File size: 71,197 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
{
    "paper_id": "W07-0104",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T04:38:02.111836Z"
    },
    "title": "Active Learning for the Identification of Nonliteral Language *",
    "authors": [
        {
            "first": "Julia",
            "middle": [],
            "last": "Birke",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Simon Fraser University Burnaby",
                "location": {
                    "postCode": "V5A 1S6",
                    "region": "BC",
                    "country": "Canada"
                }
            },
            "email": "[email protected]"
        },
        {
            "first": "Anoop",
            "middle": [],
            "last": "Sarkar",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Simon Fraser University Burnaby",
                "location": {
                    "postCode": "V5A 1S6",
                    "region": "BC",
                    "country": "Canada"
                }
            },
            "email": ""
        },
        {
            "first": "Bill",
            "middle": [],
            "last": "Dolan",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Simon Fraser University Burnaby",
                "location": {
                    "postCode": "V5A 1S6",
                    "region": "BC",
                    "country": "Canada"
                }
            },
            "email": ""
        },
        {
            "first": "Fred",
            "middle": [],
            "last": "Popowich",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Simon Fraser University Burnaby",
                "location": {
                    "postCode": "V5A 1S6",
                    "region": "BC",
                    "country": "Canada"
                }
            },
            "email": ""
        },
        {
            "first": "Dan",
            "middle": [],
            "last": "Fass",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Simon Fraser University Burnaby",
                "location": {
                    "postCode": "V5A 1S6",
                    "region": "BC",
                    "country": "Canada"
                }
            },
            "email": ""
        },
        {
            "first": "Katja",
            "middle": [],
            "last": "Markert",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Simon Fraser University Burnaby",
                "location": {
                    "postCode": "V5A 1S6",
                    "region": "BC",
                    "country": "Canada"
                }
            },
            "email": ""
        },
        {
            "first": "Yudong",
            "middle": [],
            "last": "Liu",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Simon Fraser University Burnaby",
                "location": {
                    "postCode": "V5A 1S6",
                    "region": "BC",
                    "country": "Canada"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In this paper we present an active learning approach used to create an annotated corpus of literal and nonliteral usages of verbs. The model uses nearly unsupervised word-sense disambiguation and clustering techniques. We report on experiments in which a human expert is asked to correct system predictions in different stages of learning: (i) after the last iteration when the clustering step has converged, or (ii) during each iteration of the clustering algorithm. The model obtains an f-score of 53.8% on a dataset in which literal/nonliteral usages of 25 verbs were annotated by human experts. In comparison, the same model augmented with active learning obtains 64.91%. We also measure the number of examples required when model confidence is used to select examples for human correction as compared to random selection. The results of this active learning system have been compiled into a freely available annotated corpus of literal/nonliteral usage of verbs in context.",
    "pdf_parse": {
        "paper_id": "W07-0104",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In this paper we present an active learning approach used to create an annotated corpus of literal and nonliteral usages of verbs. The model uses nearly unsupervised word-sense disambiguation and clustering techniques. We report on experiments in which a human expert is asked to correct system predictions in different stages of learning: (i) after the last iteration when the clustering step has converged, or (ii) during each iteration of the clustering algorithm. The model obtains an f-score of 53.8% on a dataset in which literal/nonliteral usages of 25 verbs were annotated by human experts. In comparison, the same model augmented with active learning obtains 64.91%. We also measure the number of examples required when model confidence is used to select examples for human correction as compared to random selection. The results of this active learning system have been compiled into a freely available annotated corpus of literal/nonliteral usage of verbs in context.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "In this paper, we propose a largely automated method for creating an annotated corpus of literal vs. nonliteral usages of verbs. For example, given the verb \"pour\", we would expect our method to identify the sentence \"Custom demands that cognac be poured from a freshly opened bottle\" as literal, and the sentence \"Salsa and rap music pour out of the windows\" as nonliteral, which, indeed, it does.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We reduce the problem of nonliteral language recognition to one of word-sense disambiguation (WSD) by redefining literal and nonliteral as two different senses of the same word, and we adapt an existing similarity-based word-sense disambiguation method to the task of separating usages of verbs into literal and nonliteral clusters. Note that treating this task as similar to WSD only means that we use features from the local context around the verb to identify it as either literal or non-literal. It does not mean that we can use a classifier trained on WSD annotated corpora to solve this issue, or use any existing WSD classification technique that relies on supervised learning. We do not have any annotated data to train such a classifier, and indeed our work is focused on building such a dataset. Indeed our work aims to first discover reliable seed data and then bootstrap a literal/nonliteral identification model. Also, we cannot use any semi-supervised learning algorithm for WSD which relies on reliably annotated seed data since we do not possess any reliably labeled data (except for our test data set). However we do exploit a noisy source of seed data in a nearly unsupervised approach augmented with active learning. Noisy data containing example sentences of literal and nonliteral usage of verbs is used in our model to cluster a particular instance of a verb into one class or the other. This paper focuses on the use of active learning using this model. We suggest that this approach produces a large saving of effort compared to creating such an annotated corpus manually.