File size: 142,320 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
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
{
    "paper_id": "2021",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T13:27:31.921189Z"
    },
    "title": "Resolving Implicit References in Instructional Texts",
    "authors": [
        {
            "first": "Talita",
            "middle": [],
            "last": "Rani",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Stuttgart Institute for Natural Language Processing",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Anthonio",
            "middle": [
                "Michael"
            ],
            "last": "Roth",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Stuttgart Institute for Natural Language Processing",
                "location": {}
            },
            "email": "rothml@ims.uni-stuttgart.de"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "The usage of (co-)referring expressions in discourse contributes to the coherence of a text. However, text comprehension can be difficult when referring expressions are non-verbalized and have to be resolved in the discourse context. In this paper, we propose a novel dataset of such implicit references, which we automatically derive from insertions of references in collaboratively edited how-to guides. Our dataset consists of 6,014 instances, making it one of the largest datasets of implicit references and a useful starting point to investigate misunderstandings caused by underspecified language. We test different methods for resolving implicit references in our dataset based on the Generative Pre-trained Transformer model (GPT) and compare them to heuristic baselines. Our experiments indicate that GPT can accurately resolve the majority of implicit references in our data. Finally, we investigate remaining errors and examine human preferences regarding different resolutions of an implicit reference given the discourse context.",
    "pdf_parse": {
        "paper_id": "2021",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "The usage of (co-)referring expressions in discourse contributes to the coherence of a text. However, text comprehension can be difficult when referring expressions are non-verbalized and have to be resolved in the discourse context. In this paper, we propose a novel dataset of such implicit references, which we automatically derive from insertions of references in collaboratively edited how-to guides. Our dataset consists of 6,014 instances, making it one of the largest datasets of implicit references and a useful starting point to investigate misunderstandings caused by underspecified language. We test different methods for resolving implicit references in our dataset based on the Generative Pre-trained Transformer model (GPT) and compare them to heuristic baselines. Our experiments indicate that GPT can accurately resolve the majority of implicit references in our data. Finally, we investigate remaining errors and examine human preferences regarding different resolutions of an implicit reference given the discourse context.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Implicit language phenomena can be challenging for both human and machine processing. For example, references play a crucial role in instructional texts as they provide answers to questions such as Which objects need to be used? If such references are not made explicitly, they might be clear to readers who have task-specific knowledge, but for others they might cause problems or misunderstandings. Resolving such implicit references could improve clarity and prevent problems in discourse processing when multiple interpretations exist.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In natural language processing, implicit references have been handled as part of existing tasks such as semantic role labeling of implicit arguments (Gerber and Chai, 2012, cf. \u00a73) . Implicit arguments are generally hard to model computationally because they do not show up in easy to",
                "cite_spans": [
                    {
                        "start": 149,
                        "end": 180,
                        "text": "(Gerber and Chai, 2012, cf. \u00a73)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "(1) Place the ground beef in a microwave.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "(2) Microwave until it finishes thawing.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "(3) Use within 1 or 2 days.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "(3') Use the beef within 1 or 2 days.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "(3\") Use the microwave within 1 or 2 days. Table 1 : Simplified example based on the wikiHow article \"Defrost Ground Beef\": sentences (1-3) show the original version of a text. Sentence (3') and (3\") show revised versions that include a manually inserted or automatically generated reference (see \u00a75), respectively.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 43,
                        "end": 50,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "learn surface patterns (Ruppenhofer et al., 2009) . The use of role-semantic formalisms further complicates progress in this direction because manual annotation requires trained annotators and previous training datasets have been comparatively small. For example, most datasets of implicit arguments consist of just hundreds of instances or only predicate-specific annotations (Moor et al., 2013) . We propose a task and dataset of implicit references which we obtain without manual annotation. Specifically, we create a dataset by extracting insertions of references in the revision history of collaboratively edited how-to guides. Previous work has shown that revisions in instructional texts are typically made to improve a text (Anthonio and Roth, 2020) . Based on this observation, we assume that explicit references are inserted when an implicit reference is perceived as problematic in discourse context. A simplified example from our dataset and an illustration of our task are provided in Table 1 .",
                "cite_spans": [
                    {
                        "start": 23,
                        "end": 49,
                        "text": "(Ruppenhofer et al., 2009)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 377,
                        "end": 396,
                        "text": "(Moor et al., 2013)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 732,
                        "end": 757,
                        "text": "(Anthonio and Roth, 2020)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 998,
                        "end": 1005,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "As shown by the example, our data consists of insertions of single references and the task is to predict these inserted references. As a benefit over existing work, the task does not depend on any formalism of role semantics, which means that models can be evaluated in an end-to-end setting.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "As a dataset for the proposed task, we provide 6,014 instances of implicit references, which we extracted automatically by comparing different versions of articles in wikiHow 1 . In practice, we make use of an existing resource of wikiHow sentences and revisions called wikiHowToImprove (Anthonio et al., 2020) , from which we select specifically those cases in which a referring expression was inserted that refers to an entity mentioned in the preceding context. Based on this dataset, we set up a cloze task in which we evaluate the ability of computational models to generate references for insertions that occur naturally in publicly available texts. Finally, we analyze predictions of different modeling approaches as well as differences between model-generated and human-inserted references, which provide useful insights regarding potential weaknesses of existing models and potential causes of human misunderstandings.",
                "cite_spans": [
                    {
                        "start": 287,
                        "end": 310,
                        "text": "(Anthonio et al., 2020)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "In sum, we make the following contributions:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "\u2022 We propose a new task that requires NLP models to generate explicit references to resolve cases of implicit language ( \u00a72).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "\u2022 We provide a dataset of 6,014 texts that involve the insertion of an explicit reference according to the text's revision history ( \u00a74).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "\u2022 We show that methods based on the Generative Pre-trained Transformer model (GPT) present a strong baseline for this task ( \u00a75).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "\u2022 We conduct two analyses that shed light on the strengths of GPT and reveal potential avenues for future research ( \u00a76).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Defrost Ground Beef",
                "sec_num": null
            },
            {
                "text": "Task definition. We formally define the task of resolving implicit references as a generation task that requires the prediction of a reference S, given:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Implicit Reference Resolution",
                "sec_num": "2"
            },
            {
                "text": "1. The original/revised sentence and its preceding context C p , which includes at least one mention that co-refers to the correct reference (for the example shown in Table 1 : Place the ground beef in a microwave. Microwave until it finishes thawing. Use ).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 167,
                        "end": 174,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Implicit Reference Resolution",
                "sec_num": "2"
            },
            {
                "text": "2. The number of tokens L of the reference to be generated according to the final version of a sentence (in case of the example: 2).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Implicit Reference Resolution",
                "sec_num": "2"
            },
            {
                "text": "Performing this task requires a model to generate the sequence of tokens s 1 . . . s L for the reference S conditioned on the context C p , C f . In practice, the full task can be approached by first sampling candidate reference tokens r 1 . . . r L from a conditional probability distribution P (r i |C p , r 1 . . . r i\u22121 ) and then re-ranking the highest scoring candidates according to the full sequence probability",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Implicit Reference Resolution",
                "sec_num": "2"
            },
            {
                "text": "P (C p , r 1 . . . r L , C f ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Implicit Reference Resolution",
                "sec_num": "2"
            },
            {
                "text": "Formulating the task in this way enables a direct application of language models and we demonstrate suitable baselines based on an auto-regressive language model in Section 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Implicit Reference Resolution",
                "sec_num": "2"
            },
            {
                "text": "The task of resolving implicit references can be viewed as a modified version of implicit argument labeling. First studies on implicit argument labeling were conducted by Gerber and Chai (2010) and Ruppenhofer et al. (2009) . Gerber and Chai (2010) collected a dataset by manually labeling implicit arguments of 10 different nominal predicates in NomBank (Meyers et al., 2004) , yielding about 1,000 instances. Ruppenhofer et al. (2009) created a dataset through manual annotation of fictional text. Their dataset contains more different predicates than previous studies, but is smaller in size. More recent studies make use of the two datasets and attempted to create additional training data artificially (Silberer and Frank, 2012; Laparra and Rigau, 2013a,b; Chiarcos and Schenk, 2015) . Many of them are based on co-reference and discourse salience, which we also use for our baselines. Schenk and Chiarcos (2016) propose an unsupervised approach by aligning implicit arguments to semantic role labeling annotated data. Erk (2019, 2018) generated large amounts of training data automatically using co-reference resolution. They also build a neural model based on argument fillers that occur multiple times in a narrative event chain. Finally, there are also datasets with domain-specific annotations such as geographic-event roles (Ebner et al., 2020) and on recipes (Jiang et al., 2020) .",
                "cite_spans": [
                    {
                        "start": 171,
                        "end": 193,
