File size: 129,589 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
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
{
    "paper_id": "P13-1045",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T09:36:16.827414Z"
    },
    "title": "Parsing with Compositional Vector Grammars",
    "authors": [
        {
            "first": "Richard",
            "middle": [],
            "last": "Socher",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Stanford University",
                "location": {
                    "postCode": "94305",
                    "settlement": "Stanford",
                    "region": "CA",
                    "country": "USA"
                }
            },
            "email": "richard@socher.org"
        },
        {
            "first": "John",
            "middle": [],
            "last": "Bauer",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Stanford University",
                "location": {
                    "postCode": "94305",
                    "settlement": "Stanford",
                    "region": "CA",
                    "country": "USA"
                }
            },
            "email": ""
        },
        {
            "first": "Christopher",
            "middle": [
                "D"
            ],
            "last": "Manning",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Stanford University",
                "location": {
                    "postCode": "94305",
                    "settlement": "Stanford",
                    "region": "CA",
                    "country": "USA"
                }
            },
            "email": "manning@stanford.edu"
        },
        {
            "first": "Andrew",
            "middle": [
                "Y"
            ],
            "last": "Ng",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Stanford University",
                "location": {
                    "postCode": "94305",
                    "settlement": "Stanford",
                    "region": "CA",
                    "country": "USA"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Natural language parsing has typically been done with small sets of discrete categories such as NP and VP, but this representation does not capture the full syntactic nor semantic richness of linguistic phrases, and attempts to improve on this by lexicalizing phrases or splitting categories only partly address the problem at the cost of huge feature spaces and sparseness. Instead, we introduce a Compositional Vector Grammar (CVG), which combines PCFGs with a syntactically untied recursive neural network that learns syntactico-semantic, compositional vector representations. The CVG improves the PCFG of the Stanford Parser by 3.8% to obtain an F1 score of 90.4%. It is fast to train and implemented approximately as an efficient reranker it is about 20% faster than the current Stanford factored parser. The CVG learns a soft notion of head words and improves performance on the types of ambiguities that require semantic information such as PP attachments.",
    "pdf_parse": {
        "paper_id": "P13-1045",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Natural language parsing has typically been done with small sets of discrete categories such as NP and VP, but this representation does not capture the full syntactic nor semantic richness of linguistic phrases, and attempts to improve on this by lexicalizing phrases or splitting categories only partly address the problem at the cost of huge feature spaces and sparseness. Instead, we introduce a Compositional Vector Grammar (CVG), which combines PCFGs with a syntactically untied recursive neural network that learns syntactico-semantic, compositional vector representations. The CVG improves the PCFG of the Stanford Parser by 3.8% to obtain an F1 score of 90.4%. It is fast to train and implemented approximately as an efficient reranker it is about 20% faster than the current Stanford factored parser. The CVG learns a soft notion of head words and improves performance on the types of ambiguities that require semantic information such as PP attachments.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Syntactic parsing is a central task in natural language processing because of its importance in mediating between linguistic expression and meaning. For example, much work has shown the usefulness of syntactic representations for subsequent tasks such as relation extraction, semantic role labeling (Gildea and Palmer, 2002) and paraphrase detection (Callison-Burch, 2008) .",
                "cite_spans": [
                    {
                        "start": 299,
                        "end": 324,
                        "text": "(Gildea and Palmer, 2002)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 350,
                        "end": 372,
                        "text": "(Callison-Burch, 2008)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Syntactic descriptions standardly use coarse discrete categories such as NP for noun phrases or PP for prepositional phrases. However, recent work has shown that parsing results can be greatly improved by defining more fine-grained syntactic Discrete Syntactic -Continuous Semantic Representations in the Compositional Vector Grammar Figure 1 : Example of a CVG tree with (category,vector) representations at each node. The vectors for nonterminals are computed via a new type of recursive neural network which is conditioned on syntactic categories from a PCFG. categories, which better capture phrases with similar behavior, whether through manual feature engineering (Klein and Manning, 2003a) or automatic learning (Petrov et al., 2006) . However, subdividing a category like NP into 30 or 60 subcategories can only provide a very limited representation of phrase meaning and semantic similarity. Two strands of work therefore attempt to go further. First, recent work in discriminative parsing has shown gains from careful engineering of features (Taskar et al., 2004; Finkel et al., 2008) . Features in such parsers can be seen as defining effective dimensions of similarity between categories. Second, lexicalized parsers (Collins, 2003; Charniak, 2000) associate each category with a lexical item. This gives a fine-grained notion of semantic similarity, which is useful for tackling problems like ambiguous attachment decisions. However, this approach necessitates complex shrinkage estimation schemes to deal with the sparsity of observations of the lexicalized categories.",
                "cite_spans": [
                    {
                        "start": 670,
                        "end": 696,
                        "text": "(Klein and Manning, 2003a)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 719,
                        "end": 740,
                        "text": "(Petrov et al., 2006)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 1052,
                        "end": 1073,
                        "text": "(Taskar et al., 2004;",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 1074,
                        "end": 1094,
                        "text": "Finkel et al., 2008)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 1229,
                        "end": 1244,
                        "text": "(Collins, 2003;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 1245,
                        "end": 1260,
                        "text": "Charniak, 2000)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 334,
