File size: 128,322 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
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
{
    "paper_id": "D13-1007",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T16:42:30.010292Z"
    },
    "title": "A Log-Linear Model for Unsupervised Text Normalization",
    "authors": [
        {
            "first": "Yi",
            "middle": [],
            "last": "Yang",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Georgia Institute of Technology",
                "location": {}
            },
            "email": "yiyang@gatech.edu"
        },
        {
            "first": "Jacob",
            "middle": [],
            "last": "Eisenstein",
            "suffix": "",
            "affiliation": {},
            "email": "jacobe@gatech.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "We present a unified unsupervised statistical model for text normalization. The relationship between standard and non-standard tokens is characterized by a log-linear model, permitting arbitrary features. The weights of these features are trained in a maximumlikelihood framework, employing a novel sequential Monte Carlo training algorithm to overcome the large label space, which would be impractical for traditional dynamic programming solutions. This model is implemented in a normalization system called UNLOL, which achieves the best known results on two normalization datasets, outperforming more complex systems. We use the output of UNLOL to automatically normalize a large corpus of social media text, revealing a set of coherent orthographic styles that underlie online language variation.",
    "pdf_parse": {
        "paper_id": "D13-1007",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "We present a unified unsupervised statistical model for text normalization. The relationship between standard and non-standard tokens is characterized by a log-linear model, permitting arbitrary features. The weights of these features are trained in a maximumlikelihood framework, employing a novel sequential Monte Carlo training algorithm to overcome the large label space, which would be impractical for traditional dynamic programming solutions. This model is implemented in a normalization system called UNLOL, which achieves the best known results on two normalization datasets, outperforming more complex systems. We use the output of UNLOL to automatically normalize a large corpus of social media text, revealing a set of coherent orthographic styles that underlie online language variation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Social media language can differ substantially from other written text. Many of the attempts to characterize and overcome this variation have focused on normalization: transforming social media language into text that better matches standard datasets (Sproat et al., 2001; Liu et al., 2011) . Because there is little available training data, and because social media language changes rapidly (Eisenstein, 2013b) , fully supervised training is generally not considered appropriate for this task. However, due to the extremely high-dimensional output space -arbitrary sequences of words across the vocabulary -it is a very challenging problem for unsupervised learning. Perhaps it is for these reasons that the most successful systems are pipeline architectures that cobble together a diverse array of techniques and resources, including statistical language models, dependency parsers, string edit distances, off-the-shelf spellcheckers, and curated slang dictionaries (Liu et al., 2011; Han and Baldwin, 2011; Han et al., 2013) .",
                "cite_spans": [
                    {
                        "start": 251,
                        "end": 272,
                        "text": "(Sproat et al., 2001;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 273,
                        "end": 290,
                        "text": "Liu et al., 2011)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 392,
                        "end": 411,
                        "text": "(Eisenstein, 2013b)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 968,
                        "end": 986,
                        "text": "(Liu et al., 2011;",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 987,
                        "end": 1009,
                        "text": "Han and Baldwin, 2011;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 1010,
                        "end": 1027,
                        "text": "Han et al., 2013)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We propose a different approach, performing normalization in a maximum-likelihood framework. There are two main sources of information to be exploited: local context, and surface similarity between the observed strings and normalization candidates. We treat the local context using standard language modeling techniques; we treat string similarity with a log-linear model that includes features for both surface similarity and word-word pairs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Because labeled examples of normalized text are not available, this model cannot be trained in the standard supervised fashion. Nor can we apply dynamic programming techniques for unsupervised training of locally-normalized conditional models (Berg-Kirkpatrick et al., 2010) , as their complexity is quadratic in the size of label space; in normalization, the label space is the vocabulary itself, with at least 10 4 elements. Instead, we present a new training approach using Monte Carlo techniques to compute an approximate gradient on the feature weights. This training method may be applicable in other unsupervised learning problems with a large label space.",
                "cite_spans": [
                    {
                        "start": 243,
                        "end": 274,
                        "text": "(Berg-Kirkpatrick et al., 2010)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "This model is implemented in a normalization system called UNLOL (unsupervised normalization in a LOg-Linear model). It is a lightweight proba-bilistic approach, relying only on a language model for the target domain; it can be adapted to new corpora text or new domains easily and quickly. Our evaluations show that UNLOL outperforms the state-of-the-art on standard normalization datasets. In addition, we demonstrate the linguistic insights that can be obtained from normalization, using UNLOL to identify classes of orthographic transformations that form coherent linguistic styles.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The text normalization task was introduced by Sproat et al. (2001) , and attained popularity in the context of SMS messages (Choudhury et al., 2007b) . It has become still more salient in the era of widespread social media, particularly Twitter. Han and Baldwin (2011) formally define a normalization task for Twitter, focusing on normalizations between single tokens, and excluding multi-word tokens like lol (laugh out loud). The normalization task has been criticized by Eisenstein (2013b) , who argues that it strips away important social meanings. In recent work, normalization has been shown to yield improvements for part-of-speech tagging (Han et al., 2013) , parsing (Zhang et al., 2013) , and machine translation (Hassan and Menezes, 2013) . As we will show in Section 7, accurate automated normalization can also improve our understanding of the nature of social media language.",
                "cite_spans": [
                    {
                        "start": 46,
                        "end": 66,
                        "text": "Sproat et al. (2001)",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 124,
                        "end": 149,
                        "text": "(Choudhury et al., 2007b)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 474,
                        "end": 492,
                        "text": "Eisenstein (2013b)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 647,
                        "end": 665,
                        "text": "(Han et al., 2013)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 676,
                        "end": 696,
                        "text": "(Zhang et al., 2013)",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 723,
                        "end": 749,
                        "text": "(Hassan and Menezes, 2013)",
                        "ref_id": "BIBREF22"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background",
                "sec_num": "2"
            },
            {
                "text": "Supervised methods Early work on normalization focused on labeled SMS datasets, using approaches such as noisy-channel modeling (Choudhury et al., 2007a) and machine translation (Aw et al., 2006) , as well as hybrid combinations of spelling correction and speech recognition (Kobus et al., 2008; Beaufort et al., 2010) . This work sought to balance language models (favoring words that fit in context) with transformation models (favoring words that are similar to the observed text). Our approach can also be seen as a noisy channel model, but unlike this prior work, no labeled data is required.",
                "cite_spans": [
                    {
                        "start": 128,
                        "end": 153,
                        "text": "(Choudhury et al., 2007a)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 178,
                        "end": 195,
                        "text": "(Aw et al., 2006)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 275,
                        "end": 295,
                        "text": "(Kobus et al., 2008;",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 296,
                        "end": 318,
                        "text": "Beaufort et al., 2010)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background",
                "sec_num": "2"
            },
            {
                "text": "Unsupervised methods Cook and Stevenson (2009) manually identify several word formation types within a noisy channel framework. They parametrize each formation type with a small num-ber of scalar values, so that all legal transformations of a given type are equally likely. The scalar parameters are then estimated using expectation maximization. This work stands apart from most of the other unsupervised models, which are pipelines. Contractor et al. (2010) use string edit distance to identify closely-related candidate orthographic forms and then decode the message using a language model. Gouws et al. (2011) refine this approach by mining an \"exception dictionary\" of stronglyassociated word pairs such as you/u. Like Contractor et al. 2010, we apply string edit distance, and like Gouws et al. (2011) , we capture strongly related word pairs. However, rather than applying these properties as filtering steps in a pipeline, we add them as features in a unified log-linear model. Recent approaches have sought to improve accuracy by bringing more external resources and complex architectures to bear. Han and Baldwin (2011) begin with a set of string similarity metrics, and then apply dependency parsing to identify contextuallysimilar words. Liu et al. (2011) extract noisy training pairs from the search snippets that result from carefully designed queries to Google, and then train a conditional random field (Lafferty et al., 2001) to estimate a character-based translation model. They later extend this work by adding a model of visual priming, an off-the-shelf spell-checker, and local context (Liu et al., 2012a) . Hassan and Menezes (2013) use a random walk framework to capture contextual similarity, which they then interpolate with an edit distance metric. Rather than seeking additional external resources or designing more complex metrics of context and similarity, we propose a unified statistical model, which learns feature weights in a maximum-likelihood framework.",
