File size: 78,174 Bytes
6fa4bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
{
    "paper_id": "I05-1048",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:25:51.765112Z"
    },
    "title": "A Lexicon-Constrained Character Model for Chinese Morphological Analysis",
    "authors": [
        {
            "first": "Yao",
            "middle": [],
            "last": "Meng",
            "suffix": "",
            "affiliation": {},
            "email": "mengyao@frdc.fujitsu.com"
        },
        {
            "first": "Hao",
            "middle": [],
            "last": "Yu",
            "suffix": "",
            "affiliation": {},
            "email": "yu@frdc.fujitsu.com"
        },
        {
            "first": "Fumihito",
            "middle": [],
            "last": "Nishino",
            "suffix": "",
            "affiliation": {},
            "email": "nishino@frdc.fujitsu.com"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This paper proposes a lexicon-constrained character model that combines both word and character features to solve complicated issues in Chinese morphological analysis. A Chinese character-based model constrained by a lexicon is built to acquire word building rules. Each character in a Chinese sentence is assigned a tag by the proposed model. The word segmentation and partof-speech tagging results are then generated based on the character tags. The proposed method solves such problems as unknown word identification, data sparseness, and estimation bias in an integrated, unified framework. Preliminary experiments indicate that the proposed method outperforms the best SIGHAN word segmentation systems in the open track on 3 out of the 4 test corpora. Additionally, our method can be conveniently integrated with any other Chinese morphological systems as a post-processing module leading to significant improvement in performance.",
    "pdf_parse": {
        "paper_id": "I05-1048",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This paper proposes a lexicon-constrained character model that combines both word and character features to solve complicated issues in Chinese morphological analysis. A Chinese character-based model constrained by a lexicon is built to acquire word building rules. Each character in a Chinese sentence is assigned a tag by the proposed model. The word segmentation and partof-speech tagging results are then generated based on the character tags. The proposed method solves such problems as unknown word identification, data sparseness, and estimation bias in an integrated, unified framework. Preliminary experiments indicate that the proposed method outperforms the best SIGHAN word segmentation systems in the open track on 3 out of the 4 test corpora. Additionally, our method can be conveniently integrated with any other Chinese morphological systems as a post-processing module leading to significant improvement in performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Chinese morphological analysis is a fundamental problem that has been studied extensively [1] , [2] , [3] , [4] , [5] , [6] , [7] , [8] . Researchers make use of word or character features to cope with this problem. However, neither of them seems completely satisfactory.",
                "cite_spans": [
                    {
                        "start": 90,
                        "end": 93,
                        "text": "[1]",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 96,
                        "end": 99,
                        "text": "[2]",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 102,
                        "end": 105,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 108,
                        "end": 111,
                        "text": "[4]",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 114,
                        "end": 117,
                        "text": "[5]",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 120,
                        "end": 123,
                        "text": "[6]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 126,
                        "end": 129,
                        "text": "[7]",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 132,
                        "end": 135,
                        "text": "[8]",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In general, a simple word-based approach can achieve about 90% accuracy for segmentation with a medium-size dictionary. However, since no dictionary includes every Chinese word, the unknown word (or Out Of Vocabulary, OOV) problem [9] , [10] can severely affect the performance of word-based approaches. Furthermore, word-based models have an estimation bias when faced with segmentation candidates with different numbers of words. For , is related to the number of the words in the segmentation candidates. As such, a candidate with fewer words is preferred over those with more words in the selection process. Therefore, most wordbased models are likely to fail when a combinational ambiguity 1 sequence is separated into multiple words.",
                "cite_spans": [
                    {
                        "start": 231,
                        "end": 234,
                        "text": "[9]",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 237,
                        "end": 241,
                        "text": "[10]",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 432,
                        "end": 435,
                        "text": "For",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Compared with Chinese words, Chinese characters are relatively less unambiguous. The Chinese character set is very limited. Therefore, unknown characters occur rarely in a sentence. The grammatical advantages of characters have inspired researchers to adopt character features in Chinese morphology and parsing [5] , [6] , [11] , [12] . However, it is difficult to incorporate necessary word features, such as the form of a Chinese word and its fixed part-of-speech tags, in most character-based approaches. For this reason, character-based approaches have not achieved satisfactory performance in large-scale open tests.",
                "cite_spans": [
                    {
                        "start": 311,
                        "end": 314,
                        "text": "[5]",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 317,
                        "end": 320,
                        "text": "[6]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 323,
                        "end": 327,
                        "text": "[11]",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 330,
                        "end": 334,
