File size: 76,992 Bytes
6fa4bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
{
    "paper_id": "R13-1013",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T14:56:24.834808Z"
    },
    "title": "Automatic Extraction of Contextual Valence Shifters",
    "authors": [
        {
            "first": "No\u00e9mi",
            "middle": [],
            "last": "Boubel",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "ILC (Universit\u00e9 catholique de Louvain",
                "location": {
                    "settlement": "Cental"
                }
            },
            "email": "noemi.boubel@uclouvain.be"
        },
        {
            "first": "Thomas",
            "middle": [],
            "last": "Fran\u00e7ois",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "ILC (Universit\u00e9 catholique de Louvain",
                "location": {
                    "settlement": "Cental"
                }
            },
            "email": "thomas.francois@uclouvain.be"
        },
        {
            "first": "Hubert",
            "middle": [],
            "last": "Naets",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "ILC (Universit\u00e9 catholique de Louvain",
                "location": {
                    "settlement": "Cental"
                }
            },
            "email": "hubert.naets@uclouvain.be"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In opinion mining, many linguistic structures, called contextual valence shifters, may modify the prior polarity of items. Some systems of sentiment analysis have tried to take these shifters into account, but few studies have focused on the identification of all these structures and their impact on polarized words. In this paper, we describe a method that automatically identifies contextual valence shifters. It relies on a chi-square test applied to the contingency table representing the distribution of a candidate shifter in a corpus of reviews of various opinions. The system depends on two resources in French-a corpus of reviews and a lexicon of valence terms-to build a list of French contextual valence shifters. We also introduce a set of rules used to classify the extracted contextual valence shifters according to their impact on polarized words. They make use of the Pearson residuals in contingency tables to filter candidate shifters and classify them. We show that the technique reaches an F-measure of either 0.56 or 0.66, depending on how the categories of shifters are defined.",
    "pdf_parse": {
        "paper_id": "R13-1013",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In opinion mining, many linguistic structures, called contextual valence shifters, may modify the prior polarity of items. Some systems of sentiment analysis have tried to take these shifters into account, but few studies have focused on the identification of all these structures and their impact on polarized words. In this paper, we describe a method that automatically identifies contextual valence shifters. It relies on a chi-square test applied to the contingency table representing the distribution of a candidate shifter in a corpus of reviews of various opinions. The system depends on two resources in French-a corpus of reviews and a lexicon of valence terms-to build a list of French contextual valence shifters. We also introduce a set of rules used to classify the extracted contextual valence shifters according to their impact on polarized words. They make use of the Pearson residuals in contingency tables to filter candidate shifters and classify them. We show that the technique reaches an F-measure of either 0.56 or 0.66, depending on how the categories of shifters are defined.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Most opinion mining systems rely on the extraction of sentiment words to detect opinions. These words, which we will rather refer to as polarized words, convey useful information about the semantic orientation (positive or negative) of a text. However, the context in which these words appear may modify their valence in many ways. Although being of importance, this issue has been investigated only recently and is now the object of an increasing attention. Polanyi and Zaenen (2004) first postulated the existence of contextual valence shifters, which are contextual phenomena altering the prior polarity of a term. Afterwards, some of these phenomena (such as negative or conditional syntactic structures) were dealt with on a case by case basis (Das and Chen, 2001; Na et al., 2004; Popescu and Etzioni, 2005; Pang et al., 2002; Wilson et al., 2005; Wilson et al., 2006; Councill et al., 2010) . Studies addressing the phenomenon as a whole flourished later. They aimed at best modelling the expression of opinions (Polanyi and Zaenen, 2004; Taboada et al., 2011; Hatzivassiloglou and Wiebe, 2000; Morsy and Rafea, 2012; Musat and Trausan-Matu, 2010) , before embedding those in a classification system. The main purposes of these studies are to determine a list of contextual valence shifters that impact the polarity of a term as well as to define the nature of this impact. However, these lists are often manually built from linguistic intuitions and not learned from language data. Works relying on a corpus of texts to develop resources that best reflect the actual role played by the linguistic context for opinion mining are few. Li et al. (2010) suggested a technique to automatically select polarityshifting features in order to improve a sentiment classification system based on a machine-learning approach.",
                "cite_spans": [
                    {
                        "start": 459,
                        "end": 484,
                        "text": "Polanyi and Zaenen (2004)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 749,
                        "end": 769,
                        "text": "(Das and Chen, 2001;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 770,
                        "end": 786,
                        "text": "Na et al., 2004;",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 787,
                        "end": 813,
                        "text": "Popescu and Etzioni, 2005;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 814,
                        "end": 832,
                        "text": "Pang et al., 2002;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 833,
                        "end": 853,
                        "text": "Wilson et al., 2005;",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 854,
                        "end": 874,
                        "text": "Wilson et al., 2006;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 875,
                        "end": 897,
                        "text": "Councill et al., 2010)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 1019,
                        "end": 1045,
