File size: 73,511 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
{
    "paper_id": "O12-1006",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:02:55.078003Z"
    },
    "title": "Associating Collocations with WordNet Senses Using Hybrid Models",
    "authors": [
        {
            "first": "Yi-Chun",
            "middle": [],
            "last": "Chen",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "National Tsing Hua University",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Tzu-Xi",
            "middle": [],
            "last": "Yen",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "National Tsing Hua University",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Jason",
            "middle": [
                "S"
            ],
            "last": "Chang",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "National Tsing Hua University",
                "location": {}
            },
            "email": "jason.jschang@gmail.com"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In this paper, we introduce a hybrid method to associate English collocations with sense class members chosen from WordNet. Our combinational approach includes a learning-based method, a paraphrase-based method and a sense frequency ranking method. At training time, a set of collocations with their tagged senses is prepared. We use the sentence information extracted from a large corpus and cross-lingual information to train a learning-based model. At run time, the corresponding senses of an input collocation will be decided via majority voting. The three outcomes participated in voting are as follows: 1. the result from a learning-based model; 2. the result from a paraphrase-based model; 3. the result from sense frequency ranking method. The sense with most votes will be associated with the input collocation. Evaluation shows that the hybrid model achieves significant improvement when comparing with the other method described in evaluation time. Our method provides more reliable result on associating collocations with senses that can help lexicographers in",
    "pdf_parse": {
        "paper_id": "O12-1006",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In this paper, we introduce a hybrid method to associate English collocations with sense class members chosen from WordNet. Our combinational approach includes a learning-based method, a paraphrase-based method and a sense frequency ranking method. At training time, a set of collocations with their tagged senses is prepared. We use the sentence information extracted from a large corpus and cross-lingual information to train a learning-based model. At run time, the corresponding senses of an input collocation will be decided via majority voting. The three outcomes participated in voting are as follows: 1. the result from a learning-based model; 2. the result from a paraphrase-based model; 3. the result from sense frequency ranking method. The sense with most votes will be associated with the input collocation. Evaluation shows that the hybrid model achieves significant improvement when comparing with the other method described in evaluation time. Our method provides more reliable result on associating collocations with senses that can help lexicographers in",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "compilation of collocations dictionaries and assist learners to understand collocation usages.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "A collocation is a pair of words that co-occur with more frequency than random. A collocation usually contains a base word (e.g., \"oil\" in fuel oil) and a collocate (e.g., \"fuel\" in fuel oil). In a collocations dictionary, we can find many collocates of a base word (e.g., fuel oil, motor oil, peanut oil, salad oil). Some collocations dictionaries show the collocates for all senses, while other collocations dictionaries present the collocates by senses of a base word so learners can better grasp the usage of a collocation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Determining the set of broad senses to classify collocations is not an easy task. Researches have used thesaurus topics such as Roget's (Yarowsky, 1992) or arbitrarily top-level WordNet senses as classes. There are 44 semantic classes called lexicographer-files and each synset in WordNet is assigned to one lexicographer-file. There are 26 lexicographer-files (or supersenses), which can be used to tag common nouns. Consider the word \"oil\" which can be used as fuel/to make machines work smoothly, or as belonging to the noun.substance supersense and used in cooking could be seen as belonging to the noun.food supersense.",
                "cite_spans": [
                    {
                        "start": 136,
                        "end": 152,
                        "text": "(Yarowsky, 1992)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we present a hybrid model that automatically associated a given collocation with the corresponding supersense. The hybrid model is composed of a learning-based method, a paraphrase-based method and a sense frequency ranking method. The output supersense of a collocation is decided via majority vote of the above three methods.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "At training time, we need some collocations tagged with supersenses as seeds. There are a huge number of collocations in WordNet, so we can use those collocation and supersense pairs to train the model. Sentences containing the input collocations extracted from a large corpus and Chinese translation of the collocations are used as features of the model. We will descript the training process in more details in Chapter 3. An example procedure for associating the collocation of fuel oil with a supersense noun.substance is shown in Figure 1 . We extract sentences containing the input collocation from a corpus and take the sentences and Chinese translation as features. Then, we use the pre-trained machine learning model to predict the supersense. Second, we use the words similarity and words dependency relations to paraphrase the base word. Then, we calculate the WordNet similarity of base word and the paraphrases to identify the supersense. Third, we simply list the lexicographer-files of the input collocation base word and choose the first one as the supersense since the order of the list corresponds to the sense frequency of that word. At last, a relative majority vote for the three results determines the final output.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 534,