                        "text": "Gerber and Chai (2010)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 198,
                        "end": 223,
                        "text": "Ruppenhofer et al. (2009)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 226,
                        "end": 248,
                        "text": "Gerber and Chai (2010)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 355,
                        "end": 376,
                        "text": "(Meyers et al., 2004)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 411,
                        "end": 436,
                        "text": "Ruppenhofer et al. (2009)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 707,
                        "end": 733,
                        "text": "(Silberer and Frank, 2012;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 734,
                        "end": 761,
                        "text": "Laparra and Rigau, 2013a,b;",
                        "ref_id": null
                    },
                    {
                        "start": 762,
                        "end": 788,
                        "text": "Chiarcos and Schenk, 2015)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 891,
                        "end": 917,
                        "text": "Schenk and Chiarcos (2016)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 1024,
                        "end": 1040,
                        "text": "Erk (2019, 2018)",
                        "ref_id": null
                    },
                    {
                        "start": 1335,
                        "end": 1355,
                        "text": "(Ebner et al., 2020)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 1371,
                        "end": 1391,
                        "text": "(Jiang et al., 2020)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "3"
            },
            {
                "text": "Another closely related task is zero anaphora resolution, which has been extensively studied in pro-drop languages such as Chinese (Yeh and Chen, 2003) and Japanese (Taira et al., 2008; Isozaki and Hirao, 2003; Seki et al., 2002; Nakaiwa, 1997; Imamura et al., 2009) . A closely related study to ours is Imamura et al. (2009) , who used language model probabilities as features.",
                "cite_spans": [
                    {
                        "start": 131,
                        "end": 151,
                        "text": "(Yeh and Chen, 2003)",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 165,
                        "end": 185,
                        "text": "(Taira et al., 2008;",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 186,
                        "end": 210,
                        "text": "Isozaki and Hirao, 2003;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 211,
                        "end": 229,
                        "text": "Seki et al., 2002;",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 230,
                        "end": 244,
                        "text": "Nakaiwa, 1997;",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 245,
                        "end": 266,
                        "text": "Imamura et al., 2009)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 304,
                        "end": 325,
                        "text": "Imamura et al. (2009)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "3"
            },
            {
                "text": "As a commonality, previous work addresses semantic arguments of predicates that are realized outside a local syntactic scope. Our definition of implicit references subsumes such arguments, with the main difference that our task does not require the type of an argument or its semantic role to be specified. As a consequence, references in our task can fill one, none or multiple roles of different predicates. Once the correct reference has been identified, our task additionally requires the generation of a referring expression. This task has been addressed separately in previous work, for instance, using rule-based approaches (Reiter and Dale, 2000) , feature-based machine learning (Nenkova and McKeown, 2003; Greenbacker and McCoy, 2009; Same and van Deemter, 2020; Kibrik et al., 2016) , and deep neural networks (Castro Ferreira et al., 2016; Cao and Cheung, 2019) .",
                "cite_spans": [
                    {
                        "start": 631,
                        "end": 654,
                        "text": "(Reiter and Dale, 2000)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 688,
                        "end": 715,
                        "text": "(Nenkova and McKeown, 2003;",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 716,
                        "end": 744,
                        "text": "Greenbacker and McCoy, 2009;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 745,
                        "end": 772,
                        "text": "Same and van Deemter, 2020;",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 773,
                        "end": 793,
                        "text": "Kibrik et al., 2016)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 821,
                        "end": 851,
                        "text": "(Castro Ferreira et al., 2016;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 852,
                        "end": 873,
                        "text": "Cao and Cheung, 2019)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "3"
            },
            {
                "text": "The starting point for our data are revision histories from wikiHow, in which we can find insertions of references that were implicit in earlier versions of a sentence. We use wikiHowToImprove (Anthonio et al., 2020) , a resource derived from wikiHow that consists of approximately 2.7 million sentences and their revisions. For our purpose, we extract sentences in which a reference was inserted during revision. Most of the sentences in wikiHow are only edited once (about 83%). In other cases, intermediate versions are mostly the result of stylistic refinements or typo corrections. Therefore, we only make use of the final version of a sentence (henceforth revised sentence), which includes an inserted reference, and the original sentence, in which the reference is assumed to be implicit. As a result, each data point in our collection consists of a pair of two versions of a sentence, henceforth original-revised sentence pair. We describe our selection of implicit references in Section 4.1 and present the data statistics in Section 4.2.",
                "cite_spans": [
                    {
                        "start": 193,
                        "end": 216,
                        "text": "(Anthonio et al., 2020)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data",
                "sec_num": "4"
            },
            {
                "text": "In order to find pairs with an implicit reference in the original sentence that is explicit in the revised sentence, we take the instances where the revised sentence was created by inserting a word or contiguous set of words in the original sentence. In other words, eliminating the insertion from the revised sentence yields the original sentence. This is a logical starting point, as the implicit reference in the original sentence can be verbalized through insertion. We find cases with contiguous insertions in wikiHowToImprove by computing the differences between the original and revised sentence using difflib. 2 As a result, we found 336,129 sentence pairs in which the original sentence was only modified by a contiguous insertion.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection",
                "sec_num": "4.1"
            },
            {
                "text": "In the next step, we identify the subset of insertions that are referential and resolvable in context: that is, we identify words and phrases that refer to a discourse entity. Our study focuses on insertions of single references (i.e., referring expressions that refer to exactly one discourse entity), which are usually not verbalized by sequences exceeding three tokens. Therefore, we only consider insertions that consist of one, two or three word tokens (i.e., unigram, bigram and trigram insertion). We identify references by obtaining co-reference chains on the paragraph level using the Stanza 3 coreference parser. More specifically, by using a combination of the revised sentence and the original context, we can identify referring expressions that are explicit in the revised sentence and coreferent with discourse entities in the original context. Therefore, we parse the revised sentence and the preceding sentences from the original context, within the same paragraph.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection",
                "sec_num": "4.1"
            },
            {
                "text": "We add the corresponding original-revised sentence pair to our collection if the full span of the insertion (full insertion) or parts of it refer to an entity in the discourse context. In other words, the insertion can contain tokens in addition to the referring expression. However, we only keep insertions that include additional tokens if the additional tokens are required grammatically, given their position in the sentence (e.g., of you, of the shoe). In particular, we keep the insertions that consist of a reference and specific types of function words (determiners, prepositions) or punctuation. 4 We excluded cases with conjunctions and non-function words as these insertions mainly add or extend factual information. Examples of different insertions",
                "cite_spans": [
                    {
                        "start": 605,
                        "end": 606,
                        "text": "4",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection",
                "sec_num": "4.1"
            },
            {
                "text": "Reference Example",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "unigram (N =2,599) unigram",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "This treatment can be performed by a dermatologist but it is quite expensive.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "bigram (N =1,837) unigram (N =700)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "If you are using the mobile app, tap the \"More\" button and then tap your name. Select the photo's of you tab. bigram (N =1,137) It's not pleasant to read a book that has been \"personalized\" by someone else. If it rains or the book gets lost, you'll have to pay to replace it.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "trigram (N =1,578) unigram (N =118)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "Bend your left knee and lift it ( as close as you can get it).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "bigram (N =1,370)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "1. Clean canvas shoes by spot washing using a mild detergent and soft toothbrush. Test the spray on the tongue of the shoe to make sure it won't stain.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Insertion",
                "sec_num": null
            },
            {
                "text": "Check the outer labelling on the ham shank to see if its fully cooked. If it isn't, use the other method instead. Remove the wrapping and place the ham shank in a roasting pan. Table 2 : Examples of from our dataset: underlined tokens mark an insertion, tokens in bold highlight references to the same entity. Tokens that are underlined and highlighted are the reference tokens to be predicted in our task.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 177,
                        "end": 184,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "trigram (N =90)",
                "sec_num": null
            },
            {
                "text": "Note that the span of the reference can differ from the insertion because of additional tokens (e.g., punctuation).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "trigram (N =90)",
                "sec_num": null
            },
            {
                "text": "and inserted references are shown in Table 2 . Note that some sentences contain grammar/spelling related errors, which were not corrected in the shown versions of the wikiHow articles.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 37,
                        "end": 44,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "trigram (N =90)",
                "sec_num": null
            },
            {