                        "text": "(Polanyi and Zaenen, 2004;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 1046,
                        "end": 1067,
                        "text": "Taboada et al., 2011;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 1068,
                        "end": 1101,
                        "text": "Hatzivassiloglou and Wiebe, 2000;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 1102,
                        "end": 1124,
                        "text": "Morsy and Rafea, 2012;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 1125,
                        "end": 1154,
                        "text": "Musat and Trausan-Matu, 2010)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 1641,
                        "end": 1657,
                        "text": "Li et al. (2010)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction and State of the Art",
                "sec_num": "1"
            },
            {
                "text": "All these studies agree that contextual valence shifters can have diverse impacts on polarized words. They classify them according to the nature of this impact (Polanyi and Zaenen, 2004; Quirk et al., 1985; Kennedy and Inkpen, 2006) : inversers invert the polarity of a polarized item, intensifiers intensify it and attenuators diminish it.",
                "cite_spans": [
                    {
                        "start": 160,
                        "end": 186,
                        "text": "(Polanyi and Zaenen, 2004;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 187,
                        "end": 206,
                        "text": "Quirk et al., 1985;",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 207,
                        "end": 232,
                        "text": "Kennedy and Inkpen, 2006)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction and State of the Art",
                "sec_num": "1"
            },
            {
                "text": "This study, based on a French corpus, focuses on the issue of contextual valence shifters and pursues two main objectives: (1) propose an automatic method that efficiently models contextual valence shifters, with the aim of improving performance of opinion mining systems (especially those based on a term-counting method); (2) clarify the linguistic structures constituting a hindrance to current classification systems. From these two perspectives, our approach differs from the work of Li et al. (2010) . Moreover, we are interested in describing the effect of all kind of modifiers (inversers, but also intensifiers and attenuators). We restricted our study to all lexicosyntactic patterns located in the immediate context of a polarized term and impacting the valence of this term. This restriction means dealing with individual words. However, it should be noted that contextual shifters may sometimes be phrases too. Our approach also relies on the assumption that contextual shifters are in direct syntactic relation with the polarized word, which has to be confirmed.",
                "cite_spans": [
                    {
                        "start": 489,
                        "end": 505,
                        "text": "Li et al. (2010)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction and State of the Art",
                "sec_num": "1"
            },
            {
                "text": "Based on the results of previous works (Boubel, 2012; Boubel and Bestgen, 2011) , we propose here a system that automatically extracts modifiers (in the form of lexico-syntactic patterns) and classifies them according to their semantic impact. The general methodology is detailed in Section 2 and we report the evaluation of the method in Section 3. The paper concludes with Section 4, discussing some issues we faced, in particular the problem of the attenuating valence shifters.",
                "cite_spans": [
                    {
                        "start": 39,
                        "end": 53,
                        "text": "(Boubel, 2012;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 54,
                        "end": 79,
                        "text": "Boubel and Bestgen, 2011)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction and State of the Art",
                "sec_num": "1"
            },
            {
                "text": "In order to identify valence shifters along with their semantic impact on polarized words, we propose to exploit two different pieces of information regarding the expression of polarity in a text: (1) the overall polarity t of the text, i.e. the score assigned to it on a scale from very negative to very positive, and (2) the polarity p (positive or negative) of a polarized word which appears in the text. We noticed that the distribution of the patterns related to polarized words (i.e. potential modifiers) is influenced by the values of p and t. Intuitively, we can consider three cases:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Key principle",
                "sec_num": "2.1"
            },
            {
                "text": "\u2022 patterns in which p is of opposite polarity than t will mitigate or reverse the valence of their associated term;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Key principle",
                "sec_num": "2.1"
            },
            {
                "text": "\u2022 patterns that reinforce the polarity of a word will appear especially when p and t share the same polarity;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Key principle",
                "sec_num": "2.1"
            },
            {
                "text": "\u2022 finally, a larger number of expressions having an attenuating effect on p will be found when t is around the middle of its scale (texts presenting a nuanced view).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Key principle",
                "sec_num": "2.1"
            },
            {
                "text": "Based on this principle, we developed a system able to automatically detect and classify modifiers. It relies on two resources: (1) a corpus containing evaluative texts whose global polarity t is known and (2) a lexicon of terms whose polarity p is also known.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The system",
                "sec_num": "2.2"
            },
            {
                "text": "Our system performs a two-fold process. First, applying a parser to a corpus, we extract all syntactic dependency relationships that links a polarized term with another term (see Section 2.3). A statistical analysis is then performed to detect, among those, valence shifter candidates (see Section 2.3).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The system",
                "sec_num": "2.2"
            },
            {