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "An active learning approach to machine learning is one in which the learner has the ability to influence the selection of at least a portion of its training data. In our approach, a clustering algorithm for literal/nonliteral recognition tries to annotate the examples that it can, while in each iteration it sends a small set of examples to a human expert to annotate, which in turn provides additional benefit to the bootstrapping process. Our active learn-ing method is similar to the Uncertainty Sampling algorithm of (Lewis & Gale, 1994) but in our case interacts with iterative clustering. As we shall see, some of the crucial criticisms leveled against uncertainty sampling and in favor of Committee-based sampling (Engelson & Dagan, 1996) do not apply in our case, although the latter may still be more accurate in our task.",
                "cite_spans": [
                    {
                        "start": 722,
                        "end": 746,
                        "text": "(Engelson & Dagan, 1996)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "For the purposes of this paper we will take the simplified view that literal is anything that falls within accepted selectional restrictions (\"he was forced to eat his spinach\" vs. \"he was forced to eat his words\") or our knowledge of the world (\"the sponge absorbed the water\" vs. \"the company absorbed the loss\"). Nonliteral is then anything that is \"not literal\", including most tropes, such as metaphors, idioms, as well as phrasal verbs and other anomalous expressions that cannot really be seen as literal. We aim to automatically discover the contrast between the standard set of selectional restrictions for the literal usage of verbs and the non-standard set which we assume will identify the nonliteral usage.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "Our identification model for literal vs. nonliteral usage of verbs is described in detail in a previous publication (Birke & Sarkar, 2006 ). Here we provide a brief description of the model so that the use of this model in our proposed active learning approach can be explained.",
                "cite_spans": [
                    {
                        "start": 116,
                        "end": 137,
                        "text": "(Birke & Sarkar, 2006",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "Since we are attempting to reduce the problem of literal/nonliteral recognition to one of word-sense disambiguation, we use an existing similarity-based word-sense disambiguation algorithm developed by (Karov & Edelman, 1998) , henceforth KE. The KE algorithm is based on the principle of attraction: similarities are calculated between sentences containing the word we wish to disambiguate (the target word) and collections of seed sentences (feedback sets). It requires a target set -the set of sentences containing the verbs to be classified into literal or nonliteral -and the seed sets: the literal feedback set and the nonliteral feedback set. A target set sentence is considered to be attracted to the feedback set containing the sentence to which it shows the highest similarity. Two sentences are similar if they contain similar words and two words are similar if they are contained in similar sentences. The resulting transitive similarity allows us to defeat the knowledge acquisition bottleneck -i.e. the low likelihood of finding all possible usages of a word in a single corpus. Note that the KE algorithm concentrates on similarities in the way sentences use the target literal or nonliteral word, not on similarities in the meanings of the sentences themselves.",
                "cite_spans": [
                    {
                        "start": 202,
                        "end": 225,
                        "text": "(Karov & Edelman, 1998)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "Algorithms 1 and 2 summarize our approach. Note that p(w, s) is the unigram probability of word w in sentence s, normalized by the total number of words in s. We omit some details about the algorithm here which do not affect our discussion about active learning. These details are provided in a previous publication (Birke & Sarkar, 2006) .",
                "cite_spans": [
                    {
                        "start": 316,
                        "end": 338,
                        "text": "(Birke & Sarkar, 2006)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "As explained before, our model requires a target set and two seed sets: the literal feedback set and the nonliteral feedback set. We do not explain the details of how these feedback sets were constructed in this paper, however, it is important to note that the feedback sets themselves are noisy and not carefully vetted by human experts. The literal feedback set was built from WSJ newswire text, and for the nonliteral feedback set, we use expressions from various datasets such as the Wayne Magnuson English Idioms Sayings & Slang and George Lakoff's Conceptual Metaphor List, as well as example sentences from these sources. These datasets provide lists of verbs that may be used in a nonliteral usage, but we cannot explicitly provide only those sentences that contain nonliteral use of that verb in the nonliteral feedback set. In particular, knowing that an expression can be used nonliterally does not mean that you can tell when it is being used nonliterally. In fact even the literal feedback set has noise from nonliteral uses of verbs in the news articles. To deal with this issue (Birke & Sarkar, 2006) provides automatic methods to clean up the feedback sets during the clustering algorithm. Note that the feedback sets are not cleaned up by human experts, however the test data is carefully annotated by human experts (details about inter-annotator agreement on the test set are provided below). The test set is not large enough to be split up into a training and test set that can support learning using a supervised learning method.",
                "cite_spans": [
                    {
                        "start": 1093,
                        "end": 1115,
                        "text": "(Birke & Sarkar, 2006)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "The sentences in the target set and feedback sets were augmented with some shallow syntactic information such as part of speech tags provided Algorithm 1 KE-train: (Karov & Edelman, 1998) algorithm adapted to literal/nonliteral identification Require: S: the set of sentences containing the target word (each sentence is classified as literal/nonliteral) Require: L: the set of literal seed sentences Require: N : the set of nonliteral seed sentences Require: W: the set of words/features, w \u2208 s means w is in sentence s, s w means s contains w Require: : threshold that determines the stopping condition",