                        "end": 542,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The experimental results show that our hybrid method can automatically associate collocations with supersenses with a higher performance than the baseline method. The results can also be used to help lexicographers in compilation of a collocations dictionary. Furthermore, learners could understand the usage of collocations in a specific sense.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Associating collocations with supersenses in WordNet is similar to Word Sense Disambiguation (WSD), the process of identifying the meaning of a specific word in a given context. In this paper, we address a special case of disambiguating the headword of a given collocation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Previous work in WSD is mostly based on some kind of machine learning models. Hearst (1991) uses a set of orthographic, syntactic and lexical features to train large text corpora and disambiguates noun homographs. Yarowsky (1992) uses Na\u00efve Bayesian model to train large corpora to disambiguate words to Roget's Thesaurus categories. Leacock, Towell and Voorhees (1993) bases on Bayesian decision theory, neural networks and content vectors to train the knowledge about patterns of words co-occurrences and disambiguates words to WordNet senses. The main disadvantage is that the demand of annotated training data which are time-consuming and labor intensive to obtain.",
                "cite_spans": [
                    {
                        "start": 214,
                        "end": 229,
                        "text": "Yarowsky (1992)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "In a work more closely to our research, Inumella, Kilgarriff and Kovar (2009) try to assign the collocations for a word that automatically identified from a large corpus to its distinct senses. Their short term goal is to generate a new English collocations dictionary (Macmillan Collocation Dictionary). Most of the previous works focus on words level, while this research focuses on collocations. We describe two of their automatic approaches: Thesaurus method and Yarowsky's method (1995) . The thesaurus method works on the promise that a sense shares its collocates with its thesaurus class members. For example, consider a thesaurus class with six members {cricket, butterfly, leech, worm, bee, queen}, they extract collocates such as young, fly, feed, breed that at least appear in two class members and insert them to that sense. Another method is Yarowsky's method, which relies on the heuristic of \"one sense per collocation\" (Yarowsky, 1993) and \"one sense per discourse\" (Gale, Church and Yarowsky, 1992) . The algorithm first collects some seed collocations with senses by dictionary-parsing and uses supervised classification algorithms for training and labeling. Then they add new labeled collocations to training set and repeat labeling. Finally, they use a decision list algorithm to terminate.",
                "cite_spans": [
                    {
                        "start": 485,
                        "end": 491,
                        "text": "(1995)",
                        "ref_id": null
                    },
                    {
                        "start": 936,
                        "end": 952,
                        "text": "(Yarowsky, 1993)",
                        "ref_id": null
                    },
                    {
                        "start": 983,
                        "end": 1016,
                        "text": "(Gale, Church and Yarowsky, 1992)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "In contrast to previous works in Word Sense Disambiguation and semantic classification, we present a hybrid system that automatically associates collocations to supersenses using a learning-based method, a paraphrase-based method and a sense frequency ranking method, with the goal to help lexicographers in compilation of collocation dictionaries and help learners to better grasp the usage of a collocation. We describe the method in more details in the next chapter.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Associating collocations (e.g., required course) with dictionary senses often does not work very well. To obtain a better performance, we introduce a learning-based method using context and cross-lingual features, a paraphrase-based method using words similarity relation and dependency relation, and a sense frequency ranking method.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Method",
                "sec_num": "3"
            },
            {
                "text": "We focus on automatically associating collocations with corresponding supersenses. The output senses could be used by lexicographers to save effort in compile collocations dictionaries and learners can better grasp the usage of a collocation. Supersenses are 26 lexicographer-files in WordNet noun hierarchy chosen by lexicographers and are believed to be general enough for sense allocation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Statement",
                "sec_num": "3.1"
            },
            {
                "text": "In this section, we explain our approaches to find the supersense including a learning-based method, a paraphrase-based method and a sense frequency ranking method. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Sense Assignment Models",
                "sec_num": "3.2"
            },
            {
                "text": "In the first stage (Step (1) in Figure 2 ), we attempt to find a set of collocations and their pre-tagged supersenses pairs as seeds collocations to train a machine learning model \u202b\u072f\u202c from WordNet. For example, the supersense for a collocation fuel oil is noun.substance. Examples of collocation and supersense pairs extracted from WordNet are shown in Figure 3 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 32,
                        "end": 40,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    },
                    {
                        "start": 353,
                        "end": 361,
                        "text": "Figure 3",
                        "ref_id": "FIGREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Generating Collocation and Supersense Pairs",
                "sec_num": "3.2.1"
            },
            {
                "text": "We use two heuristics to achieve this goal. First, we go through each hyponyms of noun synsets and examine their lemma names to find collocations. For example, consider a synset Synset('discharge.n.01'), one of its lemma name is discharge and one of its hyponyms is Synset('electrical_discharge.n.01') with a lemma name electrical_discharge. Since the base word of electrical_discharge matches Synset('discharge.n.01')'s lemma name discharge, we can take electrical discharge as a collocation and the lexicographer-file of Synset('discharge.n.01') noun.event as a supersense to form the < \u202b,\u074a\u0745\u0750\u073d\u073f\u0748\u0748\u073f\u202c \u202b\u0741\u074f\u074a\u0741\u074f\u074e\u0741\u0751\u074f\u202c > pair, (electronic discharge, noun.event).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generating Collocation and Supersense Pairs",
                "sec_num": "3.2.1"
            },
            {