                "cite_spans": [
                    {
                        "start": 21,
                        "end": 46,
                        "text": "Cook and Stevenson (2009)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 435,
                        "end": 459,
                        "text": "Contractor et al. (2010)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 594,
                        "end": 613,
                        "text": "Gouws et al. (2011)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 788,
                        "end": 807,
                        "text": "Gouws et al. (2011)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 1107,
                        "end": 1129,
                        "text": "Han and Baldwin (2011)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 1250,
                        "end": 1267,
                        "text": "Liu et al. (2011)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 1419,
                        "end": 1442,
                        "text": "(Lafferty et al., 2001)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 1607,
                        "end": 1626,
                        "text": "(Liu et al., 2012a)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 1629,
                        "end": 1654,
                        "text": "Hassan and Menezes (2013)",
                        "ref_id": "BIBREF22"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background",
                "sec_num": "2"
            },
            {
                "text": "Our approach is motivated by the following criteria:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Unsupervised. We want to be able to train a model without labeled data. At present, labeled data for Twitter normalization is available only in small quantities. Moreover, as social media language is undergoing rapid change (Eisenstein, 2013b) , labeled datasets may become stale and increasingly ill-suited to new spellings and words.",
                "cite_spans": [
                    {
                        "start": 226,
                        "end": 245,
                        "text": "(Eisenstein, 2013b)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Low-resource.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "Other unsupervised approaches take advantage of resources such as slang dictionaries and spell checkers (Han and Baldwin, 2011; Liu et al., 2011) . Resources that characterize the current state of internet language risk becoming outdated; in this paper we investigate whether high-quality normalization is possible without any such resources.",
                "cite_spans": [
                    {
                        "start": 104,
                        "end": 127,
                        "text": "(Han and Baldwin, 2011;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 128,
                        "end": 145,
                        "text": "Liu et al., 2011)",
                        "ref_id": "BIBREF27"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Featurized. The relationship between any pair of words can be characterized in a number of different ways, ranging from simple characterlevel rules (e.g., going/goin) to larger substitutions (e.g., someone/sum1), and even to patterns that are lexically restricted (e.g., you/u, to/2). For these reasons, we seek a model that permits many overlapping features to describe candidate word pairs. These features may include simple string edit distance metrics, as well as lexical features that memorize specific pairs of standard and nonstandard words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Context-driven. Learning potentially arbitrary word-to-word transformations without supervision would be impossible without the strong additional cue of local context. For example, in the phrase",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "give me suttin to believe in, even a reader who has never before seen the word suttin may recognize it as a phonetic transcription of something. The relatively high string edit distance is overcome by the strong contextual preference for the word something over orthographically closer alternatives such as button or suiting. We can apply an arbitrary target language model, leveraging large amounts of unlabeled data and catering to the desired linguistic characteristics of the normalized content.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Holistic. While several prior approachessuch as normalization dictionaries -operate at the token level, our approach reasons over the scope of the entire message. The necessity for such holistic, joint inference and learning can be seen by changing the example above to:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "gimme suttin 2 beleive innnn.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "None of these tokens are standard (except 2, which appears in a nonstandard sense here), so without joint inference, it would not be possible to use context to help normalize suttin.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "Only by jointly reasoning over the entire message can we obtain the correct normalization.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "These desiderata point towards a featurized sequence model, which must be trained without labeled examples. While there is prior work on training sequence models without supervision (Smith and Eisner, 2005; Berg-Kirkpatrick et al., 2010) , there is an additional complication not faced by models for tasks such as part-of-speech tagging and named entity recognition: the potential label space of standard words is large, on the order of at least 10 4 . Naive application of Viterbi decoding -which is a component of training for both Contrastive Estimation (Smith and Eisner, 2005) and the locally-normalized sequence labeling model of Berg-Kirkpatrick et al. (2010) -will be stymied by Viterbi's quadratic complexity in the dimension of the label space. While various pruning heuristics may be applied, we instead look to Sequential Monte Carlo (SMC), a randomized algorithm which approximates the necessary feature expectations through weighted samples.",
                "cite_spans": [
                    {
                        "start": 182,
                        "end": 206,
                        "text": "(Smith and Eisner, 2005;",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 207,
                        "end": 237,
                        "text": "Berg-Kirkpatrick et al., 2010)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 557,
                        "end": 581,
                        "text": "(Smith and Eisner, 2005)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Approach",
                "sec_num": "3"
            },
            {
                "text": "Given a set of source-language sentences S = {s 1 , s 2 , . . .} (e.g., Tweets), our goal is to transduce them into target-language sentences T = {t 1 , t 2 , . . .} (standard English). We are given a target language model P (t), which can be estimated from some large set of unlabeled target-language sentences. We denote the vocabularies of source language and target language as \u03bd S and \u03bd T respectively.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "We define a log-linear model that scores source and target strings, with the form",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "P (s|t; \u03b8) \u221d exp \u03b8 T f (s, t) .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "(1)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "The desired conditional probability P (t|s) can be obtained by combining this model with the target language model, P (t|s) \u221d P (s|t; \u03b8)P (t). Since no labeled data is available, the parameters \u03b8 must be estimated by maximizing the log-likelihood of the source-language data. We define the log-likelihood \u03b8 (s) for a source-language sentence s as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "\u03b8 (s) = log P (s) = log t P (s|t; \u03b8)P (t)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "We would like to maximize this objective by making gradient-based updates.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u2202 \u03b8 (s) \u2202\u03b8 = 1 P (s) t P (t) \u2202 \u2202\u03b8 P (s|t; \u03b8) = t P (t|s) f (s, t) \u2212 s P (s |t)f (s , t) = E t|s [f (s, t) \u2212 E s |t [f (s , t)]]",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "We are left with a difference in expected feature counts, as is typical in log-linear models. However, unlike the supervised case, here both terms are expectations: the outer expectation is over all target sequences (given the observed source sequence), and the nested expectation is over all source sequences, given the target sequence. As the space of possible target sequences t grows exponentially in the length of the source sequence, it will not be practical to compute this expectation directly.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "Dynamic programming is the typical solution for computing feature expectations, and can be applied to sequence models when the feature function decomposes locally. There are two reasons this will not work in our case. First, while the forwardbackward algorithm would enable us to compute E t|s , it would not give us the nested expectation E t|s [E s |t ]; this is the classic challenge in training globally-normalized log-linear models without labeled data (Smith and Eisner, 2005) . Second, both forward-backward and the Viterbi algorithm have time complexity that is quadratic in the dimension of the label space, at least 10 4 or 10 5 . As we will show, Sequential Monte Carlo (SMC) algorithms have a number of advantages in this setting: they permit the efficient computation of both the outer and inner expectations, they are trivially parallelizable, and the number of samples provides an intuitive tuning tradeoff between accuracy and speed.",
                "cite_spans": [
                    {
                        "start": 458,
                        "end": 482,
                        "text": "(Smith and Eisner, 2005)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model",