                "text": "In total, our collection procedure yields 6,014 instances. More specifically, it contains 2,599 unigram (43.22%), 1,837 bigram (30.50%) and 1,578 trigram (26.24%) insertions. Table 2 shows examples of references and insertions and how they are distributed over different lengths. The numbers indicate that a majority of references are unigrams (N = 3, 854) and that only a small proportion are trigrams. The table also shows that most references consist of the full insertion (56%), which are 2,599 unigrams, 700 bigrams and 90 trigrams. Figure 1 indicates the positions of the closest antecedent to resolve an implicit reference. The distribution shows that most references refer to an entity in the same sentence (46.33%, N = 2, 786) or to an entity in the previous sentence (25.21%, N = 1, 517). The remaining instances can be resolved within 3 up to 75 sentences. Finally, we observe that in the majority of the 6,014 instances, the reference is mentioned only once (43.15%, N = 2, 595), twice (18.12%, N = 1, 090) or three times (9.38%, N = 564) in the original context.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 175,
                        "end": 182,
                        "text": "Table 2",
                        "ref_id": null
                    },
                    {
                        "start": 538,
                        "end": 546,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Statistics",
                "sec_num": "4.2"
            },
            {
                "text": "In the remainder of this paper, we conduct experiments with our collection of 6,014 implicit references, which we split into a train (81.09%, N = 4, 877), development (9.83%, N = 591) and test (9.08%, N = 546) set, following the original split by article of wikiHowToImprove. 5",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statistics",
                "sec_num": "4.2"
            },
            {
                "text": "In this section, we describe a set of experiments in which we investigate the use of a transformer-based language model for the task of resolving implicit references. In particular, we aim to answer the following research questions: Can we find the manually inserted reference among the top completions predicted by a language model, and is it possible to select a correct prediction based on its fit in the sentence or paragraph context? We describe our experimental set-up in Section 5.1 and our results in Section 5.2. Further analyses are provided in Section 6.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language Model Experiments",
                "sec_num": "5"
            },
            {
                "text": "Data. The starting point for our investigation are the 6,014 instances of original-revised sentence pairs described in Section 4. Each revision involves the insertion of a reference into the given sentence. That is, the revised sentence always contains a reference that was not explicitly present in the original sentence. The full insertion may consist of one, two or three tokens, and it may contain function words in addition to the reference itself.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental setting",
                "sec_num": "5.1"
            },
            {
                "text": "Method. Resolving implicit references of varying length requires a generative model. We chose the Generative Pre-trained Transformer model (GPT) described by Radford et al. (2018) as a benchmark model because it fulfils this requirement and because it is pre-trained on data that does not overlap with our development and test sets. 6 Since GPT is an auto-regressive language model, which means that predictions are made word-by-word (unidirectional), we apply an additional re-ranking procedure over the top-100 generated sequences and their full (left and right) context. For re-ranking, we use the same GPT model and compute its perplexity score for whole sequences on two levels of context: (a) full sentence (+S-perplexity) and (b) sentence plus preceding paragraph context (+P-perplexity). Finally, we also fine-tune the GPT model on our training set to improve its fit to how-to guides (+fine-tuning).",
                "cite_spans": [
                    {
                        "start": 158,
                        "end": 179,
                        "text": "Radford et al. (2018)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 333,
                        "end": 334,
                        "text": "6",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental setting",
                "sec_num": "5.1"
            },
            {
                "text": "Upper bound and baselines. We approximate an upper bound on our data by assessing the performance of a model that has access to the inserted reference itself, namely the coreference parser used during data creation (see Section 4.1). We further compare GPT to the following baselines: Most-Frequent always selects the most frequent refer-ring expression of the most frequent entity in the context, ClosestRef selects referring expression(s) from the preceding context by how close they are to the point of the insertion, and TF-IDF ranks possible n-grams (where n equals the number of tokens in the manually inserted reference) by their tf-idf score (Jones, 1972) , for which we take into account all training and development documents.",
                "cite_spans": [
                    {
                        "start": 650,
                        "end": 663,
                        "text": "(Jones, 1972)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental setting",
                "sec_num": "5.1"
            },
            {
                "text": "Evaluation. We evaluate each model by its ability to generate the tokens that are part of the reference inserted in the revised version of a sentence. 7 We count a generated reference as correct if all generated tokens match the tokens in the humanproduced reference. To allow for minor variations in spelling, we ignore case when measuring recall (i.e., the relative number of correctly retrieved references) among the top-1 (R@1), top-10 (R@10) and top-100 (R@100) generated sequences.",
                "cite_spans": [
                    {
                        "start": 151,
                        "end": 152,
                        "text": "7",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experimental setting",
                "sec_num": "5.1"
            },
            {
                "text": "We first address our initial question: Can we find manually inserted references among the top completions predicted by a language model? The scores listed in Table 3 indicate the proportion of exact matches that are found within the top-1, top-10 and top-100 references generated by the pre-trained GPT model. The numbers show a similar performance of GPT on the development and test set: In about 37% of the cases, the first-best generated reference is identical to the manually inserted reference. In about 83% of the cases, the manually inserted reference can be found within the top-100 generated references. This result is close to our approximated upper bound: in a random sample from the development set, we found a coreference model, which has access to the manually inserted reference itself, to predict the correct co-reference chain in 86 out of 100 cases (for details, see Appendix A in the supplementary material).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 158,
                        "end": 165,
                        "text": "Table 3",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.2"
            },
            {
                "text": "Model comparison. We next attempt to answer our second question, namely is it possible to select the correct reference based on its fit in the sentence or paragraph context? We evaluate two additional steps for selecting references based on the top sequences generated by GPT: model fine-tuning and re-ranking based on sentence-level or paragraphlevel perplexity. We observe that GPT substantially outperforms all three baselines. Combining GPT with finetuning and paragraph-level perplexity re-ranking leads to an accurate top-1 prediction of the inserted reference in 57.4% of the cases on the test set. In 80.8% of the cases, the inserted reference can be found within the top-10 re-ranked sequences. In ablation experiments on the development set, we find that a combination of finetuning and perplexity-based re-ranking is necessary to achieve such high results. Fine-tuning and re-ranking based on sentence-level perplexity only improve R@1 by 2.5 to 3 percentage points, respectively. Without fine-tuning, re-ranking on the sentence level even reduces the chance of finding the correct reference within the top-10 sequences (R@10). Only re-ranking on the paragraph level consistently improves results, up to 8.2 and 16.8 percentage points in R@10 and R@1, respectively.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.2"
            },
            {
                "text": "Discussion. We qualitatively analyzed the top-1 predictions of each method on the development set and observed the following trends: re-ranking on the sentence level generally helps in selecting grammatically suitable candidates when the top generated sequences by the original or fine-tuned model does not fit syntactically, for example, due to number or case disagreements. Fine-tuning GPT seems to adapt the scoring of generated references to better match their occurrences in how-to guides: for example, the pronouns you and them are more frequent in this genre than I and we. Finally, we observe that re-ranking on the paragraph level considerably improves the selection of noun phrases that resemble references to entities in the discourse. Whereas the sentence-level method often produces generic references (underlined) that make sense superficially (e.g., Clean off the surface of the glass), top candidates in the paragraph-level method plausibly fit also in the specific context (e.g., Clean off the surface of your typewriter). We discuss the top predictions of both re-ranking methods in more detail in Section 6.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.2"
            },
            {
                "text": "In this section, we aim to answer two questions evoked by the results in Section 5. First, we ask how fine-tuning and perplexity-based re-ranking improved the scoring of the top-100 generated sequences and what differences can be seen among the re-ranked top-10 sequences (Section 6.1). Secondly, we investigate the plausibility of the two highest ranked fillers generated by the model (Section 6.2). The latter analysis provides us with insights regarding the existence of a single most plausible filler (or whether none/two fillers can be plausible) given the discourse context. In cases where the human-inserted reference is among the top-2, the analysis also makes it possible for us to find if/when the human-inserted reference is identified as the most plausible.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Analysis",
                "sec_num": "6"
            },
            {
                "text": "We compare the generated top-10 sequences of the fine-tuned GPT model and the re-ranked variants on the development set. Predictions for three example sentences are shown in Table 5 . The examples indicate that sequences generated by the fine-tuned GPT model often lead to ungrammatical sentences (highlighted in italics). For the re-ranked variants, we observe that such sequences are scored lower and no longer appear among the top ranks. The examples based on sentence-level reranking reveal two unfortunate side-effects: The first is that non-referential candidates may end up in higher positions, simply because they lead to a grammatical sequence. This is particularly visible in Example (2), where the correct sequence was generated by GPT+FT, but ended up outside the top-10 when using GPT+FT+S-perplexity. The second caveat is that sentence-level perplexity increases the rank of entities that are plausible within the sentence but unrelated to the activity described in the article. This is especially visible in Example (3) in Table 5 : the phrases the number, the office, a friend, the person all seem reasonable candidates in the context of calling someone, but none of them directly correspond to the salon mentioned in the article. The same applies to Example (1), in which the context mentioned a container but contains no references to a pot, bowl, cage or bag. It seems that these candidates simply mimic the usage of common knowledge, especially because none of the candidates occur in the preceding context. The aforementioned reasons could explain why sentence-level perplexity without fine-tuning decreased the recall of the manually-inserted reference among the top-10 candidates (see Section 5).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 174,
                        "end": 181,
                        "text": "Table 5",
                        "ref_id": null
                    },
                    {
                        "start": 1038,
                        "end": 1045,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Sentence vs. Paragraph Perplexity",
                "sec_num": "6.1"
            },
            {