                "text": "In the second step (see Section 2.4), a rulebased classifier further removes bad candidates and assigns a label to remaining modifiers that should correspond to their impact on polarized terms.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The system",
                "sec_num": "2.2"
            },
            {
                "text": "In order to identify valence shifter candidates using statistical tests, the initial corpus -made up of evaluative texts whose polarity t is known -is first processed by a syntactic parser to obtain the list of all syntactic dependency relationships including a polarized term. Such relationships take the form of a pair of words (the polarized term and the candidate modifier), along with the nature of this relation (e.g. NP(<NOM:d\u00e9ception>,<ADJ:total>)). For each element of the list, three pieces of information are available: (1) the pattern itself, (2) the valence p of the term included in the structure, and (3) the score t of the text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statistical processing",
                "sec_num": "2.3"
            },
            {
                "text": "Then, we generalize over the relationships extracted, removing the polarized term and keeping only the valence shifter candidate and the syntactic relation linking it to its polarized term (e.g. NP(<NOM:>,<ADJ:total>)). This allows us to determine the frequency of each of these patterns in our corpus, in relation to two variables: the type of the pattern and the score t of the text. Based on these two variables, we build a contingency table for the patterns associated with positive terms and a second table for patterns in the context of a negative term 1 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statistical processing",
                "sec_num": "2.3"
            },
            {
                "text": "Then, for a given pattern g, we compute a chi-square test (Agresti, 2002) 2 where the distribution of g over the five possible values of t is compared with the distribution of all patterns except g. The chi-square value obtained is then used to decide whether the distribution of pattern g in the evaluative texts (t) is independent from the type of pattern. When the chi-square score is significant (based on a threshold \u03b1 1 ), we consider the pattern as a valuable valence shifter candidate. Table 1 examplifies this analysis for the adjective total modifying a positive noun (e.g. \"C'est une r\u00e9ussite totale.\", it is a total success.). This pattern gets a chi-square of 139.67 (p < 0.001) and it stands out even more clearly when associated to a negative noun (\u03c7 2 = 741.35 ; p < 0.001), which confirms its interest as a good valence shifter candidate (e.g. \"d\u00e9ception totale.\", a total disappointment.).",
                "cite_spans": [
                    {
                        "start": 58,
                        "end": 73,
                        "text": "(Agresti, 2002)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 494,
                        "end": 501,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Statistical processing",
                "sec_num": "2.3"
            },
            {
                "text": "At the end of our first step, we obtain a list of valence shifter candidates, selected on the basis of their chi-square score. In the second phase of our method, we apply rules primarily to identify the impact of each candidate on valence terms, but also to further filter the candidate list.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "The idea is to rely on the adjusted residuals (Agresti, 2002) , computed for the two contingency tables available for a candidate pattern (with negative and positive terms). Adjusted residuals corresponds to a z-score, and high values (based on a threshold \u03b1 2 ) means that the pattern g is either over-represented in texts with a given value of t, or is under-represented. These residuals can sometimes display specific and interesting patterns of under-representation or overrepresentation throughout the range of scores t possible for the texts. In previous work (Boubel, 2011) , we analyzed the distributions of the adjusted residuals and we identified three typical profiles. Then, we were able to connect these profiles with their semantic role in the language, distinguishing three groups of modifiers: (1) \"intensifiers\", (2) \"inversers\", and (3) \"concessive structures\".",
                "cite_spans": [
                    {
                        "start": 46,
                        "end": 61,
                        "text": "(Agresti, 2002)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 566,
                        "end": 580,
                        "text": "(Boubel, 2011)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "These findings were translated into a set of rules that automatically classify valence shifter candidates according to their impact on polarized terms.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "Rules are based on the patterns of over-/underrepresentation and assign a score for each of the three classes of modifiers described above. At this stage, it is possible to apply a filtering threshold f s to remove the patterns that received a low score for all classes.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "We can summary the whole set of rules as the three following trends :",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "1. Structures that are over-represented in situations where the valence of p is similar to that of t, regardless of the nature of the term polarity p (positive or negative), obtain a high score in the intensification category;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "2. Structures that are over-represented in situations where p is the opposite of t obtain a high score in the inversion category (attenuating or an inversing role);",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "3. Finally, structures over-represented in reviews reporting a nuanced view (e.g. when t = 3 for texts rated on a scale from 1 to 5) obtain a high score in the concession category.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "Following this method, the adjective \"total\" modifying a noun phrase is given a score of 8 as an \"intensifier\", 0 as an \"inverser\" and 2 as a \"concessive\". It is indeed under-represented with a positive noun while the text is negative and overrepresented while it is positive (see Table 1 ). As a consequence, this pattern is classified as an intensifier.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 281,