                "sec_num": "4"
            },
            {
                "text": "Sequential Monte Carlo algorithms are a class of sampling-based algorithms in which latent vari-ables are sampled sequentially (Cappe et al., 2007) . They are particularly well-suited to sequence models, though they can be applied more broadly. SMC algorithms maintain a set of weighted hypotheses; the weights correspond to probabilities, and in our case, the hypotheses correspond to target language word sequences. Specifically, we approximate the conditional probability,",
                "cite_spans": [
                    {
                        "start": 127,
                        "end": 147,
                        "text": "(Cappe et al., 2007)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "P (t 1:n |s 1:n ) \u2248 K k=1 \u03c9 k n \u03b4 t k 1:n (t 1:n ),",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "where \u03c9 k n is the normalized weight of sample k at word n (\u03c9 k n is the unnormalized weight), and \u03b4 t k 1:n is a delta function centered at t k 1:n . At each step, and for each hypothesis k, a new target word is sampled from a proposal distribution, and the weight of the hypothesis is then updated. We maintain feature counts for each hypothesis, and approximate the expectation by taking a weighted average using the hypothesis weights. The proposal distribution will be described in detail later.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "We make a Markov assumption, so that the emission probability P (s|t) decomposes across the elements of the sentence P (s|t) = N n P (s n |t n ). This means that the feature functions f (s, t) must decompose on each s n , t n pair. We can then rewrite (1) as",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "P (s|t; \u03b8) = N n exp \u03b8 T f (s n , t n ) Z(t n )",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Z(t n ) = s exp \u03b8 T f (s, t n ) .",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "In addition, we assume that the target language model P (t) can be written as an N-gram language model, P (t) = n P (t n |t n\u22121 , . . . t n\u2212k+1 ). With these assumptions, we can view normalization as a finite state-space model in which the target language model defines the prior distribution of the process and Equation 3 defines the likelihood function. We are able to compute the the posterior probability P (t|s) using sequential importance sampling, a member of the SMC family. The crucial idea in sequential importance sampling is to update the hypotheses t k 1:n and their weights \u03c9 k n so that they approximate the posterior distribution at the next time step, P (t 1:n+1 |s 1:n+1 ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "Assuming the proposal distribution has the form Q(t k 1:n |s 1:n ), the importance weights are given by",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u03c9 k n \u221d P (t k 1:n |s 1:n ) Q(t k 1:n |s 1:n )",
                        "eq_num": "(5)"
                    }
                ],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "In order to update the hypotheses recursively, we rewrite P (t 1:n |s 1:n ) as:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "P (t 1:n |s 1:n ) = P (s n |t 1:n , s 1:n\u22121 )P (t 1:n |s 1:n\u22121 ) P (s n |s 1:n\u22121 ) = P (s n |t n )P (t n |t 1:n\u22121 , s 1:n\u22121 )P (t 1:n\u22121 |s 1:n\u22121 ) P (s n |s 1:n\u22121 ) \u221dP (s n |t n )P (t n |t n\u22121 )P (t 1:n\u22121 |s 1:n\u22121 ),",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "assuming a bigram language model. We further assume the proposal distribution Q can be factored as:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Q(t 1:n |s 1:n ) =Q(t n |t 1:n\u22121 , s 1:n )Q(t 1:n\u22121 |s 1:n\u22121 ) =Q(t n |t n\u22121 , s n )Q(t 1:n\u22121 |s 1:n\u22121 ).",
                        "eq_num": "(6)"
                    }
                ],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "Then the unnormalized importance weights simplify to a recurrence:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u03c9 k n = P (s n |t k n )P (t k n |t k n\u22121 )P (t k 1:n\u22121 |s 1:n\u22121 ) Q(t n |t n\u22121 , s n )Q(t k 1:n\u22121 |s 1:n\u22121 ) (7) =\u03c9 k n\u22121 P (s n |t k n )P (t k n |t k n\u22121 ) Q(t n |t n\u22121 , s n )",
                        "eq_num": "(8)"
                    }
                ],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "Therefore, we can approximate the posterior distribution P (t n |s 1:n ) \u2248 K k=1 \u03c9 k n \u03b4 t k n (t n ), and compute the outer expectation as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "E t|s [f (s, t)] = K k=1 \u03c9 k N N n=1 f (s n , t k n )",
                        "eq_num": "(9)"
                    }
                ],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "We compute the nested expectation using a nonsequential Monte Carlo approximation, assuming we can draw s ,k \u223c P (s|t k n ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "E s|t k [f (s, t k )] = 1 L N n=1 L =1 f (s ,k n , t k n )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "This gives the overall gradient computation:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "E t|s [f (s, t) \u2212 E s |t [f (s , t)]] = 1 K k=1\u03c9 k N K k=1\u03c9 k N \u00d7 N n=1 f (s n , t k n ) \u2212 1 L L =1 f (s ,k n , t k n )",
                        "eq_num": "(10)"
                    }
                ],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "where we sample t k n and update \u03c9 k n while moving from left-to-right, and sample s ,k n at each n. Note that although the sequential importance sampler moves left-to-right like a filter, we use only the final weights \u03c9 N to compute the expectation. Thus, the resulting expectation is based on the distribution P (s 1:N |t 1:N ), so that no backwards \"smoothing\" pass (Godsill et al., 2004) is needed to eliminate bias. Other applications of sequential Monte Carlo make use of resampling (Cappe et al., 2007) to avoid degeneration of the hypothesis weights, but we found this to be unnecessary due to the short length of Twitter messages.",
                "cite_spans": [
                    {
                        "start": 489,
                        "end": 509,
                        "text": "(Cappe et al., 2007)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sequential Monte Carlo approximation",
                "sec_num": "4.1"
            },
            {
                "text": "The major computational challenge for dynamic programming approaches to normalization is the large label space, equal to the size of the target vocabulary. It may appear that all we have gained by applying sequential Monte Carlo is to convert a computational problem into a statistical one: a naive sampling approach will have little hope of finding the small high-probability region of the highdimensional label space. However, sequential importance sampling allows us to address this issue through the proposal distribution, from which we sample the candidate words t n . Careful design of the proposal distribution can guide sampling towards the high-probability space. In the asymptotic limit of an infinite number of samples, any non-pathological proposal distribution will ultimately arrive at the desired estimate, but a good proposal distribution can greatly reduce the number of samples needed. Doucet et al. (2001) note that the optimal proposal -which minimizes the variance of the importance weights conditional on t 1:n\u22121 and s 1:nhas the following form:",
                "cite_spans": [
                    {
                        "start": 904,
                        "end": 924,
                        "text": "Doucet et al. (2001)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "Q(t k n |s n , t k n\u22121 ) = P (s n |t k n )P (t k n |t k n\u22121 )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "Sampling from this proposal requires computing the normalized distribution P (s n |t k n ); similarly, the update of the hypothesis weights (Equation 8) requires the calculation of Q in its normalized form. In each case, the total cost is the product of the vocabulary sizes, O(#|\u03bd T |#|\u03bd S |), which is not tractable as the vocabularies become large.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "In low-dimensional settings, a convenient solution is to set the proposal distribution equal to the transition distribution, Q(t k n |s n , t k n\u22121 ) = P (t k n |t k n\u22121 , . . . , t k n\u2212k+1 ). This choice is called the \"bootstrap filter,\" and it has the advantage that the weights \u03c9 (k) are exactly identical to the product of emission likelihoods n P (s n |t k n ). The complexity of computing the hypothesis weights is thus O(#|\u03bd S |). However, because this proposal ignores the emission likelihood, the bootstrap filter has very little hope of finding a high-probability sample in high-entropy contexts.",
                "cite_spans": [
                    {
                        "start": 283,
                        "end": 286,
                        "text": "(k)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "We strike a middle ground between efficiency and accuracy, using a proposal distribution that is closely related to the overall likelihood, yet is tractable to sample and compute:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Q(t k n |s n , t k n\u22121 ) def = P (s n |t k n )Z(t k n )P (t k n |t k n\u22121 ) t P (s n |t )Z(t )P (t |t k n\u22121 ) = exp \u03b8 T f (s n , t n ) P (t k n |t k n\u22121 ) t exp \u03b8 T f (s n , t ) P (t |t k n\u22121 )",
                        "eq_num": "(12)"
                    }
                ],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "Here, we simply replace the likelihood distribution in (11) by its unnormalized version.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "To update the unnormalized hypothesis weights \u03c9 k n , we hav\u1ebd",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u03c9 k n =\u03c9 k n\u22121 t exp \u03b8 T f (s n , t ) P (t |t k n\u22121 ) Z(t k n )",
                        "eq_num": "(13)"
                    }
                ],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "The numerator requires summing over all elements in \u03bd T and the denominator Z(t k n ) requires summing over all elements in \u03bd S , for a total cost of O(#|\u03bd T | + #|\u03bd S |).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Proposal distribution",
                "sec_num": "4.2"
            },
            {
                "text": "Given an input source sentence s, the decoding problem is to find a target sentence t that maximizes P (t|s) \u221d P (s|t)P (t) = N n P (s n |t n )P (t n |t n\u22121 ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Decoding",
                "sec_num": "4.3"
            },
            {
