File size: 136,828 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
2557
2558
2559
2560
2561
2562
{
    "paper_id": "C18-1007",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T13:08:22.950192Z"
    },
    "title": "Low-resource Cross-lingual Event Type Detection in Documents via Distant Supervision with Minimal Effort",
    "authors": [
        {
            "first": "Aldrian",
            "middle": [
                "Obaja"
            ],
            "last": "Muis",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Language Technology Institute Carnegie Mellon University Pittsburgh",
                "location": {
                    "region": "PA"
                }
            },
            "email": "amuis@andrew.cmu.edu"
        },
        {
            "first": "Naoki",
            "middle": [],
            "last": "Otani",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Language Technology Institute Carnegie Mellon University Pittsburgh",
                "location": {
                    "region": "PA"
                }
            },
            "email": "notani@andrew.cmu.edu"
        },
        {
            "first": "Nidhi",
            "middle": [],
            "last": "Vyas",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Language Technology Institute Carnegie Mellon University Pittsburgh",
                "location": {
                    "region": "PA"
                }
            },
            "email": "nkvyas@andrew.cmu.edu"
        },
        {
            "first": "Ruochen",
            "middle": [],
            "last": "Xu",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Language Technology Institute Carnegie Mellon University Pittsburgh",
                "location": {
                    "region": "PA"
                }
            },
            "email": "ruochenx@andrew.cmu.edu"
        },
        {
            "first": "Yiming",
            "middle": [],
            "last": "Yang",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Language Technology Institute Carnegie Mellon University Pittsburgh",
                "location": {
                    "region": "PA"
                }
            },
            "email": "yiming@andrew.cmu.edu"
        },
        {
            "first": "Teruko",
            "middle": [],
            "last": "Mitamura",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Language Technology Institute Carnegie Mellon University Pittsburgh",
                "location": {
                    "region": "PA"
                }
            },
            "email": "teruko@andrew.cmu.edu"
        },
        {
            "first": "Eduard",
            "middle": [],
            "last": "Hovy",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Language Technology Institute Carnegie Mellon University Pittsburgh",
                "location": {
                    "region": "PA"
                }
            },
            "email": "ehovy@andrew.cmu.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "The use of machine learning for NLP generally requires resources for training. Tasks performed in a low-resource language usually rely on labeled data in another, typically resource-rich, language. However, there might not be enough labeled data even in a resource-rich language such as English. In such cases, one approach is to use a hand-crafted approach that utilizes only a small bilingual dictionary with minimal manual verification to create distantly supervised data. Another is to explore typical machine learning techniques, for example adversarial training of bilingual word representations. We find that in event-type detection task-the task to classify [parts of] documents into a fixed set of labels-they give about the same performance. We explore ways in which the two methods can be complementary and also see how to best utilize a limited budget for manual annotation to maximize performance gain.",
    "pdf_parse": {
        "paper_id": "C18-1007",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "The use of machine learning for NLP generally requires resources for training. Tasks performed in a low-resource language usually rely on labeled data in another, typically resource-rich, language. However, there might not be enough labeled data even in a resource-rich language such as English. In such cases, one approach is to use a hand-crafted approach that utilizes only a small bilingual dictionary with minimal manual verification to create distantly supervised data. Another is to explore typical machine learning techniques, for example adversarial training of bilingual word representations. We find that in event-type detection task-the task to classify [parts of] documents into a fixed set of labels-they give about the same performance. We explore ways in which the two methods can be complementary and also see how to best utilize a limited budget for manual annotation to maximize performance gain.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "For most languages of the world, few or no language processing tools or resources exist (Baumann and Pierrehumbert, 2014) . This hinders efforts to apply certain language technologies enjoyed by languages like English, in which much current research is done.",
                "cite_spans": [
                    {
                        "start": 88,
                        "end": 121,
                        "text": "(Baumann and Pierrehumbert, 2014)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "To perform natural language processing tasks in resource-poor languages, one way to overcome data scarcity is to tap on resources from another resource-rich language. Assuming that there are already good resources and tools to solve the same tasks in the more resource-rich language (henceforth, auxiliary language), the only remaining challenge is to transfer the learning process into the resource-poor language (henceforth, target language) and adapt it to the specifics of that language. One way to do this is to build a shared word representation across the two languages and train an NL engine on this shared representation, perhaps using an adversarial domain adaptation approach to handle the domain (language) shift (Chen et al., 2017) . Usually, these approaches assume the availability of large labeled data in the auxiliary language, on the order of hundred thousands to millions.",
                "cite_spans": [
                    {
                        "start": 725,
                        "end": 744,
                        "text": "(Chen et al., 2017)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "However, for some more complex or specialized tasks, there might not be enough available training data even in a resource-rich language such as English. A case in point is the event-type classification task over the publicly available datasets, such as ACE 2005 1 and TAC KBP 2 datasets, which usually contain only a few hundred to a few thousand documents. The situation frame (SF) detection task is one example of event-type classification task, where the objective is to extract from each document one or more situation frames with their corresponding arguments. A situation frame (SF) is either an issue being described in the articles, such as civil unrest or terrorism, or a need situation such as the need for water or medical aid. In our task there are 11 situation frame types, each associated with a set of arguments, namely the location, status, relief, and urgency. For example, an article titled \"Millions of people are at the risk of starvation due to the food shortage in South Sudan\", with content describing the details and the cause of the food shortage, including a mention of difficulty accessing certain regions, can be classified as describing a food need and an infrastructure need.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "As described below, we have tried two approaches: (1) a simple keyword-matching system utilizing only a small bilingual dictionary and minimal manual verification and (2) a sophisticated neural adversarial network that learns bilingual word representations for cross-lingual transfer. We find that the methods have similar performance. We therefore explore ways in which few keyword-based models can create additional, distantly supervised data to improve the performance of a neural cross-lingual event type detection system. Our contributions are: (1) an evaluation of a state-of-the-art method in a different task showing its similar result against a simple baseline, (2) ways to improve performance of such models, and (3) an analysis of the result, with insights to practitioners as to where to focus the available, yet limited, budget for manual annotation work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "This paper is organized as follows: we first describe the related work on cross-lingual NLP tasks in low-resource settings, specifically how the available resources are used. Based on previous work, we then apply our proposed training data augmentation methods and run experiments to show the effectiveness of our methods. We then analyze the results, and follow with a few suggestions on how to best utilize the available annotation effort for maximum gain.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Keyword-based Models A keyword-based heuristic model is a simple yet effective approach to extract specific information such as events (Keane et al., 2015) , because keywords often indicate a strong presence of important information contained in documents (Marujo et al., 2015) . Such methods have been used in different tasks like text categorization (\u00d6zg\u00fcr et al., 2005) and information retrieval (Marujo et al., 2013) to extract the required information. Keyword heuristics have also been used to overcome language and domain barriers using bilingual dictionaries (Szarvas, 2008; Tran et al., 2013) . However, a weak bilingual dictionary could result in low coverage with this method. Hence, to overcome the limiting bilingual dictionary people employ bootstrapping methods to improve the coverage (Knopp, 2011; Ebrahimi et al., 2016) .",
                "cite_spans": [
                    {
                        "start": 135,
                        "end": 155,
                        "text": "(Keane et al., 2015)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 256,
                        "end": 277,
                        "text": "(Marujo et al., 2015)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 352,
                        "end": 372,
                        "text": "(\u00d6zg\u00fcr et al., 2005)",
                        "ref_id": null
                    },
                    {
                        "start": 399,
                        "end": 420,
                        "text": "(Marujo et al., 2013)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 567,
                        "end": 582,
                        "text": "(Szarvas, 2008;",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 583,
                        "end": 601,
                        "text": "Tran et al., 2013)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 801,
                        "end": 814,
                        "text": "(Knopp, 2011;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 815,
                        "end": 837,
                        "text": "Ebrahimi et al., 2016)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Cross-Lingual Text Classification Cross-lingual event type detection is closely related to crosslingual text classification (CLTC), which aims to classify text in a target language using training data from an auxiliary language (Bel et al., 2003) .",
                "cite_spans": [
                    {
                        "start": 228,
                        "end": 246,
                        "text": "(Bel et al., 2003)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "To bridge the language gap, early approaches of CLTC relied on a comprehensive bilingual dictionary to translate documents between languages (Bel et al., 2003; Shi et al., 2010; Mihalcea et al., 2007) . However, in resource-poor languages, bilingual dictionaries may be small and sparse. Therefore, the performance of direct word translation will be unsatisfactory. Some researchers utilized the bilingual dictionary to translate the models instead (Xu et al., 2016; Littell et al., 2017) .",
                "cite_spans": [
                    {
                        "start": 141,
                        "end": 159,
                        "text": "(Bel et al., 2003;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 160,
                        "end": 177,
                        "text": "Shi et al., 2010;",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 178,
