File size: 95,849 Bytes
6fa4bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
{
    "paper_id": "D09-1017",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T16:39:19.284725Z"
    },
    "title": "Review Sentiment Scoring via a Parse-and-Paraphrase Paradigm",
    "authors": [
        {
            "first": "Jingjing",
            "middle": [],
            "last": "Liu",
            "suffix": "",
            "affiliation": {
                "laboratory": "Artificial Intelligence Laboratory",
                "institution": "MIT Computer Science",
                "location": {
                    "addrLine": "32 Vassar Street",
                    "postCode": "02139",
                    "settlement": "Cambridge",
                    "region": "MA"
                }
            },
            "email": ""
        },
        {
            "first": "Stephanie",
            "middle": [],
            "last": "Seneff",
            "suffix": "",
            "affiliation": {
                "laboratory": "Artificial Intelligence Laboratory",
                "institution": "MIT Computer Science",
                "location": {
                    "addrLine": "32 Vassar Street",
                    "postCode": "02139",
                    "settlement": "Cambridge",
                    "region": "MA"
                }
            },
            "email": "seneff@csail.mit.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This paper presents a parse-and-paraphrase paradigm to assess the degrees of sentiment for product reviews. Sentiment identification has been well studied; however, most previous work provides binary polarities only (positive and negative), and the polarity of sentiment is simply reversed when a negation is detected. The extraction of lexical features such as unigram/bigram also complicates the sentiment classification task, as linguistic structure such as implicit long-distance dependency is often disregarded. In this paper, we propose an approach to extracting adverb-adjective-noun phrases based on clause structure obtained by parsing sentences into a hierarchical representation. We also propose a robust general solution for modeling the contribution of adverbials and negation to the score for degree of sentiment. In an application involving extracting aspect-based pros and cons from restaurant reviews, we obtained a 45% relative improvement in recall through the use of parsing methods, while also improving precision.",
    "pdf_parse": {
        "paper_id": "D09-1017",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This paper presents a parse-and-paraphrase paradigm to assess the degrees of sentiment for product reviews. Sentiment identification has been well studied; however, most previous work provides binary polarities only (positive and negative), and the polarity of sentiment is simply reversed when a negation is detected. The extraction of lexical features such as unigram/bigram also complicates the sentiment classification task, as linguistic structure such as implicit long-distance dependency is often disregarded. In this paper, we propose an approach to extracting adverb-adjective-noun phrases based on clause structure obtained by parsing sentences into a hierarchical representation. We also propose a robust general solution for modeling the contribution of adverbials and negation to the score for degree of sentiment. In an application involving extracting aspect-based pros and cons from restaurant reviews, we obtained a 45% relative improvement in recall through the use of parsing methods, while also improving precision.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Online product reviews have provided an extensive collection of free-style texts as well as product ratings prepared by general users, which in return provide grassroots contributions to users interested in a particular product or service as assistance. Yet, valuable as they are, free-style reviews contain much noisy data and are tedious to read through in order to reach an overall conclusion. Thus, we conducted this study to automatically process and evaluate product reviews in order to generate both numerical evaluation and textual summaries of users' opinions, with the ultimate goal of adding value to real systems such as a restaurant-guide dialogue system. Sentiment summarization has been well studied in the past decade (Turney, 2002; Pang et al., 2002; Dave et al., 2003; Liu, 2004a, 2004b; Carenini et al., 2006; Liu et al., 2007) . The polarity of users' sentiments in each segment of review texts is extracted, and the polarities of individual sentiments are aggregated among all the sentences/segments of texts to give a numerical scaling on sentiment orientation.",
                "cite_spans": [
                    {
                        "start": 734,
                        "end": 748,
                        "text": "(Turney, 2002;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 749,
                        "end": 767,
                        "text": "Pang et al., 2002;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 768,
                        "end": 786,
                        "text": "Dave et al., 2003;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 787,
                        "end": 805,
                        "text": "Liu, 2004a, 2004b;",
                        "ref_id": null
                    },
                    {
                        "start": 806,
                        "end": 828,
                        "text": "Carenini et al., 2006;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 829,
                        "end": 846,
                        "text": "Liu et al., 2007)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Most of the work done for sentiment analysis so far has employed shallow parsing features such as part-of-speech tagging. Frequent adjectives and nouns/noun phrases are extracted as opinion words and representative product features. However, the linguistic structure of the sentence is usually not taken into consideration. High level linguistic features, if well utilized and accurately extracted, can provide much insight into the semantic meaning of user opinions and contribute to the task of sentiment identification.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Furthermore, in addition to adjectives and nouns, adverbials and negation also play an important role in determining the degree of the orientation level. For example, \"very good\" and \"good\" certainly express different degrees of positive sentiment. Also, in previous studies, when negative expressions are identified, the polarity of sentiment in the associated segment of text is simply reversed. However, semantic expressions are quite different from the absolute opposite values in mathematics. For example, \"not bad\" does not express the opposite meaning of \"bad\", which would be highly positive. Simply reversing the polarity of sentiment on the appearance of negations may result in inaccurate interpretation of sentiment expressions. Thus, a system which attempts to quantify sentiment while ignoring adverbials is missing a significant component of the sentiment score, especially if the adverbial is a negative word.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Another challenging aspect of negation is proper scoping of the negative reference over the right constituent, which we argue, can be handled quite well with careful linguistic analysis. Take the sentence \"I don't think the place is very clean\" as example. A linguistic approach associating long-distance elements with semantic relations can identify that the negation \"not\" scopes over the complement clause, thus extracting \"not very clean\" instead of \"very clean\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Our goal in modeling adverbials is to investigate whether a simple linear correction model can capture the polarity contribution of all adverbials. Furthermore, is it also appropriate to adjust for multiple adverbs, including negation, via a linear additive model? I.e., can \"not very good\" be modeled as not(very(good))? The fact that \"not very good\" seems to be less negative than \"not good\" suggests that such an algorithm might work well. From these derivations we have developed a model which treats negations in the exact same way as modifying adverbs, via an accumulative linear offset model. This yields a very generic and straightforward solution to modeling the strength of sentiment expression.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper we utilize a parse-and-paraphrase paradigm to identify semantically related phrases in review texts, taking quantifiers (e.g., modifying adverbs) and qualifiers (e.g., negations) into special consideration. The approach makes use of a lexicalized probabilistic syntactic grammar to identify and extract sets of adverb-adjectivenoun phrases that match review-related patterns. Such patterns are constructed based on wellformed linguistic structure; thus, relevant phrases can be extracted reliably.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We also propose a cumulative linear offset model to calculate the degree of sentiment for joint adjectives and quantifiers/qualifiers. The proposed sentiment prediction model takes modifying adverbs and negations as universal scales on strength of sentiment, and conducts cumulative calculation on the degree of sentiment for the associated adjective. With this model, we can provide not only qualitative textual summarization such as \"good food\" and \"bad service\", but also a numerical scoring of sentiment, i.e., \"how good the food is\" and \"how bad the service is.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "There have been many studies on sentiment classification and opinion summarization Lee, 2004, 2005; Gamon et al., 2005; Popescu and Etzioni, 2005; Liu et al., 2005; Zhuang et al., 2006; Kim and Hovy, 2006) . Specifically, aspect rating as an interesting topic has also been widely studied (Titov and McDonald, 2008a; Snyder and Barzilay, 2007; Goldberg and Zhu, 2006) . Recently, Baccianella et. al. (2009) conducted a study on multi-facet rating of product reviews with special emphasis on how to generate vectorial representations of the text by means of POS tagging, sentiment analysis, and feature selection for ordinal regression learning. Titov and McDonald (2008b) proposed a joint model of text and aspect ratings which utilizes a modified LDA topic model to build topics that are representative of ratable aspects, and builds a set of sentiment predictors. Branavan et al. (2008) proposed a method for leveraging unstructured annotations in product reviews to infer semantic document properties, by clustering user annotations into semantic properties and tying the induced clusters to hidden topics in the text.",