                "text": "Feature name Description word-word pair A set of binary features for each source/target word pair s, t string similarity A set of binary features indicating whether s is one of the top N string similar nonstandard words of t, for N \u2208 {5, 10, 25, 50, 100, 250, 500, 1000} As with learning, we cannot apply the usual dynamic programming algorithm (Viterbi), because of its quadratic cost in the size of the target language vocabulary. This must be multiplied by the cost of computing the normalized probability P (s n |t n ), resulting in a prohibitive time complexity of O(#|\u03bd",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Decoding",
                "sec_num": "4.3"
            },
            {
                "text": "S |#|\u03bd T | 2 N ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Decoding",
                "sec_num": "4.3"
            },
            {
                "text": "We consider two approximate decoding algorithms. The first is to simply apply the proposal distribution, with linear complexity in the size of the two vocabularies. However, this decoder is not identical to P (t|s), because of the extra factor of Z(t) in the numerator. Alternatively, we can apply the proposal distribution for selecting target word candidates, then apply the Viterbi algorithm only within these candidates. The total cost is O(#|\u03bd S |T 2 N ), where T is the number of target word candidates we consider; this will asymptotically approach P (t|s) as T \u2192 #|\u03bd T |. Our evaluations use the more expensive proposal+Viterbi decoding, but accuracy with the more efficient proposal-based decoding is very similar.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Decoding",
                "sec_num": "4.3"
            },
            {
                "text": "Our system uses the feature types described in Table 1. The word pair features are designed to capture lexical conventions, e.g. you/u. We only consider word pair features that fired during training. The string similarity features rely on the similarity function proposed by Contractor et al. (2010) , which has proven effective for normalization in prior work. We bin this similarity to create binary features indicating whether a string s is in the top-N most similar strings to t; this binning yields substantial speed improvements without negatively impacting accuracy.",
                "cite_spans": [
                    {
                        "start": 275,
                        "end": 299,
                        "text": "Contractor et al. (2010)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Features",
                "sec_num": "4.4"
            },
            {
                "text": "The model and inference described in the previous section are implemented in a software system for normalizing text on twitter, called UNLOL: unsupervised normalization in a LOg-Linear model. The final system can process roughly 10,000 Tweets per hour. We now describe some implementation details.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Implementation and data",
                "sec_num": "5"
            },
            {
                "text": "Most tokens in tweets do not require normalization. The question of how to identify which words are to be normalized is still an open problem. Following Han and Baldwin (2011), we build a dictionary of words which are permissible in the target domain, and make no attempt to normalize source strings that match these words. As with other comparable approaches, we are therefore unable to normalize strings like ill into I'll. Our set of \"in-vocabulary\" (IV) words is based on the GNU aspell dictionary (v0.60.6), containing 97,070 words. From this dictionary, we follow Liu et al. (2012a) and remove all the words with a count of less than 20 in the Edinburgh Twitter corpus (Petrovi\u0107 et al., 2010 ) -resulting in a total of 52,449 target words. All single characters except a and i are excluded, and rt is treated as in-vocabulary. For all in-vocabulary words, we define P (s n |t n ) = \u03b4(s n , t n ), taking the value of zero when s n = t n . This effectively prevents our model from attempting to normalize these words.",
                "cite_spans": [
                    {
                        "start": 570,
                        "end": 588,
                        "text": "Liu et al. (2012a)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 675,
                        "end": 697,
                        "text": "(Petrovi\u0107 et al., 2010",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Normalization candidates",
                "sec_num": "5.1"
            },
            {
                "text": "In addition to words that are in the target vocabulary, there are many other strings that should not be normalized, such as names and multiword shortenings (e.g. going to/gonna). 1 We follow prior work and assume that the set of normalization candidates is known in advance during test set decoding (Han et al., 2013) . However, the unlabeled training data has no such information. Thus, during training we attempt to normalize all tokens that (1) are not in our lexicon of IV words, and (2) are composed of letters, numbers and the apostrophe. This set includes contractions like \"gonna\" and \"gotta\", which would not appear in the test set, but are nonetheless normalized during training. For each OOV token, we conduct a pre-normalization step by reducing any repetitions of more than two letters in the nonstandard words to exactly two letters (e.g., cooool \u2192 cool).",
                "cite_spans": [
                    {
                        "start": 179,
                        "end": 180,
                        "text": "1",
                        "ref_id": null
                    },
                    {
                        "start": 299,
                        "end": 317,
                        "text": "(Han et al., 2013)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Normalization candidates",
                "sec_num": "5.1"
            },
            {
                "text": "The Kneser-Ney smoothed trigram target language model is estimated with the SRILM toolkit Stolcke (2002) , using Tweets from the Edinburgh Twitter corpus that contain no OOV words besides hashtags and username mentions (following (Han et al., 2013) ). We use this language model for both training and decoding. We occasionally find training contexts in which the trigram t n , t n\u22121 , t n\u22122 is unobserved in the language model data; features resulting from such trigrams are not considered when computing the weight gradients.",
                "cite_spans": [
                    {
                        "start": 90,
                        "end": 104,
                        "text": "Stolcke (2002)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 230,
                        "end": 248,
                        "text": "(Han et al., 2013)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language modeling",
                "sec_num": "5.2"
            },
            {
                "text": "The Monte Carlo approximations require two parameters: the number of samples for sequential Monte Carlo (K), and the number of samples for the non-sequential sampler of the nested expectation (L, from Equation 10). The theory of Monte Carlo approximation states that the quality of the approximation should only improve as the number of samples increases; we obtained good results with K = 10 and L = 1, and found relatively little improvement by increasing these values. The number of hypotheses considered by the decoder is set to T = 10; again, the performance should only improve with T , as we more closely approximate full Viterbi decoding.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parameters",
                "sec_num": "5.3"
            },
            {
                "text": "Datasets We use two existing labeled Twitter datasets to evaluate our approach. The first dataset -which we call LWWL11, based on the names of its authors Liu et al. (2011) -contains 3,802 individual \"nonstandard\" words (i.e., words that are not in the target vocabulary) and their normalized forms. The rest of the message in which the words is appear is not available. As this corpus does not provide linguistic context, its decoding must use a unigram target language model. The second dataset -which is called LexNorm1.1 by its authors Han and Baldwin (2011) -contains 549 complete tweets with 1,184 nonstandard tokens (558 unique word types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6"
            },
            {
                "text": "Precision Recall F-measure (Liu et al. 2011 In this corpus, we can decode with a trigram language model.",
                "cite_spans": [
                    {
                        "start": 27,
                        "end": 43,
                        "text": "(Liu et al. 2011",
                        "ref_id": "BIBREF27"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": null
            },
            {
                "text": "Close analysis of LexNorm1.1 revealed some inconsistencies in annotation (for example, y'all and 2 are sometimes normalized to you and to, but are left unnormalized in other cases). In addition, several annotations disagree with existing resources on internet language and dialectal English. For example, smh is normalized to somehow in LexNorm1.1, but internetslang.com and urbandictionary.com assert that it stands for shake my head, and this is evident from examples such as smh at this girl. Similarly, finna is normalized to finally in LexNorm1.1, but from the literature on African American English (Green, 2002) , it corresponds to fixing to (e.g., i'm finna go home). To address these issues, we have produced a new version of this dataset, which we call LexNorm1.2 (after consulting with the creators of LexNorm1.1). LexNorm1.2 differs from version 1.1 in the annotations for 172 of the 2140 OOV words. We evaluate on LexNorm1.1 to compare with prior work, but we also present results on LexNorm1.2 in the hope that it will become standard in future work on normalization in English. The dataset is available at http://www.cc.gatech.edu/ jeisenst/lexnorm.v1.2.tgz.",
                "cite_spans": [
                    {
                        "start": 605,
                        "end": 618,
                        "text": "(Green, 2002)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": null
            },
            {
                "text": "To obtain unlabeled training data, we randomly sample 50 tweets from the Edinburgh Twitter corpus Petrovi\u0107 et al. (2010) for each OOV word. Some OOV words appear less than 50 times in the corpus, so we obtained more training tweets for them through the Twitter search API.",
                "cite_spans": [
                    {
                        "start": 98,
                        "end": 120,
                        "text": "Petrovi\u0107 et al. (2010)",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": null
            },
            {