                        "text": "[12]",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we propose a lexicon-constrained character model to combine the merits of both approaches. We explore how to capture the Chinese word building rules using a statistical method, which reflects the regularities in the word formation process. First, a character hidden Markov method assigns the candidate tags to each character. Next, a large-size word list combined with linguistic information is used to filter out erroneous candidates. Finally, segmentation and part-of-speech tagging for the sentence are provided based on the character tags.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The proposed model solves the problems of unknown word detection, word segmentation and part-of-speech tagging using both word and character features. Additionally, our module is a post-processing module, which can be coupled to any existing Chinese morphological system; and it can readily recall some of the unknown words omitted by the system, and as a result, significantly improves the overall performance. Evaluations of the proposed system on SIGHAN open test sets indicate that our method outperforms the best bakeoff results on 3 test sets, and ranks 2 nd in the 4 th test set [9] .",
                "cite_spans": [
                    {
                        "start": 586,
                        "end": 589,
                        "text": "[9]",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "It is recognized that there are some regularities in the process of forming words from Chinese characters. This in general can be captured by word building rules. In this paper, we explore a statistical model to acquire such rules. The following are some definitions used in the proposed model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "[Def. 1] character position feature We use four notations to denote the position of a character in a Chinese word. 'F' means the first character of the word, 'L' the last character, 'M' is a character within it and 'S' the word itself.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "[Def. 2] character tag set It is the product of the set of character position features and the set of part-ofspeech tags.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "Character tag set ={xy|",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "set word x POS \u2208 , } , , , { L M F S y \u2208 },",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "where, x denotes one part-of-speech (POS) tag and y a character position feature. Together they are used to define the rules of Chinese word formation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "[Def. 3] character tagging Given a Chinese sentence; character tagging is the process for assigning a character tag to each character in the sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "Word building rules are acquired based on the relation between the character and the corresponding character tag. Word segmentation and part-of-speech tagging can be achieved easily based on the result of character tagging. For example, a character with 'xS' is a single character word with the part-of-speech tag 'x'; a character sequence starting with 'xF' and ending with 'xL' is a multiple character word with the part-of-speech tag 'x'.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "The elementary model adopts the character bi-gram hidden Markov model. In hidden Markov model, given the sentence, ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "\u220f = \u2212 \u2212 \u00d7 = n i i i i i i xy c p xy xy xy p s t p , 1 1 2 ) | ( ) | ( ) | ( (1)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "The best character tagging result for the sentences is given by equation 2:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "\u220f = \u2212 \u2212 \u00d7 = n i i i i i i t xy c p xy xy xy p t , 1 1 2 * ) | ( ) | ( max arg (2)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "We used the People's Daily Corpus of 1998 [13] to train this model. Also we adopted a 100,000-word dictionary listing all valid part-of-speech tags for each Chinese word in the training phase to solve the data sparseness problem. The training data are converted into character tagging data through the following steps: a single character word with 'x' is converted into the character marked with tag 'xS'; a two-character word with 'x' is converted into a first character with 'xF' and a second character with 'xL'; a word with more than two characters with 'x' are converted into a first character with 'xF', middle characters with 'xM' and last character with 'xL'. We adopt the POS tag set from the People's Daily Corpus, which consists of 46 tags. Taking into account of the four position features, the final character tag set is comprised of 184 tags.",
                "cite_spans": [
                    {
                        "start": 42,
                        "end": 46,
                        "text": "[13]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "The emitted probability and transition probability of the model are estimated by the maximum likelihood method. The emitted probability is counted by the training Corpus and the dictionary, where the Chinese words in the dictionary are counted one time. The transition probability is trained from the training Corpus only.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Elementary Model to Describe Chinese Word Building Rules",
                "sec_num": "2.1"
            },
            {
                "text": "We tested the above model based on the SIGHAN open test set [9] . The average precision for word segmentation was more than 88%. This means that most of the word building rules in Chinese have been obtained by the elementary model. However, the performance was relatively inferior to other word segmentation systems. It indicated that the model needed more features to learn word building rules. In error analysis, we found that the elementary model was so flexible that it produced many pseudo-words and invalid part-of-speech tags. In practice, a Chinese word is a stable sequence of Chinese characters, whose formation and part-of-speech tags are fixed by long-term usage. It seemed that only character position and meaning cannot describe a word building rule effectively.",
                "cite_spans": [
                    {
                        "start": 60,
                        "end": 63,
                        "text": "[9]",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "We also observed that word segmentation systems based on a simple dictionary matching algorithm and a few linguistic rules could achieve about 90% accuracy [14] . This suggested that a lexicon may have contribution to word building rules. Thus, we tried to incorporate a lexicon to the model to improve the performance.",
                "cite_spans": [
                    {
                        "start": 156,
                        "end": 160,