                "text": "Second, we search the collocations from definitions and example sentences of each noun synset. We utilize a parser to generate part-of-speech and lemma for and . For a noun synset , one of its lemma name is , and the definition or example \u0745 as one of our selected < \u202b,\u074a\u0745\u0750\u073d\u073f\u0748\u0748\u073f\u202c \u202b\u0741\u074f\u074a\u0741\u074f\u074e\u0741\u0751\u074f\u202c > pair. For example, a synset Synset('day.n.05') has one lemma name day and one example sentence \"it was a busy day on the stock exchange\". So we can take busy day as a collocation and the lexicographer-file noun.time as a supersense to form the < \u202b,\u074a\u0745\u0750\u073d\u073f\u0748\u0748\u073f\u202c \u202b\u0741\u074f\u074a\u0741\u074f\u074e\u0741\u0751\u074f\u202c > pair, (busy day, noun.time).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generating Collocation and Supersense Pairs",
                "sec_num": "3.2.1"
            },
            {
                "text": "In the second stage (Step (2) in Figure 2 ), we use the collocation and supersense pairs obtained in section 3.2.1 to find sentences to train a sense classifier. First, a parser is used for generating part-of-speech tag and lemma form for all sentences in the monolingual corpus \u202b\u202c and search from on-line machine translating system \u202b\u0736\u072f\u202c for Chinese collocation translation.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 33,
                        "end": 41,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Training Machine Learning Model",
                "sec_num": "3.2.2"
            },
            {
                "text": "For example, consider the collocation required course and its supersense noun.act. We can find sentence such as \"A required course for all students, to be completed before the end of the third year, and to be examined by individual colleges\" from \u202b\u0725\u072f\u202c and its Chinese collocation translation \"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Machine Learning Model",
                "sec_num": "3.2.2"
            },
            {
                "text": "\" from on-line translation resource. The base word course has 6 different supersenses, but the words like students, third year, examined, colleges are highly related to the collocation required course and the supersense noun.act rather than other supersenses such as noun.food, noun.artifact or noun.object. The Chinese translation provides cross-lingual information like \" \" to disambiguate the sense of course. The other translation for course like \" \" or \" \" would lead to different supersenses.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Machine Learning Model",
                "sec_num": "3.2.2"
            },
            {
                "text": "The input to this stage is a set of features. The above example required course showed that context words of a collocation may contain some words highly related to the corresponding supersense and cross-lingual information for a collocation also helps to disambiguate the supersense. So the features we use for one training event are ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Machine Learning Model",
                "sec_num": "3.2.2"
            },
            {
                "text": "In the third stage (Step (3) in Figure 2 ), we attempt to predict the supersense for the input collocation \u202b,\u0725(\u202c \u202b)\u0724\u202c using the machine learning model \u202b\u072f\u202c described in Section 3.2.2. The runtime procedure is similar to the training algorithm.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 32,
                        "end": 40,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Obtaining supersense using machine learning model",
                "sec_num": "3.2.3"
            },
            {
                "text": "First, we extract sentences containing from as \u202b\u074f\u0741\u073f\u074a\u0741\u0750\u074a\u0741\u0735\u202c (Step(1a) in Figure 4 ) and use on-line machine translation system \u202b\u0736\u072f\u202c to obtain Chinese translation of \u202b,\u0725(\u202c \u202b)\u0724\u202c as \u202b\u074a\u073d\u074e\u0736\u202c (Step(1b)). For associating \u202b,\u0725(\u202c \u202b)\u0724\u202c with a supersense, we only consider \u0745 in \u202b,\u074f\u0741\u073f\u074a\u0741\u0750\u074a\u0741\u0735\u202c we extract unigram \u0737\u074a\u0745 (Step (2a)) and bigram \u202b\u0745\u0724\u202c (Step (2b)) from \u202b.\u0750\u074a\u0741\u0735\u202c Stopwords are filtered for both \u0737\u074a\u0745 and \u202b\u0745\u0724\u202c similar to what is done at training time. Then, \u0737\u074a\u0745, \u202b\u0745\u0724\u202c and \u202b\u074a\u073d\u074e\u0736\u202c are combined together to predict the supersenses using \u202b.\u072f\u202c The output of \u202b\u072f\u202c is a supersense probability list \u202b\u0750\u074f\u0745\u072e\u0750\u073f\u0745\u0740\u0741\u074e\u202c that contains all supersenses and the probability for \u202b,\u0725(\u202c \u202b)\u0724\u202c (Step 3)). Step (4a and 4b) ). Then we choose maximum as , the corresponding as (Step (5a) and (5b)). A dictionary is used to store the probability sum of each distinct",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 72,
                        "end": 80,
                        "text": "Figure 4",
                        "ref_id": "FIGREF5"
                    },
                    {
                        "start": 673,
                        "end": 684,
                        "text": "(4a and 4b)",
                        "ref_id": "FIGREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Obtaining supersense using machine learning model",
                "sec_num": "3.2.3"
            },
            {
                "text": ", and another dictionary is used to store the frequency of each distinct (",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Obtaining supersense using machine learning model",
                "sec_num": "3.2.3"
            },
            {
                "text": "Step (5c) and (5d)). For each \u202b\u0741\u074f\u074a\u0741\u074f\u202c in \u202b,\u073e\u074e\u0732\u0748\u073d\u0750\u0750\u202c we store the frequency \u202b]\u0741\u074f\u074a\u0741\u074f[\u074d\u0741\u074e\u0742\u202c and the average sense probability \u202b\u073e\u074e\u0732\u0748\u073d\u0750\u0750\u202c \u202b\u0741\u074f\u074a\u0741\u074f\u202c \u202b]\u0741\u074f\u074a\u0741\u074f[\u074d\u0741\u074e\u0742/\u202c to \u202b\u0741\u0749\u073f\u0750\u0751\u202c (Step (6)). Then, we sort \u202b\u0741\u0749\u073f\u0750\u0751\u202c in decreasing order of frequency as \u202b.\u074f\u0741\u074f\u074a\u0741\u0735\u0740\u0741\u0747\u074a\u073d\u074e\u202c If there is more than one sense in \u202b\u0741\u0749\u073f\u0750\u0751\u202c that have the same frequency, they would be sorted in decreasing order of the average sense probability. Finally, we output \u202b\u0724\u072e\u202c (Step (7) ).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 425,
                        "end": 434,
                        "text": "(Step (7)",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Obtaining supersense using machine learning model",
                "sec_num": "3.2.3"
            },
            {