                "cite_spans": [
                    {
                        "start": 164,
                        "end": 187,
                        "text": "(Karov & Edelman, 1998)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "1: w-sim 0 (w x , w y ) := 1 if w x = w y , 0 otherwise 2: s-sim I 0 (s x , s y ) := 1, for all s x , s y \u2208 S \u00d7 S where s x = s y , 0 otherwise 3: i := 0 4: while (true) do 5: s-sim L i+1 (s x , s y ) := wx\u2208sx p(w x , s x ) max wy\u2208sy w-sim i (w x , w y ), for all s x , s y \u2208 S \u00d7 L 6: s-sim N i+1 (s x , s y ) := wx\u2208sx p(w x , s x ) max wy\u2208sy w-sim i (w x , w y ), for all s x , s y \u2208 S \u00d7 N 7: for w x , w y \u2208 W \u00d7 W do 8: w-sim i+1 (w x , w y ) := i = 0 sx wx p(w x , s x ) max sy wy s-sim I i (s x , s y ) else sx wx p(w x , s x ) max sy wy {s-sim L i (s x , s y ), s-sim N i (s x , s y )} 9:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "end for 10:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "if \u2200w x , max wy {w-sim i+1 (w x , w y ) \u2212 w-sim i (w x , w y )} \u2264 then 11:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "break # algorithm converges in 1 steps.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "12:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "end if 13:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "i := i + 1 14: end while by a statistical tagger (Ratnaparkhi, 1996) and Su-perTags (Bangalore & Joshi, 1999) .",
                "cite_spans": [
                    {
                        "start": 49,
                        "end": 68,
                        "text": "(Ratnaparkhi, 1996)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 84,
                        "end": 109,
                        "text": "(Bangalore & Joshi, 1999)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "This model was evaluated on 25 target verbs:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "absorb, assault, die, drag, drown, escape, examine, fill, fix, flow, grab, grasp, kick, knock, lend, miss, pass, rest, ride, roll, smooth, step, stick, strike, touch",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "The verbs were carefully chosen to have varying token frequencies (we do not simply learn on frequently occurring verbs). As a result, the target sets contain from 1 to 115 manually annotated sentences for each verb to enable us to measure accuracy. The annotations were not provided to the learning algorithm: they were only used to evaluate the test data performance. The first round of annotations was done by the first annotator. The second annotator was given no instructions besides a few examples of literal and nonliteral usage (not covering all target verbs). The authors of this paper were the annotators. Our inter-annotator agreement on the annotations used as test data in the experiments in this paper is quite high. \u03ba (Cohen) and \u03ba (S&C) on a random sample of 200 annotated examples annotated by two different annotators was found to be 0.77. As per ((Di Eugenio & Glass, 2004) , cf. refs therein), the standard assessment for \u03ba values is that tentative conclusions on agreement exists when .67 \u2264 \u03ba < .8, and a definite conclusion on agreement exists when \u03ba \u2265 .8.",
                "cite_spans": [
                    {
                        "start": 880,
                        "end": 892,
                        "text": "Glass, 2004)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "In the case of a larger scale annotation effort, having the person leading the effort provide one or two examples of literal and nonliteral usages for each target verb to each annotator would almost certainly improve inter-annotator agreement.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "The algorithms were evaluated based on how accurately they clustered the hand-annotated sentences. Sentences that were attracted to neither cluster or were equally attracted to both were put in the opposite set from their label, making a failure to cluster a sentence an incorrect clustering. tag s x as nonliteral 6: end if precision \u2022 recall) / (precision + recall). Nonliteral precision and recall are defined similarly. Average precision is the average of literal and nonliteral precision; similarly for average recall. For overall performance, we take the f-score of average precision and average recall.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "We calculated two baselines for each word. The first was a simple majority-rules baseline (assign each word to the sense which is dominant which is always literal in our dataset). Due to the imbalance of literal and nonliteral examples, this baseline ranges from 60.9% to 66.7% for different verbs with an average of 63.6%. Keep in mind though that using this baseline, the f-score for the nonliteral set will always be 0% -which is the problem we are trying to solve in this work. We calculated a second baseline using a simple attraction algorithm. Each sentence in the target set is attracted to the feedback set with which it has the most words in common. For the baseline and for our own model, sentences attracted to neither, or equally to both sets are put in the opposite cluster to which they belong. This second baseline obtains a f-score of 29.36% while the weakly supervised model without active learning obtains an f-score of 53.8%. Results for each verb are shown in Figure 1 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 981,
                        "end": 989,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Literal vs. Nonliteral Identification",
                "sec_num": "2"
            },
            {
                "text": "The model described thus far is weakly supervised. The main proposal in this paper is to push the results further by adding in an active learning component, which puts the model described in Section 2 in the position of helping a human expert with the literal/nonliteral clustering task. The two main points to consider are: what to send to the human annotator, and when to send it.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Active Learning",
                "sec_num": "3"
            },
            {