                        "text": "[14]",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "The major errors in the elementary model were pseudo words and invalid part-ofspeech (POS) tags. We proposed two constraints based on the lexicon to deal with these errors:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "1. If a possible word produced from the elementary model is in the worddictionary, the character tag of the characters forming this word should be consistent with the part-of-speech tag of the word in the dictionary. 2. If a possible word produced is not in the dictionary, it must include one or more single characters, and none of which may be subsumed by any word in the dictionary in the current context.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "The first constraint eliminates invalid character tags. For example, the character '\u660e ' has six character tags: 'aF' (first in adjective) , 'dF' (first in adverb), 'nF' (first in noun), 'nrF' (first in person name), 'tF' (first in time), and 'vF' (first in verb). The character '\u5929' has five character tags: 'dL', 'nL', 'nrL', 'tL', and 'vL'. The combination of the two characters produces the possible word '\u660e\u5929', which includes five possible word part-of-speech tags: 'd', 'n', 'nr', 't', and 'v' based on these character tags. But '\u660e\u5929' is a word in the dictionary, which only has two valid part-of-speech tags, namely, 'time' and 'person name'. Obviously, the part-of-speech tags: 'd', 'n' and 'v' of '\u660e\u5929' are invalid. Accordingly, the tags 'aF', 'dF', 'nF' , 'vF' on '\u660e' and the tags 'dL', 'nL', 'vL' on '\u5929' are also invalid. So they should be pruned from the candidates of the character tagging.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "The second constraint prunes pseudo words in the elementary model. Many studies in dictionary-based segmentation treat unknown words as sequences of single characters [1] , [14] . The second constraint ensures that the new word produced by the elementary model must have one or more 'unattached' single characters (not subsumed by any other words). For example, the sequence '\u7a0b\u5e8f\u9519\u8bef' (program error) will combine the pseudo word '\u5e8f\u9519' because of the tag 'nF' on '\u5e8f' and the tag 'nL' on '\u9519'. The second constraint will prune '\u5e8f\u9519' since '\u7a0b\u5e8f' (program) and '\u9519\u8bef' (error) are already in the dictionary and there is no \"unattached\" single character in it. Accordingly, the tag 'nF' on '\u5e8f' and the tag 'nL' on '\u9519' will be deleted from the candidates of character tagging.",
                "cite_spans": [
                    {
                        "start": 167,
                        "end": 170,
                        "text": "[1]",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 173,
                        "end": 177,
                        "text": "[14]",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "The following experiments show the lexicon-based constraints are very effective in eliminating error cases. The elementary model faces an average of 9.3 character tags for each character. The constraints will prune 70% of these error tags from it. As a result, the performance of character tagging is improved.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "It is worth noting that the lexicon in the elementary model cannot distort the probability of the character tagging results in the model. The pruned cases are invalid cases which cannot occur in the training data because all the words and POS tags in the training data are valid. Thus, the model built from the training data is not affected by the pruning process.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "An Improved Character-Based Model Using Lexicon Constraints",
                "sec_num": "2.2"
            },
            {
                "text": "In this subsection, we illustrate the advantages of the proposed method for Chinese morphology with an example. Example:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Case Study",
                "sec_num": "2.3"
            },
            {
                "text": "\u5c0f \u660e \u660e \u5929 \u5c06 \u5c31 \u7a0b \u5e8f \u9519 \u8bef \u8fdb \u884c \u5206 \u6790",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Case Study",
                "sec_num": "2.3"
            },
            {
                "text": "(Xiaoming will analyze the program errors tomorrow).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Case Study",
                "sec_num": "2.3"
            },
            {
                "text": "Where, '\u5c0f\u660e' is an unknown word (person name), and the sequence '\u5c06\u5c31' is a combinational ambiguity (either '\u5c06\u5c31' (put up with) or '\u5c06'+ '\u5c31' (will)). Here is how our approach works.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Case Study",
                "sec_num": "2.3"
            },
            {
                "text": "Step 1: List all the character tags for each character. Figure 1 shows the character tags in the sequence '\u5c0f\u660e\u660e\u5929' . In this step we are able to find possible unknown words based on character position features. For example, the character tags in '\u5c0f\u660e\u660e\u5929' combine four possible unknown words: '\u5c0f\u660e', '\u5c0f\u660e\u660e', '\u660e\u660e\u5929' , and '\u5c0f\u660e\u660e\u5929'.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 56,
                        "end": 64,
                        "text": "Figure 1",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Case Study",
                "sec_num": "2.3"
            },
            {
                "text": "Step 2: Prune the invalid candidates using constraints. The first constraint prunes some invalid character tags. For example, '\u660e\u660e' can be either an adverb (d) or a personal name (nr); '\u660e\u5929' is a time (t) word. The other partof-speech tags of these two words will be deleted. With the second constraint, we can delete '\u660e\u660e\u5929' because '\u660e\u660e' and '\u660e\u5929' are words in the dictionary. However, '\u5c0f \u660e ' , '\u5c0f\u660e\u660e', and '\u5c0f\u660e\u660e\u5929' will be kept because '\u5c0f' is a \"unattached\" single character. The remaining candidates are shown in figure 2. Fig. 2 . Remaining Candidates for the sequence '\u5c0f\u660e\u660e\u5929'",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 518,
                        "end": 524,
                        "text": "Fig. 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Case Study",
                "sec_num": "2.3"
            },
            {
                "text": "Step 3: Choose the best character tagging result based on the proposed character hidden Markov model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "The best character tagging result is chosen using equation 2 in Section 2.1. The ambiguities in segmentation and word POS tagging are solved in the character tagging process.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "Consider the combinational ambiguity '\u5c06\u5c31' in the following 2 candidates:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "Candidate 1: '\u5c0f\u660e/nr \u660e \u5929 /t \u5c06 /d \u5c31 /d \u7a0b \u5e8f /n \u9519 \u8bef /n \u8fdb \u884c /v \u5206 \u6790 /v' Candidate 2: '\u5c0f\u660e/nr \u660e \u5929 /t \u5c06 \u5c31 /v \u7a0b \u5e8f /n \u9519 \u8bef /n \u8fdb \u884c /v \u5206 \u6790 /v'",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "In word-based linear model, the erroneous candidate 2 will be prior to the correct candidate 1 since the model counts 9 nodes in candidate 1 but 8 nodes in candidate 2. However, there is no such bias in the character model because the number of characters does not change. The combinational ambiguity '\u5c06\u5c31' will be denoted as '\u5c06/dS \u5c31 /dS' or '\u5c06/vF \u5c31 /vL'. The number of nodes in all candidates of character tagging is the same.