                        "end": 288,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "It is worth noting that the classification underlying this approach does not match the one commonly used in the field, which draws a distinction between intensifiers, shifters, and diminishers. Our second category \"inversers\" includes both shifters and diminishers, since these two classes have similar statistical properties according to our method. On the contrary, the analysis of the statistical behavior of some valence shifter candidates highlights a particular semantic behavior which is not dealt with as such in the literature: it corresponds to patterns connecting several polarized terms of different polarities and having an impact on the polarity value of the whole expression. These are the patterns gathered in the third category: the \"concessive structures\". We observe that using statistical properties from the contingency tables to identify categories of valence shifters has limitations in terms of qualitative approach of the ,047  other patterns with positive noun  283,069  588,073  1,507,934  5,454,541  4,188,908  12,022,525  283,090  588,097  1,508,000  5,454,941  4,189,444  12,023,572   Table 1: A contingency table for the adj. total. The adjusted residuals are significant for \u03b1 2 = 0.05 task, but also helps to uncover interesting phenomena. We will come back to the insightful of this classification further in the paper.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 947,
                        "end": 1156,
                        "text": ",047  other patterns with positive noun  283,069  588,073  1,507,934  5,454,541  4,188,908  12,022,525  283,090  588,097  1,508,000  5,454,941  4,189,444  12,023,572   Table 1: A contingency table for the adj.",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Validation and classification of the candidates",
                "sec_num": "2.4"
            },
            {
                "text": "The evaluation of our technique was carried out according to three steps. First, we collected the resources required by the approach, namely a corpus of evaluative texts classified according to their judgment (t), a valence lexicon, and a list of dependencies relationships in which modifiers have been annotated (our gold standard). They are further described in Section 3.1. Then, we carried out a quantitative evaluation of the technique, comparing its predictions to our gold standard (see Section 3.2). Finally, in Section 3.3, we conducted a qualitative analyse of the results, in order to better understand the way our technique works.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "3"
            },
            {
                "text": "To implement our approach, the first resource needed is a corpus of texts ranked according to the opinion they express (t). The corpus we used was provided by the NOMAO company 3 , which proposes a web and mobile application helping people to find, share and discover new places. It is made of 2,200,000 internet user reviews in French relative to restaurants or hotels (7,571,730 sentences). Every text has been given a score from 1 (very bad) to 5 (very good) by the author of the text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Resources",
                "sec_num": "3.1"
            },
            {
                "text": "The second resource needed is a valence lexicon, in which the polarities p of words are labelled. NOMAO also provided us with a such lexicon. It has been manually built and it includes 3,683 polarized French words relative to the domain of restaurant reviews (2,425 negative words and 1,258 positive words).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Resources",
                "sec_num": "3.1"
            },
            {
                "text": "Finally, for evaluation purposes, a gold standard \"corpus\" was required, in which dependencies relationships containing a polarized words and a contextual valence shifter have been annotated. Since, there was no such corpus available, we randomly selected 500 sentences from the whole NO-3 http://fr.nomao.com/ MAO corpus and discarded them from this corpus, that was therefore considered as the training corpus. The 500 sentences contained abount 2,000 dependency relationships including a polarized word 4 . These relationships were manually annotated with a two-fold procedure: (1) decide whether the term associated to a polarized word is a contextual valence shifter or not, and (2) describe its impact on the polarized word, according to one of the available categories.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Resources",
                "sec_num": "3.1"
            },
            {
                "text": "Regarding the categories, we decided to use a finer-grained system than the one based on statistical properties (see Section 2.4), because the category of attenuators, introduced in previous studies, intuitively stood out. This allowed us to discuss in Section 4 the relevance of the concession category we had statiscally identified. We therefore defined the four following classes: (1) intensifiers (INT) emphasize the valence of their associated term; (2) inversers (INV) inverse the valence of their associated term; (3) attenuators (ATT) mitigate the valence of their associated term; and (4) concessives (CONC) articulate terms or phrases of opposite polarities.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Resources",
                "sec_num": "3.1"
            },
            {
                "text": "The list of dependency relationships were annotated by two experts in accordance with these four categories. In order to estimate their interrater agreement, we computed the Fleiss' kappa (Fleiss, 1971) and obtained a substantial agreement (kappa = 0.716) for the annotations. Finally, this corpus was equally divided into a development set -used to select the best set of parameters -and a test set, to assess the performance of the best model.",
                "cite_spans": [
                    {
                        "start": 188,
                        "end": 202,
                        "text": "(Fleiss, 1971)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Resources",
                "sec_num": "3.1"
            },
            {
                "text": "Regarding the evaluation, the first issue was to define an adequate evaluation metric, since the task is a multiclass case. We opted for two different approaches commonly used in the literature. The first split the problem into a detection problem and a classification problem. It computes classic measures such as precision, recall, and F-measure (to which we will refer to as the F-measure 1) regarding the model's ability to detect a modifier, whatever its label. Then, the classification rate is computed through conditional accuracy (Abney, 2008) . The second approach consists in computing the precision, recall and F-measure for each category independently, before averaging them to obtain a global estimation (we will refer to as the F-measure 2).",
                "cite_spans": [
                    {
                        "start": 538,
                        "end": 551,