                "cite_spans": [
                    {
                        "start": 83,
                        "end": 99,
                        "text": "Lee, 2004, 2005;",
                        "ref_id": null
                    },
                    {
                        "start": 100,
                        "end": 119,
                        "text": "Gamon et al., 2005;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 120,
                        "end": 146,
                        "text": "Popescu and Etzioni, 2005;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 147,
                        "end": 164,
                        "text": "Liu et al., 2005;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 165,
                        "end": 185,
                        "text": "Zhuang et al., 2006;",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 186,
                        "end": 205,
                        "text": "Kim and Hovy, 2006)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 289,
                        "end": 316,
                        "text": "(Titov and McDonald, 2008a;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 317,
                        "end": 343,
                        "text": "Snyder and Barzilay, 2007;",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 344,
                        "end": 367,
                        "text": "Goldberg and Zhu, 2006)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 380,
                        "end": 406,
                        "text": "Baccianella et. al. (2009)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 645,
                        "end": 671,
                        "text": "Titov and McDonald (2008b)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 866,
                        "end": 888,
                        "text": "Branavan et al. (2008)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Our review summarization task is to extract sets of descriptor-topic pairs (e.g., \"excellent service\") from a set of reviews (e.g., for a particular restaurant), and to cluster the extracted phrases into representative aspects on a set of dimensions (e.g., \"food\", \"service\" and \"atmosphere\"). Driven by this motivation, we propose a three-stage system that automatically processes reviews. A block diagram is given in Figure 1 . The first stage is sentence-level data filtering. Review data published by general users is often in free-style, and a large fraction of the data is either ill-formed or not relevant to the task. We classify these as out of domain sentences. To filter out such noisy data, we collect unigram statistics on all the relevant words in the corpus, and select high frequency adjectives and nouns. Any sentence that contains none of the highfrequency nouns or adjectives is rejected from further analysis. The remaining in-domain sentences are subjected to the second stage, parse analysis and semantic understanding, for topic extraction.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 419,
                        "end": 427,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "System Overview",
                "sec_num": "3"
            },
            {
                "text": "From the parsable sentences we extract descriptor-topic phrase patterns based on a carefully-designed generation grammar. We then apply LM (language model) based topic clustering to group the extracted phrases into representative aspects. The third stage scores the degree of sentiment for adjectives, as well as the strength of sentiment for modifying adverbs and negations, which further refine the degree of sentiment of the associated adjectives. We then run a linear additive model to assign a combined sentiment score for each extracted phrase.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Overview",
                "sec_num": "3"
            },
            {
                "text": "The rest of the paper is structured as follows: In Section 4, we explain the linguistic analysis. In Section 5, we describe the cumulative model for assessing the degree of sentiment. Section 6 provides a systematic evaluation, conducted on real data in the restaurant review domain harvested from the Web. Section 7 provides a discussion analyzing the results. Section 8 summarizes the paper as well as pointing to future work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Overview",
                "sec_num": "3"
            },
            {
                "text": "Our linguistic analysis is based on a parse-andparaphrase paradigm. Instead of the flat structure of a surface string, the parser provides a hierarchical representation, which we call a linguistic frame (Xu et al., 2008) . It preserves linguistic structure by encoding different layers of semantic dependencies. The grammar captures syntactic structure through a set of carefully constructed context free grammar rules, and employs a feature-passing mechanism to enforce long distance constraints. The grammar is lexicalized, and uses a statistical model to rank order competing hypotheses. It knows explicitly about 9,000 words, with all unknown words being interpreted as nouns. The grammar probability model was trained automatically on the corpus of review sentences.",
                "cite_spans": [
                    {
                        "start": 203,
                        "end": 220,
                        "text": "(Xu et al., 2008)",
                        "ref_id": "BIBREF22"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parse-and-Paraphrase",
                "sec_num": "4.1"
            },
            {
                "text": "To produce the phrases, a set of generation rules is carefully constructed to only extract sets of related adverbs, adjectives and nouns. The adjective-noun relationships are captured from the following linguistic patterns: (1) all adjectives attached directly to a noun in a noun phrase, (2) adjectives embedded in a relative clause modifying a noun, and (3) adjectives related to nouns in a subject-predicate relationship in a clause. These patterns are compatible, i.e., if a clause contains both a modifying adjective and a predicate adjective related to the same noun, two adjective-noun pairs are generated by different patterns. As in, \"The efficient waitress was nonetheless very courteous.\" It is a \"parse-andparaphrase-like\" paradigm: the paraphrase tries to preserve the original words intact, while reordering them and/or duplicating them into multiple NP units. Since they are based on syntactic structure, the generation rules can also be applied in any other domain involving opinion mining.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parse-and-Paraphrase",
                "sec_num": "4.1"
            },
            {
                "text": "An example linguistic frame is shown in Figure 2, which encodes the sentence \"The caesar with salmon or chicken is really quite good.\" In this example, for the adjective \"good\", the nearby noun \"chicken\" would be associated with it if only proximity is considered. From the linguistic frame, however, we can easily associate \"caesar\" with \"good\" by extracting the head of the topic sub-frame and the head of the predicate subframe, which are encoded in the same layer (root layer) of the linguistic frame. Also, we can tell from the predicate sub-frame that there is an adverb \"quite\" modifying the head word \"good\". The linguistic frame also encodes an adverb \"really\" in the upstairs layer. A well-constructed generation grammar can create customized adverb-adjective-noun phrases such as \"quite good caesar\" or \"really quite good caesar\".",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 40,
                        "end": 46,
                        "text": "Figure",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Parse-and-Paraphrase",
                "sec_num": "4.1"
            },
            {
                "text": ":topic {q caesar :quantifier \"def\" :pred {p with :topic {q salmon :pred {p conjunction :or {q chicken }}}} :adv \"really\" :pred {p adj_complement :pred {p adjective :adv \"quite\" :pred {p quality :topic \"good\" }}}} Figure 2 . Linguistic frame for \"The caesar with salmon or chicken is really quite good.\"",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 213,
                        "end": 221,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "{c cstatement",
                "sec_num": null
            },
            {