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "At last, the correct result '\u5c0f/nrF",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "\u660e /nrL \u660e /tF \u5929 /tL \u5c06 /dS \u5c31 /dS\u7a0b/nF \u5e8f /nL \u9519 /nF \u8bef /nL \u8fdb /vF \u884c /vL \u5206 /vF \u6790 /",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "vL' is selected, and the corresponding morphological result is:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "'\u5c0f\u660e/nr \u660e \u5929 /t \u5c06 /d \u5c31 /d \u7a0b \u5e8f /n \u9519 \u8bef /n \u884c \u8fdb /v \u5206 \u6790 /v '.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "The above steps show the proposed approach solves the various issues related to Chinese morphology by a concise character tagging process where word building is revealed.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u5c0f",
                "sec_num": null
            },
            {
                "text": "We evaluated the proposed character method using the SIGHAN Backoff data, i.e. the one-month People's Daily Corpus of 1998, and the first version of Penn Chinese Treebank [15] . We compared our approach against two state-of-the-art systems: one is based on a bi-gram word segmentation model [7] , and the other based on a wordbased hidden Markov model [3] . For simplicity, we only considered three kinds of unknown words (personal name, location name, and organization name) in the all methods.",
                "cite_spans": [
                    {
                        "start": 171,
                        "end": 175,
                        "text": "[15]",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 291,
                        "end": 294,
                        "text": "[7]",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 352,
                        "end": 355,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments and Discussion",
                "sec_num": "3"
            },
            {
                "text": "The same corpus and word-dictionary were used to train the above three systems. The training data set was the 5-month People's Daily Corpus of 1998, which contained approximately 6,300,000 words and 46 word part-of-speech tags. The system dictionary contained 100,000 words and the valid part-of-speech tag(s) of each word. On average, there were 1.3 part-of-speech tags for a word in the dictionary.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments and Discussion",
                "sec_num": "3"
            },
            {
                "text": "In the following, chr-HMM refers to the proposed elementary model; chr-HMM+Dic refers to the character model improved by integrating linguistic information. W-Bigram is the word-based bi-gram system, and W-HMM is the word-based hidden Markov system.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments and Discussion",
                "sec_num": "3"
            },
            {
                "text": "We examined the performance of our model in comparison against W-Bigram and W-HMM. Table 1 compares the segmentation performance of our model against that of other models. Table 2 shows the accuracy in unknown word identification. Table 3 illustrates the performance of the part-of-speech tagging. The experiments in Table 1 and Table 2 were examined using the SIGHAN open test corpora. The experiments in Table 3 were performed again on the one-month People's Daily Corpus (PD corpus) and 4,000 sentences in the Penn Chinese Treebank (Penn CTB). We only examined 4 major word categories in the Penn Chinese Treebank due to inconsistency in the partof-speech tag sets between the two corpora. The 4 major word categories were: noun (shown as NN, NR in Penn CTB; n, nr, ns, nz in PD corpus), verb (VV in Penn CTB; v, vd, vn in PD corpus), adjective (JJ in Penn CTB; a, ad, an in PD corpus) and adverb (AD in Penn CTB; d in PD corpus).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 83,
                        "end": 90,
                        "text": "Table 1",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 172,
                        "end": 179,
                        "text": "Table 2",
                        "ref_id": "TABREF5"
                    },
                    {
                        "start": 231,
                        "end": 238,
                        "text": "Table 3",
                        "ref_id": null
                    },
                    {
                        "start": 317,
                        "end": 324,
                        "text": "Table 1",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 329,
                        "end": 336,
                        "text": "Table 2",
                        "ref_id": "TABREF5"
                    },
                    {
                        "start": 406,
                        "end": 413,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Morphological Experimental Results",
                "sec_num": "3.1"
            },
            {
                "text": "Segmentation and word POS tagging performance is measured in precision (P%), recall (R%) and F-score (F). Unknown words (NW) are those words not found in our word-dictionary, which include named entities and other new words. The unknown word rate (NW-Rate), the precision on unknown words (NW-Precision) and recall on total unknown words (NW-Recall) are given by: Table 1 shows that the above three systems achieve similar performances on the PK testing corpus. All of them were trained by the People's Daily corpus. For this reason, their performances were similar when the testing data had similar styles. But for other texts, the proposed character model performed much better than the wordbased models in both recall and precision. This indicated that our approach performed better for unseen data. Table 2 shows that our method for unknown word identification also outperforms the word-based method. We notice that word-based approaches and character-based approaches have similar precision on unknown word identification, however wordbased approaches have much lower recall than character-based ones. The main reason for this is that word-based systems focus only on unknown words with proper word structures, but cannot recognize newly generated words, rare words, and other new words unlisted in the dictionary. A very high proportion of these types of unknown word in the SIGHAN testing data affects the recall of the word-based methods on unknown words. The experiments reveal that our method could effectively identify all kinds of new words. This is because our model has defined word building rules for all kinds of words.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 364,
                        "end": 371,
                        "text": "Table 1",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 803,
                        "end": 810,