                "text": "Corpus-based machine learning for associating collocations with supersenses can reduce the sense dominance problem, since context words of different supersenses are generally different and translations of a same base word in different senses tend to be different, too. With this in mind, we use sentences of a collocation extracted from a corpus and the collocation translation to disambiguate the supersenses of the base word of a given collocation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Obtaining supersense using machine learning model",
                "sec_num": "3.2.3"
            },
            {
                "text": "In the fourth stage (Step (4) in Figure 2 ), we use a paraphrase-based strategy to determine the supersense. A paraphrase is a restatement of the meaning of a text or passage using another form. By calculating the similarity between a collocation and its paraphrases, we can determine its supersense. This method is based on the assumption that original collocation shares the same supersense with its paraphrases. Table 1 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 33,
                        "end": 41,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    },
                    {
                        "start": 415,
                        "end": 422,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Obtain supersense using similarity & dependency information",
                "sec_num": "3.2.4"
            },
            {
                "text": "Subsequently, we compare the synsets similarity for both (coat, sheet) and (cloth, sheet). The top-ranked similarity of (coat, sheet) is ((Synset('coating.n.01'), Synset('sheet.n.06')), 0.769) and the lexicographer-file of Synset('sheet.n.06') is noun.artifact; the top-ranked similarity of (cloth, sheet) is ((Synset('fabric.n.01'), Synset('sail.n.01')), 0.857) and the lexicographer-file of Synset('sail.n.01') is noun.artifact. So the frequency of noun.artifact is 2, while other supersenses are all 0. We then output noun.artifact as the supersense of input collocation fitted sheet.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Obtain supersense using similarity & dependency information",
                "sec_num": "3.2.4"
            },
            {
                "text": "By using paraphrase-based method, words that related to the input collocation can be the extracted. The collocation could be disambiguated since most of the words with other senses tend not to share the paraphrases. So we can find the sense relation between input collocation and extracted words to obtain the supersense.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Obtain supersense using similarity & dependency information",
                "sec_num": "3.2.4"
            },
            {
                "text": "In the last stage (Step (5) in Figure 2 ), we use the sense frequency to identify the supersense. In many previous works on WSD, sense frequency plays an important role to indicate the sense. A word may have different senses, but most of time, it tends to associated with the dominant sense. So for disambiguating word senses, choosing the most frequent sense is often used as a baseline.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 31,
                        "end": 39,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Obtaining supersense using sense frequency ranking",
                "sec_num": "3.2.5"
            },
            {
                "text": "Many sense frequency methods are based on sense estimation in a corpus. But here we use the sense frequency information in WordNet. For any word in WordNet, there are one or more synsets and the synsets are listed in decreasing order of frequency. So we can simply return the first synset as the supersense. Sense frequency ranking method has the highest coverage, and that is important since our goal is to disambiguate all collocations. We also use this method as the baseline method to compare with our results. We will describe the details of evaluation in Chapter 4.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Obtaining supersense using sense frequency ranking",
                "sec_num": "3.2.5"
            },
            {
                "text": "Once the learning-based procedure, the paraphrase-based procedure and the sense frequency ranking procedure produce the supersenses, a relative majority vote is carried out to \u202b\u0734\u0728\u202c are the three predicted supersense described in sections 3.2.2 to 3.2.4. Each supersense has one vote and the supersense with the most votes is the final output \u0735. As shown in Figure  2 , after running the three procedures for collocation fuel oil, we obtain noun.substance, noun.artifact and noun.substance. The supersense noun.substance has 2 votes and noun.artifact has 1, so the final output \u0735 is noun.substance.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 357,
                        "end": 366,
                        "text": "Figure  2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "The Runtime Hybrid Process",
                "sec_num": "3.2.6"
            },
            {
                "text": "Sometimes the three procedures produce 3 different supersenses without an agreement. Moreover, the learning-based procedure or the paraphrase-based procedure produce no results, because either the sentences containing the input collocation cannot be found in corpus \u202b,\u0725\u072f\u202c or the paraphrases of the input collocation cannot be found and leads the voting has no agreement. In this case, we use back-off to find the supersense. When there is no agreement \u202b.\u0734\u0728\u202c As long as the base word of the input collocation exists in WordNet, we can produce an output.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Runtime Hybrid Process",
                "sec_num": "3.2.6"
            },
            {
                "text": "We have proposed a hybrid model to associate collocations with broad sense classes, with the goal of helping lexicographers in compilation of collocation dictionaries. The evaluation focuses on the intended supersenses of a set of collocations produced by the proposed system. We extracted a set of collocation and supersense pairs from WordNet, so the evaluation could be done automatically.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental Setting",
                "sec_num": "4"
            },
            {
                "text": "In our experiment, we used WordNet, a large lexical database of English which contains approximately 117,000 synsets and 155,000 sense-disambiguated words and collocations, to generate the collocations for training, developing and testing. As we have described in Section 3.2.1, collocations are extracted from WordNet using two heuristics:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data set",
                "sec_num": "4.1"
            },
            {
                "text": "(1) extract collocations from hyponyms of noun synsets",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data set",
                "sec_num": "4.1"
            },
            {