                        "text": "(Abney, 2008)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "Another issue was the slight discrepancy between the set of labels from the manual annotation and the models. Manual annotation uses INT, ATT, INV, CONC, while the automatic classification uses INT, INV, CONC. For evaluation purposes, we had to project the four-label system onto the three-class one, considering that the category ATT (attenuator) was included into the category INV (inverser) (as it is already supposed in Section 2.4).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "Once these two problems were sorted out, we had to perform an optimization step. Three metaparameters can indeed be manipulated: \u03b1 1 , \u03b1 2 , and f s. \u03b1 1 is the criterion for the selection of candidate modifiers, since it determines the significance level of the chi-square test. \u03b1 2 is the significance threshold for the residuals; decreasing it makes it more difficult for a given structure to match a classification rule. f s is the filtering score assigned for each structure.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "In order to limit the number of experiments, the following values were tested for both \u03b1 1 , \u03b1 2 : 0.0001, 0.0005, 0.001, 0.005, 0.01, and 0.05, while f s was kept constant (f s\u2265 5). Once the best model according to \u03b1 1 , \u03b1 2 was selected, values ranging from 5 to 9 were experimented for f s. The evaluation metric for all models were computed as follows: a list of modifiers included in a dependency relationship were extracted from the training corpus and used to classify the relationships from the development set. It appeared that the optimal parameters are \u03b1 1 = 0.05 and \u03b1 2 = 0.005, as long as we want to exploit the whole training corpus.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "These optimal parameters were used to select 10,503 patterns, whose chi-square scores were significant among a total of 328,308 patterns. Then, the application of our classification rules further filtered those patterns, yielding a list of 6,612 contextual valence shifter candidates: 2,607 were labeled as INT, 2,677 were identified as INV, 1,328 were classified as CONC, and 216 were assigned to more than one categories 5 . However, among those candidates, only 1,147 structures received a score of 5 or higher. More strikingly, if we set f s to 9, then no more than 113 patterns are selected, among which are 66 INT and 47 INV, but no CONC.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "Manipulating the filtering score f s reveals that the number of extracted valence shifters largely varies. We used the test corpus from Section 3.1, which contains 171 valence shifters (102 INT, 16 CONC and 23 INV or ATT) , to estimate the recall, the precision, the conditional accuracy, and the two F-measures for our model trained on the training corpus (see Table 2 ).",
                "cite_spans": [
                    {
                        "start": 185,
                        "end": 194,
                        "text": "(102 INT,",
                        "ref_id": null
                    },
                    {
                        "start": 195,
                        "end": 221,
                        "text": "16 CONC and 23 INV or ATT)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 362,
                        "end": 369,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "The F-measure 1 (which represents the capacity of the model to rightly detect shifters) starts from 0.49 for patterns with a score of 5 or higher and reaches 0.64 when f s \u2265 9. This corresponds to a recall of 0.86 and a precision of 0.37. It is obvious that our system considers too many patterns as valence shifters. This F-measure can however be improved if we use a stricter filtering score. It appears that the chi-square is less efficient than the classification rules to filter valence shifters.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "The F-measure 2 is globally better than Fmeasure 1 and reaches 0.57 when filtering the patterns with intermediate scores. Interestingly, it decreases strongly for f s \u2265 9. This can be explained by the fact that the system extracts less \"concessive structure\" and globally assigns a lower score to that type of structure. Only 6 CONC patterns are correctly classified for f s \u2265 5 and the system does not detect any patterns of this type when f s \u2265 9. As a result, the recall and precision for this category equals 0.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "Finally, it is worth noting that the system obtains a very good conditional accuracy (85.9 for f s \u2265 5 and 97.6 for f s \u2265 9). This is a very interesting finding, since it shows that the classification rules we developed are relevant.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "3.2"
            },
            {
                "text": "To further analyze the efficiency of our extraction method, we submitted the list of the 260 shifters with a score of 8 or higher to a qualitative evalua-",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative analysis",
                "sec_num": "3.3"
            },
            {
                "text": "Score (f s) \u2265 5",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative analysis",
                "sec_num": "3.3"
            },
            {
                "text": "\u2265 6 \u2265 7 \u2265 8 \u2265 9 F-Measure 1 (recall, prec.) 0.49 (.86, .34 tion. The analysis confirms the conclusions drawn above: the system tends to consider too many patterns as shifters, but most of the actual shifters get the correct label, according to experts judgment. After cleaning manually the list, it appears that the system has correctly classified 85 patterns among 260, most of them being incorrectly recognized as valence shifters. Some limitations of our method could explain these errors. First, it happens that the object of the judgment, also associated with polarized words, is extracted (e.g. NP(<ADJ:>,<NOM:accueil>)).",
                "cite_spans": [
                    {
                        "start": 49,
                        "end": 58,
                        "text": "(.86, .34",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative analysis",
                "sec_num": "3.3"
            },
            {
                "text": "Second, grammatical words, such as articles, auxiliary verbs, etc. tend to be captured by the system because they are very frequent in texts. Most of these patterns are not relevant, but some others are important to extract because they can negate or reverse the valence of a polarized word (e.g. NP(<NOM:>,<_DET:aucun>)).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative analysis",
                "sec_num": "3.3"
            },
            {