                        "end": 200,
                        "text": "Mihalcea et al., 2007)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 449,
                        "end": 466,
                        "text": "(Xu et al., 2016;",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 467,
                        "end": 488,
                        "text": "Littell et al., 2017)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Another line of work focuses on the use of automatic machine translation as an oracle. The various learning algorithms treated the translations as a second view of document and facilitate cross-lingual learning with co-training (Wan, 2009) , majority learning (Amini et al., 2009) , matrix completion (Xiao and Guo, 2013) and multi-view co-regularization (Guo and Xiao, 2012a) .",
                "cite_spans": [
                    {
                        "start": 228,
                        "end": 239,
                        "text": "(Wan, 2009)",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 260,
                        "end": 280,
                        "text": "(Amini et al., 2009)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 301,
                        "end": 321,
                        "text": "(Xiao and Guo, 2013)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 355,
                        "end": 376,
                        "text": "(Guo and Xiao, 2012a)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Instead of word-level or sentence-level translation, various other approaches seek some cross-lingual mapping between document representation (Littman et al., 1998; Vinokourov et al., 2002; Platt et al., 2010; Jagarlamudi et al., 2011; De Smet et al., 2011; Guo and Xiao, 2012b; Zhou et al., 2016a; Zhou et al., 2016b; Chen et al., 2017) or label distribution (Xu and Yang, 2017) .",
                "cite_spans": [
                    {
                        "start": 142,
                        "end": 164,
                        "text": "(Littman et al., 1998;",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 165,
                        "end": 189,
                        "text": "Vinokourov et al., 2002;",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 190,
                        "end": 209,
                        "text": "Platt et al., 2010;",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 210,
                        "end": 235,
                        "text": "Jagarlamudi et al., 2011;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 236,
                        "end": 257,
                        "text": "De Smet et al., 2011;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 258,
                        "end": 278,
                        "text": "Guo and Xiao, 2012b;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 279,
                        "end": 298,
                        "text": "Zhou et al., 2016a;",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 299,
                        "end": 318,
                        "text": "Zhou et al., 2016b;",
                        "ref_id": "BIBREF39"
                    },
                    {
                        "start": 319,
                        "end": 337,
                        "text": "Chen et al., 2017)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 360,
                        "end": 379,
                        "text": "(Xu and Yang, 2017)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Bilingual Word Embedding The most recent method for sharing document representation between languages is bilingual word embedding (Mikolov et al., 2013a; Faruqui and Dyer, 2014; Luong et al., 2015) . The goal is to learn a shared embedding space between words in two languages. With the shared embedding, we are able to project all documents into a shared space. The model trained in one language can, therefore, be used in the other language.",
                "cite_spans": [
                    {
                        "start": 130,
                        "end": 153,
                        "text": "(Mikolov et al., 2013a;",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 154,
                        "end": 177,
                        "text": "Faruqui and Dyer, 2014;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 178,
                        "end": 197,
                        "text": "Luong et al., 2015)",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "To see how well recent state-of-the-art methods for CLTC work in our task, we implemented a convolutional neural classifier. We compare this against a simple keyword-based method. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Models",
                "sec_num": "3"
            },
            {
                "text": "The first step is to train a bilingual word embedding as a shared feature representation space between the two languages. We trained our bilingual word embedding for English and the incident language using the method proposed in XlingualEmb (Duong et al., 2016) . This method is a cross-lingual extension from word2vec model (Mikolov et al., 2013b) to bilingual text using two large monolingual corpora and a bilingual dictionary.",
                "cite_spans": [
                    {
                        "start": 241,
                        "end": 261,
                        "text": "(Duong et al., 2016)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 325,
                        "end": 348,
                        "text": "(Mikolov et al., 2013b)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Adversarial Convolutional Network",
                "sec_num": "3.1"
            },
            {
                "text": "Based on the shared representation, we then used a convolutional neural network (CNN) (Kim, 2014) to perform the classification. There are two main advantages of choosing a deep neural classifier over a shallow one. First, CNN outperforms shallow models like SVM or Logistic Regression in various text classification benchmark datasets (Kim, 2014; Lai et al., 2015; Johnson and Zhang, 2015; Xu and Yang, 2017) . Second, CNN takes dense word vector representations as input, allowing one to incorporate the state-of-the-art bilingual word embedding methods into the pipeline.",
                "cite_spans": [
                    {
                        "start": 86,
                        "end": 97,
                        "text": "(Kim, 2014)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 336,
                        "end": 347,
                        "text": "(Kim, 2014;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 348,
                        "end": 365,
                        "text": "Lai et al., 2015;",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 366,
                        "end": 390,
                        "text": "Johnson and Zhang, 2015;",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 391,
                        "end": 409,
                        "text": "Xu and Yang, 2017)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Adversarial Convolutional Network",
                "sec_num": "3.1"
            },
            {
                "text": "The CNN model takes a sequence of word embeddings as input and applies 1-D convolutional operation on the input to extract semantic features. The features are then passed through a fully-connected layer before reaching the final soft-max layer. The model is trained in English using the ReliefWeb dataset (Littell et al., 2017, Sec 2. 3), which is annotated at sentence level with disaster relief needs and emergency situations. Thanks to the bilingual word embedding, which maps the words from the two languages to the same distributional semantic space, the model trained in English can be applied to documents in the target language.",
                "cite_spans": [
                    {
                        "start": 305,
                        "end": 334,
                        "text": "(Littell et al., 2017, Sec 2.",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Adversarial Convolutional Network",
                "sec_num": "3.1"
            },
            {
                "text": "Ideally, if the bilingual word embedding captures the ground-truth mapping between two languages, a classifier learned from English training documents should generalize well on the target language. However, in practice, we can observe obvious domain gaps between documents in different languages when we represent them with bilingual word embeddings. In order to close the domain (language) gap between training and testing, we adapt our learned model in English to the target language with similar adversarial training techniques used in (Xu and Yang, 2017; Chen et al., 2017) . In order to alleviate the domain mismatch, we are essentially looking for a feature extractor that only captures the semantics of the event types but not the difference in language usage between English and the target language. In other words, we want the features captured by CNN to be informative for the event type classification and to be language-invariant at the same time. To achieve this goal, we include an auxiliary classifier that takes the features extracted by CNN and predicts the language that the input belongs to. During training, we update our parameter to simultaneously minimize the loss of the event type classification and maximize the loss of language classification through Gradient Reversal Layer (Ganin and Lempitsky, 2015) .",
                "cite_spans": [
                    {
                        "start": 539,
                        "end": 558,
                        "text": "(Xu and Yang, 2017;",
                        "ref_id": "BIBREF34"
                    },
                    {
                        "start": 559,
                        "end": 577,
                        "text": "Chen et al., 2017)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 1302,
                        "end": 1329,
                        "text": "(Ganin and Lempitsky, 2015)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Adversarial Convolutional Network",
                "sec_num": "3.1"
            },
            {
                "text": "As mentioned previously, a keyword-based model is a simple, quick, yet effective approach to perform text classification without much training data. In our case, we do this in two steps: (1) build a list of keywords for each SF type in English, then (2) translate the keywords into the target language automatically Instances Distribution of Situation Frames (%) Visualization terror violence regime food water med infra shelter evac utils search none Eng-Orig ( O ) 82,096 2.4 1.8 3.9 14.0 33.0 8.2 6.0 9.2 3.7 4.3 8.6 4.9 Eng-KW ( E ) 1,356,425 17.5 29.2 6.0 11.3 12.6 9.9 2.7 4.0 3.6 1.6 1.5 0.3 Tigrinya Tgt-Boot ( T ) 98 12.2 33.7 10.2 4.1 6.1 20.4 0.0 11.2 1.0 0.0 1.0 0.0 Tgt-Ann A1,012 6.7 14.3 17.6 2.0 0.8 6.6 2.2 3.3 2.4 0.9 2.7 40.6 Test Data 2,991* 3.2 6.2 0.7 2.5 0.9 3.2 0.4 0.3 0.8 0.3 0.8 80.7 Oromo Tgt-Boot T92 3.3 13.0 25.0 21.7 8.7 10.9 1.1 6.5 3.3 1.1 5.4 0.0 Tgt-Ann ( A ) 652 3.2 4.0 3.5 0.8 0.2 0.6 0.2 0.6 0.0 0.0 0.5 86.5 Test Data 2,810* 0.6 11.4 2.3 1.4 0.5 2.1 1.7 1.2 0.7 0.5 1.5 76.2 Table 1 : Statistics of the various sources of training data. Eng-Orig and Eng-KW refer to training data in English described in Littell et al. (2017, Sec 2.3) and from our English keyword model's output on ReliefWeb corpus, respectively, while Tgt-Boot and Tgt-Ann refer to training data in the target language obtained from bootstrapped keyword-spotting and from annotation, respectively. The \"none\" class signifies negative examples in the data. The last column shows a visualization of the SF types, excluding \"none\". Note: for Test Data, the instances refer to documents, while for the rest, instances refer to sentences.",
                "cite_spans": [
                    {
                        "start": 1145,
                        "end": 1175,
                        "text": "Littell et al. (2017, Sec 2.3)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 1016,
                        "end": 1023,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Keyword-based Model",
                "sec_num": "3.2"
            },
            {