                "text": "(2) extract collocations from definitions and examples sentences of noun synsets",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data set",
                "sec_num": "4.1"
            },
            {
                "text": "We extracted 18,586 collocation and supersense pairs from (1) and 1,784 pairs from (2). The extracted collocations were filtered through a collocation list. The collocation we used is a list of base word/collocates pairs for the top 60,000 lemmas from the Corpus of Contemporary American English (COCA) (Davies, 2008) which contains 4,200,000 collocations. After this step, the total number of collocations was reduced to 7,489. With heuristic (2), we used GENIA tagger (Tsuruoka, 2005 ) which analyzes English sentences and outputs the base forms, part-of-speech tags, chunk tags, and named entity tags to tag the definitions and example sentences.",
                "cite_spans": [
                    {
                        "start": 303,
                        "end": 317,
                        "text": "(Davies, 2008)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 470,
                        "end": 485,
                        "text": "(Tsuruoka, 2005",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data set",
                "sec_num": "4.1"
            },
            {
                "text": "We randomly selected 829 collocations as development set and 6,660 for training and testing from the collocation and supersense pairs. For training and testing, we split the 6,660 collocations into 10 parts that each part contains 666 collocations and we ran ten-fold validation to evaluate the performance of each part.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data set",
                "sec_num": "4.1"
            },
            {
                "text": "In learning-based procedure, we employed Maximum Entropy (ME) model to associate input collocations with supersenses. ME is a flexible statistical learning model that aims to maximize the entropy when characterizing some unknown events. The model estimates outcomes according to a set of features with least possible bias. The ME model we used for training and testing is Maximum Entropy Modeling Toolkit for Python and C++ (Zhang, 2004) . The features we used for the ME model is extracted from British National Corpus (BNC), a 100 million word collection of samples of written and spoken language from a wide range of sources. We use GENIA tagger to tag the sentences in BNC and filtered the stopwords in the sentences using Natural Language Tool Kit (NLTK), a suite of open source program modules written in Python (Loper and Bird, 2002) . More specifically, we used the stopwords in nltk.corpus and obtained the English stopwords list. Another feature, the Chinese translation of the collocations, was obtained from Google Translate.",
                "cite_spans": [
                    {
                        "start": 424,
                        "end": 437,
                        "text": "(Zhang, 2004)",
                        "ref_id": null
                    },
                    {
                        "start": 818,
                        "end": 840,
                        "text": "(Loper and Bird, 2002)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data set",
                "sec_num": "4.1"
            },
            {
                "text": "In the paraphrase-based procedure, we use a set of words with similar words which contains 100,000 words and about 24,000,000 similar words and words with dependency relations which contains 20,000,000 dependency relations. The data is obtained using MINIPAR (Lin, 1993) , a broad-coverage parser for the English language. The similarity comparison algorithm for words used in this stage is JCN similarity (Jiang and Conrath, 1997). JCN similarity bases on the information content (IC, a measure of the specific of a concept) of the Least Common Subsumer (LCS, most specific ancestor node). According to (Sinha and Mihalcea, 2007) , JCN similarity tends to work best for nouns.",
                "cite_spans": [
                    {
                        "start": 259,
                        "end": 270,
                        "text": "(Lin, 1993)",
                        "ref_id": null
                    },
                    {
                        "start": 604,
                        "end": 630,
                        "text": "(Sinha and Mihalcea, 2007)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data set",
                "sec_num": "4.1"
            },
            {
                "text": "Our approach starts with an adjective-noun or noun-noun collocation given by a user, and determines the corresponding sense to the input collocation using external resources related to the input collocation. The output of our system is a supersense in WordNet associated with the input collocation that can be used to help lexicographers in compiling collocation dictionaries, or shown to English learners directly.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "In this paper, we have proposed a hybrid model for associating collocations with supersenses, in which we used a learning-based model, a paraphrase-based similarity comparison, and a sense frequency ranking method. Therefore, we compare the results based on each method and combination of the above methods for evaluating the system performance in more details.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "We compare different methods to associating the collocation with supersense using the test set described in Section 4.1. The methods evaluated for the comparison are listed as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "-FR: Sense frequency ranking method as we described in Section 3.2.5, using the sense frequency information to determine the supersense of a collocation. This method is also the baseline method in our experiment.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "-LB: Learning-based method as we described in Section 3.2.3, using learning-based method to determine the supersense of a collocation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "-LB+FR: Combinational method of learning-based method and sense frequency ranking method, using FR as a back-off if LB cannot be applied.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "-PB: Paraphrase-based method as we described in Section 3.2.4, using similarity and dependency relations of a collocation to determine the sense of that collocation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "-PB+FR: Combinational method of paraphrase-based method and sense frequency ranking method, using FR as a back-off if PB cannot be applied.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "-LB+PB: Combinational method of learning-based method and paraphrase-based method, using PB as a back-off if LB cannot be applied. -LB+PB+FR: Hybrid method of all methods we proposed. The running sequence is",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "LB\u013aPB\u013aFR that LB determines all the test set, then PB determines those LB cannot solve, then FR determines those PB cannot solve.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "-MV+BO: The most complete version of the system we proposed. First, we run the test set using all three methods LB, PB and FR and use relative majority vote to rank supersense results. The rest of collocations that cannot be determined run in the following sequence LB\u013aPB\u013aFR.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Methods compared",