                "text": "Interpreting negation in English is not straightforward, and it is often impossible to do correctly without a deep linguistic analysis. Xuehui Wu (2005) wrote: \"The scope of negation is a complex linguistic phenomenon. It is easy to perceive but hard to be defined from a syntactic point of view. Misunderstanding or ambiguity may occur when the negative scope is not understood clearly and correctly.\" The majority rule for negation is that it scopes over the remainder of its containing clause, and this works well for most cases. For example, Figure 3 shows the linguistic frame for the sentence \"Their menu was a good one that didn't try to do too much.\" {c cstatement :topic {q menu :poss \"their\" } } :complement {q pronoun :name \"one\" :adj_clause {c cstatement :conjn \"that\" :negate \"not\" :pred {p try :to_clause {p do :topic {q object :adv \"too\" :quant \"much\" }}}} :pred {p adjective :pred {p quality :topic \"good\" }}} Figure 3 . Linguistic frame for \"Their menu was a good one that didn't try to do too much.\"",
                "cite_spans": [
                    {
                        "start": 136,
                        "end": 152,
                        "text": "Xuehui Wu (2005)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 546,
                        "end": 554,
                        "text": "Figure 3",
                        "ref_id": null
                    },
                    {
                        "start": 926,
                        "end": 934,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "{c cstatement",
                "sec_num": null
            },
            {
                "text": "Traditional approaches which do not consider the linguistic structure would treat the appearance of \"not\" as a negation and simply reverse the sentiment of the sentence to negative polarity, which is wrong as the sentence actually expresses positive opinion for the topic \"menu\". In our approach, the negation \"not\" is identified as under the sub-frame of the complement clause, instead of in the same or higher layer of the adjective sub-frame; thus it is considered as unrelated to the adjective \"good\". In this way we can successfully predict the scope of the reference of the negation over the correct constituent of a sentence and create proper association between negation and its modified words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "{c cstatement",
                "sec_num": null
            },
            {
                "text": "To categorize the extracted phrases into representative aspects, we automatically group the identified topics into a set of clusters based on LM probabilities. The LM-based algorithm assumes that topics which are semantically related have high probability of co-occurring with similar descriptive words. For example, \"delicious\" might co-occur frequently with both \"pizza\" and \"dessert\". By examining the distribution of bigram probability of these topics with corresponding descriptive words, we can group \"pizza\" and \"dessert\" into the same cluster of \"food\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LM-based Topic Clustering",
                "sec_num": "4.2"
            },
            {
                "text": "We select a small set of the most common topics, i.e., topics with the highest frequency counts, and put them into an initial set I. Then, for each candidate topic \u202b\u0750\u202c \u073f outside set I, we calculate its probability given each topic \u202b\u0750\u202c \u0745 within the initial set I, given by:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LM-based Topic Clustering",
                "sec_num": "4.2"
            },
            {
                "text": "\u0732\u123a\u202b\u0750\u202c | \u202b\u0750\u202c \u123b \u0d4c \u2211 \u0732\u123a\u202b\u0750\u202c |\u073d\u123b \u2022 \u0732\u123a\u073d|\u202b\u0750\u202c \u123b \u202b\u05d0\u202c \u0d4c \u2211 \u123a,\u0be7 \u123b \u123a\u123b \u2022 \u123a,\u0be7 \u123b \u123a\u0be7 \u123b \u202b\u05d0\u202c \u0d4c \u0b35 \u123a\u0be7 \u123b \u2211 \u0b35 \u123a\u123b \u2022 \u0732\u123a\u073d, \u202b\u0750\u202c \u123b \u2022 \u0732\u123a\u073d, \u202b\u0750\u202c \u123b \u202b\u05d0\u202c (1)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LM-based Topic Clustering",
                "sec_num": "4.2"
            },
            {
                "text": "where A represents the set of all the adjectives in the corpus. For each candidate topic \u202b\u0750\u202c \u073f , we choose the cluster of the initial topic \u202b\u0750\u202c \u0745 with which it has the highest probability score.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LM-based Topic Clustering",
                "sec_num": "4.2"
            },
            {
                "text": "There are also cases where a meaningful adjective occurs in the absence of an associated topic, e.g., \"It is quite expensive.\" We call such cases the \"widow-adjective\" case. Without hardcoded ontology matching, it is difficult to identify \"expensive\" as a price-related expression. To discover such cases and associate them with related topics, we propose a \"surrogate topic\" matching approach based on bigram probability.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LM-based Topic Clustering",
                "sec_num": "4.2"
            },
            {
                "text": "As aforementioned, the linguistic frame organizes all adjectives into separate clauses. Thus, we create a \"surrogate topic\" category in the linguistic frames for widow-adjective cases, which makes it easy to detect descriptors that are affiliated with uninformative topics like the pronoun \"it\". We then have it generate phrases such as \"expensive surrogate_topic\" and use bigram probability statistics to automatically map each sufficiently strongly associated adjective to its most common topic among our major classes, e.g., mapping \"expensive\" with its surrogate topic \"price\". Therefore, we can generate sets of additional phrases in which the topic is \"hallucinated\" from the widow-adjective.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "LM-based Topic Clustering",
                "sec_num": "4.2"
            },
            {
                "text": "Given the sets of adverb-adjective-noun phrases extracted by linguistic analysis, our goal is to assign a score for the degree of sentiment to each phrase and calculate an average rating for each aspect. An example summary is given in Table 1 . To calculate the numerical degree of sentiment, there are three major problems to solve: 1) how to associate numerical scores with textual sentiment; 2) whether to calculate sentiment scores for adjectives and adverbs jointly or separately; 3) whether to treat negations as special cases or in the same way as modifying adverbs.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 235,
                        "end": 242,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "5.1"
            },
            {
                "text": "There have been studies on building sentiment lexicons to define the strength of sentiment of words. Esuli and Sebastiani (2006) constructed a lexical resource, SentiWordNet, a WordNet-like lexicon emphasizing sentiment orientation of words and providing numerical scores of how objective, positive and negative these words are. However, lexicon-based methods can be tedious and inefficient and may not be accurate due to the complex cross-relations in dictionaries like WordNet. Instead, our primary approach to sentiment scoring is to make use of collective data such as user ratings. In product reviews collected from online forums, the format of a review entry often consists of three parts: pros/cons, free-style text and user rating. We assume that user rating is normally consistent with the tone of the review text published by the same user. By associating user ratings with each phrase extracted from review texts, we can easily associate numerical scores with textual sentiment.",
                "cite_spans": [
                    {
                        "start": 101,
                        "end": 128,
                        "text": "Esuli and Sebastiani (2006)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "5.1"
            },
            {
                "text": "A simple strategy of rating assignment is to take each extracted adverb-adjective pair as a composite unit. However, this method is likely to lead to a large number of rare combinations, thus suffering from sparse data problems. Therefore, an interesting question to ask is whether it is feasible to assign to each adverb a perturbation score, which adjusts the score of the associated adjective up or down by a fixed scalar value. This approach thus hypothesizes that \"very expensive\" is as much worse than \"expensive\" as \"very romantic\" is better than \"romantic\". This allows us to pool all instances of a given adverb regardless of which adjective it is associated with, in order to compute the absolute value of the perturbation score for that adverb. Therefore, we consider adverbs and adjectives separately when calculating the sentiment score, treating each modifying adverb as a universal quantifier which consistently scales up/down the strength of sentiment for the adjectives it modifies.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "5.1"
            },
            {
                "text": "Furthermore, instead of treating negation as a special case, the universal model works for all adverbials. The model hypothesizes that \"not bad\" is as much better than \"bad\" as \"not good\" is worse than \"good\", i.e., negations push positive/negative adjectives to the other side of sentiment polarity by a universal scale. This again, allows us to pool all instances of a given negation and compute the absolute value of the perturbation score for that negation, in the same way as dealing with modifying adverbs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "5.1"