                "text": "Also, the choice of using syntactic dependency relationships entails some limitations: the expression acting as the valence shifter is sometimes not extracted as a wole. Moreover, parsing errors frequently happen, extracting wrong patterns.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative analysis",
                "sec_num": "3.3"
            },
            {
                "text": "Finally, it happens that some words incorrectly recognized as valence shifters are actually polarized words missing from the valence lexicon.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative analysis",
                "sec_num": "3.3"
            },
            {
                "text": "To conclude this analysis, some characteristics emerge out of the correctly-classified structures. On the one hand, intensifiers (mostly adverbs and adjectives) often have a direct semantic impact on the polarized word to which they are related. On the other hand, the patterns belonging to the INV and CONC categories are more complex and heterogeneous (e.g. AP(<ADJ:loin de>,<ADV:>)) and often impact a phrase or a whole sentence, not directly a lexical item. As a consequence, the effect can be hard to model and it is sometimes difficult to distinguish between the patterns from these two classes, either manually or automatically.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Qualitative analysis",
                "sec_num": "3.3"
            },
            {
                "text": "In this paper, a new methodology for the automatic extraction and classification of valence shifters has been proposed. It reaches a very good accuracy for the classification, although it tends to extract too many structures. An interesting side of the method lies in its ability to identify relevant structures that are often not considered in other studies. In further work, it will be necessary to integrate the lexicon we obtained into a sentiment analysis system to check whether or not taking modifiers into may improve the performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion and conclusion",
                "sec_num": "4"
            },
            {
                "text": "Beyond this applicative goal, our methodology also stressed issues in the categories used to organize contextual valence shifters. The class of diminishers (or downtoners), as it is commonly referred to in the opinion mining domain, is difficult to capture in an automatic way. In our system, we defined three classes of shifters on the basis of three different statistical profiles. The INV class includes both diminishers and inversers, since their statistic profiles are very similar. The CONC class contains structures that often relates terms with different polarities. However, it is worth considering that diminishers are often used in concessive or rhetorical structures and assign them to the class CONC rather than to the class INV. The F-measure 2 for our model in this condition is interestingly better than the one reported above: 0.66 instead of 0.56 for the structures kept when f s \u2265 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion and conclusion",
                "sec_num": "4"
            },
            {
                "text": "In view of these results, it appears that ATT can belong either to the INV class or to the CONC. Our assumption on this matter is that there is actually two types of diminishers: (1) diminishers modifying the valence of a single lexical item, that have statistical profiles closer to the INV category, and (2) diminishers used in concessive structure to attenuate the overall polarity of a phrase or a sentence, which should be included in the CONC class. This hypothesis will be tested in further work, through the analysis of the statistical profiles of manually annotated diminishers.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion and conclusion",
                "sec_num": "4"
            },
            {
                "text": "We only keep patterns with a frequency higher than 20.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "We used chi-square test as a first approach. However, it would be valuable to try other statistical tests in the future.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "It is worth noting that each relationship was considered in the context of the sentence it was extracted from. Therefore, a pattern repeated in the gold standard could be annotated in more than one way. Moreover, since we only dealt with the structures that our methodology can extract, modifiers not syntactically related with a polarized word were not annotated.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "When the score used for filtering is low, a few structures can receive a same score for two classes. However, these cases disappear as soon as we filter with a score of 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We thank the NOMAO company who kindly provide their resources. This research is supported by Wallonie-Bruxelles International.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Semisupervised Learning for Computational Linguistics",
                "authors": [
                    {
                        "first": "S",
                        "middle": [
                            "P"
                        ],
                        "last": "Abney",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S.P. Abney. 2008. Semisupervised Learning for Com- putational Linguistics. Chapman and Hall/CRC, Ann Arbor, U.S.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Categorical Data Analysis",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Agresti",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A. Agresti. 2002. Categorical Data Analysis. 2nd edi- tion. Wiley-Interscience, New York.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Une proc\u00e9dure pour identifier les modifieurs de la valence affective d'un mot dans des textes",
                "authors": [
                    {
                        "first": "No\u00e9mi",
                        "middle": [],
                        "last": "Boubel",
                        "suffix": ""
                    },
                    {
                        "first": "Yves",
                        "middle": [],
                        "last": "Bestgen",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Actes de TALN11",
                "volume": "2",
                "issue": "",
                "pages": "137--142",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "No\u00e9mi Boubel and Yves Bestgen. 2011. Une proc\u00e9- dure pour identifier les modifieurs de la valence affective d'un mot dans des textes. In Actes de TALN11, volume 2, pages 137-142, Montpellier.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Extraction automatique de modifieurs de valence affective dans un texte",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Boubel",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Travaux du Cercle belge de Linguistique",