                "text": "Metrics Prior work on these datasets has assumed perfect detection of words requiring normalization, and has focused on finding the correct normalization for these words (Han and Baldwin, 2011; Han et al., 2013) . Recall has been defined as the proportion of words requiring normalization which are normalized correctly; precision is defined as the proportion of normalizations which are correct.",
                "cite_spans": [
                    {
                        "start": 170,
                        "end": 193,
                        "text": "(Han and Baldwin, 2011;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 194,
                        "end": 211,
                        "text": "Han et al., 2013)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": null
            },
            {
                "text": "We run our training algorithm for two iterations (pass the training data twice). The results are presented in Table 2 . Our system, UNLOL, achieves the highest published F-measure on both datasets. Performance on LexNorm1.2 is very similar to LexNorm1.1, despite the fact that roughly 8% of the examples were relabeled.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 110,
                        "end": 117,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": null
            },
            {
                "text": "In the normalization task that we consider, the tokens to be normalized are specified in advance. This is the same task specification as in the prior work against which we compare. At test time, our system attempts normalizes all such tokens; every error is thus both a false positive and false negative, so precision equals to recall for this task; this is also true for Han and Baldwin (2011) and Liu et al. (2011) .",
                "cite_spans": [
                    {
                        "start": 372,
                        "end": 394,
                        "text": "Han and Baldwin (2011)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 399,
                        "end": 416,
                        "text": "Liu et al. (2011)",
                        "ref_id": "BIBREF27"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": null
            },
            {
                "text": "It is possible to trade recall for precision by refusing to normalize words when the system's confidence falls below a threshold. A good setting of this threshold can improve the F-measure, but we did not report these results because we have no development set for parameter tuning.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": null
            },
            {
                "text": "Regularization One potential concern is that the number of non-zero feature weights will continually increase until the memory cost becomes overwhelming. Although we did not run up against mem- Figure 1: Effect of L1 regularization on the F-measure and the number of features with non-zero weights ory limitations in the experiments producing the results in Table 2 , this issue can be addressed through the application of L1 regularization, which produces sparse weight vectors by adding a penalty of \u03bb||\u03b8|| 1 to the log-likelihood. We perform online optimization of the L1-regularized log-likelihood by applying the truncated gradient method (Langford et al., 2009) . We use an exponential decreasing learning rate \u03b7 k = \u03b7 0 \u03b1 k/N , where k is the iteration counter and N is the size of training data. We set \u03b7 0 = 1 and \u03b1 = 0.5. Experiments were run until 300,000 training instances were observed, with a final learning rate of less than 1/32. As shown in Figure 1 , a small amount of regularization can dramatically decrease the number of active features without harming performance.",
                "cite_spans": [
                    {
                        "start": 644,
                        "end": 667,
                        "text": "(Langford et al., 2009)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 358,
                        "end": 365,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 959,
                        "end": 967,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Results",
                "sec_num": null
            },
            {
                "text": "We apply our normalization system to investigate the orthographic processes underlying language variation in social media. Using a dataset of 400,000 English language tweets, sampled from the month of August in each year from 2009 to 2012, we apply UNLOL to automatically normalize each token.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Analysis",
                "sec_num": "7"
            },
            {
                "text": "We then treat these normalizations as labeled training data, and examine the Levenshtein alignment between the source and target tokens. This alignment gives approximate character-level transduction rules to explain each OOV token. We then examine which rules are used by each author, constructing a matrix of authors and rules. 2 Factorization of the author-rule matrix reveals sets of rules that tend to be used together; we might call these rulesets \"orthographic styles.\" We apply non-negative matrix factorization (Lee and Seung, 2001) , which characterizes each author by a vector of k style loadings, and simultaneously constructs k style dictionaries, which each put weight on different orthographic rules. Because the loadings are constrained to be non-negative, the factorization can be seen as sparsely assigning varying amounts of each style to each author. We choose the factorization that minimizes the Frobenius norm of the reconstruction error, using the NIMFA software package (http://nimfa.biolab.si/).",
                "cite_spans": [
                    {
                        "start": 519,
                        "end": 540,
                        "text": "(Lee and Seung, 2001)",
                        "ref_id": "BIBREF26"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Analysis",
                "sec_num": "7"
            },
            {
                "text": "The resulting styles are shown in Table 3 , for k = 10; other values of k give similar overall results with more or less detail. The styles incorporate a number of linguistic phenomena, including: expressive lengthening (styles 7-9; see Brody and Diakopoulos, 2011); g-and t-dropping (style 5, see Eisenstein 2013a) ; th-stopping (style 6); and the dropping of several word-final vowels (styles 1-3). Some of these styles, such as t-dropping and th-stopping, have direct analogues in spoken language varieties (Tagliamonte and Temple, 2005; Green, 2002) , while others, like expressive lengthening, seem more unique to social media. The relationships between these orthographic styles and social variables such as geography and demograph-style rules examples 1. you; o-dropping y/_ ou/_u *y/ * _ o/_ u, yu, 2day, knw, gud, yur, wud, yuh, u've, toda, everthing, everwhere, ourself 2. e-dropping, u/o be/b_ e/_ o/u e*/_ * b, r, luv, cum, hav, mayb, bn, remembr, btween, gunna, gud 3. a-dropping a/_ *a/ * _ re/r_ ar/_r r, tht, wht, yrs, bck, strt, gurantee, elementry, wr, rlly, wher, rdy, preciate, neway 4. g-dropping g*/_ * ng/n_ g/_ goin, talkin, watchin, feelin, makin 5. t-dropping t*/_ * st/s_ t/_ jus, bc, shh, wha, gota, wea, mus, firts, jes, subsistutes 6. th-stopping h/_ *t/ * d th/d_ t/d dat, de, skool, fone, dese, dha, shid, dhat, dat's 7. (kd)-lengthening i_/id _/k _/d _*/k * idk, fuckk, okk, backk, workk, badd, andd, goodd, bedd, elidgible, pidgeon 8. o-lengthening o_/oo _*/o * _/o soo, noo, doo, oohh, loove, thoo, helloo 9. e-lengthening _/i e_/ee _/e _*/e * mee, ive, retweet, bestie, lovee, nicee, heey, likee, iphone, homie, ii, damnit 10. a-adding _/a __/ma _/m _*/a * ima, outta, needa, shoulda, woulda, mm, comming, tomm, boutt, ppreciate ics must be left to future research, but they offer a promising generalization of prior work that has focused almost exclusively on exclusively on lexical variation (Argamon et al., 2007; Eisenstein et al., 2010; , with a few exceptions for character-level features (Brody and Diakopoulos, 2011; Burger et al., 2011) . Note that style 10 is largely the result of mistaken normalizations. The tokens ima, outta, and needa all refer to multi-word expressions in standard English, and are thus outside the scope of the normalization task as defined by Han et al. (2013) . UNLOL has produced incorrect single-token normalizations for these terms: i/ima, out/outta, and need/needa. But while these normalizations are wrong, the resulting style nonetheless captures a coherent orthographic phenomenon.",
                "cite_spans": [
                    {
                        "start": 510,
                        "end": 540,
                        "text": "(Tagliamonte and Temple, 2005;",
                        "ref_id": "BIBREF34"
                    },
                    {
                        "start": 541,
                        "end": 553,
                        "text": "Green, 2002)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 1930,
                        "end": 1952,
                        "text": "(Argamon et al., 2007;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 1953,
                        "end": 1977,
                        "text": "Eisenstein et al., 2010;",
                        "ref_id": null
                    },
                    {
                        "start": 2031,
                        "end": 2060,
                        "text": "(Brody and Diakopoulos, 2011;",
                        "ref_id": null
                    },
                    {
                        "start": 2061,
                        "end": 2081,
                        "text": "Burger et al., 2011)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 2314,
                        "end": 2331,
                        "text": "Han et al. (2013)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 34,
                        "end": 41,
                        "text": "Table 3",
                        "ref_id": "TABREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Analysis",
                "sec_num": "7"
            },
            {
                "text": "We have presented a unified, unsupervised statistical model for normalizing social media text, attaining the best reported performance on the two standard normalization datasets. The power of our approach comes from flexible modeling of word-to-word relationships through features, while exploiting contextual regularity to train the corresponding feature weights without labeled data. The primary technical challenge was overcoming the large label space of the normalization task; we accomplish this using sequential Monte Carlo. Future work may consider whether sequential Monte Carlo can offer similar advantages in other unsupervised NLP tasks. An additional benefit of our joint statistical approach is that it may be combined with other downstream language processing tasks, such as part-of-speech tagging (Gimpel et al., 2011) and named entity resolution (Liu et al., 2012b) .",