            },
            {
                "text": "For each adjective, we average all its ratings given by:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "\u0735\u073f\u202b\u0741\u074e\u202c\u123a\u073d\u0740\u0746\u123b \u0d4c \u2211 \u0cbf \u2022 \u202b\u05d0\u202c\u0cc1 \u2211 \u0cbf (2)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "where \u0732 represents the set of appearances of adjective \u073d\u0740\u0746, \u202b\u074e\u202c represents the associated user rating in each appearance of \u073d\u0740\u0746, \u0730 represents the number of entities (e.g., restaurants) in the entire data set, and \u074a represents the number of entities with rating \u202b\u074e\u202c . The score is averaged over all the appearances, weighted by the frequency count of each category of rating to remove bias towards any category.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "As for adverbs, using a slightly modified version of equation 2, we can get a rating table for all adverb-adjective pairs. For each adverb adv, we get a list of all its possible combinations with adjectives. Then, for each adj in the list, we calculate the distance between the rating of adv-adj and the rating of the adj alone. We then aggregate the distances among all the pairs of adv-adj and adj in the list, weighted by the frequency count of each adv-adj pair:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u202b\u202a\u123b\u202c\u0752\u0740\u073d\u202a\u0741\u123a\u202c\u074e\u073f\u0735\u202c \u0d4c \u2211 \u0be8\u0be7\u123a\u0bd7\u0be9,\u0bd7 \u123b \u2211 \u0be8\u0be7\u0d6b\u0bd7\u0be9,\u0bd7 \u0cd5 \u0d6f \u0cd5\u202b\u05d0\u202c\u0cb2 \u0be7\u202b\u05d0\u202c \u2022 \u0732\u202b\u0748\u202c\u123a\u073d\u0740\u0746 \u0be7 \u123b \u2022 \u123a\u202b\u074e\u202c\u123a\u073d\u0740\u202b,\u0752\u202c \u073d\u0740\u0746 \u0be7 \u123b \u0d46 \u202b\u074e\u202c\u123a\u073d\u0740\u0746 \u0be7 \u123b\u123b",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "where \u202b,\u0752\u0740\u073d\u202a\u123a\u202c\u0750\u074a\u0751\u073f\u202c \u073d\u0740\u0746 \u0be7 \u123b represents the count of the combination \u202b\u0752\u0740\u073d\u202c \u0d46 \u073d\u0740\u0746 \u0be7 , \u202b\u0723\u202c represents the set of adjectives that co-occur with \u202b\u0752\u0740\u073d\u202c , \u202b,\u0752\u0740\u073d\u202a\u123a\u202c\u074e\u202c \u073d\u0740\u0746 \u0be7 \u123b represents the sentiment rating of the combination \u202b\u0752\u0740\u073d\u202c \u0d46 \u073d\u0740\u0746 \u0be7 , and \u202b\u074e\u202c\u123a\u073d\u0740\u0746 \u0be7 \u123b represents the sentiment rating of the adjective \u073d\u0740\u0746 \u0be7 alone. \u0732\u202b\u0748\u202c\u123a\u073d\u0740\u0746 \u0be7 \u123b represents the polarity of \u073d\u0740\u0746 \u0be7 , assigned as 1 if \u073d\u0740\u0746 \u0be7 is positive, and -1 if negative. Specifically, negations are well handled by the same scoring strategy, treated exactly the same as modifying adverbs, except that they get such strong negative scores that the sentiment of the associated adjectives is pushed to the other side of the polarity scale.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "After obtaining the strength rating for adverbs and the sentiment rating for adjectives, the next step is to assign the strength of sentiment to each phrase (negation-adverb-adjective-noun) extracted by linguistic analysis, as given by: negative. For example, given the ratings <good: 4.5>, <bad: 1.5>, <very: 0.5> and <not: -3.0>, we would assign \"5.0\" to \"very good\" (score(very(good))=4.5+0.5), \"1.0\" to \"very bad\" (score(very(bad))=1.5-0.5), and \"2.0\" to \"not very good\" (score(not(very(good)))= 4.5+0.5-3.0). The corresponding sequence of different degrees of sentiment is: \"very good: 5.0\" > \"good: 4.5\" > \"not very good: 2.0\" > \"bad: 1.5\" > \"very bad: 1.0\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u202b\u0741\u074e\u073f\u0735\u202c \u1240\u074a\u0741\u0743\u0d6b\u073d\u0740\u202b\u0752\u202c\u123a\u073d\u0740\u0746\u123b\u0d6f\u1241 \u0d4c \u202b\u074e\u202c\u123a\u073d\u0740\u0746\u123b \u0732\u202b\u0748\u202c\u123a\u073d\u0740\u0746\u123b \u2022 \u202b\u074e\u202c\u123a\u073d\u0740\u202b\u0752\u202c\u123b \u0732\u202b\u0748\u202c\u123a\u073d\u0740\u0746\u123b \u2022 \u202b\u074e\u202c\u123a\u074a\u0741\u0743\u123b",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Linear Additive Model",
                "sec_num": "5.2"
            },
            {
                "text": "In this section we present a systematic evaluation of the proposed approaches conducted on real data. We crawled a data collection of 137,569 reviews on 24,043 restaurants in 9 cities in the U.S. from an online restaurant evaluation website 1 . Most of the reviews have both pros/cons and free-style text. For the purpose of evaluation, we take those reviews containing pros/cons as the experimental set, which is 72.7% (99,147 reviews) of the original set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6"
            },
            {
                "text": "Based on the experimental set, we first filtered out-of-domain sentences based on frequency count, leaving a set of 857,466 in-domain sentences (67.5%). This set was then subjected to parse analysis; 78.6% of them are parsable. Given the parsing results in the format of linguistic frame, we used a set of language generation rules to extract relevant adverb-adjectivenoun phrases. We then selected the most frequent 6 topics that represented appropriate dimensions for the restaurant domain (\"place\", \"food\", \"service\", \"price\", \"atmosphere\" and \"portion\") as the initial set, and clustered the extracted topic mentions into different aspect categories by creating a set of topic mappings with the LMbased clustering method. Phrases not belonging to any category are filtered out.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Extraction",
                "sec_num": "6.1"
            },
            {
                "text": "To evaluate the performance of the proposed approach (LING) to topic extraction, we compare it with a baseline method similar to Liu, 2004a, 2004b; Liu et al., 2005) . We performed part-of-speech tagging on both parsable and unparsable sentences, extracted each pair of noun and adjective that has the smallest proximity, and filtered out those with low frequency counts. Adverbs and negation words that are adjacent to the identified adjectives were also extracted along with the adjective-noun pairs. We call this the \"neighbor baseline\" (NB).",
                "cite_spans": [
                    {
                        "start": 129,
                        "end": 147,
                        "text": "Liu, 2004a, 2004b;",
                        "ref_id": null
                    },
                    {
                        "start": 148,
                        "end": 165,
                        "text": "Liu et al., 2005)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Extraction",
                "sec_num": "6.1"
            },
            {
                "text": "The proposed method is unable to make use of the non-parsable sentences, which make up over 20% of the data. Hence, it seems plausible to utilize a back-off mechanism for these sentences via a combined system (COMB) incorporating NB only for the sentences that fail to parse.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Extraction",
                "sec_num": "6.1"
            },
            {
                "text": "In considering how to construct the \"ground truth\" set of pros and cons for particular aspects, our goal was to minimize error as much as possible without requiring exorbitant amounts of manual labeling. We also wanted to assure that the methods were equally fair to both systems (LING and NB). To these ends, we decided to pool together all of the topic mappings and surrogate topic hallucinations obtained automatically from both systems, and then to manually edit the resulting list to eliminate any that were deemed unreasonable. We then applied these edited mappings in an automatic procedure to the adjective-noun pairs in the user-provided pros and cons of all the restaurant reviews. The resulting aspect-categorized phrase lists are taken as the ground truth. Each system then used its own (unedited) set of mappings in processing the associated review texts.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Extraction",
                "sec_num": "6.1"
            },
            {
                "text": "We also needed an algorithm to decide on a particular set of reviews for consideration, again, with the goal of omitting bias towards either of the two systems. We decided to retain as the evaluation set all reviews which obtained at least one topic extraction from both systems. Thus the two systems processed exactly the same data with exactly the same definitions of \"ground truth\". Performance was evaluated on this set of 62,588 reviews in terms of recall (percentage of topics in the ground truth that are also identified from the review body) and precision (percentage of extracted topics that are also in the ground truth). These measures are computed separately for each review, and then averaged over all reviews.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Topic Extraction",
                "sec_num": "6.1"
            },
            {