                        "text": "Table 2",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Morphological Experimental Results",
                "sec_num": "3.1"
            },
            {
                "text": "NW-Rate=",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Experimental Results",
                "sec_num": "3.1"
            },
            {
                "text": "Without a widely recognized testing standard, it is very hard to evaluate the performance on part-of-speech tagging. The results in Penn Chinese Treebank was better than that in the People's Daily Corpus since we examined all 42 POS tags in the People's Daily Corpus, but we only tested four major POS tags in Penn Chinese Treebank. Our approach is better than the word-based method for two test data sets. However, we could not conclude that our method was superior to the word-based method because of the limited testing approaches and testing data. A thorough empirical comparison among different approaches should be investigated in the future. From Table 1 and Table 3 , we notice that chr-HMM achieved 88% accuracy in word segmentation and 80% in part-of-speech tagging without a word-dictionary. Chr-HMM is a state-of-the-art Chinese morphology system without a worddictionary. Its performance is comparable to some dictionary-based approaches (e.g., forward-maximum). This result indicates that our model has effectively captured most of the Chinese word building rules.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 654,
                        "end": 661,
                        "text": "Table 1",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 666,
                        "end": 673,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Morphological Experimental Results",
                "sec_num": "3.1"
            },
            {
                "text": "The results also show that chr-HMM+Dic outperformed the best SIGHAN word segmentation system on 3 out of the 4 SIGHAN open track test corpora, and achieved top 2 in the case of HK testing corpus.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Experimental Results",
                "sec_num": "3.1"
            },
            {
                "text": "The advantage of the proposed model is proficiency in describing word building rules and since many existing NLP application systems are weak in identifying new words, it is intuitive to integrate our model to existing systems and serves as a postprocessing subsystem. In this subsection, we show how existing word segmentation systems could be improved using chr-HMM.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "Given a segmentation result, we assume that unidentified new words may be a sequence of unattached characters. That is, all multiple-character words in the given result are considered correct, while single words, which might include unidentified new words will be rechecked by the chr-HMM. The entire process involves 3 steps:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "1. Only character tags that are consistent with the position of the character in the word are listed for multi-character words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "Consider an original result:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "\u4e54 \u4e39 [\u6628 \u65e5 \u4ece \u8c37 \u5e95 ] \u5f3a \u529b [\u53cd \u5f39",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "] (Jordan bounced back strongly from the bottom yesterday)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "The parts in brackets are the sequence of single characters where the new words may appear. The chr-HMM will list all possible character tags for these \"unattached\" characters. The parts outside the brackets are multiple-character words identified by the original system. They are assumed correct and maintain also positional information. Only the character tags, which are consistent with the positions of the character in the word are listed. The character tagging candidates for the above sample is given in Figure 3 : Chr-HMM is then applied to the character tagging candidates and the best character tagging selected based on the probability of the candidates is output as the result. In this example, the result is:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 511,
                        "end": 519,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "\u2026 \u2026",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "'\u4e54\u4e39 (Jordan) \u6628 \u65e5 (yesterday) \u4ece (from) \u8c37 \u5e95 (earth) \u5f3a \u529b (strongly) \u53cd \u5f39 (bound)'.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "The three missing new words in the original system are identified by this post-processing subsystem.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "We re-assigned the word segmentation results for all participants who have given permission to release data from the SIGHAN site (available for download from http://www.sighan.org/bakeoff2003 ). Table 4 enlists the performance of SIGHAN open test with and without chr-HMM. The participant numbers correspond to the sites listed in [9] . From Table 4 , it is obvious that word segmentation precision increases significantly, and at the same time, the corresponding recall remains the same or slightly declined. This implies that the chr-HMM retains the correct words by the original system and concurrently decreases significantly its errors.",
                "cite_spans": [
                    {
                        "start": 331,
                        "end": 334,
                        "text": "[9]",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 195,
                        "end": 202,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    },
                    {
                        "start": 342,
                        "end": 349,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    }
                ],
                "eq_spans": [],
                "section": "Incorporation with Other Systems",
                "sec_num": "3.2"
            },
            {
                "text": "Although character features are very important in Chinese morphology, research in character-based approach is unpopular. Chooi-Ling Goh et al. [16] , Jianfeng Gao et al. [8] and Huaping Zhang [3] adopted character information to handle unknown words; X. Luo [11] , Yao Meng [12] and Shengfen Luo [17] each presented characterbased parsing models for Chinese parsing or new-word extraction. T. Nakagawa used word-level information and character-level information for word segmentation [6] . Hwee Tou Ng et al. [5] investigated word-based and character-based approaches and proposed a maximum entropy character-based POS analyzer. Although the character tags proposed in this paper are essentially similar to some of the previous work mentioned above, here our focus is to integrate various word features with the characterbased model in such a way that the probability of the model is undistorted. The proposed model is effective in acquiring word building rules. To our knowledge, our work is the first character-based approach, which outperforms the word-based approaches for SIGHAN open test. Also, our approach is versatile and can be easily integrated with existing morphological systems to achieve improved performance.",