                "cite_spans": [
                    {
                        "start": 812,
                        "end": 833,
                        "text": "(Gimpel et al., 2011)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 862,
                        "end": 881,
                        "text": "(Liu et al., 2012b)",
                        "ref_id": "BIBREF29"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "8"
            },
            {
                "text": "t P (s n |t )P (t |t k n\u22121 )(11)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Whether multiword shortenings should be normalized is arguable, but they are outside the scope of current normalization datasets(Han and Baldwin, 2011).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "We tried adding these rules as features and retraining the normalization system, but this hurt performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We thank the reviewers for thoughtful comments on our submission. This work also benefitted from discussions with Timothy Baldwin, Paul Cook, Frank Dellaert, Arnoud Doucet, Micha Elsner, and Sharon Goldwater. It was supported by NSF SOCS-1111142.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Mining the blogosphere: age, gender, and the varieties of self-expression",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Argamon",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Koppel",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Pennebaker",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Schler",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "First Monday",
                "volume": "",
                "issue": "9",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Argamon, M. Koppel, J. Pennebaker, and J. Schler. 2007. Mining the blogosphere: age, gender, and the varieties of self-expression. First Monday, 12(9).",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "A phrase-based statistical model for SMS text normalization",
                "authors": [
                    {
                        "first": "Aiti",
                        "middle": [],
                        "last": "Aw",
                        "suffix": ""
                    },
                    {
                        "first": "Min",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Juan",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    },
                    {
                        "first": "Jian",
                        "middle": [],
                        "last": "Su",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "33--40",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "AiTi Aw, Min Zhang, Juan Xiao, and Jian Su. 2006. A phrase-based statistical model for SMS text normaliza- tion. In Proceedings of ACL, pages 33-40.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "A hybrid rule/model-based finite-state framework for normalizing sms messages",
                "authors": [
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Beaufort",
                        "suffix": ""
                    },
                    {
                        "first": "Sophie",
                        "middle": [],
                        "last": "Roekhaut",
                        "suffix": ""
                    },
                    {
                        "first": "Louise-Am\u00e9lie",
                        "middle": [],
                        "last": "Cougnon",
                        "suffix": ""
                    },
                    {
                        "first": "C\u00e9drick",
                        "middle": [],
                        "last": "Fairon",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "770--779",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Richard Beaufort, Sophie Roekhaut, Louise-Am\u00e9lie Cougnon, and C\u00e9drick Fairon. 2010. A hybrid rule/model-based finite-state framework for normaliz- ing sms messages. In Proceedings of ACL, pages 770- 779.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Painless unsupervised learning with features",
                "authors": [
                    {
                        "first": "Taylor",
                        "middle": [],
                        "last": "Berg-Kirkpatrick",
                        "suffix": ""
                    },
                    {
                        "first": "Alexandre",
                        "middle": [],
                        "last": "Bouchard-C\u00f4t\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Denero",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of NAACL",
                "volume": "",
                "issue": "",
                "pages": "582--590",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Taylor Berg-Kirkpatrick, Alexandre Bouchard-C\u00f4t\u00e9, John DeNero, and Dan Klein. 2010. Painless un- supervised learning with features. In Proceedings of NAACL, pages 582-590.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "using word lengthening to detect sentiment in microblogs",
                "authors": [
                    {
                        "first": "",
                        "middle": [],
                        "last": "Cooooooooooooooollllllllllllll!!!!!!!!!!!!!!",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Cooooooooooooooollllllllllllll!!!!!!!!!!!!!!: using word lengthening to detect sentiment in microblogs. In Proceedings of EMNLP.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Discriminating gender on twitter",
                "authors": [
                    {
                        "first": "John",
                        "middle": [
                            "D"
                        ],
                        "last": "Burger",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [
                            "C"
                        ],
                        "last": "Henderson",
                        "suffix": ""
                    },
                    {
                        "first": "George",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Guido",
                        "middle": [],
                        "last": "Zarrella",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John D. Burger, John C. Henderson, George Kim, and Guido Zarrella. 2011. Discriminating gender on twit- ter. In Proceedings of EMNLP.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "An overview of existing methods and recent advances in sequential monte carlo",
                "authors": [
                    {
                        "first": "Olivier",
                        "middle": [],
                        "last": "Cappe",
                        "suffix": ""
                    },
                    {
                        "first": "Simon",
                        "middle": [
                            "J"
                        ],
                        "last": "Godsill",
                        "suffix": ""
                    },
                    {
                        "first": "Eric",
                        "middle": [],
                        "last": "Moulines",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of the IEEE",
                "volume": "95",
                "issue": "5",
                "pages": "899--924",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Olivier Cappe, Simon J. Godsill, and Eric Moulines. 2007. An overview of existing methods and recent ad- vances in sequential monte carlo. Proceedings of the IEEE, 95(5):899-924, May.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Investigation and modeling of the structure of texting language",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Choudhury",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Saraf",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Jain",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Mukherjee",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Sarkar",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Basu",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "International Journal on Document Analysis and Recognition",
                "volume": "10",
                "issue": "3",
                "pages": "157--174",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Choudhury, R. Saraf, V. Jain, A. Mukherjee, S. Sarkar, and A. Basu. 2007a. Investigation and model- ing of the structure of texting language. Interna- tional Journal on Document Analysis and Recognition, 10(3):157-174.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Investigation and modeling of the structure of texting language",
                "authors": [
                    {
                        "first": "Monojit",
                        "middle": [],
                        "last": "Choudhury",
                        "suffix": ""
                    },
                    {
                        "first": "Rahul",
                        "middle": [],
                        "last": "Saraf",
                        "suffix": ""
                    },
                    {
                        "first": "Vijit",
                        "middle": [],
                        "last": "Jain",
                        "suffix": ""
                    },
                    {
                        "first": "Animesh",
                        "middle": [],
                        "last": "Mukherjee",
                        "suffix": ""
                    },
                    {
                        "first": "Sudeshna",
                        "middle": [],
                        "last": "Sarkar",
                        "suffix": ""
                    },
                    {
                        "first": "Anupam",
                        "middle": [],
                        "last": "Basu",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "International Journal of Document Analysis and Recognition (IJDAR)",
                "volume": "10",
                "issue": "3-4",
                "pages": "157--174",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Monojit Choudhury, Rahul Saraf, Vijit Jain, Animesh Mukherjee, Sudeshna Sarkar, and Anupam Basu. 2007b. Investigation and modeling of the structure of texting language. International Journal of Document Analysis and Recognition (IJDAR), 10(3-4):157-174.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Unsupervised cleansing of noisy text",
                "authors": [
                    {
                        "first": "Danish",
                        "middle": [],
                        "last": "Contractor",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Tanveer",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Faruquie",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Venkata",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Subramaniam",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of COLING",
                "volume": "",
                "issue": "",