                "text": "As shown in Table 2 , without clustering, the LING approach gets 4.6% higher recall than the NB baseline. And the recall from the COMB approach is 3.9% higher than that from the LING approach and 8.5% higher than that from the NB baseline. With topic clustering, the COMB approach also gets the highest recall, with a 4.9% and 17.5% increase from the LING approach and the NB baseline respectively. The precision is quite close among the different approaches, around 60%. Table 2 also shows that the topic clustering approach increases the recall by 4.8% for the NB baseline, 12.8% for the LING approach, and 13.8% for the COMB approach. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 12,
                        "end": 19,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 472,
                        "end": 479,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Topic Extraction",
                "sec_num": "6.1"
            },
            {
                "text": "To score the degree of sentiment for each extracted phrase, we built a table of sentiment score (<adjective: score>) for adjectives and a table of strength score (<adverb: score>) for adverbs. The pros/cons often contain short and wellstructured phrases, and have better parsing quality than the long and complex sentences in freestyle texts; pros/cons also have clear sentiment orientations. Thus, we use pros/cons to score the sentiment of adjectives, which requires strong polarity association. To obtain reliable ratings, we associate the adjectives in the \"pros\" of review entries which have a user rating 4 or 5, and associate the adjectives in the \"cons\" of review entries with user ratings 1 or 2 (the scale of user rating is 1 to 5). Reviews with rating 3 are on the boundary of sentiment, so we associate both sides with the overall rating. On the other hand, the frequencies of adverbs in free-style texts are much higher than those in pros/cons, as pros/cons mostly contain adjective-noun patterns. Thus, we use free-style texts instead of pros/cons to score the strength of adverbs. Partial results of the sentiment scoring are shown in Tables 3 and 4 . As shown in Table 3 , the polarity of sentiment as well as the degree of polarity of an adjective can be distinguished by its score. The higher the sentiment score is, the more positive the adjective is. Table 4 gives the scores of strength for most common adverbs. The higher the strength score is, the more the adverb scales up/down the degree of sentiment of the adjective it modifies. While \"not\" gets a strong negative score, some adverbs such as \"a little\" (-0.65) and \"a bit\" (-0.83) also get negative scores, indicating slightly less sentiment for the associated adjectives. To evaluate the performance of sentiment scoring, we randomly selected a subset of 1,000 adjective-noun phrases and asked two annotators to independently rate the sentiment of each phrase on a scale of 1 to 5. We compared the sentiment scoring between our system and the annotations in a measurement of mean distance:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 1150,
                        "end": 1164,
                        "text": "Tables 3 and 4",
                        "ref_id": "TABREF3"
                    },
                    {
                        "start": 1179,
                        "end": 1186,
                        "text": "Table 3",
                        "ref_id": "TABREF3"
                    },
                    {
                        "start": 1371,
                        "end": 1378,
                        "text": "Table 4",
                        "ref_id": "TABREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Sentiment Scoring",
                "sec_num": "6.2"
            },
            {
                "text": "\u202b\u0741\u073f\u074a\u073d\u0750\u074f\u0745\u0740\u202c \u0d4c \u0b35 |\u0bcc| \u2211 \u202b\u074e|\u202c \u0c22\u0bcc \u0d46 \u202b\u074e\u202c | (5)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sentiment Scoring",
                "sec_num": "6.2"
            },
            {
                "text": "where \u0735 represents the set of phrases, \u202b\u202c represents each phrase in the set \u0735, \u202b\u074e\u202c represents the rating on phrase \u202b\u202c from our sentiment scoring system, and \u202b\u074e\u202c represents the annotated rating on phrase \u202b.\u202c As shown in Table 5 , the obtained mean distance between the scoring from our approach and that from each annotation set is 0.46 and 0.43 respectively, based on the absolute rating scale from 1 to 5. This shows that the scoring of sentiment from our system is quite close to human annotation. The kappa agreement between the two annotation sets is 0.68, indicating high consistency between the annotators. The reliability of these results gives us sufficient confidence to make use of the scores of sentiments for summarization.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 219,
                        "end": 226,
                        "text": "Table 5",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Sentiment Scoring",
                "sec_num": "6.2"
            },
            {
                "text": "To examine the prediction of sentiment polarity, for each annotation set, we pooled the phrases with rating 4/5 into \"positive\", rating 1/2 into \"negative\", and rating 3 into \"neutral\". Then we rounded up the sentiment scores from our system to integers and pooled the scores into three polar-ity sets (\"positive\", \"negative\" and \"neutral\") in the same way. As shown in Table 5 , the obtained kappa agreement between the result from our system and that from each annotation set is 0.55 and 0.60 respectively. This shows reasonably high agreement on the polarity of sentiment between our system and human evaluation. To evaluate the combination of topic extraction and sentiment identification, we repeated the topic extraction experiments presented in Table 2 , but this time requiring as well a correct polarity assignment to obtain a match with the pros/cons ground truth. As shown in Table 6 , the COMB approach gets the highest recall both with and without topic clustering, and the recall from the LING approach is higher than that from the NB baseline in both cases as well, indicating the superiority of the proposed approach. The precision is stable among the different approaches, consistent with the case without the consideration of sentiment polarity.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 370,
                        "end": 377,
                        "text": "Table 5",
                        "ref_id": "TABREF5"
                    },
                    {
                        "start": 752,
                        "end": 759,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 887,
                        "end": 894,
                        "text": "Table 6",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Sentiment Scoring",
                "sec_num": "6.2"
            },
            {
                "text": "It is surprising that the parse-and-paraphrase method performs so well, despite the fact that it utilizes less than 80% of the data (parsable set). In this section, we will discuss two experiments that were done to tease apart the contributions of different variables. In both experiments, we compared the change in relative improvement in recall between NB and LING, relative to the values in Table 6 , in the with-clustering condition. In the table, LING obtains a score of 49.7% for recall, which is a 33% relative increase from the score for NB (37.4%). Three distinct factors could play a role in the improvement: the widowadjective topic hallucinations, the topic mapping for clustering, and the extracted phrases them-selves. An experiment involving omitting topic hallucinations from widow adjectives determined that these account for 12% of the relative increase. To evaluate the contribution of clustering, we replaced the mapping tables used by both systems with the edited one used by the ground truth computation. Thus, both systems made use of the same mapping table, removing this variable from consideration. This improved the performance of both systems (NB and LING), but resulted in a decrease of LING's relative improvement by 17%. This implies that LING's mapping table is superior. Since both systems use the same sentiment scores for adjectives and adverbs, the remainder of the difference (71%) must be due simply to higher quality extracted phrases.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 394,
                        "end": 401,
                        "text": "Table 6",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "7"
            },
            {
                "text": "We suspected that over-generated phrases (the 40% of phrases that find no mappings in the pros/cons) might not really be a problem. To test this hypothesis, we selected 100 reviews for their high density of extracted phrases, and manually evaluated all the over-generated phrases. We found that over 80% were well formed, correct, and informative. Therefore, a lower precision here does not necessarily mean poor performance, but instead shows that the pros/cons provided by users are often incomplete. By extracting summaries from review texts we can recover additional valuable information.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "7"
            },
            {