                "cite_spans": [
                    {
                        "start": 143,
                        "end": 147,
                        "text": "[16]",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 170,
                        "end": 173,
                        "text": "[8]",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 192,
                        "end": 195,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 258,
                        "end": 262,
                        "text": "[11]",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 274,
                        "end": 278,
                        "text": "[12]",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 296,
                        "end": 300,
                        "text": "[17]",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 484,
                        "end": 487,
                        "text": "[6]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 509,
                        "end": 512,
                        "text": "[5]",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "4"
            },
            {
                "text": "A lexicon-constrained character model is proposed to capture word building rules using word features and character features. The combination of word and character features improves the performance of word segmentation and part-of-speech tagging. The proposed model can solve complicated issues in Chinese morphological analysis. The Chinese morphological analysis is generalized into a process of specific character tagging and word filtering. A lexicon supervises the character-based model to eliminate invalid character tagging candidates.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Works",
                "sec_num": "5"
            },
            {
                "text": "Our system outperformed the best SIGHAN word segmentation system in 3 out of the 4 SIGHAN open test sets. To our knowledge, our work is the first character-based approach, which performs better than word-based approaches for SIGHAN open test. In addition, the proposed method is versatile and can be easily integrated to any existing Chinese morphological system as a post-processing subsystem leading to enhanced performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Works",
                "sec_num": "5"
            },
            {
                "text": "In this paper, we focused on word features in character-based mode, and adopted HMM as the statistical model to identify the rules. Other statistical models, such as maximum entropy, boosting, support vector machine, etc., may also be suitable for this application. They are worth investigating. The data sparseness problem is practically non-existent in the character-based model for the Chinese character set is limited. However, odd characters are occasionally found in Chinese personal or place names. Some rules using named entity identification technique may help smoothen this. In a broader view, the word building rules proposed in our model is simple enough for linguistic studies to better understand for example formation of Chinese words or even the Chinese language itself.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Works",
                "sec_num": "5"
            },
            {
                "text": "\u5317 \u4eac \u5927 \u5b66 \u73b0 \u4ee3 \u6c49 \u8bed \u8bed \u6599 \u5e93 \u57fa \u672c \u52a0 \u5de5 \u89c4 \u8303 . \u4e2d \u6587 \u4fe1 \u606f \u5b66 \u62a5",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Works",
                "sec_num": "5"
            },
            {
                "text": "A typical segmentation ambiguity, it refers to a situation in which the same Chinese sequence may be one word or several words in different contexts.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": ". The unattached characters are tagged with all possible character tags. In this way, the original segmentation result is converted into a group of character tagging candidates. 3. We then input these character tagging candidates into the chr-HMM to select the best one.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Chinese Word Segmentation in MSR-NLP",
                "authors": [
                    {
                        "first": "Andi",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. of SIGHAN Workshop",
                "volume": "",
                "issue": "",
                "pages": "127--175",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andi Wu. Chinese Word Segmentation in MSR-NLP. In Proc. of SIGHAN Workshop, Sapporo, Japan, (2003) 127-175",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "A Chinese Efficient Analyzer Integrating Word Segmentation, Part-Of-Speech Tagging, Partial Parsing and Full Parsing",
                "authors": [
                    {
                        "first": "Guodong",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Jian",
                        "middle": [],
                        "last": "Su",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. Of SIGHAN Workshop",
                "volume": "",
                "issue": "",
                "pages": "78--83",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "GuoDong Zhou and Jian Su. A Chinese Efficient Analyzer Integrating Word Segmenta- tion, Part-Of-Speech Tagging, Partial Parsing and Full Parsing. In Proc. Of SIGHAN Workshop, Sapporo, Japan, (2003) 78-83",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "HHMM-based Chinese Lexical Analyzer ICTCLAS",
                "authors": [
                    {
                        "first": "Huaping",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Hong-Kui",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. Of SIGHAN Workshop",
                "volume": "",
                "issue": "",
                "pages": "184--187",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Huaping Zhang, Hong-Kui Yu et al.. HHMM-based Chinese Lexical Analyzer ICTCLAS. In Proc. Of SIGHAN Workshop, Sapporo, Japan, (2003) 184-187",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Chinese Word Segmentation as LMR Tagging",
                "authors": [
                    {
                        "first": "Nianwen",
                        "middle": [],
                        "last": "Xue",
                        "suffix": ""
                    },
                    {
                        "first": "Libin",
                        "middle": [],
                        "last": "Shen",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. Of SIGHAN Workshop",
                "volume": "",
                "issue": "",
                "pages": "176--179",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nianwen Xue and Libin Shen. Chinese Word Segmentation as LMR Tagging. In Proc. Of SIGHAN Workshop, Sapporo, Japan, (2003) 176-179",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Chinese Part-of-Speech Tagging: One-at-a-Time or All-at-Once? Word-Based or Character-Based?",