                "text": "using a bilingual dictionary. We also asked native speakers of the target language to refine the translation, especially for domain-specific keywords which are not adequately captured by the bilingual dictionary. 3 Building English keywords is again a two-step process. First, we use the ReliefWeb dataset to generate a list of 100 candidate keywords for each class by taking the top-100 words with the highest TF.IDF scores. Similar to the keyword generation method described by Littell et al. (2017) , we manually refined the keyword list by pruning based on world knowledge. For each candidate keyword, we added 30 most similar words using the English word2vec model trained on the Google News corpus 4 . We retained only the words that have cosine-similarity greater than 70%. For each candidate keyword in this extended list, we computed a label affinity score with each SF class label (e.g., water, evacuation) using cosinesimilarity between their word2vec embeddings. Candidate keywords with similarity above a certain threshold th 1 were retained and used as keywords for the corresponding classes 5 .",
                "cite_spans": [
                    {
                        "start": 480,
                        "end": 501,
                        "text": "Littell et al. (2017)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Keyword-based Model",
                "sec_num": "3.2"
            },
            {
                "text": "4 Method: Training Data Augmentation Chen et al. (2017) assumed the auxiliary language contains a large amount of labeled data for the task, about 700k Yelp reviews. For our case, the original training data, which was built semi-automatically by Littell et al. (2017, Sec 2. 3), contains only about 80k sentences (Table 1, first row). To improve the performance of the neural model, therefore, we propose to utilize the keyword-based system to automatically augment the original training data. We also explore using additional training data obtained via manual annotation for comparison. Figure 2 is a summary of the various training data sources we compare in this paper: the original training data ( O ), keyword-spotting in the auxiliary language ( E ), keyword-spotting with bootstrapping in the target language ( T ), and annotated data in the target language ( A ). The additional training data from keyword-spotting in English ( E ) can be directly obtained by using the keyword list in English that we used for the keyword model (Section 3.2) to label a larger ReliefWeb corpus. We describe the other two ways ( T and A ) to obtain additional training data in the following sections. Figure 2 : A summary of the various training data sources that we compare in this paper ( O , E , T , A ).",
                "cite_spans": [
                    {
                        "start": 37,
                        "end": 55,
                        "text": "Chen et al. (2017)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 246,
                        "end": 274,
                        "text": "Littell et al. (2017, Sec 2.",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 588,
                        "end": 596,
                        "text": "Figure 2",
                        "ref_id": null
                    },
                    {
                        "start": 1192,
                        "end": 1200,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Keyword-based Model",
                "sec_num": "3.2"
            },
            {
                "text": "We note that using simple keyword matching can result in low coverage due to missing keywords in the bilingual dictionary or word-variations in the target language. To overcome this, we developed an iterative bootstrapping algorithm that takes into account the newly labeled documents from keywordspotting and generates additional language-specific keywords in a two-step process ( T in Figure 2 ).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 387,
                        "end": 395,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Bootstrapping Language-specific Keywords",
                "sec_num": "4.1"
            },
            {
                "text": "Clustering: In the first step, we collected labeled documents from each class, and generated clusters for them (D = {D c 1 ,..,D cm }, where D cp is the cluster of the class c p ). For each cluster D cp and nonkeyword w i in it, we then computed the label affinity score S p (w i ), defined as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bootstrapping Language-specific Keywords",
                "sec_num": "4.1"
            },
            {
                "text": "S p (w i ) = tfidf (w i ) + w j \u2208Wp cos(w i , w j ) |W p |",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bootstrapping Language-specific Keywords",
                "sec_num": "4.1"
            },
            {
                "text": "where W p was the set of keywords present in D cp . We then appended the words which exceed a certain threshold th 2 to the keywords list of class c p .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bootstrapping Language-specific Keywords",
                "sec_num": "4.1"
            },
            {
                "text": "The rationale for this step is that the keywords that were missed in the initial keyword list (due to an incomplete bilingual dictionary, the keywords being language-specific, or incident-specific) may appear more frequently in the document cluster, and the second term in S p (w i ) will capture this.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bootstrapping Language-specific Keywords",
                "sec_num": "4.1"
            },
            {
                "text": "Labeling: With the updated set of keywords for each class, we relabeled the documents to obtain a new set of labeled documents and again executed the clustering step to get more keywords. We can repeat this two-step process n times until we have the desired coverage or until this process no longer gives useful new keywords. In our experiments, we found that n = 10 generally gives good coverage. To generate the training data, we ran this procedure on the test set and took the top-100 most confident predictions. 6",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bootstrapping Language-specific Keywords",
                "sec_num": "4.1"
            },
            {
                "text": "When we have the budget and annotators to do so, we can also annotate documents in the target language with class labels of interest. Given the limited budget and the rarity of documents with SFs (14-18% in our dataset), however, one question remains: how to best pick the documents to be annotated to maximize the gain from the additional training data? Seeing that the number of documents with at least one positive class is much less common compared to the number of documents without any positive class (see Table 1 ), simply taking a randomly sampled document from the unlabeled documents will likely give a document with no class, which is less useful compared to document with at least one positive class. Thus, we opt for a simpler method of asking annotators to make a binary decision on a ) . In addition to the native speakers, we also had non-speaker linguists annotate another separate (359) sentences in Tigrinya, assisted by grapheme-tophoneme conversion, morphological glossing, and machine translation (MT) output. 7 This results in 1,012 sentences annotated in Tigrinya ( A in Figure 2 and Table 1 ). Overall, we spent less than 12 manhours with native speakers of the target language to do the keyword translation and the annotation, with the larger amount of time spent on keyword translation.",
                "cite_spans": [
                    {
                        "start": 1032,
                        "end": 1033,
                        "text": "7",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 512,
                        "end": 519,
                        "text": "Table 1",
                        "ref_id": null
                    },
                    {
                        "start": 799,
                        "end": 800,
                        "text": ")",
                        "ref_id": null
                    },
                    {
                        "start": 1095,
                        "end": 1103,
                        "text": "Figure 2",
                        "ref_id": null
                    },
                    {
                        "start": 1108,
                        "end": 1115,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Annotation in Target Language",
                "sec_num": "4.2"
            },
            {
                "text": "For the purpose of the experiments and analysis, we used the dataset from the LoReHLT 2017 8 shared task, which consists of news articles in two Ethiopian languages: Tigrinya and Oromo. 9 The statistics of the dataset is shown in Table 2 . The available resources that we used for this experiment consist of:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 230,
                        "end": 237,
                        "text": "Table 2",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": "5.1"
            },
            {
                "text": "1. Monolingual articles in Tigrinya and Oromo in various genres (news, discussion, social media). 2. Bilingual word dictionary (English-Tigrinya and English-Oromo). 3. A few hours of interaction with volunteers who are native speakers of Tigrinya or Oromo. 4. English documents about disaster recovery from ReliefWeb 10 and CrisisNet 11 annotated semiautomatically with disaster type and theme (Littell et al., 2017, Sec 2. 3). 12",
                "cite_spans": [
                    {
                        "start": 394,
                        "end": 423,
                        "text": "(Littell et al., 2017, Sec 2.",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": "5.1"
            },
            {
                "text": "We summarize more details about the experiment setup.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Setup",
                "sec_num": "5.2"
            },
            {
                "text": "Sentence-level prediction: Although the model we used can be applied to produce document-level predictions directly, working at sentence-level provided more training data for the model and made it easier to train. Doing so also enabled some insight on which sentences contain the information about the document-level predictions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Setup",
                "sec_num": "5.2"
            },
            {
                "text": "We then aggregate sentence-level predictions to a document-level prediction by assigning to each SF type the maximum confidence score of that type across all sentences in the document. Based on these scores, we calculate the mean confidence score \u00b5 cp of each SF type c p . We then took the top-k (k = 3 in our experiments) SF types as our document-level prediction and filter out the predicted SF types which have confidence scores below \u00b5 cp . In the absence 7 The MT model was also trained in a low-resource setting, with BLEU score around 12 for Tigrinya. 8 https://www.nist.gov/itl/iad/mig/lorehlt-evaluations#lorehlt17 9 For Oromo, the original dataset includes one annotator (out of 4) which annotated most of the documents with a single class. We did not consider this outlier annotator in our experiments.",
                "cite_spans": [
                    {
                        "start": 461,
                        "end": 462,
                        "text": "7",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Document-level aggregation:",
                "sec_num": null
            },
            {
                "text": "10 https://reliefweb. Table 3 : Performance of the neural model (NN) with various sources of training data, averaged over 3 runs. O is the original training data, E is the additional training data in English from keyword-spotting, A is the additional training data in target language from annotation, and T is the additional training data in target language from bootstrapping. The result on keyword model (KW) is also shown for comparison. of labeled data in the target language to be used as development set, this is one method that we can use without much tuning. In later sections we show how different document-level aggregation procedures may affect the performance.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 22,
                        "end": 29,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Document-level aggregation:",
                "sec_num": null
            },
            {
                "text": "We followed the metric defined in LoReHLT 2017 guidelines 13 , which is occurrence-weighted scores, defined as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metric:",
                "sec_num": null
            },
            {
                "text": "P occ = \u03b1 tp \u03b1 tp + \u03b1 f p , R occ = \u03b1 tp \u03b1 tp + \u03b1 f n , F occ = 2 \u2022 P occ \u2022 R occ P occ + R occ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metric:",