                        "end": 342,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In many natural language systems, single words and n-grams are usefully described by their distributional similarities (Brown et al., 1992) , among many others. But, even with large corpora, many n-grams will never be seen during training, especially when n is large. In these cases, one cannot simply use distributional similarities to represent unseen phrases. In this work, we present a new solution to learn features and phrase representations even for very long, unseen n-grams.",
                "cite_spans": [
                    {
                        "start": 119,
                        "end": 139,
                        "text": "(Brown et al., 1992)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We introduce a Compositional Vector Grammar Parser (CVG) for structure prediction. Like the above work on parsing, the model addresses the problem of representing phrases and categories. Unlike them, it jointly learns how to parse and how to represent phrases as both discrete categories and continuous vectors as illustrated in Fig. 1 . CVGs combine the advantages of standard probabilistic context free grammars (PCFG) with those of recursive neural networks (RNNs). The former can capture the discrete categorization of phrases into NP or PP while the latter can capture fine-grained syntactic and compositional-semantic information on phrases and words. This information can help in cases where syntactic ambiguity can only be resolved with semantic information, such as in the PP attachment of the two sentences: They ate udon with forks. vs. They ate udon with chicken.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 329,
                        "end": 335,
                        "text": "Fig. 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Previous RNN-based parsers used the same (tied) weights at all nodes to compute the vector representing a constituent (Socher et al., 2011b) . This requires the composition function to be extremely powerful, since it has to combine phrases with different syntactic head words, and it is hard to optimize since the parameters form a very deep neural network. We generalize the fully tied RNN to one with syntactically untied weights. The weights at each node are conditionally dependent on the categories of the child constituents. This allows different composition functions when combining different types of phrases and is shown to result in a large improvement in parsing accuracy.",
                "cite_spans": [
                    {
                        "start": 118,
                        "end": 140,
                        "text": "(Socher et al., 2011b)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Our compositional distributed representation allows a CVG parser to make accurate parsing decisions and capture similarities between phrases and sentences. Any PCFG-based parser can be improved with an RNN. We use a simplified version of the Stanford Parser (Klein and Manning, 2003a) as the base PCFG and improve its accuracy from 86.56 to 90.44% labeled F1 on all sentences of the WSJ section 23. The code of our parser is available at nlp.stanford.edu.",
                "cite_spans": [
                    {
                        "start": 258,
                        "end": 284,
                        "text": "(Klein and Manning, 2003a)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The CVG is inspired by two lines of research: Enriching PCFG parsers through more diverse sets of discrete states and recursive deep learning models that jointly learn classifiers and continuous feature representations for variable-sized inputs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "As mentioned in the introduction, there are several approaches to improving discrete representations for parsing. Klein and Manning (2003a) use manual feature engineering, while Petrov et al. (2006) use a learning algorithm that splits and merges the syntactic categories in order to maximize likelihood on the treebank. Their approach splits categories into several dozen subcategories. Another approach is lexicalized parsers (Collins, 2003; Charniak, 2000) that describe each category with a lexical item, usually the head word. More recently, combine several such annotation schemes in a factored parser. We extend the above ideas from discrete representations to richer continuous ones. The CVG can be seen as factoring discrete and continuous parsing in one model. Another different approach to the above generative models is to learn discriminative parsers using many well designed features (Taskar et al., 2004; Finkel et al., 2008) . We also borrow ideas from this line of research in that our parser combines the generative PCFG model with discriminatively learned RNNs.",
                "cite_spans": [
                    {
                        "start": 114,
                        "end": 139,
                        "text": "Klein and Manning (2003a)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 178,
                        "end": 198,
                        "text": "Petrov et al. (2006)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 428,
                        "end": 443,
                        "text": "(Collins, 2003;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 444,
                        "end": 459,
                        "text": "Charniak, 2000)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 898,
                        "end": 919,
                        "text": "(Taskar et al., 2004;",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 920,
                        "end": 940,
                        "text": "Finkel et al., 2008)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Improving Discrete Syntactic Representations",
                "sec_num": null
            },
            {
                "text": "Early attempts at using neural networks to describe phrases include Elman (1991) , who used recurrent neural networks to create representations of sentences from a simple toy grammar and to analyze the linguistic expressiveness of the resulting representations. Words were represented as one-on vectors, which was feasible since the grammar only included a handful of words. Collobert and Weston (2008) showed that neural networks can perform well on sequence labeling language processing tasks while also learning appropriate features. However, their model is lacking in that it cannot represent the recursive structure inherent in natural language. They partially circumvent this problem by using either independent window-based classifiers or a convolutional layer. RNN-specific training was introduced by Goller and K\u00fcchler (1996) to learn distributed representations of given, structured objects such as logical terms. In contrast, our model both predicts the structure and its representation. Henderson (2003) was the first to show that neural networks can be successfully used for large scale parsing. He introduced a left-corner parser to estimate the probabilities of parsing decisions conditioned on the parsing history. The input to Henderson's model consists of pairs of frequent words and their part-of-speech (POS) tags. Both the original parsing system and its probabilistic interpretation (Titov and Henderson, 2007) learn features that represent the parsing history and do not provide a principled linguistic representation like our phrase representations. Other related work includes (Henderson, 2004) , who discriminatively trains a parser based on synchrony networks and (Titov and Henderson, 2006) , who use an SVM to adapt a generative parser to different domains. Costa et al. (2003) apply recursive neural networks to re-rank possible phrase attachments in an incremental parser. Their work is the first to show that RNNs can capture enough information to make correct parsing decisions, but they only test on a subset of 2000 sentences. Menchetti et al. (2005) use RNNs to re-rank different parses. For their results on full sentence parsing, they rerank candidate trees created by the Collins parser (Collins, 2003) . Similar to their work, we use the idea of letting discrete categories reduce the search space during inference. We compare to fully tied RNNs in which the same weights are used at every node. Our syntactically untied RNNs outperform them by a significant margin. The idea of untying has also been successfully used in deep learning applied to vision (Le et al., 2010) .",