                "text": "We always send sentences from the undecided cluster -i.e. those sentences where attraction to either feedback set, or the absolute difference of the two attractions, falls below a given threshold. The number of sentences falling under this threshold varies considerably from word to word, so we additionally impose a predetermined cap on the number of sentences that can ultimately be sent to the human. Based on an experiment on a held-out set separate from our target set of sentences, sending a maximum of 30% of the original set was determined to be optimal in terms of eventual accuracy obtained. We impose an order on the candidate sentences using similarity values. This allows the original sentences with the least similarity to either feedback set to be sent to the human first. Further, we alternate positive similarity (or absolute difference) values and values of zero. Note that sending examples that score zero to the human may not help attract new sentences to either of the feedback sets (since scoring zero means that the sentence was not attracted to any of the sentences). However, human help may be the only chance these sentences have to be clustered at all. After the human provides an identification for a particular example we move the sentence not only into the correct cluster, but also into the corresponding feedback set so that other sentences might be attracted to this certifiably correctly classified sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Active Learning",
                "sec_num": "3"
            },
            {
                "text": "The second question is when to send the sentences to the human. We can send all the examples after the first iteration, after some intermediate iteration, distributed across iterations, or at the end. Sending everything after the first iteration is best for counteracting false attractions before they become entrenched and for allowing future iterations to learn from the human decisions. Risks include sending sentences to the human before our model has had a chance to make potentially correct decision about them, counteracting any saving of effort. (Karov & Edelman, 1998) state that the results are not likely to change much after the third iteration and we have confirmed this independently: similarity values continue to change until convergence, but cluster allegiance tends not to. Sending everything to the human after the third iteration could therefore entail some of the damage control of sending everything after the first iteration while giving the model a chance to do its best. Another possibility is to send the sentences in small doses in order to gain some bootstrapping benefit at each iteration i.e. the certainty measures will improve with each bit of human input, so at each iteration more appropriate sentences will be sent to the human. Ideally, this would produce a compounding of benefits. On the other hand, it could produce a compounding of risks. A final possibility is to wait until the last iteration in the hope that our model has correctly clustered everything else and those correctly labeled examples do not need to be examined by the human. This immediately destroys any bootstrapping possibilities for the current run, although it still provides benefits for iterative augmentation runs (see Section 4).",
                "cite_spans": [
                    {
                        "start": 554,
                        "end": 577,
                        "text": "(Karov & Edelman, 1998)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Active Learning",
                "sec_num": "3"
            },
            {
                "text": "A summary of our results in shown in Figure 1 . The last column in the graph shows the average across all the target verbs. We now discuss the various active learning experiments we performed using our model and a human expert annotator.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 37,
                        "end": 45,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Active Learning",
                "sec_num": "3"
            },
            {
                "text": "Experiments were performed to determine the best time to send up to 30% of the sentences to the human annotator. Sending everything after the first iteration produced an average accuracy of 66.8%; sending everything after the third iteration, 65.2%; sending a small amount at each iteration, 60.8%; sending everything after the last iteration, 64.9%. Going just by the average accuracy, the first iteration option seems optimal. However, several of the individual word results fell catastrophically below the baseline, mainly due to original sentences having been moved into a feedback set too early, causing false attraction. This risk was compounded in the distributed case, as predicted. The third iteration option gave slightly better results (0.3%) than the last iteration option, but since the difference was minor, we opted for the stability of sending everything after the last iteration. These results show an improvement of 11.1% over the model from Section 2. Individual results for each verb are given in Figure 1 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 1017,
                        "end": 1025,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Experiment 1",
                "sec_num": "3.1"
            },
            {
                "text": "In a second experiment, rather than letting our model select the sentences to send to the human, we selected them randomly. We found no significant difference in the results. For the random model to out-perform the non-random one it would have to select only sentences that our model would have clustered incorrectly; to do worse it would have to select only sentences that our model could have handled on its own. The likelihood of the random choices coming exclusively from these two sets is low.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiment 2",
                "sec_num": "3.2"
            },
            {
                "text": "Our third experiment considers the effort-savings of using our literal/nonliteral identification model. The main question must be whether the 11.1% accuracy gain of active learning is worth the effort the human must contribute. In our experiments, the human annotator is given at most 30% of the sentences to classify manually. It is expected that the human will classify these correctly and any additional accuracy gain is contributed by the model. Without semi-supervised learning, we might expect that if the human were to manually classify 30% of the sentences chosen at random, he would have 30% of the sentences classified correctly. However, in order to be able to compare the human-only scenario to the active learning scenario, we must find what the average f-score of the manual process is. The f-score depends on the distribution of literal and nonliteral sentences in the original set. For example, in a set of 100 sentences, if there are exactly 50 of each, and of the 30 chosen for manual annotation, half come from the literal set and half come from the nonliteral set, the f-score will be exactly 30%. We could compare our performance to this, but that would be unfair to the manual process since the sets on which we did our evaluation were by no means balanced. We base a hypothetical scenario on the heavy imbalance often seen in our evaluation sets, and suggest a situation where 96 of our 100 sentences are literal and only 4 are nonliteral. If it were to happen that all 4 of the nonliteral sentences were sent to the human, we would get a very high f-score, due to a perfect recall score for the nonliteral cluster and a perfect precision score for the literal cluster. If none of the four nonliteral sentences were sent to the human, the scores for the nonliteral cluster would be disastrous. This situation is purely hypothetical, but should account for the fact that 30 out of 100 sentences annotated by a human will not necessarily result in an average f-score of 30%: in fact, averaging the results of the three sitatuations described above results Figure 1 : Active Learning evaluation results. Baseline refers to the second baseline from Section 2. Semisupervised: Trust Seed Data refers to the standard KE model that trusts the seed data. Optimal Semisupervised refers to the augmented KE model described in (Birke & Sarkar, 2006) . Active Learning refers to the model proposed in this paper.",