                "text": "This paper presents a parse-and-paraphrase approach to assessing the degree of sentiment for product reviews. A general purpose context free grammar is employed to parse review sentences, and semantic understanding methods are developed to extract representative negation-adverbadjective-noun phrases based on well-defined semantic rules. A language modeling-based method is proposed to cluster topics into respective categories. We also introduced in this paper a cumulative linear offset model for supporting the assessment of the strength of sentiment in adjectives and quantifiers/qualifiers (including negations) on a numerical scale. We demonstrated that the parse-and-paraphrase method can perform substantially better than a neighbor baseline on topic extraction from reviews even with less data. The future work focuses in two directions:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions & Future Work",
                "sec_num": "8"
            },
            {
                "text": "(1) building a relational database from the summaries and ratings and using it to enhance users' experiences in a multimodal spoken dialogue system; and (2) applying our techniques to other domains to demonstrate generality.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions & Future Work",
                "sec_num": "8"
            },
            {
                "text": "http://www.citysearch.com",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Multi-facet Rating of Product Reviews",
                "authors": [
                    {
                        "first": "Stefano",
                        "middle": [],
                        "last": "Baccianella",
                        "suffix": ""
                    },
                    {
                        "first": "Andrea",
                        "middle": [],
                        "last": "Esuli",
                        "suffix": ""
                    },
                    {
                        "first": "Fabrizio",
                        "middle": [],
                        "last": "Sebastiani",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of European Conference on Information Retrieval",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Stefano Baccianella, Andrea Esuli, and Fabrizio Se- bastiani. 2009. Multi-facet Rating of Product Re- views. In Proceedings of European Conference on Information Retrieval.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Learning document-level semantic properties from free-text annotations",
                "authors": [
                    {
                        "first": "S",
                        "middle": [
                            "R K"
                        ],
                        "last": "Branavan",
                        "suffix": ""
                    },
                    {
                        "first": "Harr",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Eisenstein",
                        "suffix": ""
                    },
                    {
                        "first": "Regina",
                        "middle": [],
                        "last": "Barzilay",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the Annual Conference of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S.R.K. Branavan, Harr Chen, Jacob Eisenstein, and Regina Barzilay. 2008. Learning document-level semantic properties from free-text annotations. In Proceedings of the Annual Conference of the As- sociation for Computational Linguistics.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Multi-Document Summarization of Evaluative Text",
                "authors": [
                    {
                        "first": "Giuseppe",
                        "middle": [],
                        "last": "Carenini",
                        "suffix": ""
                    },
                    {
                        "first": "Raymond",
                        "middle": [],
                        "last": "Ng",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Pauls",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the Conference of the European Chapter of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Giuseppe Carenini, Raymond Ng, and Adam Pauls 2006. Multi-Document Summarization of Evalua- tive Text. In Proceedings of the Conference of the European Chapter of the Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Mining the peanut gallery: opinion extraction and semantic classification of product reviews",
                "authors": [
                    {
                        "first": "Kushal",
                        "middle": [],
                        "last": "Dave",
                        "suffix": ""
                    },
                    {
                        "first": "Steve",
                        "middle": [],
                        "last": "Lawrence",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [
                            "M"
                        ],
                        "last": "Pennock",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of the International Conference on World Wide Web",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kushal Dave, Steve Lawrence, and David M. Pen- nock. 2003. Mining the peanut gallery: opinion ex- traction and semantic classification of product re- views. In Proceedings of the International Confe- rence on World Wide Web.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "SENTI-WORDNET: A Publicly Available Lexical Resource for Opinion Mining",
                "authors": [
                    {
                        "first": "Andrea",
                        "middle": [],
                        "last": "Esuli",
                        "suffix": ""
                    },
                    {
                        "first": "Fabrizio",
                        "middle": [],
                        "last": "Sebastiani",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the 5th Conference on Language Resources and Evaluation",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andrea Esuli and Fabrizio Sebastiani. 2006. SENTI- WORDNET: A Publicly Available Lexical Re- source for Opinion Mining. In Proceedings of the 5th Conference on Language Resources and Evalu- ation.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Pulse: Mining customer opinions from free text",
                "authors": [
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Gamon",
                        "suffix": ""
                    },
                    {
                        "first": "Anthony",
                        "middle": [],
                        "last": "Aue",
                        "suffix": ""
                    },
                    {
                        "first": "Simon",
                        "middle": [],
                        "last": "Corston-Oliver",
                        "suffix": ""
                    },
                    {
                        "first": "Eric",
                        "middle": [],
                        "last": "Ringger",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the 6th International Symposium on Intelligent Data Analysis",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Michael Gamon, Anthony Aue, Simon Corston-Oliver, and Eric Ringger . 2005. Pulse: Mining customer opinions from free text. In Proceedings of the 6th International Symposium on Intelligent Data Anal- ysis.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Seeing stars when there aren't many stars: Graph-based semi-supervised learning for sentiment categorization",
                "authors": [
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaojin",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "HLT-NAACL 2006 Workshop on Textgraphs: Graph-based Algorithms for Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andrew Goldberg and Xiaojin Zhu. 2006. Seeing stars when there aren't many stars: Graph-based semi-supervised learning for sentiment categoriza- tion. In HLT-NAACL 2006 Workshop on Text- graphs: Graph-based Algorithms for Natural Lan- guage Processing.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Mining and summarizing customer reviews",
                "authors": [
                    {
                        "first": "Minqing",
                        "middle": [],
                        "last": "Hu",
                        "suffix": ""
                    },
                    {
                        "first": "Bing",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of the 2004 ACM SIGKDD international conference on Knowledge Discovery and Data mining",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Minqing Hu and Bing Liu. 2004a. Mining and sum- marizing customer reviews. In Proceedings of the 2004 ACM SIGKDD international conference on Knowledge Discovery and Data mining.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Mining Opinion Features in Customer Reviews",
                "authors": [
                    {
                        "first": "Minqing",
                        "middle": [],
                        "last": "Hu",
                        "suffix": ""
                    },
                    {
                        "first": "Bing",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of Nineteenth National Conference on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Minqing Hu and Bing Liu. 2004b. Mining Opinion Features in Customer Reviews. In Proceedings of Nineteenth National Conference on Artificial Intel- ligence.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Automatic identification of pro and con reasons in online reviews",
                "authors": [
                    {
                        "first": "Min",
                        "middle": [],
                        "last": "Soo",
                        "suffix": ""
                    },
                    {