                "text": "A final interesting observation is that the caveats caused by the sentence-based re-ranking are less present when applying paragraph-level perplexity. After re-ranking on the paragraph level, we find many of the top candidates to be either repetitions of words and phrases from the context or to be closely related to the manually-inserted sequence. This is illustrated by all examples in Table 5 . A quantitative analysis further confirmed this insight: based on paragraph-level perplexity, over 20% of the top-10 ranked bigrams and trigrams appear literally in the preceding context, compared to 11% when using sentence-level perplexity.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 389,
                        "end": 396,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Sentence vs. Paragraph Perplexity",
                "sec_num": "6.1"
            },
            {
                "text": "In the first analysis, we discussed the impact of re-ranking the top-100 sequences generated by the fine-tuned model. However, even the best reranking procedure is insufficient to avoid errors in our approach when the manually inserted reference from a revised sentence does not appear among the generated sequences. In this section, we address two questions. First, we investigate whether the human-inserted reference is always the one that best fits the sentence given the context. Secondly, we assess the plausibility of the top-2 completions by the model in case both are different from the human-inserted reference. These questions are motivated by the results from an internal analysis, in which we found a few instances where an annotator preferred a model-generated reference over the human-inserted reference or had no preference between the two options (a full report is provided in the supplementary material, Appendix B).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "Set-up. We take 100 instances from the development set with their top-2 completions provided by the fine-tuned model and ask a student assistant with a background in Computational Linguistics to provide annotations. We randomly select 50 instances where the human-inserted reference is identical with the best generated sequence (\"humaninsertion among top-2\") and 50 where this is not the case (\"human-insertion not among top-2\"). We show an annotator two versions of the sentence in randomized order: one with the highest ranked sequence and one with the second-highest ranked sequence generated by the model. We show each version together with the preceding sentences from the paragraph and highlight the generated reference. We ask the annotator to indicate the sentence that fits the context better, whether they both fit or whether neither fits. We discuss the main findings of this experiment below, and provide additional examples in the supplementary material (Appendix C). In the examples below, we underline the human-inserted reference.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "Human-insertion among top-2 (N = 50). The annotator indicated a strong preference for the human-inserted reference in most cases (N = 34). However, there were also cases in which the annotator indicated no preference (N = 13). This is likely the case because many generated sequences involved paraphrases of the humaninserted reference in the given context (e.g., the button/this button, the party/your party). The same holds for the remaining instances (N = 3), in which the annotator preferred the other generated sequence over the human-inserted reference (e.g., the sequence/the process). In these cases, it seems like annotations simply reflect personal preferences. Finally, we found no cases in which the annotator indicated that neither insertion was fitting. correct sequence GPT+FT GPT+FT+S-perplexity GPT+FT+P-perplexity",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "(1) Put the grasshoppers in the container.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "the container, it and, a container, the bottom, the lid, it. , it, , a plastic, a large, the plastic the water, the pot, the basket, a bowl, the cage, a bag, a pot, the bag, the refrigerator, a jar the container,. 5, the lid, a container, the bag, the bottom, the box, the water, a bag, the hole",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "(2) Rinse the parts before assembling.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "the parts, out the, off the, them off, and dry, your ke, them thoroughly, them with, them in, them out them out, them thoroughly, it out, them off, it off, the area, and rinse, each piece, and dry, each section the parts, them thoroughly, them off, them out, all parts, each part, the components, the pipes, both parts, and dry",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "(3) Call the salon and ask questions the salon, your salon, a salon, a local, the sal, a sal, up the, a hair, the hair, them and the number, the office, a friend, this number, them up, the person, a number, the store, the owner, them in the salon, your salon, a salon, their website, the spa, the stylist, them in, your stylist, each salon, a stylist Table 5 : The top-10 predictions for the fine-tuned GPT model and the reranked predictions using sentence-level and paragraph-level perplexity. Bold sequences represent the correct sequence, italic sequences are ungrammatical in context, and underlined sequences are references to entities that are not mentioned in the context.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 351,
                        "end": 358,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "Human-insertion not among top-2 (N = 50).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "In a majority of these cases, we found the annotator to select one of the model-generated sequences as best fitting, confirming that completions other than the human-inserted reference can be plausible (N = 29). A number of the selected sequences are paraphrases (N = 11) of the human-inserted reference, such as: the form/the application, the school/this school or semantically related sequences (N = 8) that differ from the humaninserted reference in terms of specificity (e.g., let the paint/the nails dry, wipe and dry the wood/the floors, remove the pan/the vegetables from the heat, your laptop/your mac).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "The remaining instances (N = 10) involve generated sequences that the annotator indicated as best fitting, although they are incompatible with the human-inserted reference (e.g., Islam/Christianity, the microwave/the freezer). We take these findings as an indicator that implicit references can be resolved incorrectly and therefore lead to misunderstandings, which could be modelled and anticipated by a language model. Finally, the annotator judged both top-2 sequences to be fitting in 17 cases and none to be fitting in 4 cases. In case of the latter, the completions led to an ungrammatical sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Plausibility of Generated Fillers",
                "sec_num": "6.2"
            },
            {
                "text": "In this paper, we introduced the task of resolving implicit references in instructional texts, which might be problematic for readers without prior knowledge of the instructed task. We approached the resolution of implicit references as a generation problem for which we leveraged original-revised sentence pairs from wikiHow. The considered pairs contained an explicit reference in the revised sentence which was non-verbalized in the original sentence. Our dataset is one of the largest datasets with implicit references and contains texts from the multiple different domains covered in wikiHow.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions",
                "sec_num": "7"
            },
            {
                "text": "We showed that a pre-trained language model is capable of predicting the human-produced insertion in a majority of cases. The best-performing method, which combines a fine-tuned GPT model and perplexity-based re-ranking, achieved results up to 57.4% (top-1) and 80.8% (top-10). Even without fine-tuning and re-ranking, 71.6% of the human-inserted references appeared in the top-10.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions",
                "sec_num": "7"
            },
            {
                "text": "We found sentence-level re-ranking useful to eliminate generated sequences that cause ungrammatical sentences and paragraph-level re-ranking to prioritize sequences that also occur in the preceding discourse. Our analysis revealed that the human-inserted reference is commonly found to fit the discourse better than a model-generated alter-native. However, we also found cases where other completions were plausible. In the future, we will extend this study and take it as a starting point for examining potential sources of misunderstanding.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions",
                "sec_num": "7"
            },
            {
                "text": "Human-inserted reference co-refers 1. Take your pregnant cat to the vet . As soon as you know , or suspect , that your cat might be pregnant , you should take her to the vets to get her checked over .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "1. On the sheet of foam, draw an outline like the one in the image. Make sure it can wrap around your bottle around once, and the strip pointing down can fold under the bottom of the bottle. 2. Cut the foam shape out. 3. Cut a piece of duct tape off the roll, so it will cover part of the bottom of the foam and about on each side. Take a new piece and start rolling it around the sides of the foam . Table 6 : Examples where the human-inserted reference co-refers with an entity in the context according to Stanza. The human-inserted reference is highlighted and underlined. The referring expressions within the same co-reference chain are highlighted.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 401,
                        "end": 408,
                        "text": "Table 6",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "Human-inserted reference does not co-refer",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "1. Know what kind of games you like (strategy, action, adventure, racing, rpg, simulators, etc. 2. Strategy games are good for you when thinking. They help you reinforce what you learned. Some strategy games are : Age Of Empires, The Settlers, City Of Heroes. Action games: these are liked by many people.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "1. Go get an ' on the floor ' cat scratch pad from anywhere. This one cost $6. My cats wo n't use it flat on the floor, like this. 2. Remove the corrugated cardboard from the box. They usually glue it down, probably because they do n't want you to flip it and re-use the back -side. (You can.) 3. Using the edge of a counter or table or other sturdy piece of furniture, break it to fit. Table 7 : Examples where the human-inserted reference did not co-refer with the referring expressions within the same co-reference chain according to Stanza. The human-inserted reference is highlighted and underlined, whereas the referring expressions within the same co-reference chain are highlighted.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 387,
                        "end": 394,
                        "text": "Table 7",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "In this section, we describe a study that we conducted to investigate the quality of the obtained co-reference chains from Stanza.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "Method. To investigate the quality of the coreference chains, we asked an annotator to identify whether the human-inserted reference occurred earlier in the context. We specifically showed the annotator the context and the revised sentence, in which we marked the human-inserted referring expression. We additionally highlighted the referring expressions that co-referred with the humaninserted reference. We asked the annotator to annotate 100 instances, which we randomly extracted from the development set of our data (see Section 4). The annotator was a student in a Computational Linguistics program.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "Results. The annotator found 86 instances where the human-produced reference co-referred with the referring expressions indicated by the Stanza coreference parser. We show two examples of these instances in Table 6 . In the remaining instances (N = 14), the human-inserted referring expression did not co-refer with the referring expressions indicated by Stanza. Nonetheless, it is still possible to find a suitable antecedent in such cases, as shown by the examples in Table 7 . We conclude from the obtained results that implicit references in our dataset (described in Section 4) are generally coreferent with an entity mentioned in the context.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 207,
