collect n-grams that will provide data for construction of features that will provide learning basis for machine learning algorithm. A sequence, such the one illustrated in Figure 1 , can be obtained from segmented corpus, and hence the probability of word boundary with specified relation to each n-gram may be derived. The resulting table which consists of each distinct n-gram entry observed in the corpus and the probability of a word boundary defines our n-gram collection. Figure 2 shows the format of the feature vectors, or interval vectors, used in this study. We build the n-gram model up to n = 2. To allow for a more fine-grained statistical information we have decomposed an interval surrounding context into two unigrams and three bi-grams. For convenience, we can define each interval by the two characters that surround it. Then, for each interval <b,c> in a 4-character window abcd we collect two unigrams b and c and three bi-grams ab, bc, cd and compute probability of that interval being a word boundary. These five n-grams are stored in a vector, which is labeled as Type 0 (character boundary) or Type 1 (word boundary): <ab, b, bc, c, cb, 0> or <ab, b, bc, c, cb, 1>. An example of an encoding of a sample from the beginning of Bakeoff 3 AS training corpus: \"\u6642\u9593\uff1a\u4e09\u6708\u5341\u65e5\" (shijian:sanyueshiri), which would be correctly segmented as \"\u6642\u9593 \uff1a \u4e09\u6708\u5341\u65e5\" (shijian : sanyue shiri) can be seen in Table 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 95,
"end": 103,
"text": "Figure 1",
"ref_id": null
},
{
"start": 730,
"end": 738,
"text": "Figure 1",
"ref_id": null
},
{
"start": 1036,
"end": 1044,
"text": "Figure 2",
"ref_id": "FIGREF1"
},
{
"start": 1961,
"end": 1968,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "c 1, I 1, c 2, I 2, ... , c n \u2212 1, I n \u2212 1, c n",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "Set of such interval vectors provides a training corpus on which we apply machine learning algorithm, in our case logarithmic regression. Unsegmented text is prepared in the same fashion and the interval vectors are subsequently labeled by a classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3."
},
{
"text": "It is our goal to develop a segmentation system that would be able handle different types of In order to verify the effect of the size of the training data, the full ASBC (~17 million intervals) and a subset of it (1 million randomly selected intervals) are used for training separately. Furthermore, four different classifiers, i.e., logistic regression (LogReg) [9] , linear discriminative analysis (LDA) [13] , multi-layer perceptron (NNET) [13] , and support vector machine (SVM) [3] , were tested.",
"cite_spans": [
{
"start": 364,
"end": 367,
"text": "[9]",
"ref_id": "BIBREF8"
},
{
"start": 407,
"end": 411,
"text": "[13]",
"ref_id": "BIBREF12"
},
{
"start": 444,
"end": 448,
"text": "[13]",
"ref_id": "BIBREF12"
},
{
"start": 484,
"end": 487,
"text": "[3]",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training the Machine Learning Model",
"sec_num": "4."
},
{
"text": "The segmentation results are compared with the \"gold standard\" provided by the SigHAN",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training the Machine Learning Model",
"sec_num": "4."