                        "first": "Eduard",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Hovy",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the COLING/ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Soo-Min Kim and Eduard Hovy. 2006. Automatic identification of pro and con reasons in online re- views. In Proceedings of the COLING/ACL 2006",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Main Conference Poster Sessions",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "483--490",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Main Conference Poster Sessions, pages 483-490.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Opinion observer: analyzing and comparing opinions on the web",
                "authors": [
                    {
                        "first": "Bing",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Minqing",
                        "middle": [],
                        "last": "Hu",
                        "suffix": ""
                    },
                    {
                        "first": "Junsheng",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of International Conference on World Wide Web",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bing Liu, Minqing Hu, and Junsheng Cheng. 2005. Opinion observer: analyzing and comparing opi- nions on the web. In Proceedings of International Conference on World Wide Web.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Low-Quality Product Review Detection in Opinion Summarization",
                "authors": [
                    {
                        "first": "Jingjing",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Yunbo",
                        "middle": [],
                        "last": "Cao",
                        "suffix": ""
                    },
                    {
                        "first": "Chin-Yew",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "Yalou",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Ming",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jingjing Liu, Yunbo Cao, Chin-Yew Lin, Yalou Huang, and Ming Zhou. 2007. Low-Quality Prod- uct Review Detection in Opinion Summarization. In Proceedings of the Conference on Empirical Methods in Natural Language Processing.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts",
                "authors": [
                    {
                        "first": "Bo",
                        "middle": [],
                        "last": "Pang",
                        "suffix": ""
                    },
                    {
                        "first": "Lillian",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of the Annual Conference of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bo Pang and Lillian Lee. 2004. A sentimental educa- tion: Sentiment analysis using subjectivity summa- rization based on minimum cuts. In Proceedings of the Annual Conference of the Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales",
                "authors": [
                    {
                        "first": "Bo",
                        "middle": [],
                        "last": "Pang",
                        "suffix": ""
                    },
                    {
                        "first": "Lillian",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the Annual Conference of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bo Pang and Lillian Lee. 2005. Seeing stars: Exploit- ing class relationships for sentiment categorization with respect to rating scales. In Proceedings of the Annual Conference of the Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Thumbs up? Sentiment classification using machine learning techniques",
                "authors": [
                    {
                        "first": "Bo",
                        "middle": [],
                        "last": "Pang",
                        "suffix": ""
                    },
                    {
                        "first": "Lillian",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Vaithyanathan",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bo Pang, Lillian Lee, and S. Vaithyanathan. 2002. Thumbs up? Sentiment classification using ma- chine learning techniques. In Proceedings of the Conference on Empirical Methods in Natural Lan- guage Processing.",
                "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 Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A.M. Popescu and O. Etzioni. 2005. Extracting prod- uct features and opinions from reviews. In Pro- ceedings of the Conference on Empirical Methods in Natural Language Processing.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Multiple Aspect Ranking using the Good Grief Algorithm",
                "authors": [
                    {
                        "first": "Benjamin",
                        "middle": [],
                        "last": "Snyder",
                        "suffix": ""
                    },
                    {
                        "first": "Regina",
                        "middle": [],
                        "last": "Barzilay",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of the Joint Conference of the North American Chapter of the Association for Computational Linguistics and Human Language Technologies",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Benjamin Snyder and Regina Barzilay. 2007. Mul- tiple Aspect Ranking using the Good Grief Algo- rithm. In Proceedings of the Joint Conference of the North American Chapter of the Association for Computational Linguistics and Human Language Technologies.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Modeling online reviews with multi-grain topic models",
                "authors": [
                    {
                        "first": "Ivan",
                        "middle": [],
                        "last": "Titov",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Mcdonald",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the 17h International Conference on World Wide Web",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ivan Titov and Ryan McDonald. 2008a. Modeling online reviews with multi-grain topic models. In Proceedings of the 17h International Conference on World Wide Web.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "A Joint Model of Text and Aspect Ratings for Sentiment Summarization",
                "authors": [
                    {
                        "first": "Ivan",
                        "middle": [],
                        "last": "Titov",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Mcdonald",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the Annual Conference of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ivan Titov and Ryan McDonald. 2008b. A Joint Mod- el of Text and Aspect Ratings for Sentiment Sum- marization. In Proceedings of the Annual Confe- rence of the Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Thumbs up or thumbs down? Sentiment orientation applied to unsupervised classification of reviews",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Peter",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Turney",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of the Annual Conference of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Peter D. Turney. 2002. Thumbs up or thumbs down? Sentiment orientation applied to unsupervised clas- sification of reviews. In Proceedings of the Annual Conference of the Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "On the Scope of Negation in English",
                "authors": [
                    {
                        "first": "Xuehui",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Sino-US English Teaching",
                "volume": "2",
                "issue": "9",
                "pages": "53--56",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xuehui Wu, 2005. On the Scope of Negation in Eng- lish, Sino-US English Teaching, Vol. 2, No. 9, Sep. 2005. pp. 53-56.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Mandarin Language Understanding in Dialogue Context",
                "authors": [
                    {
                        "first": "Yushi",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Jingjing",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Stephanie",
                        "middle": [],
                        "last": "Seneff",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of International Symposium on Chinese Spoken Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yushi Xu, Jingjing Liu, Stephanie Seneff. 2008. Mandarin Language Understanding in Dialogue Context. In Proceedings of International Sympo- sium on Chinese Spoken Language Processing.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Movie review mining and summarization",
                "authors": [
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Zhuang",
                        "suffix": ""
                    },
                    {
                        "first": "Feng",
                        "middle": [],
                        "last": "Jing",
                        "suffix": ""
                    },
                    {
                        "first": "Xiao-Yan",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the 15th ACM international conference on Information and knowledge management",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Li Zhuang, Feng Jing, and Xiao-Yan Zhu. 2006. Movie review mining and summarization. In Pro- ceedings of the 15th ACM international conference on Information and knowledge management.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "type_str": "figure",
                "uris": null,
                "text": "Framework of review processing."