                        "end": 214,
                        "text": "Table 6",
                        "ref_id": null
                    },
                    {
                        "start": 470,
                        "end": 477,
                        "text": "Table 7",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "A Co-reference Quality Analysis",
                "sec_num": null
            },
            {
                "text": "In Section 6.1, we discussed the impact of reranking the top-100 sequences generated by the fine-tuned GPT model. However, even the best reranking procedure is insufficient to avoid errors in our approach when the manually inserted reference from a revised sentence does not appear among the generated sequences. Therefore, we perform an additional study in which we analyze cases of human-produced references that do not show up among the top-100 candidates generated by the fine-tuned GPT model. Set-up. We select all instances for which the fine-tuned GPT model did not generate the humanproduced reference among the top-100 (N = 84) and ask one annotator to provide judgements. We show the annotator two versions of the revised sentence in randomized order: one containing the human-produced reference and the other containing the top-1 generated completion by GPT. Each version is shown together with the preceding sentences from the paragraph.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "In the annotation interface, we highlight the references and ask the annotator to select the version that fits the sentence better, given the context, or whether \"both fit\". We discuss the three different outcomes below.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "Preference for human insertion (N = 57). In most cases, the annotator labeled the humanproduced reference as being a better fit than the generated sequence. In 26 of these cases, the model-generated sequence was not a reference or the reference was accompanied by function words or punctuation. Both types of sequences usually yield an ungrammatical sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "The remaining 31 instances can be categorized into three groups: The first are cases where the generated sequence does not make sense in the given sentence position (N = 12), such as Rub the dog's coat with the chamois. Another subset (N = 13) consists of sequences that are referential but the insertion yields an ungrammatical sentence, for example: It can also be unpleasant to withdraw from your. The rest (N = 6) seem to be sensible references according to our observations, such as:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "(1) Leave your diya uncovered at room temperature, and do your best to keep it away from moisture. The clay should set after 24 hours. If you put your diya on a plate or mat and notice it starts drooping, lightly grease a sheet of aluminum foil with your spatula/vegetable oil.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "Here, the generated (boldface) sequence your spatula seems to fulfil a plausible but different semantic role than the human-produced (underlined) sequence vegetable oil. Even though it remains unclear why the annotator preferred the humanproduced sequence in these cases, it is interesting to see that the model managed to generate a reference that fills a different semantic role in the given sentence (cf. summary below).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "No preference (N = 20). In 20 out of 84 instances (23.81%), the annotator marked both sequences as being equally fitting in the context. Some of the produced sequences were not references or entities (N = 3), but still plausible insertions in the sentence, for example, Open or create a new word document. The remaining cases (N = 17) contain generated references which are plausible but different from the human-produced reference, such as: Many of your answers on the subject/the regents . . . (the context here is an article on regents exams). The high relative frequency of such cases suggests that the model-generated sequences might be able to reflect alternate, plausible fillers for an implicit reference.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "Preference for model insertion (N = 7). In 7 cases (8.33%), the annotator marked the generated sequence as fitting better than the human-produced sequence. In 4 cases, the human-produced reference caused fluency or grammatical issues. 8 In two instances, the generated sequence referred to a different entity than the human-produced reference, such as:",
                "cite_spans": [
                    {
                        "start": 235,
                        "end": 236,
                        "text": "8",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "(2) Look closely and carefully at the grain pattern on the handbag. The pattern of the grain on a crocodile leather handbag will have some irregularities. If the grain pattern of the leather/the scales is very uniform , it has probably been stamped on .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "Given the annotator's preference, these examples support the finding that model-generated sequences may reflect alternate, plausible fillers of an implicit reference. We also noticed one instance where a generated sequence filled the same semantic role as the human-produced reference, but differed in terms of granularity.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "(3) Click the tab at the top left that says \"Themes\". This will take your sidebar to the theme garden, which looks like this. Make sure to choose from the type of theme you want to look at first from the sidebar/wikihow.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "In this case, the annotator might have chosen the generated sequence because it was mentioned in the previous sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "Summary. We showed that most incorrect predictions of the fine-tuned GPT model are indeed errors, as confirmed by the annotator's preference for the human-produced reference over the generated references. We further found the annotator's preference to reflect the effects of re-ranking when the correct reference can be found among the top-100 candidates: The annotator also preferred references that do not disrupt grammaticality and that 8 It might be that further revision is needed or that they refer to an external link/image. also occur in the preceding discourse. Finally, there are a fair number of cases in which the annotator had no preference or preferred the model-generated sequence, indicating that there exist plausible, alternate references. In cases like Example (1), such references can be distinguished by the semantic role they fill. However, we also find examples, such as (2) and 3, in which different references can fill the same role at varying levels of granularity. Therefore, it seems unclear whether semantic roles would be helpful in this task and what a suitable role inventory would be.",
                "cite_spans": [
                    {
                        "start": 440,
                        "end": 441,
                        "text": "8",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "C Additional Examples for Section 6.2",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "In this section, we provide additional examples for the analysis conducted in Section 6.2. In the examples, we highlight the preferred filler by the annotator and underline the human-inserted reference.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Error Analysis",
                "sec_num": null
            },
            {
                "text": "An example where the annotator preferred the human-insertion is:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.1 Human-insertion among top-2",
                "sec_num": null
            },
            {
                "text": "(1) 1. Read books or go to the Library. Kids love it when you take them there. 2. Play games with them. Little kids like games like 'Simon says', 'Hide and seek', 'Tag', etc. Older kids might play board games or video games. 3. Make up your own games. Kids have a great time doing this ! Watch a movie with friends/them. In this example, the annotator probably preferred the human-inserted reference because the top-2 completion friends does not make sense in the given context. Instead, the referring expression them seems more plausible because it can be used to refer to children.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.1 Human-insertion among top-2",
                "sec_num": null
            },
            {
                "text": "In addition, we show an example where the annotator had no preference below:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.1 Human-insertion among top-2",
                "sec_num": null
            },
            {
                "text": "(2) Has a friend of yours read something personal or embarrassing that belongs to you? Here 's some tips on how to deal with that/it. Example (2) shows a common phenomenon that we noticed for all the instances where the annotator had no preference, namely that the fillers are paraphrases in the given context. Finally, we show two examples where the annotator preferred the top-2 generated completion by the model, instead of the human-insertion.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.1 Human-insertion among top-2",
                "sec_num": null
            },
            {
                "text": "(3) The corks will retain moisture longer than traditional milch and help maintain your plant's health between waterings. 2. Use as a fire starter. When you need to start a fire, remove a cork or two and place them under the wood to be kindled before lighting the fire/a fire.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.1 Human-insertion among top-2",
                "sec_num": null
            },
            {
                "text": "(4) it's as if you're trying to brush some debris off your pants. Return to the original position. repeat the process with your left knee. 5. Practice the two -step. the two -step is a very basic dance move that can help you get into the rhythm of the music. practicing the two -step can help you form a dance routine. Repeat the process/the sequence with your left food.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.1 Human-insertion among top-2",
                "sec_num": null
            },
            {
                "text": "Both examples are instances for which we concluded that the annotations reflect personal preferences, since the human-insertion and top-2 filler are paraphrases.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.1 Human-insertion among top-2",
                "sec_num": null
            },
            {
                "text": "Annother has preference. The two examples below are instances where the annotator preferred either of the top-2 generated sequences.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.2 Human-insertion not among top-2",
                "sec_num": null
            },
            {
                "text": "(5) Use the pie within several months. While a properly frozen pecan pie will last for a while in a freezer, it won't last forever. Try to use the pie within 2 months, as after that it is at risk of developing freezer burn. * To reheat a frozen pie, let it thaw overnight in the refrigerator. Then warm it in a oven for 15 to 20 minutes. The pie will do better if it is kept at a constant temperature in the oven/the microwave.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.2 Human-insertion not among top-2",
                "sec_num": null
            },
            {
                "text": "In example (5), the human-inserted reference was the freezer, which also differs from the top-1 and top-2 completions by the fine-tuned model. This example indicates the possibility of the annotator preferring a different filler than the human-inserted reference and therefore a mismatch between the interpretation of the implicit reference of the writer",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C.2 Human-insertion not among top-2",
                "sec_num": null
            },
            {