                "cite_spans": [
                    {
                        "start": 2339,
                        "end": 2361,
                        "text": "(Birke & Sarkar, 2006)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 2077,
                        "end": 2085,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Experiment 3",
                "sec_num": "3.3"
            },
            {
                "text": "in an avarage f-score of nearly 36.9%. This is 23% higher than the 30% of the balanced case, which is 1.23 times higher. For this reason, we give the human scores a boost by assuming that whatever the human annotates in the manual scenario will result in an f-score that is 1.23 times higher. For our experiment, we take the number of sentences that our active learning method sent to the human for each word -note that this is not always 30% of the total number of sentences -and multiply that by 1.23 -to give the human the benefit of the doubt, so to speak. Still we find that using active learning gives us an avarage accuracy across all words of 64.9%, while we get only 21.7% with the manual process. This means that for the same human effort, using the weakly supervised classifier produced a threefold improvement in accuracy. Looking at this conversely, this means that in order to obtain an accuracy of 64.9%, by a purely manual process, the human would have to classify nearly 53.6% of the sentences, as opposed to the 17.7% he needs to do using active learning. This is an effort-savings of about 35%. To conclude, we claim that our model combined with active learning is a helpful tool for a literal/nonliteral clustering project. It can save the human significant effort while still producing reasonable results.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiment 3",
                "sec_num": "3.3"
            },
            {
                "text": "In this section we discuss the development of an annotated corpus of literal/nonliteral usages of verbs in context. First, we examine iterative augmentation. Then we discuss the structure and contents of the annotated corpus and the potential for expansion. After an initial run for a particular target word, we have the cluster results plus a record of the feedback sets augmented with the newly clustered sentences. ***pour*** *nonliteral cluster* wsj04:7878 N As manufacturers get bigger , they are likely to pour more money into the battle for shelf space , raising the ante for new players ./. wsj25:3283 N Salsa and rap music pour out of the windows ./. wsj06:300 U Investors hungering for safety and high yields are pouring record sums into single-premium , interest-earning annuities ./. *literal cluster* wsj59:3286 L Custom demands that cognac be poured from a freshly opened bottle ./. Each feedback set sentence is saved with a weight, with newly clustered sentences receiving a weight of 1.0. Subsequent runs may be done to augment the initial clusters. For these runs, we use the the output identification over the examples from our initial run as feedback sets. New sentences for clustering are treated like a regular target set. Running the algorithm in this way produces new clusters and a re-weighted model augmented with newly clustered sentences. There can be as many runs as desired; hence iterative augmentation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Annotated corpus built using active learning",
                "sec_num": "4"
            },
            {
                "text": "We used the iterative augmentation process to build a small annotated corpus consisting of the target words from Table 1 , as well as another 25 words drawn from the examples of previously published work (see Section 5). It is important to note that in building the annotated corpus, we used the Active Learning component as described in this paper, which improved our average f-score from 53.8% to 64.9% on the original 25 target words, and we expect also improved performance on the remainder of the words in the annotated corpus.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 113,
                        "end": 120,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Annotated corpus built using active learning",
                "sec_num": "4"
            },
            {
                "text": "An excerpt from the annotated corpus is shown in Figure 2 . Each entry includes an ID number and a Nonliteral, Literal, or Unannotated tag. Annotations are from testing or from active learning during annotated corpus construction. The corpus is available at http://www.cs.sfu.ca/\u223canoop/students/jbirke/. Further unsupervised expansion of the existing clusters as well as the production of additional clusters is a possibility.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 49,
                        "end": 57,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Annotated corpus built using active learning",
                "sec_num": "4"
            },
            {
                "text": "To our knowledge there has not been any previous work done on taking a model for literal/nonliteral language and augmenting it with an active learning approach which allows human expert knowledge to become part of the learning process.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Previous Work",
                "sec_num": "5"
            },
            {
                "text": "Our approach to active learning is similar to the Uncertainty Sampling approach of (Lewis & Gale, 1994) and (Fujii et. al., 1998) in that we pick those examples that we could not classify due to low confidence in the labeling at a particular point. We employ a resource-limited version in which only a small fixed sample is ever annotated by a human. Some of the criticisms leveled against uncertainty sampling and in favor of Committee-based sampling (Engelson & Dagan, 1996) (and see refs therein) do not apply in our case.",
                "cite_spans": [
                    {
                        "start": 92,
                        "end": 103,
                        "text": "Gale, 1994)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 108,
                        "end": 129,