                "sec_num": "4.2"
            },
            {
                "text": "In this chapter, we report the evaluation results of our experiments using methodologies and the settings we described in Chapter 4. We evaluated 8 different methods as described in Section 4.2. We ran ten-fold validation on 6,660 random selected collocations. We report the average performance of the 10 test results. For non-learning based method, we evaluated the whole 6,660 collocations. For comparison, we used the baseline of sense frequency ranking method FR with 75% precision, recall and F-measure. The learning-based method LB achieves the precision 80% and recall 62% with 5% increases in precision. But the recall decreases since no sentences containing the collocations are found in the corpus. Those collocations are not given a supersense. If we add FR to the system as LB+FR, the precision, recall and F-measure increases to 80%. The paraphrase-based method PB on development dataset has a 5% increase on precision comparing with baseline, but on test dataset, the precision decreases to 76% with a low recall of 55%. The low recall is due to the fact that many collocations paraphrases cannot be found. For this we also add FR to the system as a back-off and the precision, recall and F-measure of PB+FR increases to 76%. The experimental result on LB+PB shows that the precision maintains on 80%, and recall increases nearly 10% comparing with LB and achieves the highest recall in all the methods without FR.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Result and Discussion",
                "sec_num": "5"
            },
            {
                "text": "The performance of LB+PB+FR reaches 80%, the same as LB+FR since the performance of PB is not good as LB. We believe that using a relative majority vote to determine the supersense would lead a better performance. MV+BO confirms our hypothesis and achieves the best performance of precision, recall and F-measure 81%. The precision of majority vote that has 3 votes is 95% with recall of 33% while the majority with 2 votes is precision 79% and recall 34%. So with more than 2 votes, the precision reaches 87% with a recall of 67% and F-measure of 76%.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Result and Discussion",
                "sec_num": "5"
            },
            {
                "text": "Take a deeper look in the sense dominance problem we mentioned in Chapter 3. Previous work suffered from that the collocations are often associated with dominant senses. We show the performance of MV+BO when dealing with two different condition: (1) most frequent sense collocations, (2) non-most frequent sense collocations. We could see that when dealing with most frequent sense collocations, 93% of collocations can be correctly associated with supersenses. When dealing with non-most frequent sense collocations, we are still able to correctly associate 46% of collocations with supersenses. So we prove that the sense dominance problem can be reduced by using our hybrid algorithm.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Result and Discussion",
                "sec_num": "5"
            },
            {
                "text": "Many avenues exist for future research and improvement of our system. For example, in the learning-based method, the recall could increase by using a larger corpus or the web data to extract more sentences as collocations' features. The cases where the sentences of the input collocation are not found in a corpus could be reduced. Additionally, we could improve the quality of collocation translations to improve the performance of the learning-based method. In the paraphrase-based method, both precision and recall are not satisfactory, but we still believe that the method has potential. By generating a new similar words list and dependency relations list using a large corpus could produce better paraphrases for associating collocations with supersenses and increasing the recall. Most of the 26 supersenses are natural and reasonable. However, we still find that some supersenses are not very intuitive and may cause problems in tagging. So finding more appropriate set of classes is worth further study.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            },
            {
                "text": "In summary, we have introduced a hybrid method to automatically associate collocations with supersenses. Our goal is to help lexicographers in compilation of a collocation dictionary and help learners to better grasp the usage of a collocation. Our method is composed of a learning-based model, a paraphrase-based method, and a sense frequency ranking method. In our evaluation, we have shown that the hybrid method is significantly better compared with other methods described in this paper. And we also prove that our model can partially reduce the sense dominance problem.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            },
            {
                "text": "Proceedings of the Twenty-Fourth Conference on Computational Linguistics and Speech Processing(ROCLING 2012)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Distributional clustering of words for text classification",
                "authors": [
                    {
                        "first": "L",
                        "middle": [
                            "D"
                        ],
                        "last": "Baker",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "K"
                        ],
                        "last": "Mccallum",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Proceedings of the 21st annual international ACM SIGIR conference on research and development in information retrieval ACM",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Baker L. D. and McCallum A. K., \"Distributional clustering of words for text classification.\" Proceedings of the 21st annual international ACM SIGIR conference on research and development in information retrieval ACM, p. 96 , 1998.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Supersense tagging of unknown nouns using semantic similarity",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "R"
                        ],