                "volume": "6",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "N. Boubel. 2011. Extraction automatique de modi- fieurs de valence affective dans un texte. \u00c9tude ex- ploratoire appliqu\u00e9e au cas de l'adverbe. In Travaux du Cercle belge de Linguistique, volume 6.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Construction automatique d'un lexique de modifieurs de polarit\u00e9",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Boubel",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Actes de TALN12",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "N. Boubel. 2012. Construction automatique d'un lex- ique de modifieurs de polarit\u00e9. In Actes de TALN12, Grenoble.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "What's great and what's not: learning to classify the scope of negation for improved sentiment analysis",
                "authors": [
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Isaac",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Councill",
                        "suffix": ""
                    },
                    {
                        "first": "Leonid",
                        "middle": [],
                        "last": "Mcdonald",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Velikovich",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the workshop on negation and speculation in natural language processing",
                "volume": "",
                "issue": "",
                "pages": "51--59",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Isaac G Councill, Ryan McDonald, and Leonid Ve- likovich. 2010. What's great and what's not: learn- ing to classify the scope of negation for improved sentiment analysis. In Proceedings of the workshop on negation and speculation in natural language processing, pages 51-59. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Yahoo! for amazon: Extracting market sentiment from stock message boards",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proceedings of the 8th Asia Pacific Finance Association Annual Conference",
                "volume": "",
                "issue": "",
                "pages": "37--56",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Das and M. Chen. 2001. Yahoo! for amazon: Extracting market sentiment from stock message boards. In Proceedings of the 8th Asia Pacific Fi- nance Association Annual Conference, pages 37-56.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Measuring nominal scale agreement among many raters",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "L"
                        ],
                        "last": "Fleiss",
                        "suffix": ""
                    }
                ],
                "year": 1971,
                "venue": "Psychological bulletin",
                "volume": "76",
                "issue": "5",
                "pages": "378--382",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J.L. Fleiss. 1971. Measuring nominal scale agree- ment among many raters. Psychological bulletin, 76(5):378-382.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Effects of adjective orientation and gradability on sentence subjectivity",
                "authors": [
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Hatzivassiloglou",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Wiebe",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proceedings of the 18th conference on Computational linguistics",
                "volume": "1",
                "issue": "",
                "pages": "299--305",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "V. Hatzivassiloglou and J. M Wiebe. 2000. Effects of adjective orientation and gradability on sentence subjectivity. In Proceedings of the 18th conference on Computational linguistics-Volume 1, pages 299- 305.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Sentiment classification of movie reviews using contextual valence shifters",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Kennedy",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Inkpen",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Computational Intelligence",
                "volume": "22",
                "issue": "2",
                "pages": "110--125",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A. Kennedy and D. Inkpen. 2006. Sentiment classi- fication of movie reviews using contextual valence shifters. Computational Intelligence, 22(2):110- 125.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Sentiment classification and polarity shifting",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [
                            "Y M"
                        ],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "R"
                        ],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 23rd International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "635--643",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Li, S.Y.M. Lee, Y. Chen, C.R. Huang, and G. Zhou. 2010. Sentiment classification and polarity shifting. In Proceedings of the 23rd International Conference on Computational Linguistics, pages 635-643. As- sociation for Computational Linguistics.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Improving documentlevel sentiment classification using contextual valence shifters. Natural Language Processing and Information Systems",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Morsy",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Rafea",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "253--258",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Morsy and A. Rafea. 2012. Improving document- level sentiment classification using contextual va- lence shifters. Natural Language Processing and Information Systems, pages 253-258.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The impact of valence shifters on mining implicit economic opinions",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Musat",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Trausan-Matu",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Artificial Intelligence: Methodology, Systems, and Applications",
                "volume": "",
                "issue": "",
                "pages": "131--140",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. Musat and S. Trausan-Matu. 2010. The impact of valence shifters on mining implicit economic opin- ions. Artificial Intelligence: Methodology, Systems, and Applications, pages 131-140.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Effectiveness of simple linguistic processing in automatic sentiment classification of product reviews",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "C"
                        ],
                        "last": "Na",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Sui",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Khoo",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Chan",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Advances in Knowledge Organization",