                "cite_spans": [
                    {
                        "start": 68,
                        "end": 80,
                        "text": "Elman (1991)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 375,
                        "end": 402,
                        "text": "Collobert and Weston (2008)",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 809,
                        "end": 834,
                        "text": "Goller and K\u00fcchler (1996)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 999,
                        "end": 1015,
                        "text": "Henderson (2003)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 1405,
                        "end": 1432,
                        "text": "(Titov and Henderson, 2007)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 1602,
                        "end": 1619,
                        "text": "(Henderson, 2004)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 1691,
                        "end": 1718,
                        "text": "(Titov and Henderson, 2006)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 1787,
                        "end": 1806,
                        "text": "Costa et al. (2003)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 2062,
                        "end": 2085,
                        "text": "Menchetti et al. (2005)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 2226,
                        "end": 2241,
                        "text": "(Collins, 2003)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 2594,
                        "end": 2611,
                        "text": "(Le et al., 2010)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Deep Learning and Recursive Deep Learning",
                "sec_num": null
            },
            {
                "text": "This paper uses several ideas of (Socher et al., 2011b) . The main differences are (i) the dual representation of nodes as discrete categories and vectors, (ii) the combination with a PCFG, and (iii) the syntactic untying of weights based on child categories. We directly compare models with fully tied and untied weights. Another work that represents phrases with a dual discrete-continuous representation is (Kartsaklis et al., 2012) .",
                "cite_spans": [
                    {
                        "start": 33,
                        "end": 55,
                        "text": "(Socher et al., 2011b)",
                        "ref_id": "BIBREF34"
                    },
                    {
                        "start": 410,
                        "end": 435,
                        "text": "(Kartsaklis et al., 2012)",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Deep Learning and Recursive Deep Learning",
                "sec_num": null
            },
            {
                "text": "This section introduces Compositional Vector Grammars (CVGs), a model to jointly find syntactic structure and capture compositional semantic information.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Compositional Vector Grammars",
                "sec_num": "3"
            },
            {
                "text": "CVGs build on two observations. Firstly, that a lot of the structure and regularity in languages can be captured by well-designed syntactic patterns.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Compositional Vector Grammars",
                "sec_num": "3"
            },
            {
                "text": "Hence, the CVG builds on top of a standard PCFG parser. However, many parsing decisions show fine-grained semantic factors at work. Therefore we combine syntactic and semantic information by giving the parser access to rich syntacticosemantic information in the form of distributional word vectors and compute compositional semantic vector representations for longer phrases (Costa et al., 2003; Menchetti et al., 2005; Socher et al., 2011b ). The CVG model merges ideas from both generative models that assume discrete syntactic categories and discriminative models that are trained using continuous vectors.",
                "cite_spans": [
                    {
                        "start": 375,
                        "end": 395,
                        "text": "(Costa et al., 2003;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 396,
                        "end": 419,
                        "text": "Menchetti et al., 2005;",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 420,
                        "end": 440,
                        "text": "Socher et al., 2011b",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Compositional Vector Grammars",
                "sec_num": "3"
            },
            {
                "text": "We will first briefly introduce single word vector representations and then describe the CVG objective function, tree scoring and inference.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Compositional Vector Grammars",
                "sec_num": "3"
            },
            {
                "text": "In most systems that use a vector representation for words, such vectors are based on cooccurrence statistics of each word and its context (Turney and Pantel, 2010) . Another line of research to learn distributional word vectors is based on neural language models (Bengio et al., 2003) which jointly learn an embedding of words into an n-dimensional feature space and use these embeddings to predict how suitable a word is in its context. These vector representations capture interesting linear relationships (up to some accuracy), such as king\u2212man+woman \u2248 queen (Mikolov et al., 2013) . Collobert and Weston (2008) introduced a new model to compute such an embedding. The idea is to construct a neural network that outputs high scores for windows that occur in a large unlabeled corpus and low scores for windows where one word is replaced by a random word. When such a network is optimized via gradient ascent the derivatives backpropagate into the word embedding matrix X. In order to predict correct scores the vectors in the matrix capture co-occurrence statistics.",
                "cite_spans": [
                    {
                        "start": 139,
                        "end": 164,
                        "text": "(Turney and Pantel, 2010)",
                        "ref_id": "BIBREF40"
                    },
                    {
                        "start": 264,
                        "end": 285,
                        "text": "(Bengio et al., 2003)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 563,
                        "end": 585,
                        "text": "(Mikolov et al., 2013)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 588,
                        "end": 615,
                        "text": "Collobert and Weston (2008)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word Vector Representations",