                        "text": "(Fujii et. al., 1998)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 452,
                        "end": 476,
                        "text": "(Engelson & Dagan, 1996)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Previous Work",
                "sec_num": "5"
            },
            {
                "text": "Our similarity measure is based on two views of sentence-and word-level similarity and hence we get an estimate of appropriate identification rather than just correct classification. As a result, by embedding an Uncertainty Sampling active learning model within a two-view clustering algorithm, we gain the same advantages as other uncertainty sampling methods obtain when used in bootstrapping methods (e.g. (Fujii et. al., 1998) ). Other machine learning approaches that derive from optimal experiment design are not appropriate in our case because we do not yet have a strong predictive (or generative) model of the literal/nonliteral distinction.",
                "cite_spans": [
                    {
                        "start": 409,
                        "end": 430,
                        "text": "(Fujii et. al., 1998)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Previous Work",
                "sec_num": "5"
            },
            {
                "text": "Our machine learning model only does identification of verb usage as literal or nonliteral but it can be seen as a first step towards the use of machine learning for more sophisticated metaphor and metonymy processing tasks on larger text corpora. Rule-based systems -some using a type of interlingua (Russell, 1976) ; others using complicated networks and hierarchies often referred to as metaphor maps (e.g. (Fass, 1997; Martin, 1990; Martin, 1992) -must be largely hand-coded and generally work well on an enumerable set of metaphors or in limited domains. Dictionary-based systems use existing machine-readable dictionaries and path lengths between words as one of their primary sources for metaphor processing information (e.g. (Dolan, 1995) ). Corpus-based systems primarily extract or learn the necessary metaphor-processing information from large corpora, thus avoiding the need for manual annotation or metaphor-map construction. Examples of such systems are (Murata et. al., 2000; Nissim & Markert, 2003; Mason, 2004) . Nissim & Markert (2003) approach metonymy resolution with machine learning methods, \"which [exploit] the similarity between examples of conventional metonymy\" ( (Nissim & Markert, 2003) , p. 56). They see metonymy resolution as a classification problem between the literal use of a word and a number of pre-defined metonymy types. They use similarities between possibly metonymic words (PMWs) and known metonymies as well as context similarities to classify the PMWs. Mason (2004) presents CorMet, \"a corpus-based system for discovering metaphorical mappings between concepts\" ( (Mason, 2004) , p. 23). His system finds the selectional restrictions of given verbs in particular domains by statistical means. It then finds metaphorical mappings between domains based on these selectional preferences. By finding semantic differences between the selectional preferences, it can \"articulate the higher-order structure of conceptual metaphors\" ( (Mason, 2004) , p. 24), finding mappings like LIQUID\u2192MONEY.",
                "cite_spans": [
                    {
                        "start": 301,
                        "end": 316,
                        "text": "(Russell, 1976)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 410,
                        "end": 422,
                        "text": "(Fass, 1997;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 423,
                        "end": 436,
                        "text": "Martin, 1990;",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 437,
                        "end": 450,
                        "text": "Martin, 1992)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 733,
                        "end": 746,
                        "text": "(Dolan, 1995)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 968,
                        "end": 990,
                        "text": "(Murata et. al., 2000;",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 991,
                        "end": 1014,
                        "text": "Nissim & Markert, 2003;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1015,
                        "end": 1027,
                        "text": "Mason, 2004)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 1030,
                        "end": 1053,
                        "text": "Nissim & Markert (2003)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1191,
                        "end": 1215,
                        "text": "(Nissim & Markert, 2003)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1498,
                        "end": 1510,
                        "text": "Mason (2004)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 1609,
                        "end": 1622,
                        "text": "(Mason, 2004)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 1972,
                        "end": 1985,
                        "text": "(Mason, 2004)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Previous Work",
                "sec_num": "5"
            },
            {
                "text": "Metaphor processing has even been approached with connectionist systems storing world-knowledge as probabilistic dependencies (Narayanan, 1999) .",
                "cite_spans": [
                    {
                        "start": 126,
                        "end": 143,
                        "text": "(Narayanan, 1999)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Previous Work",
                "sec_num": "5"
            },
            {
                "text": "In this paper we presented a system for separating literal and nonliteral usages of verbs through statistical word-sense disambiguation and clustering techniques. We used active learning to combine the predictions of this system with a human expert annotator in order to boost the overall accuracy of the system by 11.1%. We used the model together with active learning and iterative augmentation, to build an annotated corpus which is publicly available, and is a resource of literal/nonliteral usage clusters that we hope will be useful not only for future research in the field of nonliteral language processing, but also as training data for other statistical NLP tasks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Supertagging: an approach to almost parsing",
                "authors": [
                    {
                        "first": "Srinivas",
                        "middle": [],
                        "last": "Bangalore",
                        "suffix": ""
                    },
                    {
                        "first": "Aravind",
                        "middle": [