},
{
"text": ". Tables 2 and 3 show the training and testing accuracies of various classifiers trained with the ASBC. All classifiers tested perform as expected, with their training errors increase with the size of the training data, and the testing errors decrease with it. Table 2 clearly shows that the training data size has little effect on the testing error while it is above 1000. This proves that once a sufficient n-gram collection is provided for preparation of the interval vectors, classifier can be trained with little input.",
"cite_spans": [],
"ref_spans": [
{
"start": 2,
"end": 16,
"text": "Tables 2 and 3",
"ref_id": "TABREF1"
},
{
"start": 261,
"end": 268,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Bakeoff2",
"sec_num": null
},
{
"text": "It is also shown in Table 2 that four classifiers give similar performance when the training data size is above 1000. However, while the training sample size drops to 100, the SVM and LDA algorithms show their strength by giving similar performance to the experiments trained with larger training data sets. ",
"cite_spans": [],
"ref_spans": [
{
"start": 20,
"end": 27,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Bakeoff2",
"sec_num": null
},
{
"text": "In our test to compare our performance objectively with other approaches, we adopt logarithmic regression as our learning algorithm as it yielded best results during our test. We apply the segmentation system to two traditional Chinese corpora, CKIP and CityU, provided",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "for SigHAN Bakeoff 3. In the first set of tests, we used training corpora provided by SigHAN Bakeoff3 for n-gram collection, training and testing. Results of these tests are presented in Table 6 .",
"cite_spans": [],
"ref_spans": [
{
"start": 187,
"end": 194,
"text": "Table 6",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "In addition, to underline the adaptability of this approach, we also tried combining both corpora and then ran training on random sample of vectors. This set of tests is designed to exclude the possibility of over-fitting and to underline the robustness of the WBD model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "Note that such tests are not performed in SigHAN Bakeoffs as many of the best performances are likely over-fitted. Results of this test are shown in Table 7 . Table 6 and 7 show that our OOV recall is comparable with our overall F-score, especially when our system is trained on selected vectors from combined corpus. This is in direct contrast with all existing systems, which typically has a much lower OOV recall than IV recall. In other words, our approach applies robustly to all textual variations with reliably good results. Table 8 shows that indeed our OOV recall rate shows over 16% improvement ",
"cite_spans": [],
"ref_spans": [
{
"start": 149,
"end": 156,
"text": "Table 7",
"ref_id": "TABREF6"
},
{
"start": 159,
"end": 166,
"text": "Table 6",
"ref_id": "TABREF5"
},
{
"start": 532,
"end": 539,
"text": "Table 8",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "We refined and implemented the WBD model for Chinese word segmentation and show that it is a robust and realistic model for Chinese language technology. Most crucially, we show that the WBD model is able to reconcile the two competitive goals of the lexicon-driven and text-driven approaches. The WBD model maintains comparable F-score level with the most recent CRF character-classification based results, yet improves substantially on the OOV recall.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "We showed that our system is robust and not over-fitted to a particular corpus, as it yields comparable and reliable results for both OOV and IV words. In addition, we show that same level of consistently high results can be achieved across different text sources. Our results",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "show that Chinese word segmentation system can be quite efficient even when using very simple model and simple set of features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "Our current system, which has not been optimized for speed, is able to segment text in less then 50 seconds. Time measurement includes preparation of testing data, but also training phase. We believe that with optimized and linked computing power, it will be easy to implement a real time application system based on our model. In the training stage, we have shown that sampling of around 1,000 vectors is enough to yield one of the best results. Again, this is a promise fact for the WBD model of segmentation to be robust. It is notable, that in case of training on combined corpora (CKIP and CityU) the results are even better than test in respective data sets, i.e. CKIP training corpus for segmenting CKIP testing text, or CityU respectively. This is undoubtedly the result of our strategy of granulation of the context around each interval. Since four characters that we use for representation of the interval context are broken up into two unigrams and three bi-grams, we let the system to get more (Levow, 2006) refined insight into the segmented area.",
"cite_spans": [
{
"start": 1006,
"end": 1019,
"text": "(Levow, 2006)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "Consequently, the system is learning morphology of Chinese with greater generality and this results in higher OOV scores. It can be argued that in our combined corpora test, the OOV recall is even higher, because the input contains two different variants of Chinese language, Taiwanese variant contained in CKIP corpus and Hong Kong variant contained in CityU corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "Text preparation and post-processing also add to overall processing time. In our current results, apart from context vector preparation there was no other preprocessing employed and neither any post-processing. This fact also shows that our system is able to handle any type of input without the need to define special rules to pre-or post-process the text. Early results applying our model to simplified Chinese corpora are also promising.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
},
{
"text": "In sum, our WBD model for Chinese word segmentation yields one of the truly robust and realistic segmentation program for language technology applications. If these experiments are treated as simulation, our results also support the linguistic hypothesis that word can be reliably discovered without a built-in/innate lexicon. We will look into developing a more complete model to allow for more explanatory account for domain specific shifts as well as for effective bootstrapping with some lexical seeds.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6."