                "sec_num": "3.1"
            },
            {
                "text": "For further details and evaluations of these embeddings, see (Turian et al., 2010; Huang et al., 2012) . The resulting X matrix is used as follows. Assume we are given a sentence as an ordered list of m words. Each word w has an index [w] = i into the columns of the embedding matrix. This index is used to retrieve the word's vector representation a w using a simple multiplication with a binary vector e, which is zero everywhere, except at the ith index. So a w = Le i \u2208 R n . Henceforth, after mapping each word to its vector, we represent a sentence S as an ordered list of (word,vector) pairs:",
                "cite_spans": [
                    {
                        "start": 61,
                        "end": 82,
                        "text": "(Turian et al., 2010;",
                        "ref_id": "BIBREF39"
                    },
                    {
                        "start": 83,
                        "end": 102,
                        "text": "Huang et al., 2012)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word Vector Representations",
                "sec_num": "3.1"
            },
            {
                "text": "x = ((w 1 , a w 1 ), . . . , (w m , a wm )).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word Vector Representations",
                "sec_num": "3.1"
            },
            {
                "text": "Now that we have discrete and continuous representations for all words, we can continue with the approach for computing tree structures and vectors for nonterminal nodes.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word Vector Representations",
                "sec_num": "3.1"
            },
            {
                "text": "The goal of supervised parsing is to learn a function g : X \u2192 Y, where X is the set of sentences and Y is the set of all possible labeled binary parse trees. The set of all possible trees for a given sentence x i is defined as Y (x i ) and the correct tree for a sentence is y i . We first define a structured margin loss \u2206(y i ,\u0177) for predicting a tree\u0177 for a given correct tree. The loss increases the more incorrect the proposed parse tree is (Goodman, 1998) . The discrepancy between trees is measured by counting the number of nodes N (y) with an incorrect span (or label) in the proposed tree:",
                "cite_spans": [
                    {
                        "start": 446,
                        "end": 461,
                        "text": "(Goodman, 1998)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u2206(y i ,\u0177) = d\u2208N (\u0177) \u03ba1{d / \u2208 N (y i )}.",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "We set \u03ba = 0.1 in all experiments. For a given set of training instances (x i , y i ), we search for the function g \u03b8 , parameterized by \u03b8, with the smallest expected loss on a new sentence. It has the following form:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "g \u03b8 (x) = arg max y\u2208Y (x) s(CVG(\u03b8, x,\u0177)),",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "where the tree is found by the Compositional Vector Grammar (CVG) introduced below and then scored via the function s. The higher the score of a tree the more confident the algorithm is that its structure is correct. This max-margin, structureprediction objective (Taskar et al., 2004; Ratliff et al., 2007; Socher et al., 2011b) trains the CVG so that the highest scoring tree will be the correct tree: g \u03b8 (x i ) = y i and its score will be larger up to a margin to other possible trees\u0177 \u2208 Y(x i ):",
                "cite_spans": [
                    {
                        "start": 264,
                        "end": 285,
                        "text": "(Taskar et al., 2004;",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 286,
                        "end": 307,
                        "text": "Ratliff et al., 2007;",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 308,
                        "end": 329,
                        "text": "Socher et al., 2011b)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "s(CVG(\u03b8, x i , y i )) \u2265 s(CVG(\u03b8, x i ,\u0177)) + \u2206(y i ,\u0177).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "This leads to the regularized risk function for m training examples:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "J(\u03b8) = 1 m m i=1 r i (\u03b8) + \u03bb 2 \u03b8 2 2 ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "where",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "r i (\u03b8) = max y\u2208Y (x i ) s(CVG(x i ,\u0177)) + \u2206(y i ,\u0177) \u2212 s(CVG(x i , y i ))",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "Intuitively, to minimize this objective, the score of the correct tree y i is increased and the score of the highest scoring incorrect tree\u0177 is decreased.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Max-Margin Training Objective for CVGs",
                "sec_num": "3.2"
            },
            {
                "text": "For ease of exposition, we first describe how to score an existing fully labeled tree with a standard RNN and then with a CVG. The subsequent section will then describe a bottom-up beam search and its approximation for finding the optimal tree. Assume, for now, we are given a labeled parse tree as shown in Fig. 2 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 308,
                        "end": 314,
                        "text": "Fig. 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "We define the word representations as (vector, POS) pairs: ((a, A), (b, B), (c, C)), where the vectors are defined as in Sec. 3.1 and the POS tags come from a PCFG. The standard RNN essentially ignores all POS tags and syntactic categories and each nonterminal node is associated with the same neural network (i.e., the weights across nodes are fully tied). We can represent the binary tree in Fig. 2 in the form of branching triplets (p \u2192 c 1 c 2 ). Each such triplet denotes that a parent node p has two children and each c k can be either a word vector or a non-terminal node in the tree. For the example in Fig. 2 , we would get the triples ((p 1 \u2192 bc), (p 2 \u2192 ap 1 )). Note that in order to replicate the neural network and compute node representations in a bottom up fashion, the parent must have the same dimensionality as the children:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 394,
                        "end": 400,
                        "text": "Fig. 2",
                        "ref_id": null
                    },
                    {
                        "start": 611,
                        "end": 617,
                        "text": "Fig. 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "p \u2208 R n .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "Given this tree structure, we can now compute activations for each node from the bottom up. We begin by computing the activation for p 1 using the children's word vectors. We first concatenate the children's representations b, c \u2208 R n\u00d71 into a vector b c \u2208 R 2n\u00d71 . Then the composition function multiplies this vector by the parameter weights of the RNN W \u2208 R n\u00d72n and applies an element-wise nonlinearity function f = tanh to the output vector. The resulting output p (1) is then given as input to compute p (2) . Figure 2 : An example tree with a simple Recursive Neural Network: The same weight matrix is replicated and used to compute all non-terminal node representations. Leaf nodes are n-dimensional vector representations of words.",