                "sec_num": null
            },
            {
                "text": "where \u03b1 tp is the number of true positives, weighted with the number of annotators that agree with it. \u03b1 f p and \u03b1 f n are similarly defined for false positives and false negatives. False negatives always have weight 1. For brevity, we drop the occ subscript when referring to these scores.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metric:",
                "sec_num": null
            },
            {
                "text": "In our neural CNN model, we used filter lengths of {3, 4, 5} and 300 filters for each length. We also applied dropout on the extracted feature by CNN at a rate of 0.2. The model was optimized in mini-batches of size 128 by Adam (Kingma and Ba, 2014) optimizer at the learning rate of 0.001. The optimization was terminated after 30 epochs or a convergence criteria was satisfied on the held-out training data. Table 3 shows the results in Tigrinya and Oromo with the varying training data described in Section 4. First, the keyword model (KW) outperforms the neural model with only original training data (NN O ). This suggests that in a low-resource setting, a keyword-based model can be used as a way to quickly get a working classifier, without the hassle of training a machine learning classifier or getting a large additional training data.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 410,
                        "end": 417,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Model hyperparameters:",
                "sec_num": null
            },
            {
                "text": "Next, the additional training data did help to significantly improve the performance of the baseline neural model in both languages. The large additional English data (+ E ) provided a large boost both in Tigrinya (+4.97) and Oromo (+4.27). Interestingly, with only about 900 examples in the target language, the additional annotation in the target language (+ A ) gave more improvements in F 1 -score in both Tigrinya (5.3 points more) and Oromo (3.27 points more), compared to the large additional training data in English. Recall that the annotation was done on a subset of the neural model's output (trained on O ). This suggests that when an annotation budget is available, using that to verify the output of a model is a good investment.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3"
            },
            {
                "text": "It is interesting to note that each source of additional training data improved a different aspect of the model. The additional training data in English ( O + E ) seemed to improve recall more, while the additional training data in target language from annotation seemed to improve precision more ( O + A ), and combining both ( O + E + A ) provided the best of both worlds, especially in Tigrinya.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3"
            },
            {
                "text": "When we included the training data in target language from the keyword model with bootstrapping together with all other training data ( O + E + T + A ), it further improved the result in Tigrinya, but not in Oromo, although when it was used alone ( O + T ) it still gave some improvements over the original training data ( O ). This could be due to the lower quality of the keyword system in Oromo. Recall that it was created by taking the top-100 most confident predictions of the keyword model. This set of predictions gave 75.9% precision in Tigrinya and 47.1% precision in Oromo. This lower quality of Oromo bootstrapping method can also be seen in the diverging SF Type distribution, as can be seen in Table 1 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 707,
                        "end": 714,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3"
            },
            {
                "text": "The best overall improvement was more pronounced in Oromo (+7.54 points in F 1 for O + E + A ) than in Tigrinya (+15.19 points in F 1 for O + E + T + A ). This could be related to the fact that the baseline score was much lower in Oromo than in Tigrinya to begin with.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3"
            },
            {
                "text": "For completeness, we also compare the results of the keyword model (KW) in target language without and with bootstrapping in the first two rows of Table 3 . As anticipated, the bootstrapping process increased recall significantly, almost doubling the recall in Oromo. Although the precision was slightly reduced, it still resulted in an overall improvement in F 1 -score for both languages.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 147,
                        "end": 154,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3"
            },
            {
                "text": "In summary, there are four main observations:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3"
            },
            {
                "text": "1. The Neural model with original training data (NN O ) gave lower performance compared to the keyword (KW) model, although the keyword model was a much simpler system. 2. With large additional training data in English (+ E ), we obtained large improvements both in Tigrinya (+4.97) and Oromo (+4.27). 3. With only small additional annotations in target language (+ A ) we obtained even better performance compared to using large English training data. 4. Getting additional training data in the target language through the keyword model can help if the quality of the keyword model is good enough.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "5.3"
            },
            {
                "text": "We hypothesize that the focused improvements on recall when using additional training data in English ( E ) could be attributed to the large number of training data in E (over 1 million sentences). This causes the model to be able to learn more unique contexts for SF types, thereby increasing recall. In contrast, the annotated dataset in the target language ( A ), being made from the output of the NN model on a separate dataset, will mostly help the model to do better recognition of false positives, thereby improving precision.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "6"
            },
            {
                "text": "Another reason for increased precision could be the similar distribution of SF types in the annotated dataset ( A ) to the true distribution, compared to the distribution of SF types in the additional training data in English ( E ). We can see this from the SF type distribution shown in Table 1 ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 288,
                        "end": 295,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "6"
            },
            {
                "text": "To analyze the differences between the various source of additional training data, we plot the cooccurrence of classes on the Tigrinya dataset in Figure 3 . Each row describes the percentage of a particular SF type co-occurring with other SF types in the same document (recall that each document might be labeled with multiple SF types), including none, in which the SF type is the single label for that document. The numbers in a row sums to unity.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 146,
                        "end": 154,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "by comparing the visualization at the last column.",
                "sec_num": null
            },
            {
                "text": "As can be seen, predictions of the NN system trained on the additional English data ( Figure 3b ) and target language data (Figure 3c ) have different co-occurrence patterns. The additional English data apparently allowed the NN model to find a strong correlation between the crime violence class and the terrorism class, which is consistent with our intuition. On the other hand, the NN model fine-tuned on the Tigrinya annotations apparently found the crime violence and terrorism classes tend to occur alone (last column of Figure 3c ).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 86,
                        "end": 95,
                        "text": "Figure 3b",
                        "ref_id": null
                    },
                    {
                        "start": 123,
                        "end": 133,
                        "text": "(Figure 3c",
                        "ref_id": null
                    },
                    {
                        "start": 527,
                        "end": 536,
                        "text": "Figure 3c",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "by comparing the visualization at the last column.",
                "sec_num": null
            },
            {
                "text": "There is also an interesting phenomenon that arises from the correlation between keywords and class labels. We found that the SF type terrorism is associated with the keyword \" \" which means 1  0  23  0  1  0  0  0  0  0  24  52  0  0  24  1.1  0  0  0  3.2  0  0  24  47  0  0  33  0  0  0  0  0  0  0  56  11  0  0  50  0  0  0  0  17  0  17  17  0  3.6  0  31 Table 4 : Impact of various aggregation strategy to the performance. \"youth\" or \"juvenile\", as in the example sentence (English translation, the word recognized as keyword in the original language underlined) \"According to the information, the Eritrean girls killed in the incident hide near the tyre of a car and was hot by the Sudanese soldiers.\" Examining the various examples in the dataset, we found that the violence inherent in terrorism is often depicted with youths as the victims. This could be related to the tendency of news outlets to focus on the suffering experienced by young people to make more emotional appeal.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 191,
                        "end": 362,
                        "text": "1  0  23  0  1  0  0  0  0  0  24  52  0  0  24  1.1  0  0  0  3.2  0  0  24  47  0  0  33  0  0  0  0  0  0  0  56  11  0  0  50  0  0  0  0  17  0  17  17  0  3.6  0  31",
                        "ref_id": "TABREF1"
                    },
                    {
                        "start": 363,
                        "end": 370,
                        "text": "Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "by comparing the visualization at the last column.",
                "sec_num": null
            },
            {
                "text": "In Section 5.2 we showed one heuristic to do document-level aggregation. One might wonder whether one can do better in the classification performance by using another aggregation strategy, such as filtering out classes with confidence scores under certain threshold, or using different k when taking top-k classes. In this section, we explore the impact of different aggregation strategies on the performance under different conditions. Assuming the more realistic case of having no development set to prefer one strategy over another, we can use the top-k strategy like we did in our experiments, or set a fixed threshold on the confidence score based on the average confidence score of each type across all documents in the test set. We also show the result when we set the threshold based on the reference annotation, to show how well the result can be in the case that we have a development set to find the best threshold. The result is shown in Table 4 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 950,
                        "end": 957,
                        "text": "Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Impact of Document-level Aggregation Strategy",
                "sec_num": "6.1"
            },
            {
                "text": "The significant gap of performance between the one tuned on the reference annotation and the rest suggests that if additional training data can be obtained in the target language, independently from the model's predictions, we should allocate a portion of them to be used for validation, since there are still large room for improvements just from tuning the thresholds (about 4 points in Tigrinya and over 17 points in Oromo). Note that in our experiments, since the annotation was done on the output of our neural model, we cannot use them as validation set, as it is biased towards the output of our model. So there is a trade-off between ease of annotation process and the amount of data that can be used as validation set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Impact of Document-level Aggregation Strategy",
                "sec_num": "6.1"
            },
            {