                "text": "http://www.wikihow.org 3. The follow-up context C f , which contains the remaining tokens of the original/revised sentence to ensure that the reference fits into the sentence grammatically (in the example, within 1 or 2 days needs to fit after Use ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://docs.python.org/3/library/ difflib.html 3 https://stanfordnlp.github.io/stanza/ 4 We rely on automatic part-of-speech tags for this additional filtering procedure.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://github.com/irshadbhat/ wikiHowToImprove",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Some models, such as GPT-2, were pre-trained on data that includes wikiHow, which could make it possible for them to make correct predictions in our data based on training memory. However, we also experimented with other models in a preliminary study (e.g., XLNet(Yang et al., 2020), Trans-formerXL(Dai et al., 2019) and BART(Lewis et al., 2019)) and did not observe any advantages over GPT.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Note that a model only needs to generate the reference part of an insertion. Additional words, as described in Section 4, are provided to all models as part of the context.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "The research presented in this paper was funded by the DFG Emmy Noether program (RO 4848/2-1).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgements",
                "sec_num": null
            },
            {
                "text": "and a reader. The second example shows differences between the top-2 in terms of granularity: (6) Wash your face/your mouth with warm water in the morning.In this case, the human-inserted reference was your lips. The annotator therefore preferred the filler that was the closest to the human-inserted reference.Annotator has no preference. Finally, we show an example from the set where the annotator had no preference. This subset consisted of paraphrases, such as: (..) It's recommended that you use the manual setting in order to manipulate the flash to produce the highest quality photos. Change the power of a flash/the flash depending on the ambient light and the subject you are shooting .The high occurrence of paraphrases in the generated fillers shows that GPT can generate several plausible fillers for a given implicit reference and is an interesting point for future research.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "annex",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "2020. wikiHowToImprove: A resource and analyses on edits in instructional texts",
                "authors": [
                    {
                        "first": "Talita",
                        "middle": [],
                        "last": "Anthonio",
                        "suffix": ""
                    },
                    {
                        "first": "Irshad",
                        "middle": [],
                        "last": "Bhat",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proceedings of the 12th Language Resources and Evaluation Conference",
                "volume": "",
                "issue": "",
                "pages": "5721--5729",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Talita Anthonio, Irshad Bhat, and Michael Roth. 2020. wikiHowToImprove: A resource and analyses on edits in instructional texts. In Proceedings of the 12th Language Resources and Evaluation Confer- ence, pages 5721-5729, Marseille, France. Euro- pean Language Resources Association.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "What can we learn from noun substitutions in revision histories?",
                "authors": [
                    {
                        "first": "Talita",
                        "middle": [],
                        "last": "Anthonio",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 28th International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "1359--1370",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.coling-main.117"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Talita Anthonio and Michael Roth. 2020. What can we learn from noun substitutions in revision histo- ries? In Proceedings of the 28th International Con- ference on Computational Linguistics, pages 1359- 1370, Barcelona, Spain (Online). International Com- mittee on Computational Linguistics.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Referring expression generation using entity profiles",
                "authors": [
                    {
                        "first": "Meng",
                        "middle": [],
                        "last": "Cao",
                        "suffix": ""
                    },
                    {
                        "first": "Jackie Chi Kit",
                        "middle": [],
                        "last": "Cheung",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "3163--3172",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D19-1312"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Meng Cao and Jackie Chi Kit Cheung. 2019. Refer- ring expression generation using entity profiles. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP), pages 3163- 3172, Hong Kong, China. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Towards more variation in text generation: Developing and evaluating variation models for choice of referential form",
                "authors": [
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Thiago Castro Ferreira",
                        "suffix": ""
                    },
                    {
                        "first": "Sander",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Wubben",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "568--577",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P16-1054"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Thiago Castro Ferreira, Emiel Krahmer, and Sander Wubben. 2016. Towards more variation in text gen- eration: Developing and evaluating variation models for choice of referential form. In Proceedings of the 54th Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 568-577, Berlin, Germany. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Implicit argument prediction with event knowledge",
                "authors": [
                    {
                        "first": "Pengxiang",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Katrin",
                        "middle": [],
                        "last": "Erk",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "",
                "issue": "",
                "pages": "831--840",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N18-1076"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Pengxiang Cheng and Katrin Erk. 2018. Implicit ar- gument prediction with event knowledge. In Pro- ceedings of the 2018 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, Vol- ume 1 (Long Papers), pages 831-840, New Orleans, Louisiana. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Implicit argument prediction as reading comprehension",
                "authors": [
                    {
                        "first": "Pengxiang",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Katrin",
                        "middle": [],
                        "last": "Erk",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the AAAI Conference on Artificial Intelligence",
                "volume": "33",
                "issue": "",
                "pages": "6284--6291",
                "other_ids": {
                    "DOI": [
                        "10.1609/aaai.v33i01.33016284"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Pengxiang Cheng and Katrin Erk. 2019. Implicit ar- gument prediction as reading comprehension. Pro- ceedings of the AAAI Conference on Artificial Intel- ligence, 33(01):6284-6291.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Memorybased acquisition of argument structures and its application to implicit role detection",
                "authors": [
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Chiarcos",
                        "suffix": ""
                    },
                    {
                        "first": "Niko",
                        "middle": [],
                        "last": "Schenk",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue",
                "volume": "",
                "issue": "",
                "pages": "178--187",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W15-4626"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Christian Chiarcos and Niko Schenk. 2015. Memory- based acquisition of argument structures and its ap- plication to implicit role detection. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 178-187, Prague, Czech Republic. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Transformer-xl: Attentive language models beyond a fixed",
                "authors": [
                    {
                        "first": "Zihang",
                        "middle": [],
                        "last": "Dai",
                        "suffix": ""
                    },
                    {
                        "first": "Zhilin",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Yiming",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Jaime",
                        "middle": [],
                        "last": "Carbonell",
                        "suffix": ""
                    },
                    {
                        "first": "Quoc",
                        "middle": [
                            "V"
                        ],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Ruslan",
                        "middle": [],
                        "last": "Salakhutdinov",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Car- bonell, Quoc V. Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Multi-sentence argument linking",
                "authors": [
                    {
                        "first": "Seth",
                        "middle": [],
                        "last": "Ebner",
                        "suffix": ""
                    },
                    {
                        "first": "Patrick",
                        "middle": [],
                        "last": "Xia",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Culkin",
                        "suffix": ""
                    },
                    {
                        "first": "Kyle",
                        "middle": [],
                        "last": "Rawlins",
                        "suffix": ""
                    },
                    {
                        "first": "Benjamin",
                        "middle": [],
                        "last": "Van Durme",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "8057--8077",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.acl-main.718"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Seth Ebner, Patrick Xia, Ryan Culkin, Kyle Rawlins, and Benjamin Van Durme. 2020. Multi-sentence ar- gument linking. In Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics, pages 8057-8077, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Semantic role labeling of implicit arguments for nominal predicates",
                "authors": [
                    {
                        "first": "Matthew",
                        "middle": [],
                        "last": "Gerber",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "Y"
                        ],
                        "last": "Chai",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Computational Linguistics",
                "volume": "38",
                "issue": "",
                "pages": "755--798",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matthew Gerber and J. Y. Chai. 2012. Semantic role la- beling of implicit arguments for nominal predicates. Computational Linguistics, 38:755-798.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Beyond Nom-Bank: A study of implicit arguments for nominal predicates",
                "authors": [
                    {
                        "first": "Matthew",
                        "middle": [],
                        "last": "Gerber",
                        "suffix": ""
                    },
                    {
                        "first": "Joyce",
                        "middle": [],
                        "last": "Chai",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "1583--1592",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matthew Gerber and Joyce Chai. 2010. Beyond Nom- Bank: A study of implicit arguments for nominal predicates. In Proceedings of the 48th Annual Meet- ing of the Association for Computational Linguis- tics, pages 1583-1592, Uppsala, Sweden. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "UDel: Generating referring expressions guided by psycholinguistc findings",
                "authors": [
                    {
                        "first": "Charles",
                        "middle": [],
                        "last": "Greenbacker",
                        "suffix": ""
                    },
                    {
                        "first": "Kathleen",
                        "middle": [],
                        "last": "Mccoy",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the 2009 Workshop on Language Generation and Summarisation",
                "volume": "",
                "issue": "",