                "cite_spans": [
                    {
                        "start": 510,
                        "end": 513,
                        "text": "(2)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 516,
                        "end": 524,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "p (1) = f W b c , p (2) = f W a p 1 (A, a= ) (B, b= ) (C, c= ) P (1) , p (1) = P (2) , p (2) = Standard Recursive Neural Network = f W b c = f W a p (1)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "In order to compute a score of how plausible of a syntactic constituent a parent is the RNN uses a single-unit linear layer for all i:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "s(p (i) ) = v T p (i) ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "where v \u2208 R n is a vector of parameters that need to be trained. This score will be used to find the highest scoring tree. For more details on how standard RNNs can be used for parsing, see Socher et al. (2011b) .",
                "cite_spans": [
                    {
                        "start": 190,
                        "end": 211,
                        "text": "Socher et al. (2011b)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "The standard RNN requires a single composition function to capture all types of compositions: adjectives and nouns, verbs and nouns, adverbs and adjectives, etc. Even though this function is a powerful one, we find a single neural network weight matrix cannot fully capture the richness of compositionality. Several extensions are possible: A two-layered RNN would provide more expressive power, however, it is much harder to train because the resulting neural network becomes very deep and suffers from vanishing gradient problems. proposed to give every single word a matrix and a vector. The matrix is then applied to the sibling node's vector during the composition. While this results in a powerful composition function that essentially depends on the words being combined, the number of model parameters explodes and the composition functions do not capture the syntactic commonalities between similar POS tags or syntactic categories.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "Based on the above considerations, we propose the Compositional Vector Grammar (CVG) that conditions the composition function at each node on discrete syntactic categories extracted from a (A, a= ) (B, b= ) (C, c= ) P (1) , p (1) = P (2) , p (2) = Syntactically Untied Recursive Neural Network = f W (B,C) b c = f W (A,P ) a p (1) (1) Figure 3 : Example of a syntactically untied RNN in which the function to compute a parent vector depends on the syntactic categories of its children which we assume are given for now.",
                "cite_spans": [
                    {
                        "start": 218,
                        "end": 221,
                        "text": "(1)",
                        "ref_id": null
                    },
                    {
                        "start": 300,
                        "end": 305,
                        "text": "(B,C)",
                        "ref_id": null
                    },
                    {
                        "start": 327,
                        "end": 330,
                        "text": "(1)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 335,
                        "end": 343,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "PCFG. Hence, CVGs combine discrete, syntactic rule probabilities and continuous vector compositions. The idea is that the syntactic categories of the children determine what composition function to use for computing the vector of their parents.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "While not perfect, a dedicated composition function for each rule RHS can well capture common composition processes such as adjective or adverb modification versus noun or clausal complementation. For instance, it could learn that an NP should be similar to its head noun and little influenced by a determiner, whereas in an adjective modification both words considerably determine the meaning of a phrase. The original RNN is parameterized by a single weight matrix W . In contrast, the CVG uses a syntactically untied RNN (SU-RNN) which has a set of such weights. The size of this set depends on the number of sibling category combinations in the PCFG. Fig. 3 shows an example SU-RNN that computes parent vectors with syntactically untied weights. The CVG computes the first parent vector via the SU-RNN:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 655,
                        "end": 661,
                        "text": "Fig. 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "p (1) = f W (B,C) b c ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "where W (B,C) \u2208 R n\u00d72n is now a matrix that depends on the categories of the two children. In this bottom up procedure, the score for each node consists of summing two elements: First, a single linear unit that scores the parent vector and second, the log probability of the PCFG for the rule that combines these two children:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s p (1) = v (B,C) T p (1) + log P (P 1 \u2192 B C),",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "where P (P 1 \u2192 B C) comes from the PCFG. This can be interpreted as the log probability of a discrete-continuous rule application with the following factorization:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "P ((P 1 , p 1 ) \u2192 (B, b)(C, c)) (5) = P (p 1 \u2192 b c|P 1 \u2192 B C)P (P 1 \u2192 B C),",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "Note, however, that due to the continuous nature of the word vectors, the probability of such a CVG rule application is not comparable to probabilities provided by a PCFG since the latter sum to 1 for all children.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "Assuming that node p 1 has syntactic category P 1 , we compute the second parent vector via:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "p (2) = f W (A,P 1 ) a p (1) .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "The score of the last parent in this trigram is computed via:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "s p (2) = v (A,P 1 ) T p (2) + log P (P 2 \u2192 A P 1 ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Scoring Trees with CVGs",
                "sec_num": "3.3"
            },
            {
                "text": "The above scores (Eq. 4) are used in the search for the correct tree for a sentence. The goodness of a tree is measured in terms of its score and the CVG score of a complete tree is the sum of the scores at each node:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parsing with CVGs",
                "sec_num": "3.4"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s(CVG(\u03b8, x,\u0177)) = d\u2208N (\u0177) s p d .",
                        "eq_num": "(6)"
                    }
                ],
                "section": "Parsing with CVGs",
                "sec_num": "3.4"
            },
            {