                "pages": "189--196",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Danish Contractor, Tanveer A. Faruquie, and L. Venkata Subramaniam. 2010. Unsupervised cleansing of noisy text. In Proceedings of COLING, pages 189-196.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "An unsupervised model for text message normalization",
                "authors": [
                    {
                        "first": "Paul",
                        "middle": [],
                        "last": "Cook",
                        "suffix": ""
                    },
                    {
                        "first": "Suzanne",
                        "middle": [],
                        "last": "Stevenson",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the Workshop on Computational Approaches to Linguistic Creativity, CALC '09",
                "volume": "",
                "issue": "",
                "pages": "71--78",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Paul Cook and Suzanne Stevenson. 2009. An unsu- pervised model for text message normalization. In Proceedings of the Workshop on Computational Ap- proaches to Linguistic Creativity, CALC '09, pages 71-78, Stroudsburg, PA, USA. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Particle filters for state estimation of jump markov linear systems",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Doucet",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [
                            "J"
                        ],
                        "last": "Gordon",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Krishnamurthy",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proceedings of EMNLP",
                "volume": "49",
                "issue": "",
                "pages": "613--624",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A. Doucet, N.J. Gordon, and V. Krishnamurthy. 2001. Particle filters for state estimation of jump markov lin- ear systems. Trans. Sig. Proc., 49(3):613-624, March. Jacob Eisenstein, Brendan O'Connor, Noah A. Smith, and Eric P. Xing. 2010. A latent variable model for ge- ographic lexical variation. In Proceedings of EMNLP.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Discovering sociolinguistic associations with structured sparsity",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Eisenstein",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "Eric",
                        "middle": [
                            "P"
                        ],
                        "last": "Xing",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jacob Eisenstein, Noah A. Smith, and Eric P. Xing. 2011. Discovering sociolinguistic associations with structured sparsity. In Proceedings of ACL.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Phonological factors in social media writing",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Eisenstein",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the NAACL Workshop on Language Analysis in Social Media",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jacob Eisenstein. 2013a. Phonological factors in social media writing. In Proceedings of the NAACL Work- shop on Language Analysis in Social Media.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "What to do about bad language on the internet",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Eisenstein",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of NAACL",
                "volume": "",
                "issue": "",
                "pages": "359--369",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jacob Eisenstein. 2013b. What to do about bad language on the internet. In Proceedings of NAACL, pages 359- 369.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Part-of-speech tagging for twitter: annotation, features, and experiments",
                "authors": [
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Gimpel",
                        "suffix": ""
                    },
                    {
                        "first": "Nathan",
                        "middle": [],
                        "last": "Schneider",
                        "suffix": ""
                    },
                    {
                        "first": "O'",
                        "middle": [],
                        "last": "Brendan",
                        "suffix": ""
                    },
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Connor",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Mills",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Eisenstein",
                        "suffix": ""
                    },
                    {
                        "first": "Dani",
                        "middle": [],
                        "last": "Heilman",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Yogatama",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Flanigan",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kevin Gimpel, Nathan Schneider, Brendan O'Connor, Dipanjan Das, Daniel Mills, Jacob Eisenstein, Michael Heilman, Dani Yogatama, Jeffrey Flanigan, and Noah A. Smith. 2011. Part-of-speech tagging for twitter: annotation, features, and experiments. In Pro- ceedings of ACL.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Monte carlo smoothing for non-linear time series",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Simon",
                        "suffix": ""
                    },
                    {
                        "first": "Arnaud",
                        "middle": [],
                        "last": "Godsill",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Doucet",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "West",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Journal of the American Statistical Association",
                "volume": "",
                "issue": "",
                "pages": "156--168",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Simon J. Godsill, Arnaud Doucet, and Mike West. 2004. Monte carlo smoothing for non-linear time series. In Journal of the American Statistical Association, pages 156-168.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Unsupervised mining of lexical variants from noisy text",
                "authors": [
                    {
                        "first": "Stephan",
                        "middle": [],
                        "last": "Gouws",
                        "suffix": ""
                    },
                    {
                        "first": "Dirk",
                        "middle": [],
                        "last": "Hovy",
                        "suffix": ""
                    },
                    {
                        "first": "Donald",
                        "middle": [],
                        "last": "Metzler",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of the First Workshop on Unsupervised Learning in NLP, EMNLP '11",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Stephan Gouws, Dirk Hovy, and Donald Metzler. 2011. Unsupervised mining of lexical variants from noisy text. In Proceedings of the First Workshop on Unsu- pervised Learning in NLP, EMNLP '11.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "African American English: A Linguistic Introduction",
                "authors": [
                    {
                        "first": "Lisa",
                        "middle": [
                            "J"
                        ],
                        "last": "Green",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lisa J. Green. 2002. African American English: A Linguistic Introduction. Cambridge University Press, September.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Lexical normalisation of short text messages: makn sens a #twitter",
                "authors": [
                    {
                        "first": "Bo",
                        "middle": [],
                        "last": "Han",
                        "suffix": ""
                    },
                    {
                        "first": "Timothy",
                        "middle": [],
                        "last": "Baldwin",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "368--378",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bo Han and Timothy Baldwin. 2011. Lexical normalisa- tion of short text messages: makn sens a #twitter. In Proceedings of ACL, pages 368-378.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Lexical normalization for social media text",
                "authors": [
                    {
                        "first": "Bo",
                        "middle": [],
                        "last": "Han",
                        "suffix": ""
                    },
                    {
                        "first": "Paul",
                        "middle": [],
                        "last": "Cook",
                        "suffix": ""
                    },
                    {
                        "first": "Timothy",
                        "middle": [],
                        "last": "Baldwin",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "ACM Transactions on Intelligent Systems and Technology",
                "volume": "4",
                "issue": "1",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bo Han, Paul Cook, and Timothy Baldwin. 2013. Lex- ical normalization for social media text. ACM Trans- actions on Intelligent Systems and Technology, 4(1):5.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Social text normalization using contextual graph random walks",
                "authors": [
                    {
                        "first": "Hany",
                        "middle": [],
                        "last": "Hassan",
                        "suffix": ""
                    },
                    {
                        "first": "Arul",
                        "middle": [],
                        "last": "Menezes",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hany Hassan and Arul Menezes. 2013. Social text nor- malization using contextual graph random walks. In Proceedings of ACL.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Normalizing sms: are two metaphors better than one?",
                "authors": [
                    {
                        "first": "Catherine",
                        "middle": [],
                        "last": "Kobus",
                        "suffix": ""
                    },
                    {
                        "first": "Fran\u00e7ois",
                        "middle": [],
                        "last": "Yvon",
                        "suffix": ""
                    },
                    {
                        "first": "G\u00e9raldine",
                        "middle": [],
                        "last": "Damnati",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of COLING",
                "volume": "",
                "issue": "",