                        "last": "Curran",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "proceedings of the 43rd annual meeting on association for computational linguistics Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Curran J. R., \"Supersense tagging of unknown nouns using semantic similarity.\" in proceedings of the 43rd annual meeting on association for computational linguistics Association for Computational Linguistics, 26 p., 2005",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "The corpus of contemporary american english (coca): 400 million Proceedings of the Twenty-Fourth Conference on Computational Linguistics and Speech Processing",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Davies",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Davies M. 2008. The corpus of contemporary american english (coca): 400 million Proceedings of the Twenty-Fourth Conference on Computational Linguistics and Speech Processing (ROCLING 2012) words, 1990-present. Available Online at http://www.Americancorpus.Org .",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Theory and Applications of Ontology: Computer Applications",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Fellbaum",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Wordnet",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "231--274",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fellbaum C. \"WordNet\" in Theory and Applications of Ontology: Computer Applications, pp. 231-43, 2010.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "One sense per discourse",
                "authors": [
                    {
                        "first": "W",
                        "middle": [
                            "A"
                        ],
                        "last": "Gale",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [
                            "W"
                        ],
                        "last": "Church",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Yarowsky",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "proceedings of the workshop on speech and natural language Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Gale W. A., Church K. W. and Yarowsky D., \"One sense per discourse,\" in proceedings of the workshop on speech and natural language Association for Computational Linguistics, p. 233, 1992.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Associating collocations with dictionary senses",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Inumella",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Kilgarriff",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kovar",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Inumella A, Kilgarriff A, Kovar. Associating collocations with dictionary senses.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Semantic similarity based on corpus statistics and lexical taxonomy",
                "authors": [
                    {
                        "first": "Jiang",
                        "middle": [],
                        "last": "Jj",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "W"
                        ],
                        "last": "Conrath",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Arxiv Preprint Cmp-lg",
                "volume": "9709008",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jiang JJ and Conrath DW., Semantic similarity based on corpus statistics and lexical taxonomy. Arxiv Preprint Cmp-lg/9709008, 1997",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Corpus-based statistical sense resolution",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Leacock",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Towell",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Voorhees",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "Proceedings of the ARPA workshop on human language technology",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Leacock C., Towell G. and Voorhees E., \"Corpus-based statistical sense resolution.\" Proceedings of the ARPA workshop on human language technology. p. 260, 1993.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Dependency-based evaluation of MINIPAR. Treebanks",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "317--346",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lin D., Dependency-based evaluation of MINIPAR. Treebanks, pp. 317-29, 2003.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "WordNet: A lexical database for English",
                "authors": [
                    {
                        "first": "G",
                        "middle": [
                            "A"
                        ],
                        "last": "Miller",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Commun ACM",
                "volume": "38",
                "issue": "11",
                "pages": "39--41",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Miller GA. WordNet: A lexical database for English. Commun ACM 38(11):39-41., 1995.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Synonymy in collocation extraction",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Pearce",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proceedings of the workshop on WordNet and other lexical resources",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Pearce D., \"Synonymy in collocation extraction.\" Proceedings of the workshop on WordNet and other lexical resources, second meeting of the north american chapter of the association for computational linguistics. 41 p., 2001.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Developing a robust part-of-speech tagger for biomedical text",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Tsuruoka",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Tateishi",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "D"
                        ],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Ohta",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Mcnaught",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Ananiadou",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Tsujii",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Advances in Informatics",
                "volume": "",
                "issue": "",
                "pages": "382--92",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tsuruoka Y, Tateishi Y, Kim JD, Ohta T, McNaught J, Ananiadou S, Tsujii J., \"Developing a robust part-of-speech tagger for biomedical text.\" In Advances in Informatics, pp. 382-92, 2005.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Unsupervised word sense disambiguation rivaling supervised methods",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Yarowsky",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Proceedings of the 33rd annual meeting on association for computational linguistics Association for Computational Linguistics. 189 p",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yarowsky D. \"Unsupervised word sense disambiguation rivaling supervised methods.\" Proceedings of the 33rd annual meeting on association for computational linguistics Association for Computational Linguistics. 189 p., 1995.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Word-sense disambiguation using statistical models of roget's categories trained on large corpora",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Yarowsky",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Proceedings of the 14th conference on computational linguistics",