                "text": "The main objective function in Eq. 3 includes a maximization over all possible trees max\u0177 \u2208Y (x) . Finding the global maximum, however, cannot be done efficiently for longer sentences nor can we use dynamic programming. This is due to the fact that the vectors break the independence assumptions of the base PCFG. A (category, vector) node representation is dependent on all the words in its span and hence to find the true global optimum, we would have to compute the scores for all binary trees. For a sentence of length n, there are Catalan(n) many possible binary trees which is very large even for moderately long sentences. One could use a bottom-up beam search, keeping a k-best list at every cell of the chart, possibly for each syntactic category. This beam search inference procedure is still considerably slower than using only the simplified base PCFG, especially since it has a small state space (see next section for details). Since each probability look-up is cheap but computing SU-RNN scores requires a matrix product, we would like to reduce the number of SU-RNN score computations to only those trees that require semantic information. We note that labeled F1 of the Stanford PCFG parser on the test set is 86.17%. However, if one used an oracle to select the best tree from the top 200 trees that it produces, one could get an F1 of 95.46%.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parsing with CVGs",
                "sec_num": "3.4"
            },
            {
                "text": "We use this knowledge to speed up inference via two bottom-up passes through the parsing chart. During the first one, we use only the base PCFG to run CKY dynamic programming through the tree. The k = 200-best parses at the top cell of the chart are calculated using the efficient algorithm of (Huang and Chiang, 2005) . Then, the second pass is a beam search with the full CVG model (including the more expensive matrix multiplications of the SU-RNN). This beam search only considers phrases that appear in the top 200 parses. This is similar to a re-ranking setup but with one main difference: the SU-RNN rule score computation at each node still only has access to its child vectors, not the whole tree or other global features. This allows the second pass to be very fast. We use this setup in our experiments below.",
                "cite_spans": [
                    {
                        "start": 294,
                        "end": 318,
                        "text": "(Huang and Chiang, 2005)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parsing with CVGs",
                "sec_num": "3.4"
            },
            {
                "text": "The full CVG model is trained in two stages. First the base PCFG is trained and its top trees are cached and then used for training the SU-RNN conditioned on the PCFG. The SU-RNN is trained using the objective in Eq. 3 and the scores as exemplified by Eq. 6. For each sentence, we use the method described above to efficiently find an approximation for the optimal tree.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training SU-RNNs",
                "sec_num": "3.5"
            },
            {
                "text": "To minimize the objective we want to increase the scores of the correct tree's constituents and decrease the score of those in the highest scoring incorrect tree. Derivatives are computed via backpropagation through structure (BTS) (Goller and K\u00fcchler, 1996) . The derivative of tree i has to be taken with respect to all parameter matrices W (AB) that appear in it. The main difference between backpropagation in standard RNNs and SU-RNNs is that the derivatives at each node only add to the overall derivative of the specific matrix at that node. For more details on backpropagation through RNNs, see Socher et al. (2010) ",
                "cite_spans": [
                    {
                        "start": 232,
                        "end": 258,
                        "text": "(Goller and K\u00fcchler, 1996)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 603,
                        "end": 623,
                        "text": "Socher et al. (2010)",
                        "ref_id": "BIBREF32"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training SU-RNNs",
                "sec_num": "3.5"
            },
            {
                "text": "The objective function is not differentiable due to the hinge loss. Therefore, we generalize gradient ascent via the subgradient method (Ratliff et al., 2007) which computes a gradient-like direction. Let \u03b8 = (X, W (\u2022\u2022) , v (\u2022\u2022) ) \u2208 R M be a vector of all M model parameters, where we denote W (\u2022\u2022) as the set of matrices that appear in the training set. The subgradient of Eq. 3 becomes:",
                "cite_spans": [
                    {
                        "start": 136,
                        "end": 158,
                        "text": "(Ratliff et al., 2007)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 215,
                        "end": 219,
                        "text": "(\u2022\u2022)",
                        "ref_id": null
                    },
                    {
                        "start": 224,
                        "end": 228,
                        "text": "(\u2022\u2022)",
                        "ref_id": null
                    },
                    {
                        "start": 294,
                        "end": 298,
                        "text": "(\u2022\u2022)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Subgradient Methods and AdaGrad",
                "sec_num": "3.6"
            },
            {
                "text": "\u2202J \u2202\u03b8 = i \u2202s(x i ,\u0177 max ) \u2202\u03b8 \u2212 \u2202s(x i , y i ) \u2202\u03b8 + \u03b8,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Subgradient Methods and AdaGrad",
                "sec_num": "3.6"
            },
            {
                "text": "where\u0177 max is the tree with the highest score. To minimize the objective, we use the diagonal variant of AdaGrad (Duchi et al., 2011) with minibatches. For our parameter updates, we first define g \u03c4 \u2208 R M \u00d71 to be the subgradient at time step \u03c4 and G t = t \u03c4 =1 g \u03c4 g T \u03c4 . The parameter update at time step t then becomes:",
                "cite_spans": [
                    {
                        "start": 113,
                        "end": 133,
                        "text": "(Duchi et al., 2011)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Subgradient Methods and AdaGrad",
                "sec_num": "3.6"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u03b8 t = \u03b8 t\u22121 \u2212 \u03b1 (diag(G t )) \u22121/2 g t ,",
                        "eq_num": "(7)"
                    }
                ],
                "section": "Subgradient Methods and AdaGrad",
                "sec_num": "3.6"
            },
            {
                "text": "where \u03b1 is the learning rate. Since we use the diagonal of G t , we only have to store M values and the update becomes fast to compute: At time step t, the update for the i'th parameter \u03b8 t,i is:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Subgradient Methods and AdaGrad",
                "sec_num": "3.6"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u03b8 t,i = \u03b8 t\u22121,i \u2212 \u03b1 t \u03c4 =1 g 2 \u03c4,i g t,i .",
                        "eq_num": "(8)"
                    }
                ],
                "section": "Subgradient Methods and AdaGrad",
                "sec_num": "3.6"
            },
            {
                "text": "Hence, the learning rate is adapting differently for each parameter and rare parameters get larger updates than frequently occurring parameters. This is helpful in our setting since some W matrices appear in only a few training trees. This procedure found much better optima (by \u22483% labeled F1 on the dev set), and converged more quickly than L-BFGS which we used previously in RNN training (Socher et al., 2011a) . Training time is roughly 4 hours on a single machine.",