                "authors": [
                    {
                        "first": "",
                        "middle": [],
                        "last": "Hwee Tou Ng",
                        "suffix": ""
                    },
                    {
                        "first": "Jin",
                        "middle": [],
                        "last": "Low",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kiat",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proc. of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "277--284",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hwee Tou Ng, Low, Jin Kiat. Chinese Part-of-Speech Tagging: One-at-a-Time or All-at- Once? Word-Based or Character-Based? In Proc. of EMNLP, Barcelona, Spain, (2004) 277-284",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Chinese and Japanese Word Segmentation Using Word-level and Character-level Information",
                "authors": [
                    {
                        "first": "Tetsuji",
                        "middle": [],
                        "last": "Nakagawa",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proc. of the 20 th COLING",
                "volume": "",
                "issue": "",
                "pages": "466--472",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tetsuji Nakagawa. Chinese and Japanese Word Segmentation Using Word-level and Character-level Information, In Proc. of the 20 th COLING, Geneva, Switzerland, (2004) 466-472",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "A Two-stage Statistical Word Segmentation System for Chinese",
                "authors": [
                    {
                        "first": "Guohong",
                        "middle": [],
                        "last": "Fu",
                        "suffix": ""
                    },
                    {
                        "first": "Kang-Kwong",
                        "middle": [],
                        "last": "Luke",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. Of SIGHAN Workshop",
                "volume": "",
                "issue": "",
                "pages": "156--157",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Guohong Fu and Kang-Kwong Luke. A Two-stage Statistical Word Segmentation System for Chinese. In Proc. Of SIGHAN Workshop, Sapporo, Japan, (2003) 156-157",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Adaptive Chinese Word Segmentation",
                "authors": [
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "Andi",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Chang-Ning",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proc. of 42 nd ACL",
                "volume": "",
                "issue": "",
                "pages": "462--469",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jianfeng Gao, Andi Wu, Chang-Ning Huang et al. Adaptive Chinese Word Segmentation. In Proc. of 42 nd ACL. Barcelona, Spain, (2004) 462-469",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "The First International Chinese Word Segmentation Bakeoff",
                "authors": [
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Sproat",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Emerson",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. Of SIGHAN Workshop",
                "volume": "",
                "issue": "",
                "pages": "133--143",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Richard Sproat and Thomas Emerson. The First International Chinese Word Segmentation Bakeoff. In Proc. Of SIGHAN Workshop, Sapporo, Japan, (2003) 133-143",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "A Maximum Entropy Chinese Character-based Parser",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Luo",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. of EMNLP. Sapporo",
                "volume": "",
                "issue": "",
                "pages": "192--199",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "X. Luo. A Maximum Entropy Chinese Character-based Parser. In Proc. of EMNLP. Sap- poro, Japan, (2003) 192-199",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "A Chinese Dictionary Construction Algorithm for Information Retrieval",
                "authors": [
                    {
                        "first": "Honglan",
                        "middle": [],
                        "last": "Jin",
                        "suffix": ""
                    },
                    {
                        "first": "Kam-Fai",
                        "middle": [],
                        "last": "Wong",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "ACM Transactions on Asian Language Information Processing",
                "volume": "1",
                "issue": "4",
                "pages": "281--296",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Honglan Jin, Kam-Fai Wong, \"A Chinese Dictionary Construction Algorithm for Informa- tion Retrieval\", ACM Transactions on Asian Language Information Processing, 1(4):281- 296, Dec. 2002.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Chinese New Word Identification Based on Character Parsing Model",
                "authors": [
                    {
                        "first": "Yao",
                        "middle": [],
                        "last": "Meng",
                        "suffix": ""
                    },
                    {
                        "first": "Hao",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    },
                    {
                        "first": "Fumihito",
                        "middle": [],
                        "last": "Nishino",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proc. of 1 st IJCNLP",
                "volume": "",
                "issue": "",
                "pages": "489--496",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yao Meng, Hao Yu and Fumihito Nishino. 2004. Chinese New Word Identification Based on Character Parsing Model. In Proc. of 1 st IJCNLP, Hainan, China, (2004) 489-496",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Ambiguity Resolution in Chinese Word Segmentation",
                "authors": [
                    {
                        "first": "Maosong",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "K. T'",
                        "middle": [],
                        "last": "Benjamin",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Sou",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Proc. of 10 th Pacific Asia Conference on Language, Information & Computation",
                "volume": "",
                "issue": "",
                "pages": "121--126",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Maosong Sun and Benjamin K. T' Sou. Ambiguity Resolution in Chinese Word Segmenta- tion. In Proc. of 10 th Pacific Asia Conference on Language, Information & Computation, (1995) 121-126",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Building a Large-scale Annotated Chinese Corpus",
                "authors": [
                    {