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Word Identification for Mandarin Chinese sentences",
"authors": [
{
"first": "K",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "S",
"middle": [
"H"
],
"last": "Liu",
"suffix": ""
}
],
"year": 1992,
"venue": "Proceedings of the 14th conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "101--107",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K.J Chen and S.H. Liu, -Word Identification for Mandarin Chinese sentences-, in Proceedings of the 14th conference on Computational Linguistics, pp.101-107, 1992.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Readings in Chinese Natural Language Processing",
"authors": [
{
"first": "T.-H",
"middle": [],
"last": "Chiang",
"suffix": ""
},
{
"first": "J.-S",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "M.-Y",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "K.-Y.",
"middle": [],
"last": "Su",
"suffix": ""
}
],
"year": 1996,
"venue": "",
"volume": "9",
"issue": "",
"pages": "147--173",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T.-H. Chiang, J.-S. Chang, M.-Y. Lin and K.-Y. Su, -Statistical Word Segmentation\u2016, in C.-R. Huang, K.-J. Chen and B.K. T'sou (eds.): Journal of Chinese Linguistics, Monograph Series, Number 9, Readings in Chinese Natural Language Processing, pp. 147-173, 1996.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Weingessel -e1071: Misc. Functions of the Department of Statistics (e1071)\u2016, TU Wien R package version 1.5-17",
"authors": [
{
"first": "E",
"middle": [],
"last": "Dimitriadou",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Hornik",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Leisch",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Meyer",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "E. Dimitriadou, K. Hornik, F. Leisch, D. Meyer and A.Weingessel -e1071: Misc. Functions of the Department of Statistics (e1071)\u2016, TU Wien R package version 1.5-17., 2007.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Proceedings of ACL-2004",
"authors": [
{
"first": "J",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "C.-N",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Xia",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Qin",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Gao, A. Wu, M. Li, C.-N. Huang, H. Li, X. Xia and H.Qin, -Adaptive Chinese Word Segmentation\u2016, in Proceedings of ACL-2004, 2004.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Which Is Essential for Chinese Word Segmentation: Character versus Word\u2016",
"authors": [
{
"first": "C.-N",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Zhao",
"suffix": ""
}
],
"year": 2006,
"venue": "The 20th Pacific Asia Conference on Language, Information and Computation (PACLIC-20)",
"volume": "",
"issue": "",
"pages": "1--12",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C.-N. Huang and H. Zhao, -Which Is Essential for Chinese Word Segmentation: Character versus Word\u2016, The 20th Pacific Asia Conference on Language, Information and Computation (PACLIC-20), pp.1-12, 2006.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "An Analytical Study of Transformational Tagging for Chinese Text\u2016",
"authors": [
{
"first": "H",
"middle": [],
"last": "Meng",
"suffix": ""
},
{
"first": "C",
"middle": [
"W"
],
"last": "Ip",
"suffix": ""
}
],
"year": 1999,
"venue": "Proceedings of ROCLING XII",
"volume": "",
"issue": "",
"pages": "101--122",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Meng and C. W. Ip, -An Analytical Study of Transformational Tagging for Chinese Text\u2016, in Proceedings of ROCLING XII, pp. 101-122, 1999.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data\u2016",
"authors": [
{
"first": "J",
"middle": [
"D"
],
"last": "Lafferty",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Mccallum",
"suffix": ""
},
{
"first": "F",
"middle": [
"C N"
],
"last": "Pereira",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "282--289",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.D. Lafferty, A. McCallum and F.C.N. Pereira, -Conditional random fields: Probabilistic models for segmenting and labeling sequence data\u2016, in ICML 2001, pp 282-289, 2001.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The Third International Chinese Language Processing Bakeoff: Word Segmentation and Named Entity Recognition\u2016",
"authors": [
{
"first": "G. -A",
"middle": [],
"last": "Levow",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing",
"volume": "",
"issue": "",