                            "K"
                        ],
                        "last": "Joshi",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Comput. Linguist",
                "volume": "25",
                "issue": "",
                "pages": "237--265",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Srinivas Bangalore and Aravind K. Joshi. 1999. Supertagging: an approach to almost parsing. Comput. Linguist. 25, 2 (Jun. 1999), 237-265.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, EACL-2006",
                "authors": [
                    {
                        "first": "Julia",
                        "middle": [],
                        "last": "Birke",
                        "suffix": ""
                    },
                    {
                        "first": "Anoop",
                        "middle": [],
                        "last": "Sarkar",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Julia Birke and Anoop Sarkar. 2006. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, EACL-2006. Trento, Italy. April 3-7.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "The kappa statistic: a second look",
                "authors": [
                    {
                        "first": "Barbara",
                        "middle": [
                            "Di"
                        ],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "Eugenio",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Glass",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Comput. Linguist",
                "volume": "30",
                "issue": "",
                "pages": "95--101",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Barbara Di Eugenio and Michael Glass. 2004. The kappa statistic: a second look. Comput. Linguist. 30, 1 (Mar. 2004), 95-101.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Metaphor as an emergent property of machine-readable dictionaries",
                "authors": [
                    {
                        "first": "William",
                        "middle": [
                            "B"
                        ],
                        "last": "Dolan",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Proceedings of Representation and Acquisition of Lexical Knowledge: Polysemy, Ambiguity, and Generativity",
                "volume": "",
                "issue": "",
                "pages": "27--29",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "William B. Dolan. 1995. Metaphor as an emergent property of machine-readable dictionaries. In Proceedings of Repre- sentation and Acquisition of Lexical Knowledge: Polysemy, Ambiguity, and Generativity (March 1995, Stanford Univer- sity, CA). AAAI 1995 Spring Symposium Series, 27-29.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Proc. of 34th Meeting of the ACL",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Sean",
                        "suffix": ""
                    },
                    {
                        "first": "Ido",
                        "middle": [],
                        "last": "Engelson",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Dagan",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "319--326",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sean P. Engelson and Ido Dagan. 1996. In Proc. of 34th Meet- ing of the ACL. 319-326.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Processing metonymy and metaphor",
                "authors": [
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Fass",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dan Fass. 1997. Processing metonymy and metaphor. Green- wich, CT: Ablex Publishing Corporation.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Selective sampling for example-based word sense disambiguation",
                "authors": [
                    {
                        "first": "Atsushi",
                        "middle": [],
                        "last": "Fujii",
                        "suffix": ""
                    },
                    {
                        "first": "Takenobu",
                        "middle": [],
                        "last": "Tokunaga",
                        "suffix": ""
                    },
                    {
                        "first": "Kentaro",
                        "middle": [],
                        "last": "Inui",
                        "suffix": ""
                    },
                    {
                        "first": "Hozumi",
                        "middle": [],
                        "last": "Tanaka",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Comput. Linguist",
                "volume": "24",
                "issue": "",
                "pages": "573--597",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Atsushi Fujii, Takenobu Tokunaga, Kentaro Inui and Hozumi Tanaka. 1998. Selective sampling for example-based word sense disambiguation. Comput. Linguist. 24, 4 (Dec. 1998), 573-597.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Similarity-based word sense disambiguation",
                "authors": [
                    {
                        "first": "Yael",
                        "middle": [],
                        "last": "Karov",
                        "suffix": ""
                    },
                    {
                        "first": "Shimon",
                        "middle": [],
                        "last": "Edelman",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Comput. Linguist",
                "volume": "24",
                "issue": "",
                "pages": "41--59",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yael Karov and Shimon Edelman. 1998. Similarity-based word sense disambiguation. Comput. Linguist. 24, 1 (Mar. 1998), 41-59.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "A sequential algorithm for training text classifiers",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "David",
                        "suffix": ""
                    },
                    {
                        "first": "William",
                        "middle": [
                            "A"
                        ],
                        "last": "Lewis",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Gale",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "Proc. of SIGIR-94",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David D. Lewis and William A. Gale. 1994. A sequential algo- rithm for training text classifiers. In Proc. of SIGIR-94.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "A computational model of metaphor interpretation",
                "authors": [
                    {
                        "first": "James",
                        "middle": [
                            "H"
                        ],
                        "last": "Martin",
                        "suffix": ""
                    }
                ],