                        "first": "Nianwen",
                        "middle": [],
                        "last": "Xue",
                        "suffix": ""
                    },
                    {
                        "first": "Fu-Dong",
                        "middle": [],
                        "last": "Chiou",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proc. of the 19 th COLING",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nianwen Xue, Fu-Dong Chiou and Martha Palmer. Building a Large-scale Annotated Chinese Corpus. In Proc. of the 19 th COLING. Taibei, Taiwan, (2002)",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Chinese Unknown Word Identification Using Character-based Tagging and Chunking",
                "authors": [
                    {
                        "first": "Goh",
                        "middle": [],
                        "last": "Chooi-Ling",
                        "suffix": ""
                    },
                    {
                        "first": "Masayuki",
                        "middle": [],
                        "last": "Asahara",
                        "suffix": ""
                    },
                    {
                        "first": "Yuji",
                        "middle": [],
                        "last": "Matsumoto",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. of the 41 st ACL, Interactive Poster/Demo Sessions",
                "volume": "",
                "issue": "",
                "pages": "197--200",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chooi-Ling GOH, Masayuki Asahara, Yuji Matsumoto. Chinese Unknown Word Identifi- cation Using Character-based Tagging and Chunking. In Proc. of the 41 st ACL, Interac- tive Poster/Demo Sessions, Sapporo, Japan, (2003) 197-200",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Two-character Chinese Word Extraction Based on Hybrid of Internal and Contextual Measure",
                "authors": [
                    {
                        "first": "Shengfen",
                        "middle": [],
                        "last": "Luo",
                        "suffix": ""
                    },
                    {
                        "first": "Maosong",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proc. of the 2 nd SIGHAN Workshop, Sapporo",
                "volume": "",
                "issue": "",
                "pages": "20--30",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shengfen Luo, Maosong Sun. 2003, Two-character Chinese Word Extraction Based on Hybrid of Internal and Contextual Measure, In Proc. of the 2 nd SIGHAN Workshop, Sap- poro, Japan, (2003) 20-30",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "uris": null,
                "num": null,
                "text": "of result t of s is estimated as:"
            },
            "FIGREF1": {
                "type_str": "figure",
                "uris": null,
                "num": null,
                "text": "Candidates for the sequence '\u5c0f\u5c0f\u5c0f\u5c0f'"
            },
            "TABREF0": {
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td>T</td><td>*</td><td>=</td><td>T max arg</td><td>T ( p</td><td>|</td><td>W )</td><td>=</td><td>\u220f = i 1 n max T arg</td><td>i w ( p</td><td>|</td><td>i t</td><td>)</td><td>i t ( p</td><td>|</td><td>i t ... 1 t</td><td>\u2212 1</td><td>)</td></tr></table>",
                "num": null,
                "text": "example, in the standard hidden Markov model, the best result,"
            },
            "TABREF4": {
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td/><td>PK</td><td/><td>CTB</td><td/><td>HK</td><td/><td>AS</td></tr><tr><td/><td>R%/ P%</td><td>F</td><td>R%/ P%</td><td>F</td><td>R%/ P%</td><td>F</td><td>R%/ P%</td><td>F</td></tr><tr><td>Chr-HMM</td><td colspan=\"5\">91.9/91.8 91.8 86.9/87.3 87.1 87.7/86.7</td><td>87.2</td><td>89.9/89.1</td><td>89.5</td></tr><tr><td colspan=\"6\">Chr-HMM+Dic 95.9/96.7 96.3 92.7/93.5 93.1 91.1/91.9</td><td>91.5</td><td>92.3/93.9</td><td>93.1</td></tr><tr><td>W-Bigram</td><td colspan=\"5\">94.7/95.4 95.1 87.4/86.8 87.1 88.7/83.7</td><td>86.3</td><td>87.9/85.1</td><td>86.5</td></tr><tr><td>W-HMM</td><td colspan=\"5\">94.6/95.1 94.9 88.6/89.2 88.9 90.7/89.1</td><td>89.9</td><td>90.7/87.2</td><td>89.0</td></tr><tr><td colspan=\"6\">Rank 1 in SIG 96.3/95.6 96.0 91.6/90.7 91.2 95.8/95.4</td><td>95.6</td><td>91.5/89.4</td><td>90.5</td></tr><tr><td colspan=\"6\">Rank 2 in SIG 96.3/94.3 95.3 91.1/89.1 90.1 90.9/86.3</td><td>88.6</td><td>89.2/85.3</td><td>87.3</td></tr></table>",
                "num": null,
                "text": "Comparison of word segmentation based on SIGHAN open test sets"
            },
            "TABREF5": {
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td/><td/><td>PK</td><td/><td>CTB</td><td/><td/><td>HK</td><td>AS</td></tr><tr><td colspan=\"8\">Chr-HMM UWR% P% R% UWR% P% R% UWR% P% R% UWR% P% R%</td></tr><tr><td colspan=\"2\">Chr-HMM+Dic 2.3</td><td colspan=\"4\">56.2 54.8 10.4 68.8 64.4</td><td>9.7</td><td>61.4 58.4</td><td>8</td><td>65.4 62.9</td></tr><tr><td>W-Bigram</td><td>2.3</td><td colspan=\"4\">54.7 53.6 10.4 53.9 23.8</td><td>9.7</td><td>53.0 29.6</td><td>8</td><td>64.6 35.3</td></tr><tr><td>W-HMM</td><td>2.3</td><td colspan=\"4\">58.1 51.3 10.4 68.3 37.2</td><td>9.7</td><td>62.3 40.7</td><td>8</td><td>68.4 41.1</td></tr><tr><td/><td/><td colspan=\"6\">Table 3. Comparison of word part-of-speech tagging</td></tr><tr><td/><td/><td/><td/><td colspan=\"2\">People Daily</td><td/><td>Penn CTB</td></tr><tr><td/><td/><td/><td>P%</td><td colspan=\"3\">R% F-score P%</td><td>R%</td><td>F-score</td></tr><tr><td/><td colspan=\"2\">Chr-HMM</td><td colspan=\"2\">82.4% 82.5%</td><td colspan=\"3\">82.5 89.7% 88.5%</td><td>89.1</td></tr><tr><td/><td colspan=\"3\">Chr-HMM+Dic 89.3</td><td>87.8</td><td>88.6</td><td colspan=\"2\">92.5 91.5</td><td>92.0</td></tr><tr><td/><td colspan=\"2\">W-HMM</td><td colspan=\"2\">86.2% 85.4%</td><td colspan=\"3\">85.7 91.1% 90.8%</td><td>91.0</td></tr></table>",
                "num": null,
                "text": "Accuracy of unknown word identification for SIGHAN open test sets"
            },
            "TABREF7": {
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td>Corpus</td><td>Site</td><td/><td>R%</td><td>P%</td><td>F</td></tr><tr><td>AS</td><td>03</td><td>Before After</td><td>89.2 90.8</td><td>85.3 92.0</td><td>87.2 91.4</td></tr><tr><td/><td>01</td><td>Before After</td><td>88.7 90.1</td><td>87.6 91.8</td><td>88.1 90.9</td></tr><tr><td>CTB</td><td>03</td><td>Before After</td><td>85.3 86.4</td><td>80.6 87.8</td><td>82.9 87.1</td></tr><tr><td/><td>10</td><td>Before After</td><td>91.1 91.0</td><td>89.1 93.5</td><td>90.1 92.3</td></tr><tr><td>HK</td><td>03</td><td>Before After</td><td>90.9 89.4</td><td>86.3 91.0</td><td>88.6 90.2</td></tr><tr><td>PK</td><td>03 10</td><td>Before After Before After</td><td>94.1 94.4 96.3 95.6</td><td>91.1 95.3 95.6 97.7</td><td>92.5 94.9 95.9 96.7</td></tr></table>",
                "num": null,
                "text": "Comparison of results with and without chr-HMM"
            }
        }
    }
}