            },
            "TABREF0": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Aspect</td><td>Extracted phrases</td><td>Rating</td></tr><tr><td>Atmosphere</td><td>very nice ambiance, outdoor patio</td><td>4.8</td></tr><tr><td>Food</td><td>not bad meal, quite authentic food</td><td>4.1</td></tr><tr><td>Place</td><td>not great place, very smoky restaurant</td><td>2.8</td></tr><tr><td>Price</td><td>so high bill, high cost, not cheap price</td><td>2.2</td></tr></table>",
                "text": "Example of review summary."
            },
            "TABREF1": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table/>",
                "text": "where \u202b\u074e\u202c\u123a\u073d\u0740\u0746\u123b represents the rating of adjective \u073d\u0740\u0746, \u202b\u074e\u202c\u123a\u073d\u0740\u202b\u0752\u202c\u123b represents the rating of adverb \u202b,\u0752\u0740\u073d\u202c and \u202b\u074e\u202c\u123a\u074a\u0741\u0743\u123b represents the rating of negation \u074a\u0741\u0743. \u0732\u202b\u0748\u202c\u123a\u073d\u0740\u0746\u123b represents the polarity of \u073d\u0740\u0746, assigned as 1 if \u073d\u0740\u0746 is positive, and -1 if negative. Thus, if \u073d\u0740\u0746 is positive, we assign a combined rating \u202b\u074e\u202c\u123a\u073d\u0740\u0746\u123b \u202b\u074e\u202c\u123a\u073d\u0740\u202b\u0752\u202c\u123b to this phrase. If it is negative, we assign \u202b\u074e\u202c\u123a\u073d\u0740\u0746\u123b \u0d46 \u202b\u074e\u202c\u123a\u073d\u0740\u202b\u0752\u202c\u123b. Specifically, if it is a negation case, we further assign a linear offset \u202b\u074e\u202c\u123a\u074a\u0741\u0743\u123b if \u073d\u0740\u0746 is positive or \u0d46\u202b\u074e\u202c\u123a\u074a\u0741\u0743\u123b if \u073d\u0740\u0746 is"
            },
            "TABREF2": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td colspan=\"4\">Experimental results of topic extraction by</td></tr><tr><td colspan=\"4\">the NB baseline, the proposed LING approach and</td></tr><tr><td colspan=\"3\">a combined system (COMB).</td><td/></tr><tr><td/><td/><td>No Clustering</td><td/></tr><tr><td/><td>NB</td><td>LING</td><td>COMB</td></tr><tr><td>Recall</td><td>39.6%</td><td>44.2%</td><td>48.1%</td></tr><tr><td>Precision</td><td>60.2%</td><td>60.0%</td><td>59.8%</td></tr><tr><td/><td/><td colspan=\"2\">With Clustering</td></tr><tr><td/><td>NB</td><td>LING</td><td>COMB</td></tr><tr><td>Recall</td><td>44.4%</td><td>57.0%</td><td>61.9%</td></tr><tr><td>Precision</td><td>56.8%</td><td>61.1%</td><td>60.8%</td></tr></table>",
                "text": ""
            },
            "TABREF3": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Adjective</td><td>Rating</td><td>Adjective</td><td>Rating</td></tr><tr><td>Excellent</td><td>5.0</td><td>Awesome</td><td>4.8</td></tr><tr><td>Easy</td><td>4.1</td><td>Great</td><td>4.4</td></tr><tr><td>Good</td><td>3.9</td><td>Limited</td><td>3.4</td></tr><tr><td>Inattentive</td><td>2.75</td><td>Overpriced</td><td>2.3</td></tr><tr><td>Rude</td><td>1.69</td><td>Horrible</td><td>1.3</td></tr></table>",
                "text": "Sentiment scoring for selected adjectives."
            },
            "TABREF4": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Adverb</td><td>Rating</td><td>Adverb</td><td>Rating</td></tr><tr><td>Super</td><td>0.58</td><td>Fairly</td><td>0.13</td></tr><tr><td>Extremely</td><td>0.54</td><td>Pretty</td><td>0.07</td></tr><tr><td>Incredibly</td><td>0.49</td><td>A little</td><td>-0.65</td></tr><tr><td>Very</td><td>0.44</td><td>A bit</td><td>-0.83</td></tr><tr><td>Really</td><td>0.39</td><td>Not</td><td>-3.10</td></tr></table>",
                "text": "Strength scoring for selected adverbs."
            },
            "TABREF5": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td/><td colspan=\"3\">Annotation 1 Annotation 2</td></tr><tr><td colspan=\"2\">Mean distance</td><td>0.46</td><td>0.43</td></tr><tr><td colspan=\"2\">Kappa agreement</td><td>0.55</td><td>0.60</td></tr><tr><td colspan=\"4\">Table 6. Experimental results of topic extraction</td></tr><tr><td colspan=\"4\">based on sentiment polarity matching.</td></tr><tr><td/><td/><td>No Clustering</td><td/></tr><tr><td/><td>NB</td><td>LING</td><td>COMB</td></tr><tr><td>Recall</td><td>34.5%</td><td>38.9%</td><td>42.2%</td></tr><tr><td>Precision</td><td>53.8%</td><td>54.0%</td><td>53.3%</td></tr><tr><td/><td/><td colspan=\"2\">With Clustering</td></tr><tr><td/><td>NB</td><td>LING</td><td>COMB</td></tr><tr><td>Recall</td><td>37.4%</td><td>49.7%</td><td>54.1%</td></tr><tr><td>Precision</td><td>48.5%</td><td>52.9%</td><td>51.4%</td></tr></table>",
                "text": "Comparison of sentiment scoring between the proposed approach and two annotation sets."
            }
        }
    }
}