                "text": "In this paper we tackled the problem of event type detection and classification in low-resource setting. We found that a neural model with adversarial training compared about the same as a simple keywordbased model using a small bilingual dictionary. Given that the problem lies with the limited amount of training data available, we proposed and compared methods to increase the amount of training data: to get significant gain in performance one can either use a very large additional semi-automatically labeled dataset in a resource-rich language, or annotate a small amount of documents in the target resource-poor language. We also showed how investing in a development set for tuning might also be a good strategy when there is a limited budget for annotation, after allocating some of them for keyword translation and additional training data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "7"
            },
            {
                "text": "One possible direction for future work is to address the mismatch of distribution of the classes between the additional training data and the actual test data as we see in Oromo. One way to mitigate the mismatch would be to make the classifier itself less prone to overfitting. In Section 6.1 we have shown how the document-level aggregation strategy may significantly affect the final result. Thus, exploring ways to effectively select the thresholds might be worthwhile. We could also incorporate the correlation between classes as evident from Figure 3 , which has proven useful in multi-label classification (Zhang and Zhang, 2010) .",
                "cite_spans": [
                    {
                        "start": 612,
                        "end": 635,
                        "text": "(Zhang and Zhang, 2010)",
                        "ref_id": "BIBREF36"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 547,
                        "end": 555,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "7"
            },
            {
                "text": "We showed the native speakers translation pairs obtained through the bilingual dictionary, and asked them to verify the translation as acceptable, or to supply a better translation.4 https://code.google.com/archive/p/word2vec/ 5 Threshold th1 = 0.9 was determined by a grid search on a held-out English dataset.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "As explained below, we used sentences as our training data, by taking the sentences which contain the keywords found.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://goo.gl/FwRCwj",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We acknowledge NIST for coordinating the SF type evaluation and providing the test data. NIST serves to coordinate the evaluations in order to support research and to help advance the state-of-the-art. NIST evaluations are not viewed as a competition, and such results reported by NIST are not to be construed, or represented, as endorsements of any participants system, or as official findings on the part of NIST or the U.S. Government. We thank Lori Levin for the inputs for an earlier version of this paper. This project was sponsored by the Defense Advanced Research Projects Agency (DARPA) Information Innovation Office (I2O), program: Low Resource Languages for Emergent Incidents (LORELEI), issued by DARPA/I2O under Contract No. HR0011-15-C-0114.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Learning from multiple partially observed views-an application to multilingual text categorization",
                "authors": [
                    {
                        "first": "Massih",
                        "middle": [],
                        "last": "Amini",
                        "suffix": ""
                    },
                    {
                        "first": "Nicolas",
                        "middle": [],
                        "last": "Usunier",
                        "suffix": ""
                    },
                    {
                        "first": "Cyril",
                        "middle": [],
                        "last": "Goutte",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Advances in Neural Information Processing Systems (NIPS)",
                "volume": "",
                "issue": "",
                "pages": "28--36",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Massih Amini, Nicolas Usunier, and Cyril Goutte. 2009. Learning from multiple partially observed views-an application to multilingual text categorization. In Advances in Neural Information Processing Systems (NIPS), pages 28-36, Vancouver, British Columbia, Canada, December.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Using resource-rich languages to improve morphological analysis of under-resourced languages",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Baumann",
                        "suffix": ""
                    },
                    {
                        "first": "Janet",
                        "middle": [],
                        "last": "Pierrehumbert",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC)",
                "volume": "",
                "issue": "",
                "pages": "3355--3359",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Peter Baumann and Janet Pierrehumbert. 2014. Using resource-rich languages to improve morphological analysis of under-resourced languages. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC), pages 3355-3359, Reykjavik, Iceland, May.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Cross-lingual text categorization",
                "authors": [
                    {
                        "first": "Nuria",
                        "middle": [],
                        "last": "Bel",
                        "suffix": ""
                    },
                    {
                        "first": "H A",
                        "middle": [],
                        "last": "Cornelis",
                        "suffix": ""
                    },
                    {
                        "first": "Marta",
                        "middle": [],
                        "last": "Koster",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Villegas",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "The 7th European Conference on Research and Advanced Technology for Digital Libraries",
                "volume": "2769",
                "issue": "",
                "pages": "126--139",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nuria Bel, Cornelis H A Koster, and Marta Villegas. 2003. Cross-lingual text categorization. The 7th European Conference on Research and Advanced Technology for Digital Libraries, 2769:126-139.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Adversarial deep averaging networks for cross-lingual sentiment classification",
                "authors": [
                    {
                        "first": "Xilun",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Yu",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Ben",
                        "middle": [],
                        "last": "Athiwaratkun",
                        "suffix": ""
                    },
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Cardie",
                        "suffix": ""
                    },
                    {
                        "first": "Kilian",
                        "middle": [],
                        "last": "Weinberger",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xilun Chen, Yu Sun, Ben Athiwaratkun, Claire Cardie, and Kilian Weinberger. 2017. Adversarial deep averaging networks for cross-lingual sentiment classification. ArXiv e-prints.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Knowledge transfer across multilingual corpora via latent topics",
                "authors": [
                    {
                        "first": "Jie",
                        "middle": [],
                        "last": "Wim De Smet",
                        "suffix": ""
                    },
                    {
                        "first": "Marie-Francine",
                        "middle": [],
                        "last": "Tang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Moens",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of the Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD)",
                "volume": "",
                "issue": "",
                "pages": "549--560",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wim De Smet, Jie Tang, and Marie-Francine Moens. 2011. Knowledge transfer across multilingual corpora via latent topics. In Proceedings of the Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD), pages 549-560, Shenzhen, China, May. Springer.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Learning crosslingual word embeddings without bilingual corpora",
                "authors": [
                    {
                        "first": "Long",
                        "middle": [],
                        "last": "Duong",
                        "suffix": ""
                    },
                    {
                        "first": "Hiroshi",
                        "middle": [],
                        "last": "Kanayama",
                        "suffix": ""
                    },
                    {
                        "first": "Tengfei",
                        "middle": [],
                        "last": "Ma",
                        "suffix": ""
                    },
                    {
                        "first": "Steven",
                        "middle": [],
                        "last": "Bird",
                        "suffix": ""
                    },
                    {
                        "first": "Trevor",
                        "middle": [],
                        "last": "Cohn",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "1285--1295",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Long Duong, Hiroshi Kanayama, Tengfei Ma, Steven Bird, and Trevor Cohn. 2016. Learning crosslingual word embeddings without bilingual corpora. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1285-1295, Austin, Texas, USA, November. Association for Computa- tional Linguistics (ACL).",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Weakly supervised tweet stance classification by relational bootstrapping",
                "authors": [
                    {
                        "first": "Javid",
                        "middle": [],
                        "last": "Ebrahimi",
                        "suffix": ""
                    },
                    {
                        "first": "Dejing",
                        "middle": [],
                        "last": "Dou",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Lowd",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "1012--1017",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Javid Ebrahimi, Dejing Dou, and Daniel Lowd. 2016. Weakly supervised tweet stance classification by relational bootstrapping. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1012-1017, Texas, USA, November. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Improving vector space word representations using multilingual correlation",
                "authors": [
                    {
                        "first": "Manaal",
                        "middle": [],
                        "last": "Faruqui",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Dyer",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics (EACL)",
                "volume": "",
                "issue": "",
                "pages": "462--471",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Manaal Faruqui and Chris Dyer. 2014. Improving vector space word representations using multilingual corre- lation. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 462-471, Gothenburg, Sweden, April. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Unsupervised domain adaptation by backpropagation",
                "authors": [
                    {
                        "first": "Yaroslav",
                        "middle": [],
                        "last": "Ganin",
                        "suffix": ""
                    },
                    {
                        "first": "Victor",
                        "middle": [],
                        "last": "Lempitsky",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 32nd International Conference on International Conference on Machine Learning (ICML)",
                "volume": "",
                "issue": "",