"pages": "108--117",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "G. -A. Levow, -The Third International Chinese Language Processing Bakeoff: Word Segmentation and Named Entity Recognition\u2016, in Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Association for Computational Linguistics, pp.108-117, 2006.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing\u2016",
"authors": [],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R Development Core Team, -R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing\u2016, Vienna, Austria, 2008.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The Universality of Simple Distributional Methods: Identifying Syntactic Categories in Mandarin Chinese\u2016",
"authors": [
{
"first": "M",
"middle": [],
"last": "Redington",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Chater",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 1995,
"venue": "Proceedings of the International Conference on Cognitive Science and Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Redington, N. Chater, C. Huang, L. Chang and K. Chen, -The Universality of Simple Distributional Methods: Identifying Syntactic Categories in Mandarin Chinese\u2016, in Proceedings of the International Conference on Cognitive Science and Natural Language Processing, 1995.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A systematic cross-comparison of sequence classifiers\u2016",
"authors": [
{
"first": "B",
"middle": [],
"last": "Rosenfeld",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Feldman",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Fresko",
"suffix": ""
}
],
"year": 2006,
"venue": "SDM 2006",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "B. Rosenfeld, R. Feldman and M. Fresko, -A systematic cross-comparison of sequence classifiers\u2016, in SDM 2006, 2006.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A Stochastic Finite-State Word-Segmentation Algorithm for Chinese\u2016, Computational Linguistics",
"authors": [
{
"first": "R",
"middle": [],
"last": "Sproat",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Shih",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Gale",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Chang",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "22",
"issue": "",
"pages": "377--404",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Sproat, C. Shih, W. Gale and N. Chang, -A Stochastic Finite-State Word-Segmentation Algorithm for Chinese\u2016, Computational Linguistics, 22(3) pp. 377-404, 1997.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Modern Applied Statistics with S\u2016, Fourth Edition",
"authors": [
{
"first": "W",
"middle": [
"N"
],
"last": "Venables",
"suffix": ""
},
{
"first": "B",
"middle": [
"D"
],
"last": "Ripley",
"suffix": ""
}
],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. N. Venables and B. D. Ripley, -Modern Applied Statistics with S\u2016, Fourth Edition, Springer New York, ISBN 0-387-95457-0, 2002.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Chinese Word Segmentation as Character Tagging\u2016, Computational Linguistics and Chinese Language Processing",
"authors": [
{
"first": "N",
"middle": [],
"last": "Xue",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "8",
"issue": "",
"pages": "29--48",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. Xue, -Chinese Word Segmentation as Character Tagging\u2016, Computational Linguistics and Chinese Language Processing. 8(1), pp. 29-48, 2003.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Rethinking Chinese Word Segmentation: Tokenization, Character Classification, or Wordbreak Identification\u2016",
"authors": [
{
"first": "C.-R",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "\u0160imon",
"suffix": ""
},
{
"first": "S.-K",
"middle": [],
"last": "Hsieh",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Pr\u00e9vot",
"suffix": ""
}
],
"year": null,
"venue": "Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions",
"volume": "",
"issue": "",
"pages": "69--72",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C.-R. Huang, P. \u0160imon, S.-K. Hsieh and L. Pr\u00e9vot, -Rethinking Chinese Word Segmentation: Tokenization, Character Classification, or Wordbreak Identification\u2016, Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, Association for Computational Linguistics,pp. 69-72, Prague, Czech Republic",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "An improved Chinese word segmentation system with conditional random field\u2016",
"authors": [
{