                "pages": "101--102",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Charles Greenbacker and Kathleen McCoy. 2009. UDel: Generating referring expressions guided by psycholinguistc findings. In Proceedings of the 2009 Workshop on Language Generation and Sum- marisation (UCNLG+Sum 2009), pages 101-102, Suntec, Singapore. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Discriminative approach to predicateargument structure analysis with zero-anaphora resolution",
                "authors": [
                    {
                        "first": "Kenji",
                        "middle": [],
                        "last": "Imamura",
                        "suffix": ""
                    },
                    {
                        "first": "Kuniko",
                        "middle": [],
                        "last": "Saito",
                        "suffix": ""
                    },
                    {
                        "first": "Tomoko",
                        "middle": [],
                        "last": "Izumi",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the ACL-IJCNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kenji Imamura, Kuniko Saito, and Tomoko Izumi. 2009. Discriminative approach to predicate- argument structure analysis with zero-anaphora res- olution. In Proceedings of the ACL-IJCNLP 2009",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Suntec, Singapore",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "85--88",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Conference Short Papers, pages 85-88, Suntec, Sin- gapore. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Japanese zero pronoun resolution based on ranking rules and machine learning",
                "authors": [
                    {
                        "first": "Hideki",
                        "middle": [],
                        "last": "Isozaki",
                        "suffix": ""
                    },
                    {
                        "first": "Tsutomu",
                        "middle": [],
                        "last": "Hirao",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "184--191",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hideki Isozaki and Tsutomu Hirao. 2003. Japanese zero pronoun resolution based on ranking rules and machine learning. In Proceedings of the 2003 Con- ference on Empirical Methods in Natural Language Processing, pages 184-191.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Recipe instruction semantics corpus (RISeC): Resolving semantic structure and zero anaphora in recipes",
                "authors": [
                    {
                        "first": "Yiwei",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    },
                    {
                        "first": "Klim",
                        "middle": [],
                        "last": "Zaporojets",
                        "suffix": ""
                    },
                    {
                        "first": "Johannes",
                        "middle": [],
                        "last": "Deleu",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Demeester",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Develder",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "821--826",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yiwei Jiang, Klim Zaporojets, Johannes Deleu, Thomas Demeester, and Chris Develder. 2020. Recipe instruction semantics corpus (RISeC): Re- solving semantic structure and zero anaphora in recipes. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Compu- tational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 821-826, Suzhou, China. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "A statistical interpretation of term specificity and its application in retrieval",
                "authors": [
                    {
                        "first": "Karen Sp\u00e4rck",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    }
                ],
                "year": 1972,
                "venue": "Journal of Documentation",
                "volume": "28",
                "issue": "",
                "pages": "11--21",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Karen Sp\u00e4rck Jones. 1972. A statistical interpretation of term specificity and its application in retrieval. Journal of Documentation, 28:11-21.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Referential choice: Predictability and its limits",
                "authors": [
                    {
                        "first": "Andrej",
                        "middle": [
                            "A"
                        ],
                        "last": "Kibrik",
                        "suffix": ""
                    },
                    {
                        "first": "Mariya",
                        "middle": [
                            "V"
                        ],
                        "last": "Khudyakova",
                        "suffix": ""
                    },
                    {
                        "first": "Grigory",
                        "middle": [
                            "B"
                        ],
                        "last": "Dobrov",
                        "suffix": ""
                    },
                    {
                        "first": "Anastasia",
                        "middle": [],
                        "last": "Linnik",
                        "suffix": ""
                    },
                    {
                        "first": "Dmitrij",
                        "middle": [
                            "A"
                        ],
                        "last": "Zalmanov",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Frontiers in Psychology",
                "volume": "7",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.3389/fpsyg.2016.01429"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Andrej A. Kibrik, Mariya V. Khudyakova, Grigory B. Dobrov, Anastasia Linnik, and Dmitrij A. Zalmanov. 2016. Referential choice: Predictability and its lim- its. Frontiers in Psychology, 7(1429).",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "ImpAr: A deterministic algorithm for implicit semantic role labelling",
                "authors": [
                    {
                        "first": "Egoitz",
                        "middle": [],
                        "last": "Laparra",
                        "suffix": ""
                    },
                    {
                        "first": "German",
                        "middle": [],
                        "last": "Rigau",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "1180--1189",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Egoitz Laparra and German Rigau. 2013a. ImpAr: A deterministic algorithm for implicit semantic role la- belling. In Proceedings of the 51st Annual Meet- ing of the Association for Computational Linguis- tics (Volume 1: Long Papers), pages 1180-1189, Sofia, Bulgaria. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Sources of evidence for implicit argument resolution",
                "authors": [
                    {
                        "first": "Egoitz",
                        "middle": [],
                        "last": "Laparra",
                        "suffix": ""
                    },
                    {
                        "first": "German",
                        "middle": [],
                        "last": "Rigau",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 10th International Conference on Computational Semantics (IWCS 2013) -Long Papers",
                "volume": "",
                "issue": "",
                "pages": "155--166",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Egoitz Laparra and German Rigau. 2013b. Sources of evidence for implicit argument resolution. In Pro- ceedings of the 10th International Conference on Computational Semantics (IWCS 2013) -Long Pa- pers, pages 155-166, Potsdam, Germany. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension",
                "authors": [
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Lewis",
                        "suffix": ""
                    },
                    {
                        "first": "Yinhan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Naman",
                        "middle": [],
                        "last": "Goyal ; Abdelrahman Mohamed",
                        "suffix": ""
                    },
                    {
                        "first": "Omer",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    },
                    {
                        "first": "Ves",
                        "middle": [],
                        "last": "Stoyanov",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mike Lewis, Yinhan Liu, Naman Goyal, Mar- jan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "The NomBank project: An interim report",
                "authors": [
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Meyers",
                        "suffix": ""
                    },
                    {
                        "first": "Ruth",
                        "middle": [],
                        "last": "Reeves",
                        "suffix": ""
                    },
                    {
                        "first": "Catherine",
                        "middle": [],
                        "last": "Macleod",
                        "suffix": ""
                    },
                    {
                        "first": "Rachel",
                        "middle": [],
                        "last": "Szekely",
                        "suffix": ""
                    },
                    {
                        "first": "Veronika",
                        "middle": [],
                        "last": "Zielinska",
                        "suffix": ""
                    },
                    {
                        "first": "Brian",
                        "middle": [],
                        "last": "Young",
                        "suffix": ""
                    },
                    {
                        "first": "Ralph",
                        "middle": [],
                        "last": "Grishman",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of the Workshop Frontiers in Corpus Annotation at HLT-NAACL 2004",
                "volume": "",
                "issue": "",
                "pages": "24--31",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Adam Meyers, Ruth Reeves, Catherine Macleod, Rachel Szekely, Veronika Zielinska, Brian Young, and Ralph Grishman. 2004. The NomBank project: An interim report. In Proceedings of the Workshop Frontiers in Corpus Annotation at HLT-NAACL 2004, pages 24-31, Boston, Massachusetts, USA. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Predicate-specific annotations for implicit role binding: Corpus annotation, data analysis and evaluation experiments",
                "authors": [
                    {
                        "first": "Tatjana",
                        "middle": [],
                        "last": "Moor",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    },
                    {
                        "first": "Anette",
                        "middle": [],
                        "last": "Frank",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 10th International Conference on Computational Semantics (IWCS 2013) -Short Papers",
                "volume": "",
                "issue": "",
                "pages": "369--375",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tatjana Moor, Michael Roth, and Anette Frank. 2013. Predicate-specific annotations for implicit role bind- ing: Corpus annotation, data analysis and evalua- tion experiments. In Proceedings of the 10th Inter- national Conference on Computational Semantics (IWCS 2013) -Short Papers, pages 369-375, Pots- dam, Germany. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Automatic extraction of rules for anaphora resolution of japanese zero pronouns from aligned sentence pairs",
                "authors": [
                    {
                        "first": "Hiromi",
                        "middle": [],
                        "last": "Nakaiwa",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Proceedings of a Workshop on Operational Factors in Practical, Robust Anaphora Resolution for Unrestricted Texts, ANARESOLUTION '97",
                "volume": "",
                "issue": "",
                "pages": "22--29",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hiromi Nakaiwa. 1997. Automatic extraction of rules for anaphora resolution of japanese zero pronouns from aligned sentence pairs. In Proceedings of a Workshop on Operational Factors in Practical, Ro- bust Anaphora Resolution for Unrestricted Texts, ANARESOLUTION '97, page 22-29, USA. Asso- ciation for Computational Linguistics.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "References to named entities: a corpus study",
                "authors": [
                    {
                        "first": "Ani",
                        "middle": [],
                        "last": "Nenkova",