                "cite_spans": [
                    {
                        "start": 391,
                        "end": 413,
                        "text": "(Socher et al., 2011a)",
                        "ref_id": "BIBREF33"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Subgradient Methods and AdaGrad",
                "sec_num": "3.6"
            },
            {
                "text": "In the absence of any knowledge on how to combine two categories, our prior for combining two vectors is to average them instead of performing a completely random projection. Hence, we initialize the binary W matrices with:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Initialization of Weight Matrices",
                "sec_num": "3.7"
            },
            {
                "text": "W (\u2022\u2022) = 0.5[I n\u00d7n I n\u00d7n 0 n\u00d71 ] + ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Initialization of Weight Matrices",
                "sec_num": "3.7"
            },
            {
                "text": "where we include the bias in the last column and the random variable is uniformly distributed: \u223c U[\u22120.001, 0.001]. The first block is multiplied by the left child and the second by the right child:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Initialization of Weight Matrices",
                "sec_num": "3.7"
            },
            {
                "text": "W (AB) \uf8ee \uf8f0 a b 1 \uf8f9 \uf8fb = W (A) W (B) bias \uf8ee \uf8f0 a b 1 \uf8f9 \uf8fb = W (A) a + W (B) b + bias.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Initialization of Weight Matrices",
                "sec_num": "3.7"
            },
            {
                "text": "We evaluate the CVG in two ways: First, by a standard parsing evaluation on Penn Treebank WSJ and then by analyzing the model errors in detail.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "We used the first 20 files of WSJ section 22 to cross-validate several model and optimization choices. The base PCFG uses simplified categories of the Stanford PCFG Parser (Klein and Manning, 2003a) . We decreased the state splitting of the PCFG grammar (which helps both by making it less sparse and by reducing the number of parameters in the SU-RNN) by adding the following options to training: '-noRightRec -dominatesV 0 -baseNP 0'. This reduces the number of states from 15,276 to 12,061 states and 602 POS tags. These include split categories, such as parent annotation categories like VP\u02c6S. Furthermore, we ignore all category splits for the SU-RNN weights, resulting in 66 unary and 882 binary child pairs. Hence, the SU-RNN has 66+882 transformation matrices and scoring vectors. Note that any PCFG, including latent annotation PCFGs (Matsuzaki et al., 2005 ) could be used. However, since the vectors will capture lexical and semantic information, even simple base PCFGs can be substantially improved. Since the computational complexity of PCFGs depends on the number of states, a base PCFG with fewer states is much faster.",
                "cite_spans": [
                    {
                        "start": 172,
                        "end": 198,
                        "text": "(Klein and Manning, 2003a)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 843,
                        "end": 866,
                        "text": "(Matsuzaki et al., 2005",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Cross-validating Hyperparameters",
                "sec_num": "4.1"
            },
            {
                "text": "Testing on the full WSJ section 22 dev set (1700 sentences) takes roughly 470 seconds with the simple base PCFG, 1320 seconds with our new CVG and 1600 seconds with the currently published Stanford factored parser. Hence, increased performance comes also with a speed improvement of approximately 20%.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Cross-validating Hyperparameters",
                "sec_num": "4.1"
            },
            {
                "text": "We fix the same regularization of \u03bb = 10 \u22124 for all parameters. The minibatch size was set to performance and were faster than 50-,100-or 200dimensional ones. We hypothesize that the larger word vector sizes, while capturing more semantic knowledge, result in too many SU-RNN matrix parameters to train and hence perform worse.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Cross-validating Hyperparameters",
                "sec_num": "4.1"
            },
            {
                "text": "The dev set accuracy of the best model is 90.93% labeled F1 on all sentences. This model resulted in 90.44% on the final test set (WSJ section 23). Table 1 compares our results to the two Stanford parser variants (the unlexicalized PCFG (Klein and Manning, 2003a) and the factored parser (Klein and Manning, 2003b) ) and other parsers that use richer state representations: the Berkeley parser (Petrov and Klein, 2007) , Collins parser (Collins, 1997) , SSN: a statistical neural network parser (Henderson, 2004) , Factored PCFGs , Charniak-SelfTrain: the self-training approach of McClosky et al. (2006) , which bootstraps and parses additional large corpora multiple times, Charniak-RS: the state of the art self-trained and discriminatively re-ranked Charniak-Johnson parser combining (Charniak, 2000; McClosky et al., 2006; Charniak and Johnson, 2005) . See Kummerfeld et al. (2012) for more comparisons. We compare also to a standard RNN 'CVG (RNN)' and to the proposed CVG with SU-RNNs.",
                "cite_spans": [
                    {
                        "start": 237,
                        "end": 263,
                        "text": "(Klein and Manning, 2003a)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 288,
                        "end": 314,
                        "text": "(Klein and Manning, 2003b)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 394,
                        "end": 418,
                        "text": "(Petrov and Klein, 2007)",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 436,
                        "end": 451,
                        "text": "(Collins, 1997)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 495,
                        "end": 512,
                        "text": "(Henderson, 2004)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 582,
                        "end": 604,
                        "text": "McClosky et al. (2006)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 788,
                        "end": 804,
                        "text": "(Charniak, 2000;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 805,
                        "end": 827,
                        "text": "McClosky et al., 2006;",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 828,
                        "end": 855,
                        "text": "Charniak and Johnson, 2005)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 862,
                        "end": 886,
                        "text": "Kummerfeld et al. (2012)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 148,
                        "end": 155,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Results on WSJ",
                "sec_num": "4.2"
            },
            {