"first": "H",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "C.-N",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2006,
"venue": "SIGHAN-5",
"volume": "",
"issue": "",
"pages": "162--165",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Zhao, C.-N. Huang and M. Li, -An improved Chinese word segmentation system with conditional random field\u2016, in SIGHAN-5, pp 162-165, 2006.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Unsupervised Segmentation Helps Supervised Learning of Character Tagging for Word Segmentation and Named Entity Recognition\u2016",
"authors": [
{
"first": "H",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Kit",
"suffix": ""
}
],
"year": 2008,
"venue": "The Sixth SIGHAN Workshop on Chinese Language Processing",
"volume": "",
"issue": "",
"pages": "106--111",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Zhao and C. Kit, -Unsupervised Segmentation Helps Supervised Learning of Character Tagging for Word Segmentation and Named Entity Recognition\u2016, The Sixth SIGHAN Workshop on Chinese Language Processing (SIGHAN-6), pp.106-111, 2008.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Especially among different Chinese speaking communities. Note that different Chinese speaking communities in PRC, Taiwan, Hong Kong Singapore etc.",
"num": null,
"uris": null,
"type_str": "figure"
},
"FIGREF1": {
"text": "The feature vectors used in this study.While C denotes a character in the sequence, Bindicates the imaginary boundary. Thus CBC denotes a bi-gram containing the interval.",
"num": null,
"uris": null,
"type_str": "figure"
},
"TABREF0": {
"num": null,
"text": "text. A large uniform training corpus is desirable for high precision of segmentation, but that would cause a specialization of the classifier to types of texts contained in the corpus and system's generality would be compromised.Furthermore, using a training data set converted from an independent corpus may give supplementary information and provide certain adaptation mechanism for the classifier during training, but leave the basic n-gram collection untouched. However, a smaller set of training data may give similar performance but with much lower cost.If the features in the n-gram collection are properly defined, the final results from different machine learning algorithms may not differ too much. On the contrary, if the available n-gram collection does not provide efficient information, classifiers with ability to adjust the feature space may be necessary.In our preliminary tests, during which we wanted to decide which machine learning algorithm would be most appropriate, the Academia Sinica Balance Corpus (ASBC) is used for the derivation of the n-gram collection and training data. The CityU corpus from the SigHAN Bakeoff2 collection is used for testing.",
"content": "<table><tr><td>Example of encoding and labeling of interval vectors in a</td></tr><tr><td>4-character window ABCD</td></tr></table>",
"html": null,
"type_str": "table"
},
"TABREF1": {
"num": null,
"text": "",
"content": "<table/>",
"html": null,
"type_str": "table"
},
"TABREF2": {
"num": null,
"text": "",
"content": "<table/>",
"html": null,
"type_str": "table"
},
"TABREF3": {
"num": null,
"text": "",
"content": "<table/>",
"html": null,
"type_str": "table"
},
"TABREF4": {
"num": null,
"text": "Performance during testing: new corpusTo further explore the effectiveness of our approach, we have modified the experiment to show the performance in model adaptation. In the modified experiments the training and testing data sets are both taken from a foreign corpus (CityU), while our n-gram collection is still from ASBC. The relation between the derived features and the true segmentation may be different from the ASBC, and hence is learned by the classifiers. The results of the modified experiments are shown inTables 4 and 5.",
"content": "<table/>",
"html": null,
"type_str": "table"
},
"TABREF5": {
"num": null,
"text": "",
"content": "<table><tr><td>: Combined results (Bakeoff 3</td></tr><tr><td>dataset): traditional Chinese</td></tr></table>",
"html": null,
"type_str": "table"
},
"TABREF6": {
"num": null,
"text": "Bakeoff3 result for CityU, and over 27% improvement over best result for CKIP data.",
"content": "<table><tr><td>: Results (Bakeoff 3 dataset):</td></tr><tr><td>traditional Chinese</td></tr></table>",
"html": null,
"type_str": "table"
},
"TABREF7": {
"num": null,
"text": "Our OOV recall results compared to best performing systems in",
"content": "<table/>",
"html": null,
"type_str": "table"
}
}
}
} |