                "volume": "2",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yarowsky D. \"Word-sense disambiguation using statistical models of roget's categories trained on large corpora.\" Proceedings of the 14th conference on computational linguistics-volume 2 Association for Computational Linguistics. 454 p., 1992.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "uris": null,
                "text": "An example procedure for associating the collocation of fuel oil with a supersense noun.substance",
                "type_str": "figure",
                "num": null
            },
            "FIGREF1": {
                "uris": null,
                "text": "",
                "type_str": "figure",
                "num": null
            },
            "FIGREF2": {
                "uris": null,
                "text": "describes the processes of our methods.(1) Generate collocation and supersense pairs from WordNet (Section 3.2.1) Train machine learning model from corpus for collocations (Section 3.2.2) Obtain supersense using machine learning model (Section 3.2.3) Obtain supersense using similarity and dependency information (Section 3.2.4) Obtain supersense using sense frequency ranking from WordNet (Section 3.2.5) Outline of the process for obtaining supersense in different",
                "type_str": "figure",
                "num": null
            },
            "FIGREF3": {
                "uris": null,
                "text": "Example of collocation and supersense pairs extracted from WordNet",
                "type_str": "figure",
                "num": null
            },
            "FIGREF4": {
                "uris": null,
                "text": "Obtaining supersense using machine learning model PROCEDURE MachineLearningEvaluateSupersense( ) (1a) Sentences = extractSentences( Candidates = getLexFiles( ) topScore = \u202b,\u202c topSense = \u202b,\u202c freq = \u202b,\u202c totalProb = \u202b,\u202c avgProb = \u202b,\u202c outcome = \u202b\u202c rankedSenses = Sort outcome in decreasing order of freq, if more than 1 sense share same frequency, sofrt those sense in decreasing order of average probability (8) Return the top rankedSenses Proceedings of the Twenty-Fourth Conference on Computational Linguistics and Speech Processing (ROCLING 2012)",
                "type_str": "figure",
                "num": null
            },
            "FIGREF5": {
                "uris": null,
                "text": "Algorithm for obtaining supersense using machine learning modelWe go through each (",
                "type_str": "figure",
                "num": null
            },
            "FIGREF6": {
                "uris": null,
                "text": "For example, consider an input collocation fitted sheet using the paraphrase method. The word sheet has four supersenses: noun.object, noun.communication, noun.artifacrt, noun.shape in WordNet. Paraphrase candidates of fitted sheet are coat, cloth, plate, pan, foil, plastic identified base on similar words list of sheet and coat,Table 1. Example similar words and dependent words of required course Similar words of sheet Dependency relation of fitted Proceedings of the Twenty-Fourth Conference on Computational Linguistics and Speech Processing (ROCLING 2012) ('plate', 0.16), ('sheeting', 0.15) ('pan', 0.14), ('steel', 0.14) ('coat', 0.13), ('tube', 0.12) ('metal', 0.12), ('paper', 0.12) ('slab', 0.11), ('pipe', 0.11) ('layer', 0.11), ('cold-rolled', 0.11) ('stainless', 0.11), ('surface', 0.11) ('glass', 0.11), ('tubing', 0.11) ('booklet', 0.11), ('cut-sheet', 0.11) ', 1), ('than', 1) cloth, jacket, suit based on dependency relations list of fitted. The intersection of the two candidate list contains coat and cloth. It means that coat and cloth are paraphrases of sheet when collocating with fitted. The example similar words and dependency relations of fitted sheet is shown in",
                "type_str": "figure",
                "num": null
            },
            "TABREF1": {
                "text": "Chinese translation of \u202b\u0748\u0725\u202c from \u202b\u0736\u072f\u202c as \u202b.\u074f\u074a\u073d\u074e\u0736\u202c Then, for each sentence \u202b\u0750\u074a\u0741\u0735\u202c in \u202b,\u074f\u0741\u073f\u074a\u0741\u0750\u074a\u0741\u0735\u202c we extract unigram \u0737\u074a\u0745 and bigram \u202b\u0745\u0724\u202c from \u202b.\u0750\u074a\u0741\u0735\u202c Note that stopwords are filtered for both \u0737\u074a\u0745 and \u202b.\u0745\u0724\u202c The next step, we use \u0737\u074a\u0745, \u202b\u0745\u0724\u202c and \u202b\u074f\u074a\u073d\u074e\u0736\u202c as features and \u0735 as the standard output supersense to append machine learning event to \u202b.\u074f\u0741\u074e\u0751\u0750\u073d\u0741\u0728\u202c Note that \u202b\u074f\u074a\u073d\u074e\u0736\u202c actually transforms to a list of unigram and bigram of Chinese words while training. The output of this stage is a probability model \u202b\u072f\u202c trained from a set of training events \u202b\u074f\u0741\u074e\u0751\u0750\u073d\u0741\u0728\u202c for predicting the collocation supersenses using a machine learning tool \u202b.\u072e\u072f\u202c",
                "type_str": "table",
                "num": null,
                "html": null,
                "content": "<table><tr><td colspan=\"3\">(1) unigram and bigram of a sentence extracted from</td><td colspan=\"2\">containing the collocation</td></tr><tr><td colspan=\"3\">(2) Chinese translation of the collocation from \u202b\u0736\u072f\u202c</td><td/><td/></tr><tr><td>For each</td><td>pairs in</td><td colspan=\"2\">, we extract sentences containing</td><td>from</td><td>as</td></tr><tr><td/><td>and obtain</td><td/><td/><td/></tr></table>"
            },
            "TABREF2": {
                "text": "shows the performance for development dataset and test dataset in 8 different methods based on precision, recall and F-measure. Performance for development dataset and test dataset in 8 different methods based on precision, recall and F-measure",
                "type_str": "table",
                "num": null,
                "html": null,
                "content": "<table><tr><td>strategy</td><td/><td colspan=\"2\">Development Set</td><td/><td>Test Set</td><td/></tr><tr><td/><td>Prec.</td><td>Rec.</td><td>F-m.</td><td>Prec.</td><td>Rec.</td><td>F-m.</td></tr><tr><td>FR (baseline)</td><td>.74</td><td>.74</td><td>.74</td><td>.75</td><td>.75</td><td>.75</td></tr><tr><td>LB</td><td>.80</td><td>.61</td><td>.69</td><td>.80</td><td>.62</td><td>.70</td></tr><tr><td>LB+FR</td><td>.78</td><td>.78</td><td>.78</td><td>.80</td><td>.80</td><td>.80</td></tr><tr><td>PB</td><td>.79</td><td>.57</td><td>.66</td><td>.76</td><td>.55</td><td>.63</td></tr><tr><td>PB+FR</td><td>.78</td><td>.78</td><td>.78</td><td>.76</td><td>.76</td><td>.76</td></tr><tr><td>LB+PB</td><td>.80</td><td>.72</td><td>.76</td><td>.80</td><td>.72</td><td>.75</td></tr><tr><td>LB+PB+FR</td><td>.80</td><td>.80</td><td>.80</td><td>.80</td><td>.80</td><td>.80</td></tr><tr><td>MV+BO</td><td>.81</td><td>.81</td><td>.81</td><td>.81</td><td>.81</td><td>.81</td></tr></table>"
            }
        }
    }
}