                "pages": "1180--1189",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yaroslav Ganin and Victor Lempitsky. 2015. Unsupervised domain adaptation by backpropagation. In Proceed- ings of the 32nd International Conference on International Conference on Machine Learning (ICML), pages 1180-1189, Lille, France, July.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Cross language text classification via subspace co-regularized multi-view learning",
                "authors": [
                    {
                        "first": "Yuhong",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Min",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1206.6481"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yuhong Guo and Min Xiao. 2012a. Cross language text classification via subspace co-regularized multi-view learning. arXiv preprint arXiv:1206.6481.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Transductive representation learning for cross-lingual text classification",
                "authors": [
                    {
                        "first": "Yuhong",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Min",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "2012 IEEE 12th International Conference on Data Mining (ICDM)",
                "volume": "",
                "issue": "",
                "pages": "888--893",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yuhong Guo and Min Xiao. 2012b. Transductive representation learning for cross-lingual text classification. In 2012 IEEE 12th International Conference on Data Mining (ICDM), pages 888-893. Institute of Electrical and Electronics Engineers (IEEE).",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Improving bilingual projections via sparse covariance matrices",
                "authors": [
                    {
                        "first": "Jagadeesh",
                        "middle": [],
                        "last": "Jagarlamudi",
                        "suffix": ""
                    },
                    {
                        "first": "Raghavendra",
                        "middle": [],
                        "last": "Udupa",
                        "suffix": ""
                    },
                    {
                        "first": "Hal",
                        "middle": [],
                        "last": "Daum\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Iii",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "Abhijit",
                        "middle": [],
                        "last": "Bhole",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "930--940",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jagadeesh Jagarlamudi, Raghavendra Udupa, Hal Daum\u00e9 III, and Abhijit Bhole. 2011. Improving bilingual projections via sparse covariance matrices. In Proceedings of the Conference on Empirical Methods in Natu- ral Language Processing (EMNLP), pages 930-940, Edinburgh, Scotland, UK. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Semi-supervised convolutional neural networks for text categorization via region embedding",
                "authors": [
                    {
                        "first": "Rie",
                        "middle": [],
                        "last": "Johnson",
                        "suffix": ""
                    },
                    {
                        "first": "Tong",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Advances in Neural Information Processing Systems (NIPS)",
                "volume": "",
                "issue": "",
                "pages": "919--927",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rie Johnson and Tong Zhang. 2015. Semi-supervised convolutional neural networks for text categorization via region embedding. In Advances in Neural Information Processing Systems (NIPS), pages 919-927, Montr\u00e9al, Canada, December.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Using topic modeling and similarity thresholds to detect events",
                "authors": [
                    {
                        "first": "Nathan",
                        "middle": [],
                        "last": "Keane",
                        "suffix": ""
                    },
                    {
                        "first": "Connie",
                        "middle": [],
                        "last": "Yee",
                        "suffix": ""
                    },
                    {
                        "first": "Liang",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the The 3rd Workshop on EVENTS: Definition, Detection, Coreference, and Representation",
                "volume": "",
                "issue": "",
                "pages": "34--42",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nathan Keane, Connie Yee, and Liang Zhou. 2015. Using topic modeling and similarity thresholds to detect events. In Proceedings of the The 3rd Workshop on EVENTS: Definition, Detection, Coreference, and Repre- sentation, pages 34-42, Denver, Colorado, USA, June. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Convolutional Neural Networks for Sentence Classification",
                "authors": [
                    {
                        "first": "Yoon",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "1746--1751",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yoon Kim. 2014. Convolutional Neural Networks for Sentence Classification. In Proceedings of the 2014 Con- ference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746-1751, Stroudsburg, PA, USA. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Adam: A Method for Stochastic Optimization",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Diederik",
                        "suffix": ""
                    },
                    {
                        "first": "Jimmy",
                        "middle": [],
                        "last": "Kingma",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1412.6980"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Extending a multilingual lexical resource by bootstrapping named entity classification using wikipedia's category system",
                "authors": [
                    {
                        "first": "Johannes",
                        "middle": [],
                        "last": "Knopp",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of the Fifth International Workshop On Cross Lingual Information Access",
                "volume": "",
                "issue": "",
                "pages": "35--43",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Johannes Knopp. 2011. Extending a multilingual lexical resource by bootstrapping named entity classification using wikipedia's category system. In Proceedings of the Fifth International Workshop On Cross Lingual Information Access, pages 35-43. Asian Federation of Natural Language Processing (AFNLP).",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Recurrent convolutional neural networks for text classification",
                "authors": [
                    {
                        "first": "Siwei",
                        "middle": [],
                        "last": "Lai",
                        "suffix": ""
                    },
                    {
                        "first": "Liheng",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Kang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Jun",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI)",
                "volume": "",
                "issue": "",
                "pages": "2267--2273",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Siwei Lai, Liheng Xu, Kang Liu, and Jun Zhao. 2015. Recurrent convolutional neural networks for text classifica- tion. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI), pages 2267-2273, Austin, Texas, USA, January. Association for the Advancement of Artificial Intelligence (AAAI).",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Yiming Yang, Teruko Mitamura, and Eduard Hovy. 2017. The ARIEL-CMU situation frame detection pipeline for LoReHLT16: a model translation approach",
                "authors": [
                    {
                        "first": "Patrick",
                        "middle": [],
                        "last": "Littell",
                        "suffix": ""
                    },
                    {
                        "first": "Tian",
                        "middle": [],
                        "last": "Tian",
                        "suffix": ""
                    },
                    {
                        "first": "Ruochen",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Zaid",
                        "middle": [],
                        "last": "Sheikh",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Mortensen",
                        "suffix": ""
                    },
                    {
                        "first": "Lori",
                        "middle": [],
                        "last": "Levin",
                        "suffix": ""
                    },
                    {
                        "first": "Francis",
                        "middle": [],
                        "last": "Tyers",
                        "suffix": ""
                    },
                    {
                        "first": "Hiroaki",
                        "middle": [],
                        "last": "Hayashi",
                        "suffix": ""
                    },
                    {
                        "first": "Graham",
                        "middle": [],
                        "last": "Horwood",
                        "suffix": ""
                    },
                    {
                        "first": "Steve",
                        "middle": [],
                        "last": "Sloto",
                        "suffix": ""
                    },
                    {
                        "first": "Emily",
                        "middle": [],
                        "last": "Tagtow",
                        "suffix": ""
                    },
                    {
                        "first": "Alan",
                        "middle": [],
                        "last": "Black",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Machine Translation",
                "volume": "",
                "issue": "",
                "pages": "1--22",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Patrick Littell, Tian Tian, Ruochen Xu, Zaid Sheikh, David Mortensen, Lori Levin, Francis Tyers, Hiroaki Hayashi, Graham Horwood, Steve Sloto, Emily Tagtow, Alan Black, Yiming Yang, Teruko Mitamura, and Eduard Hovy. 2017. The ARIEL-CMU situation frame detection pipeline for LoReHLT16: a model translation approach. Machine Translation, pages 1-22, October.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Automatic cross-language information retrieval using latent semantic indexing",
                "authors": [
                    {
                        "first": "Susan",
                        "middle": [
                            "T"
                        ],
                        "last": "Michael L Littman",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [
                            "K"
                        ],
                        "last": "Dumais",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Landauer",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Cross-Language Information Retrieval",
                "volume": "",
                "issue": "",
                "pages": "51--62",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Michael L Littman, Susan T Dumais, and Thomas K Landauer. 1998. Automatic cross-language information retrieval using latent semantic indexing. In Cross-Language Information Retrieval, pages 51-62. Springer.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Bilingual word representations with monolingual quality in mind",
                "authors": [
                    {
                        "first": "Thang",
                        "middle": [],
                        "last": "Luong",
                        "suffix": ""
                    },
                    {
                        "first": "Hieu",
                        "middle": [],
                        "last": "Pham",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher D",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 1st Workshop on Vector Space Modeling for Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "151--159",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Thang Luong, Hieu Pham, and Christopher D Manning. 2015. Bilingual word representations with monolin- gual quality in mind. In Proceedings of the 1st Workshop on Vector Space Modeling for Natural Language Processing, pages 151-159, Berlin, Germany, August. Association of Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Hourly traffic prediction of news stories",