                "volume": "9",
                "issue": "",
                "pages": "49--54",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J.C. Na, H. Sui, C. Khoo, S. Chan, and Y. Zhou. 2004. Effectiveness of simple linguistic processing in au- tomatic sentiment classification of product reviews. Advances in Knowledge Organization, 9:49-54.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Thumbs up?: sentiment classification using machine learning techniques",
                "authors": [
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Pang",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Vaithyanathan",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of the ACL-02 conference on Empirical methods in natural language processing",
                "volume": "10",
                "issue": "",
                "pages": "79--86",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "B. Pang, L. Lee, and S. Vaithyanathan. 2002. Thumbs up?: sentiment classification using machine learn- ing techniques. In Proceedings of the ACL-02 con- ference on Empirical methods in natural language processing-Volume 10, pages 79-86.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Contextual valence shifters",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Polanyi",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Zaenen",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of AAAI Spring Symposium on Exploring Attitude and Affect in Text",
                "volume": "",
                "issue": "",
                "pages": "106--111",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "L. Polanyi and A. Zaenen. 2004. Contextual valence shifters. In Proceedings of AAAI Spring Symposium on Exploring Attitude and Affect in Text, pages 106- 111.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Extracting product features and opinions from reviews",
                "authors": [
                    {
                        "first": "A",
                        "middle": [
                            "M"
                        ],
                        "last": "Popescu",
                        "suffix": ""
                    },
                    {
                        "first": "O",
                        "middle": [],
                        "last": "Etzioni",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "339--346",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A.M. Popescu and O. Etzioni. 2005. Extracting prod- uct features and opinions from reviews. In Proceed- ings of the conference on Human Language Tech- nology and Empirical Methods in Natural Language Processing, pages 339-346. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "A comprehensive grammar of the English language",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Quirk",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Greenbaum",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Leech",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Svartvik",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Crystal",
                        "suffix": ""
                    }
                ],
                "year": 1985,
                "venue": "",
                "volume": "397",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Quirk, S. Greenbaum, G. Leech, J. Svartvik, and D. Crystal. 1985. A comprehensive grammar of the English language, volume 397. Cambridge Univ Press.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Lexicon-based methods for sentiment analysis",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Taboada",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Brooke",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Tofiloski",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Voll",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Stede",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Computational Linguistics",
                "volume": "37",
                "issue": "2",
                "pages": "267--307",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Taboada, J. Brooke, M. Tofiloski, K. Voll, and M. Stede. 2011. Lexicon-based methods for sentiment analysis. Computational Linguistics, 37(2):267-307.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Recognizing contextual polarity in phrase-level sentiment analysis",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Wilson",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Wiebe",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Hoffmann",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "347--354",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "T. Wilson, J. Wiebe, and P. Hoffmann. 2005. Recog- nizing contextual polarity in phrase-level sentiment analysis. In Proceedings of the conference on Hu- man Language Technology and Empirical Methods in Natural Language Processing, pages 347-354.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Recognizing strong and weak opinion clauses",
                "authors": [
                    {
                        "first": "Theresa",
                        "middle": [],
                        "last": "Wilson",
                        "suffix": ""
                    },
                    {
                        "first": "Janyce",
                        "middle": [],
                        "last": "Wiebe",
                        "suffix": ""
                    },
                    {
                        "first": "Rebecca",
                        "middle": [],
                        "last": "Hwa",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Computational Intelligence",
                "volume": "22",
                "issue": "2",
                "pages": "73--99",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Theresa Wilson, Janyce Wiebe, and Rebecca Hwa. 2006. Recognizing strong and weak opinion clauses. Computational Intelligence, 22(2):73-99.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF1": {
                "html": null,
                "text": ") 0.51(.84, .37) 0.55(.82, .42) 0.52(.67, .43) 0.64(.60, .69)",
                "type_str": "table",
                "content": "<table><tr><td>Conditional accuracy</td><td>85.9%</td><td>85.5%</td><td>86%</td><td>92.5%</td><td>97.6%</td></tr><tr><td>F-Measure 2 (recall, prec.)</td><td>0.56(.51.62)</td><td>0.55(.50.62)</td><td>0.56(.50.63)</td><td>0.49(.38.69)</td><td>0.40(.32.56)</td></tr></table>",
                "num": null
            },
            "TABREF2": {
                "html": null,
                "text": "Evaluation measures for the model with filtering scores ranging from 5 to 9.",
                "type_str": "table",
                "content": "<table/>",
                "num": null
            }
        }
    }
}