                        "suffix": ""
                    },
                    {
                        "first": "Kathleen",
                        "middle": [],
                        "last": "Mckeown",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Companion Volume of the Proceedings of HLT-NAACL 2003 -Short Papers",
                "volume": "",
                "issue": "",
                "pages": "70--72",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ani Nenkova and Kathleen McKeown. 2003. Refer- ences to named entities: a corpus study. In Compan- ion Volume of the Proceedings of HLT-NAACL 2003 -Short Papers, pages 70-72.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Improving language understanding by generative pre-training",
                "authors": [
                    {
                        "first": "Alec",
                        "middle": [],
                        "last": "Radford",
                        "suffix": ""
                    },
                    {
                        "first": "Karthik",
                        "middle": [],
                        "last": "Narasimhan",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alec Radford, Karthik Narasimhan, Tim salimans, and Ilya Sutskever. 2018. Improving language under- standing by generative pre-training. OpenAI.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Building Natural Language Generation Systems",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Dale",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Studies in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.1017/CBO9780511519857"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter and Robert Dale. 2000. Building Natural Language Generation Systems. Studies in Natural Language Processing. Cambridge University Press.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Automatically identifying implicit arguments to improve argument linking and coherence modeling",
                "authors": [
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    },
                    {
                        "first": "Anette",
                        "middle": [],
                        "last": "Frank",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the Main Conference and the Shared Task: Semantic Textual Similarity",
                "volume": "1",
                "issue": "",
                "pages": "306--316",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Michael Roth and Anette Frank. 2013. Automatically identifying implicit arguments to improve argument linking and coherence modeling. In Second Joint Conference on Lexical and Computational Seman- tics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task: Semantic Textual Similarity, pages 306-316, Atlanta, Georgia, USA. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "SemEval-2010 task 10: Linking events and their participants in discourse",
                "authors": [
                    {
                        "first": "Josef",
                        "middle": [],
                        "last": "Ruppenhofer",
                        "suffix": ""
                    },
                    {
                        "first": "Caroline",
                        "middle": [],
                        "last": "Sporleder",
                        "suffix": ""
                    },
                    {
                        "first": "Roser",
                        "middle": [],
                        "last": "Morante",
                        "suffix": ""
                    },
                    {
                        "first": "Collin",
                        "middle": [],
                        "last": "Baker",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the Workshop on Semantic Evaluations: Recent Achievements and Future Directions (SEW-2009)",
                "volume": "",
                "issue": "",
                "pages": "106--111",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Josef Ruppenhofer, Caroline Sporleder, Roser Morante, Collin Baker, and Martha Palmer. 2009. SemEval- 2010 task 10: Linking events and their participants in discourse. In Proceedings of the Workshop on Se- mantic Evaluations: Recent Achievements and Fu- ture Directions (SEW-2009), pages 106-111, Boul- der, Colorado. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "A linguistic perspective on reference: Choosing a feature set for generating referring expressions in context",
                "authors": [
                    {
                        "first": "Fahime",
                        "middle": [],
                        "last": "Same",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kees Van Deemter",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 28th International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "4575--4586",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.coling-main.403"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Fahime Same and Kees van Deemter. 2020. A lin- guistic perspective on reference: Choosing a fea- ture set for generating referring expressions in con- text. In Proceedings of the 28th International Con- ference on Computational Linguistics, pages 4575- 4586, Barcelona, Spain (Online). International Com- mittee on Computational Linguistics.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Unsupervised learning of prototypical fillers for implicit semantic role labeling",
                "authors": [
                    {
                        "first": "Niko",
                        "middle": [],
                        "last": "Schenk",
                        "suffix": ""
                    },
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Chiarcos",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "",
                "issue": "",
                "pages": "1473--1479",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N16-1173"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Niko Schenk and Christian Chiarcos. 2016. Un- supervised learning of prototypical fillers for im- plicit semantic role labeling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technologies, pages 1473-1479, San Diego, California. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "A probabilistic method for analyzing Japanese anaphora integrating zero pronoun detection and resolution",
                "authors": [
                    {
                        "first": "Kazuhiro",
                        "middle": [],
                        "last": "Seki",
                        "suffix": ""
                    },
                    {
                        "first": "Atsushi",
                        "middle": [],
                        "last": "Fujii",
                        "suffix": ""
                    },
                    {
                        "first": "Tetsuya",
                        "middle": [],
                        "last": "Ishikawa",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "COLING 2002: The 19th International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kazuhiro Seki, Atsushi Fujii, and Tetsuya Ishikawa. 2002. A probabilistic method for analyzing Japanese anaphora integrating zero pronoun detec- tion and resolution. In COLING 2002: The 19th International Conference on Computational Linguis- tics.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Casting implicit role linking as an anaphora resolution task",
                "authors": [
                    {
                        "first": "Carina",
                        "middle": [],
                        "last": "Silberer",
                        "suffix": ""
                    },
                    {
                        "first": "Anette",
                        "middle": [],
                        "last": "Frank",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "*SEM 2012: The First Joint Conference on Lexical and Computational Semantics",
                "volume": "1",
                "issue": "",
                "pages": "1--10",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Carina Silberer and Anette Frank. 2012. Casting im- plicit role linking as an anaphora resolution task. In *SEM 2012: The First Joint Conference on Lexical and Computational Semantics -Volume 1: Proceed- ings of the main conference and the shared task, and Volume 2: Proceedings of the Sixth Interna- tional Workshop on Semantic Evaluation (SemEval 2012), pages 1-10, Montr\u00e9al, Canada. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "A Japanese predicate argument structure analysis using decision lists",
                "authors": [
                    {
                        "first": "Hirotoshi",
                        "middle": [],
                        "last": "Taira",
                        "suffix": ""
                    },
                    {
                        "first": "Sanae",
                        "middle": [],
                        "last": "Fujita",
                        "suffix": ""
                    },
                    {
                        "first": "Masaaki",
                        "middle": [],
                        "last": "Nagata",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "523--532",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hirotoshi Taira, Sanae Fujita, and Masaaki Nagata. 2008. A Japanese predicate argument structure anal- ysis using decision lists. In Proceedings of the 2008 Conference on Empirical Methods in Natu- ral Language Processing, pages 523-532, Honolulu, Hawaii. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Xlnet: Generalized autoregressive pretraining for language understanding",
                "authors": [
                    {
                        "first": "Zhilin",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Zihang",
                        "middle": [],
                        "last": "Dai",
                        "suffix": ""
                    },
                    {
                        "first": "Yiming",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Jaime",
                        "middle": [],
                        "last": "Carbonell",
                        "suffix": ""
                    },
                    {
                        "first": "Ruslan",
                        "middle": [],
                        "last": "Salakhutdinov",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Quoc",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Car- bonell, Ruslan Salakhutdinov, and Quoc V. Le. 2020. Xlnet: Generalized autoregressive pretraining for language understanding.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Using zero anaphora resolution to improve text categorization",
                "authors": [
                    {
                        "first": "Ching-Long",
                        "middle": [],
                        "last": "Yeh",
                        "suffix": ""
                    },
                    {
                        "first": "Yi-Chun",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of the 17th Pacific Asia Conference on Language, Information and Computation",
                "volume": "",
                "issue": "",
                "pages": "423--430",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ching-Long Yeh and Yi-Chun Chen. 2003. Using zero anaphora resolution to improve text categorization. In Proceedings of the 17th Pacific Asia Conference on Language, Information and Computation, pages 423-430, Sentosa, Singapore. COLIPS PUBLICA- TIONS.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "text": "The likelihood that a reference can be found within the previous x sentences in the original context.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "TABREF0": {
                "html": null,
                "num": null,
                "content": "<table><tr><td>dataset</td><td>R@1</td><td>R@10</td><td>R@100</td></tr><tr><td>develop</td><td>37.06%</td><td>67.85%</td><td>82.91%</td></tr><tr><td/><td colspan=\"3\">(219 TPs) (401 TPs) (490 TPs)</td></tr><tr><td>test</td><td>36.36%</td><td>71.61%</td><td>83.89%</td></tr><tr><td/><td colspan=\"3\">(198 TPs) (391 TPs) (458 TPs)</td></tr></table>",
                "text": "shows the results of each selection approach, combinations and baselines.",
                "type_str": "table"
            },
            "TABREF1": {
                "html": null,
                "num": null,
                "content": "<table/>",
                "text": "Relative and absolute number of exact matches among the top sequences generated by the GPT model and the manually inserted reference found in a revised sentence.",
                "type_str": "table"
            },
            "TABREF3": {
                "html": null,
                "num": null,
                "content": "<table/>",
                "text": "",
                "type_str": "table"
            }
        }
    }
}