                "authors": [
                    {
                        "first": "Lu\u00eds",
                        "middle": [],
                        "last": "Marujo",
                        "suffix": ""
                    },
                    {
                        "first": "Miguel",
                        "middle": [],
                        "last": "Bugalho",
                        "suffix": ""
                    },
                    {
                        "first": "Jo\u00e3o",
                        "middle": [],
                        "last": "Paulo Da",
                        "suffix": ""
                    },
                    {
                        "first": "Silva",
                        "middle": [],
                        "last": "Neto",
                        "suffix": ""
                    },
                    {
                        "first": "Anatole",
                        "middle": [],
                        "last": "Gershman",
                        "suffix": ""
                    },
                    {
                        "first": "Jaime",
                        "middle": [],
                        "last": "Carbonell",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1306.4608"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Lu\u00eds Marujo, Miguel Bugalho, Jo\u00e3o Paulo da Silva Neto, Anatole Gershman, and Jaime Carbonell. 2013. Hourly traffic prediction of news stories. arXiv preprint arXiv:1306.4608.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Automatic keyword extraction on twitter",
                "authors": [
                    {
                        "first": "Luis",
                        "middle": [],
                        "last": "Marujo",
                        "suffix": ""
                    },
                    {
                        "first": "Wang",
                        "middle": [],
                        "last": "Ling",
                        "suffix": ""
                    },
                    {
                        "first": "Isabel",
                        "middle": [],
                        "last": "Trancoso",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Dyer",
                        "suffix": ""
                    },
                    {
                        "first": "Alan",
                        "middle": [
                            "W"
                        ],
                        "last": "Black",
                        "suffix": ""
                    },
                    {
                        "first": "Anatole",
                        "middle": [],
                        "last": "Gershman",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Martins De Matos",
                        "suffix": ""
                    },
                    {
                        "first": "Jo\u00e3o",
                        "middle": [],
                        "last": "Neto",
                        "suffix": ""
                    },
                    {
                        "first": "Jaime",
                        "middle": [],
                        "last": "Carbonell",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (ACL-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "637--643",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Luis Marujo, Wang Ling, Isabel Trancoso, Chris Dyer, Alan W Black, Anatole Gershman, David Martins de Matos, Jo\u00e3o Neto, and Jaime Carbonell. 2015. Automatic keyword extraction on twitter. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (ACL-IJCNLP), pages 637-643, Beijing, China, July. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Learning multilingual subjective language via crosslingual projections",
                "authors": [
                    {
                        "first": "Rada",
                        "middle": [],
                        "last": "Mihalcea",
                        "suffix": ""
                    },
                    {
                        "first": "Carmen",
                        "middle": [],
                        "last": "Banea",
                        "suffix": ""
                    },
                    {
                        "first": "Janyce",
                        "middle": [],
                        "last": "Wiebe",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL)",
                "volume": "",
                "issue": "",
                "pages": "976--983",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rada Mihalcea, Carmen Banea, and Janyce Wiebe. 2007. Learning multilingual subjective language via cross- lingual projections. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL), pages 976-983, Czech Republic, June. Association of Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Efficient estimation of word representations in vector space",
                "authors": [
                    {
                        "first": "Tomas",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    },
                    {
                        "first": "Kai",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Greg",
                        "middle": [],
                        "last": "Corrado",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Dean",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1301.3781"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Distributed representations of words and phrases and their compositionality",
                "authors": [
                    {
                        "first": "Tomas",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    },
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    },
                    {
                        "first": "Kai",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Greg",
                        "middle": [
                            "S"
                        ],
                        "last": "Corrado",
                        "suffix": ""
                    },
                    {
                        "first": "Jeff",
                        "middle": [],
                        "last": "Dean",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Advances in Neural Information Processing Systems (NIPS)",
                "volume": "",
                "issue": "",
                "pages": "3111--3119",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems (NIPS), pages 3111-3119, Lake Tahoe, Nevada, United States, December.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Text categorization with class-based and corpus-based keyword selection",
                "authors": [
                    {
                        "first": "Levent\u00f6zg\u00fcr",
                        "middle": [],
                        "last": "Arzucan\u00f6zg\u00fcr",
                        "suffix": ""
                    },
                    {
                        "first": "Tunga",
                        "middle": [],
                        "last": "G\u00fcng\u00f6r",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the 20th International Symposium on Computer and Information Sciences (ISCIS)",
                "volume": "",
                "issue": "",
                "pages": "606--615",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Arzucan\u00d6zg\u00fcr, Levent\u00d6zg\u00fcr, and Tunga G\u00fcng\u00f6r. 2005. Text categorization with class-based and corpus-based keyword selection. In Proceedings of the 20th International Symposium on Computer and Information Sciences (ISCIS), pages 606-615, Istanbul, Turke, October. Springer.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Translingual document representations from discriminative projections",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "John",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Platt",
                        "suffix": ""
                    },
                    {
                        "first": "Wen-Tau",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Yih",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "251--261",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John C Platt, Kristina Toutanova, and Wen-tau Yih. 2010. Translingual document representations from discrimina- tive projections. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 251-261, Cambridge, Massachusetts, USA, October. Association for Computational Linguis- tics (ACL).",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Cross language text classification by model translation and semi-supervised learning",
                "authors": [
                    {
                        "first": "Lei",
                        "middle": [],
                        "last": "Shi",
                        "suffix": ""
                    },
                    {
                        "first": "Rada",
                        "middle": [],
                        "last": "Mihalcea",
                        "suffix": ""
                    },
                    {
                        "first": "Mingjun",
                        "middle": [],
                        "last": "Tian",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "1057--1067",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lei Shi, Rada Mihalcea, and Mingjun Tian. 2010. Cross language text classification by model translation and semi-supervised learning. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1057-1067, Cambridge, Massachusetts, USA, October. Association for Computa- tional Linguistics (ACL).",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Hedge classification in biomedical texts with a weakly supervised selection of keywords",
                "authors": [
                    {
                        "first": "Gy\u00f6rgy",
                        "middle": [],
                        "last": "Szarvas",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL-HLT)",
                "volume": "",
                "issue": "",
                "pages": "281--289",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Gy\u00f6rgy Szarvas. 2008. Hedge classification in biomedical texts with a weakly supervised selection of keywords. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL-HLT), pages 281-289, Columbus, Ohio, June. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Learning based approaches for vietnamese question classification using keywords extraction from the web",
                "authors": [
                    {
                        "first": "Dang",
                        "middle": [],
                        "last": "Tran",
                        "suffix": ""
                    },
                    {
                        "first": "Cuong",
                        "middle": [],
                        "last": "Chu",
                        "suffix": ""
                    },
                    {
                        "first": "Son",
                        "middle": [],
                        "last": "Pham",
                        "suffix": ""
                    },
                    {
                        "first": "Minh",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the Sixth International Joint Conference on Natural Language Processing (IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "740--746",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dang Tran, Cuong Chu, Son Pham, and Minh Nguyen. 2013. Learning based approaches for vietnamese question classification using keywords extraction from the web. In Proceedings of the Sixth International Joint Confer- ence on Natural Language Processing (IJCNLP), pages 740-746, Nagoya, Japan, October. Asian Federation of Natural Language Processing (AFNLP).",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Inferring a semantic representation of text via cross-language correlation analysis",
                "authors": [
                    {
                        "first": "Alexei",
                        "middle": [],
                        "last": "Vinokourov",
                        "suffix": ""
                    },
                    {
                        "first": "Nello",
                        "middle": [],
                        "last": "Cristianini",
                        "suffix": ""
                    },
                    {
                        "first": "John S Shawe-Taylor",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Advances in Neural Information Processing Systems (NIPS)",
                "volume": "",
                "issue": "",