                "pages": "441--448",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Catherine Kobus, Fran\u00e7ois Yvon, and G\u00e9raldine Damnati. 2008. Normalizing sms: are two metaphors better than one? In Proceedings of COLING, pages 441-448.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
                "authors": [
                    {
                        "first": "John",
                        "middle": [
                            "D"
                        ],
                        "last": "Lafferty",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Mccallum",
                        "suffix": ""
                    },
                    {
                        "first": "Fernando",
                        "middle": [
                            "C N"
                        ],
                        "last": "Pereira",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proceedings of ICML",
                "volume": "",
                "issue": "",
                "pages": "282--289",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilis- tic models for segmenting and labeling sequence data. In Proceedings of ICML, pages 282-289.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Sparse online learning via truncated gradient",
                "authors": [
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Langford",
                        "suffix": ""
                    },
                    {
                        "first": "Lihong",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Tong",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "The Journal of Machine Learning Research",
                "volume": "10",
                "issue": "",
                "pages": "777--801",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John Langford, Lihong Li, and Tong Zhang. 2009. Sparse online learning via truncated gradient. The Journal of Machine Learning Research, 10:777-801.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Algorithms for Non-Negative Matrix Factorization",
                "authors": [
                    {
                        "first": "D",
                        "middle": [
                            "D"
                        ],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [
                            "S"
                        ],
                        "last": "Seung",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Advances in Neural Information Processing Systems (NIPS)",
                "volume": "13",
                "issue": "",
                "pages": "556--562",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. D. Lee and H. S. Seung. 2001. Algorithms for Non- Negative Matrix Factorization. In Advances in Neural Information Processing Systems (NIPS), volume 13, pages 556-562.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Insertion, deletion, or substitution?: normalizing text messages without pre-categorization nor supervision",
                "authors": [
                    {
                        "first": "Fei",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Fuliang",
                        "middle": [],
                        "last": "Weng",
                        "suffix": ""
                    },
                    {
                        "first": "Bingqing",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "71--76",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fei Liu, Fuliang Weng, Bingqing Wang, and Yang Liu. 2011. Insertion, deletion, or substitution?: normaliz- ing text messages without pre-categorization nor su- pervision. In Proceedings of ACL, pages 71-76.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "A broadcoverage normalization system for social media language",
                "authors": [
                    {
                        "first": "Fei",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Fuliang",
                        "middle": [],
                        "last": "Weng",
                        "suffix": ""
                    },
                    {
                        "first": "Xiao",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "1035--1044",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fei Liu, Fuliang Weng, and Xiao Jiang. 2012a. A broad- coverage normalization system for social media lan- guage. In Proceedings of ACL, pages 1035-1044.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Joint inference of named entity recognition and normalization for tweets",
                "authors": [
                    {
                        "first": "Xiaohua",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Ming",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Xiangyang",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Zhongyang",
                        "middle": [],
                        "last": "Fu",
                        "suffix": ""
                    },
                    {
                        "first": "Furu",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xiaohua Liu, Ming Zhou, Xiangyang Zhou, Zhongyang Fu, and Furu Wei. 2012b. Joint inference of named entity recognition and normalization for tweets. In Proceedings of ACL.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "The edinburgh twitter corpus",
                "authors": [
                    {
                        "first": "Sa\u0161a",
                        "middle": [],
                        "last": "Petrovi\u0107",
                        "suffix": ""
                    },
                    {
                        "first": "Miles",
                        "middle": [],
                        "last": "Osborne",
                        "suffix": ""
                    },
                    {
                        "first": "Victor",
                        "middle": [],
                        "last": "Lavrenko",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the NAACL HLT Workshop on Computational Linguistics in a World of Social Media",
                "volume": "",
                "issue": "",
                "pages": "25--26",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sa\u0161a Petrovi\u0107, Miles Osborne, and Victor Lavrenko. 2010. The edinburgh twitter corpus. In Proceedings of the NAACL HLT Workshop on Computational Lin- guistics in a World of Social Media, pages 25-26.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Contrastive estimation: training log-linear models on unlabeled data",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Noah",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Eisner",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, ACL '05",
                "volume": "",
                "issue": "",
                "pages": "354--362",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Noah A. Smith and Jason Eisner. 2005. Contrastive estimation: training log-linear models on unlabeled data. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, ACL '05, pages 354-362, Stroudsburg, PA, USA. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Normalization of non-standard words",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sproat",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "W"
                        ],
                        "last": "Black",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Kumar",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Ostendorf",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Richards",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Computer Speech & Language",
                "volume": "15",
                "issue": "3",
                "pages": "287--333",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Sproat, A.W. Black, S. Chen, S. Kumar, M. Os- tendorf, and C. Richards. 2001. Normalization of non-standard words. Computer Speech & Language, 15(3):287-333.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "SRILM -an extensible language modeling toolkit",
                "authors": [
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "Stolcke",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of ICSLP",
                "volume": "",
                "issue": "",
                "pages": "901--904",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andreas Stolcke. 2002. SRILM -an extensible language modeling toolkit. In Proceedings of ICSLP, pages 901-904.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "New perspectives on an ol' variable: (t,d) in british english",
                "authors": [
                    {
                        "first": "Sali",
                        "middle": [],
                        "last": "Tagliamonte",
                        "suffix": ""
                    },
                    {
                        "first": "Rosalind",
                        "middle": [],
                        "last": "Temple",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Language Variation and Change",
                "volume": "17",
                "issue": "",
                "pages": "281--302",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sali Tagliamonte and Rosalind Temple. 2005. New perspectives on an ol' variable: (t,d) in british en- glish. Language Variation and Change, 17:281-302, September.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Adaptive parsercentric text normalization",
                "authors": [
                    {
                        "first": "Congle",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Tyler",
                        "middle": [],
                        "last": "Baldwin",
                        "suffix": ""
                    },
                    {
                        "first": "Howard",
                        "middle": [],
                        "last": "Ho",
                        "suffix": ""
                    },
                    {
                        "first": "Benny",
                        "middle": [],
                        "last": "Kimelfeld",
                        "suffix": ""
                    },
                    {
                        "first": "Yunyao",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "1159--1168",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Congle Zhang, Tyler Baldwin, Howard Ho, Benny Kimelfeld, and Yunyao Li. 2013. Adaptive parser- centric text normalization. In Proceedings of ACL, pages 1159-1168.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF0": {
                "type_str": "table",
                "text": "The feature set for our log-linear model",
                "content": "<table/>",
                "num": null,
                "html": null
            },
            "TABREF2": {
                "type_str": "table",
                "text": "Empirical results",
                "content": "<table/>",
                "num": null,
                "html": null
            },
            "TABREF4": {
                "type_str": "table",
                "text": "Orthographic styles induced from automatically normalized Twitter text",
                "content": "<table/>",
                "num": null,
                "html": null
            }
        }
    }
}