                "year": 1990,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "James H. Martin. 1990. A computational model of metaphor interpretation. Toronto, ON: Academic Press, Inc.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Computer understanding of conventional metaphoric language",
                "authors": [
                    {
                        "first": "James",
                        "middle": [
                            "H"
                        ],
                        "last": "Martin",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Cognitive Science",
                "volume": "16",
                "issue": "",
                "pages": "233--270",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "James H. Martin. 1992. Computer understanding of conven- tional metaphoric language. Cognitive Science 16, 2 (1992), 233-270.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "CorMet: a computational, corpusbased conventional metaphor extraction system",
                "authors": [
                    {
                        "first": "Zachary",
                        "middle": [
                            "J"
                        ],
                        "last": "Mason",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Comput. Linguist",
                "volume": "30",
                "issue": "",
                "pages": "23--44",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zachary J. Mason. 2004. CorMet: a computational, corpus- based conventional metaphor extraction system. Comput. Linguist. 30, 1 (Mar. 2004), 23-44.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Metonymy interpretation using x no y examples",
                "authors": [
                    {
                        "first": "Masaki",
                        "middle": [],
                        "last": "Murata",
                        "suffix": ""
                    },
                    {
                        "first": "Qing",
                        "middle": [],
                        "last": "Ma",
                        "suffix": ""
                    },
                    {
                        "first": "Atsumu",
                        "middle": [],
                        "last": "Yamamoto",
                        "suffix": ""
                    },
                    {
                        "first": "Hitoshi",
                        "middle": [],
                        "last": "Isahara",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proceedings of SNLP2000",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Masaki Murata, Qing Ma, Atsumu Yamamoto, and Hitoshi Isa- hara. 2000. Metonymy interpretation using x no y exam- ples. In Proceedings of SNLP2000 (Chiang Mai, Thailand, 10 May 2000).",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Moving right along: a computational model of metaphoric reasoning about events",
                "authors": [
                    {
                        "first": "Srini",
                        "middle": [],
                        "last": "Narayanan",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Proceedings of the 16th National Conference on Artificial Intelligence and the 11th IAAI Conference",
                "volume": "",
                "issue": "",
                "pages": "121--127",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Srini Narayanan. 1999. Moving right along: a computational model of metaphoric reasoning about events. In Proceed- ings of the 16th National Conference on Artificial Intelli- gence and the 11th IAAI Conference (Orlando, US, 1999). 121-127.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Syntactic features and word similarity for supervised metonymy resolution",
                "authors": [
                    {
                        "first": "Malvina",
                        "middle": [],
                        "last": "Nissim",
                        "suffix": ""
                    },
                    {
                        "first": "Katja",
                        "middle": [],
                        "last": "Markert",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics (ACL-03)",
                "volume": "",
                "issue": "",
                "pages": "56--63",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Malvina Nissim and Katja Markert. 2003. Syntactic features and word similarity for supervised metonymy resolution. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics (ACL-03) (Sapporo, Japan, 2003). 56-63.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "A maximum entropy part-ofspeech tagger",
                "authors": [
                    {
                        "first": "Adwait",
                        "middle": [],
                        "last": "Ratnaparkhi",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Proceedings of the Empirical Methods in Natural Language Processing Conference",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Adwait Ratnaparkhi. 1996. A maximum entropy part-of- speech tagger. In Proceedings of the Empirical Methods in Natural Language Processing Conference (University of Pennsylvania, May 17-18 1996).",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Computer understanding of metaphorically used verbs",
                "authors": [
                    {
                        "first": "Sylvia",
                        "middle": [
                            "W"
                        ],
                        "last": "Russell",
                        "suffix": ""
                    }
                ],
                "year": 1976,
                "venue": "American Journal of Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sylvia W. Russell. 1976. Computer understanding of metaphorically used verbs. American Journal of Computa- tional Linguistics, Microfiche 44.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "uris": null,
                "text": "Excerpt from our annotated corpus of literal/nonliteral usages of verbs in context.",
                "num": null
            },
            "TABREF0": {
                "content": "<table><tr><td>3:</td><td>tag s x as literal</td></tr><tr><td colspan=\"2\">4: else</td></tr><tr><td>5:</td><td/></tr></table>",
                "html": null,
                "type_str": "table",
                "text": "Evaluation results were recorded as recall, precision, and f-score values. Literal recall is defined as (correct literals in literal cluster / total correct literals). Literal precision is defined as (correct literals in literal cluster / size of literal cluster). If there are no literals, literal recall is 100%; literal precision is 100% if there are no nonliterals in the literal cluster and 0% otherwise. The f-score is defined as (2 \u2022 Algorithm 2 KE-test: classifying literal/nonliteral 1: For any sentence s x \u2208 S 2: if max sy s-sim L (s x , s y ) > max sy s-sim N (s x , s y ) then",
                "num": null
            }
        }
    }
}