                "pages": "1473--1480",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alexei Vinokourov, Nello Cristianini, and John S Shawe-Taylor. 2002. Inferring a semantic representation of text via cross-language correlation analysis. In Advances in Neural Information Processing Systems (NIPS), pages 1473-1480, Vancouver, British Columbia, Canada, December.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Co-training for cross-lingual sentiment classification",
                "authors": [
                    {
                        "first": "Xiaojun",
                        "middle": [],
                        "last": "Wan",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP (ACL-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "235--243",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xiaojun Wan. 2009. Co-training for cross-lingual sentiment classification. In Proceedings of the Joint Confer- ence of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP (ACL-IJCNLP), pages 235-243, Singapore, August. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "A novel two-step method for cross language representation learning",
                "authors": [
                    {
                        "first": "Min",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    },
                    {
                        "first": "Yuhong",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Advances in Neural Information Processing Systems (NIPS)",
                "volume": "",
                "issue": "",
                "pages": "1--9",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Min Xiao and Yuhong Guo. 2013. A novel two-step method for cross language representation learning. In Advances in Neural Information Processing Systems (NIPS), pages 1-9, Lake Tahoe, Nevada, United States, December.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Cross-lingual distillation for text classification",
                "authors": [
                    {
                        "first": "Ruochen",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Yiming",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL)",
                "volume": "",
                "issue": "",
                "pages": "1415--1425",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ruochen Xu and Yiming Yang. 2017. Cross-lingual distillation for text classification. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL), pages 1415-1425, Vancouver, British Columbia, Canada, July. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Cross-lingual text classification via model translation with limited dictionaries",
                "authors": [
                    {
                        "first": "Ruochen",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Yiming",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Hanxiao",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Hsi",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 25th ACM International on Conference on Information and Knowledge Management (CIKM)",
                "volume": "",
                "issue": "",
                "pages": "95--104",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ruochen Xu, Yiming Yang, Hanxiao Liu, and Andrew Hsi. 2016. Cross-lingual text classification via model translation with limited dictionaries. In Proceedings of the 25th ACM International on Conference on Informa- tion and Knowledge Management (CIKM), pages 95-104, Indianapolis, Indiana, USA, October. Association for Computing Machinery (ACM).",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Multi-label Learning by Exploiting Label Dependency",
                "authors": [
                    {
                        "first": "Min-Ling",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Kun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining",
                "volume": "",
                "issue": "",
                "pages": "999--1008",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Min-Ling Zhang and Kun Zhang. 2010. Multi-label Learning by Exploiting Label Dependency. In Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 999-1008, Washington DC, USA, July. Association for Computing Machinery (ACM).",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "Learning bilingual sentiment word embeddings for cross-language sentiment classification",
                "authors": [
                    {
                        "first": "Huiwei",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Long",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Fulin",
                        "middle": [],
                        "last": "Shi",
                        "suffix": ""
                    },
                    {
                        "first": "Degen",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (ACL-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "430--440",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Huiwei Zhou, Long Chen, Fulin Shi, and Degen Huang. 2015. Learning bilingual sentiment word embeddings for cross-language sentiment classification. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (ACL- IJCNLP), pages 430-440, Beijing, China, July. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Attention-based LSTM network for cross-lingual sentiment classification",
                "authors": [
                    {
                        "first": "Xinjie",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaojun",
                        "middle": [],
                        "last": "Wan",
                        "suffix": ""
                    },
                    {
                        "first": "Jianguo",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "247--256",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xinjie Zhou, Xiaojun Wan, and Jianguo Xiao. 2016a. Attention-based LSTM network for cross-lingual sentiment classification. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 247-256, Austin, Texas, USA., November. Association for Computational Linguistics (ACL).",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "Cross-lingual sentiment classification with bilingual document representation learning",
                "authors": [
                    {
                        "first": "Xinjie",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaojun",
                        "middle": [],
                        "last": "Wan",
                        "suffix": ""
                    },
                    {
                        "first": "Jianguo",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL)",
                "volume": "",
                "issue": "",
                "pages": "1403--1412",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xinjie Zhou, Xiaojun Wan, and Jianguo Xiao. 2016b. Cross-lingual sentiment classification with bilingual doc- ument representation learning. In Proceedings of the 54th Annual Meeting of the Association for Computa- tional Linguistics (ACL), pages 1403-1412, Berlin, Germany, August. Association for Computational Linguis- tics (ACL).",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "text": "Architecture of the neural classifier with adversarial domain (language) adaptation byGanin and Lempitsky (2015). Arrows show the flow of gradient.",
                "type_str": "figure",
                "uris": null
            },
            "FIGREF1": {
                "num": null,
                "text": "Tigrinya NN( O + E )terror regime crime food water search evac med utils shelter infra n/a",
                "type_str": "figure",
                "uris": null
            },
            "TABREF1": {
                "text": "Data statistics for Situation Frame (SF) type extraction task in Tigrinya and Oromo dataset. subset of our model's output on a separate dataset, different from the test set. We obtained 653 annotated sentences in Tigrinya this way (and 652 in Oromo",
                "content": "<table/>",
                "type_str": "table",
                "num": null,
                "html": null
            },
            "TABREF4": {
                "text": "Tuned on ref. 79.71 46.17 58.48 31.12 51.81 38.88",
                "content": "<table><tr><td/><td>terror</td><td>0</td><td>18</td><td>18</td><td>1.7</td><td>0</td><td>1.7</td><td>5.1</td><td>3.4</td><td>0</td><td>4.2</td><td>5.1</td><td>43</td><td/><td>terror</td><td>0</td><td>11</td><td>23</td><td>2.3</td><td>2.3</td><td>6.5</td><td>7.1</td><td>4.2</td><td>0</td><td>4.5</td><td>1.9</td><td>37</td></tr><tr><td/><td>regime</td><td>7.6</td><td>0</td><td>25</td><td>1.4</td><td>1.6</td><td>2.9</td><td>6.5</td><td>5.6</td><td>1.6</td><td>4.7</td><td>2.3</td><td>41</td><td/><td>regime</td><td>6.5</td><td>0</td><td colspan=\"3\">27 0.96 1.5</td><td colspan=\"2\">7.3 0.19</td><td>5</td><td colspan=\"2\">0.96 4.6</td><td>1.7</td><td>45</td></tr><tr><td/><td>crime</td><td>6.9</td><td>22</td><td>0</td><td>1.8</td><td>2.4</td><td>2.1</td><td>6.4</td><td>4.3</td><td>0.8</td><td>4</td><td>4.3</td><td>45</td><td/><td>crime</td><td>11</td><td>22</td><td>0</td><td colspan=\"2\">0.79 2.1</td><td colspan=\"3\">6.8 0.79 3.9</td><td>0</td><td>5.1</td><td>1.4</td><td>46</td></tr><tr><td/><td>food</td><td>4.9</td><td>9.8</td><td>13</td><td>0</td><td>11</td><td>11</td><td>4.9</td><td>4.9</td><td>8.5</td><td>3.7</td><td>2.4</td><td>26</td><td/><td>food</td><td>7.5</td><td>5.4</td><td>5.4</td><td>0</td><td>29</td><td>3.2</td><td>4.3</td><td>5.4</td><td>2.2</td><td>7.5</td><td>0</td><td>30</td></tr><tr><td>class</td><td>water search evac</td><td>0 2.3 4.3</td><td>10 9.2 13</td><td>17 7.5 14</td><td colspan=\"3\">10 5.2 1.4 0.72 18 0 5.6 2.9 0</td><td>2.2 29 0</td><td>9 4.6 5.8</td><td>5.6 1.1 0</td><td>2.2 18 25</td><td>2.2 1.1 2.5</td><td>36 19 14</td><td>class</td><td>water search evac</td><td colspan=\"4\">6.9 5.2 12 0.54 2.7 7.9 13 9.9 11 0.78 27 2.2</td><td>0 1 1.1</td><td>4 0 31</td><td>2 15 0</td><td colspan=\"3\">6.9 0.99 1.6 0.52 8.6 2 0.54 0.54 43</td><td>2 3.1 2.2</td><td>28 43 4.3</td></tr><tr><td/><td>med</td><td>4.1</td><td>16</td><td>14</td><td>2</td><td>4.1</td><td>4.1</td><td>8.1</td><td>0</td><td>1.5</td><td>4.6</td><td>2.5</td><td>40</td><td/><td>med</td><td>7.3</td><td>15</td><td>14</td><td>2.8</td><td>3.9</td><td colspan=\"2\">3.4 0.56</td><td>0</td><td>1.7</td><td colspan=\"2\">4.5 0.56 47</td></tr><tr><td/><td>utils</td><td>0</td><td>13</td><td>7</td><td>9.9</td><td>7</td><td>2.8</td><td>0</td><td>4.2</td><td>0</td><td>2.8</td><td>2.8</td><td>51</td><td/><td>utils</td><td>0</td><td>12</td><td>0</td><td>4.8</td><td>2.4</td><td>4.8</td><td>2.4</td><td>7.1</td><td>0</td><td>2.4</td><td>21</td><td>43</td></tr><tr><td/><td>shelter</td><td>4.9</td><td>13</td><td>12</td><td colspan=\"3\">1.5 0.97 15</td><td>34</td><td colspan=\"2\">4.4 0.97</td><td>0</td><td colspan=\"2\">0.97 13</td><td/><td>shelter</td><td>6.4</td><td>11</td><td>15</td><td colspan=\"3\">3.2 0.91 15</td><td>37</td><td colspan=\"2\">3.7 0.46</td><td>0</td><td>0.91 7.3</td></tr><tr><td/><td>infra</td><td>9.2</td><td>9.9</td><td>21</td><td>1.5</td><td>1.5</td><td>1.5</td><td>5.3</td><td>3.8</td><td>1.5</td><td>1.5</td><td>0</td><td>44</td><td/><td>infra</td><td>4.9</td><td>7.3</td><td>7.3</td><td>0</td><td>1.6</td><td>9.8</td><td colspan=\"3\">3.3 0.81 7.3</td><td>1.6</td><td>0</td><td>56</td></tr><tr><td/><td/><td/><td/><td/><td/><td colspan=\"4\">co-occurrence (%)</td><td/><td/><td/><td/><td/><td/><td colspan=\"11\">terror regime crime food water search evac med utils shelter infra n/a co-occurrence (%)</td></tr><tr><td/><td/><td/><td/><td colspan=\"6\">(c) Tigrinya NN( O + A )</td><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"7\">(d) Tigrinya NN( O + E + A )</td><td/></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"10\">Figure 3: Co-occurrence of classes</td><td/><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"3\">Tigrinya</td><td/><td/><td/><td/><td colspan=\"3\">Oromo</td><td/><td/></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">Method</td><td/><td/><td>P</td><td/><td>R</td><td/><td>F</td><td/><td/><td>P</td><td/><td>R</td><td/><td>F</td><td/></tr><tr><td/><td/><td/><td colspan=\"5\">NN ( O + E + A )</td><td/><td colspan=\"2\">top-1</td><td/><td colspan=\"14\">64.02 41.56 50.40 26.99 12.48 16.99</td></tr><tr><td/><td/><td/><td colspan=\"5\">NN ( O + E + A )</td><td/><td colspan=\"2\">top-2</td><td/><td colspan=\"14\">59.43 50.13 54.38 26.24 16.26 20.05</td></tr><tr><td/><td/><td/><td colspan=\"5\">NN ( O + E + A )</td><td/><td colspan=\"2\">top-3</td><td/><td colspan=\"14\">56.13 52.71 54.36 25.42 17.95 21.04</td></tr><tr><td/><td/><td/><td colspan=\"5\">NN ( O + E + A )</td><td/><td colspan=\"2\">mean</td><td/><td colspan=\"14\">45.57 55.86 50.18 15.41 29.83 20.21</td></tr><tr><td/><td/><td/><td colspan=\"5\">NN ( O + E + A )</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr></table>",
                "type_str": "table",
                "num": null,
                "html": null
            }
        }
    }
}