                "text": "Analysis of Error Types. Table 2 shows a detailed comparison of different errors. We use the code provided by Kummerfeld et al. (2012) and compare to the previous version of the Stanford factored parser as well as to the Berkeley and Charniak-reranked-self-trained parsers (defined above). See Kummerfeld et al. (2012) Analysis of Composition Matrices. An analysis of the norms of the binary matrices reveals that the model learns a soft vectorized notion of head words: Head words are given larger weights and importance when computing the parent vector: For the matrices combining siblings with categories VP:PP, VP:NP and VP:PRT, the weights in the part of the matrix which is multiplied with the VP child vector dominates. Similarly NPs dominate DTs. Fig. 5 shows example matrices. The two strong diagonals are due to the initialization described in Sec. 3.7. Semantic Transfer for PP Attachments. In this small model analysis, we use two pairs of sentences that the original Stanford parser and the CVG did not parse correctly after training on the WSJ. We then continue to train both parsers on two similar sentences and then analyze if the parsers correctly transferred the knowledge. The training sentences are He eats spaghetti with a fork. and She eats spaghetti with pork. The very similar test sentences are He eats spaghetti with a spoon. and He eats spaghetti with meat. Initially, both parsers incorrectly attach the PP to the verb in both test sentences. After training, the CVG parses both correctly, while the factored Stanford parser incorrectly attaches both PPs to spaghetti. The CVG's ability to transfer the correct PP attachments is due to the semantic word vector similarity between the words in the sentences. Fig. 4 shows the outputs of the two parsers. showing that head words dominate the composition. The model learns to not give determiners much importance. The two diagonals show clearly the two blocks that are multiplied with the left and right children, respectively.",
                "cite_spans": [
                    {
                        "start": 110,
                        "end": 134,
                        "text": "Kummerfeld et al. (2012)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 294,
                        "end": 318,
                        "text": "Kummerfeld et al. (2012)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 25,
                        "end": 32,
                        "text": "Table 2",
                        "ref_id": "TABREF3"
                    },
                    {
                        "start": 755,
                        "end": 761,
                        "text": "Fig. 5",
                        "ref_id": null
                    },
                    {
                        "start": 1736,
                        "end": 1742,
                        "text": "Fig. 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Model Analysis",
                "sec_num": "4.3"
            },
            {
                "text": "We introduced Compositional Vector Grammars (CVGs), a parsing model that combines the speed of small-state PCFGs with the semantic richness of neural word representations and compositional phrase vectors. The compositional vectors are learned with a new syntactically untied recursive neural network. This model is linguistically more plausible since it chooses different composition functions for a parent node based on the syntactic categories of its children. The CVG obtains 90.44% labeled F1 on the full WSJ test set and is 20% faster than the previous Stanford parser.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "5"
            },
            {
                "text": ". We also cross-validated on AdaGrad's learning rate which was eventually set to \u03b1 = 0.1 and word vector size. The 25-dimensional vectors provided byTurian et al. (2010) provided the best",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We thank Percy Liang for chats about the paper. Richard is supported by a Microsoft Research PhD fellowship. The authors gratefully acknowledge the support of the Defense Advanced Research Projects Agency (DARPA) Deep Exploration and Filtering of Text (DEFT) Program under Air Force Research Laboratory (AFRL) prime contract no. FA8750-13-2-0040, and the DARPA Deep Learning program under contract number FA8650-10-C-7020. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the view of DARPA, AFRL, or the US government.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "A neural probabilistic language model",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Ducharme",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Vincent",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Janvin",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Journal of Machine Learning Research",
                "volume": "3",
                "issue": "",
                "pages": "1137--1155",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Y. Bengio, R. Ducharme, P. Vincent, and C. Janvin. 2003. A neural probabilistic language model. Jour- nal of Machine Learning Research, 3:1137-1155.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Class-based n-gram models of natural language",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "F"
                        ],
                        "last": "Brown",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [
                            "V"
                        ],
                        "last": "Desouza",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [
                            "L"
                        ],
                        "last": "Mercer",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [
                            "J"
                        ],
                        "last": "Della Pietra",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "C"
                        ],
                        "last": "Lai",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Computational Linguistics",
                "volume": "18",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P. F. Brown, P. V. deSouza, R. L. Mercer, V. J. Della Pietra, and J. C. Lai. 1992. Class-based n-gram models of natural language. Computational Lin- guistics, 18.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Syntactic constraints on paraphrases extracted from parallel corpora",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Callison-Burch",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "196--205",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. Callison-Burch. 2008. Syntactic constraints on paraphrases extracted from parallel corpora. In Pro- ceedings of EMNLP, pages 196-205.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Coarse-to-fine n-best parsing and maxent discriminative reranking",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Charniak",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Johnson",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "E. Charniak and M. Johnson. 2005. Coarse-to-fine n-best parsing and maxent discriminative reranking. In ACL.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "A maximum-entropy-inspired parser",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Charniak",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "132--139",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "E. Charniak. 2000. A maximum-entropy-inspired parser. In Proceedings of ACL, pages 132-139.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Three generative, lexicalised models for statistical parsing",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Collins",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Collins. 1997. Three generative, lexicalised models for statistical parsing. In ACL.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Head-driven statistical models for natural language parsing",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Collins",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Computational Linguistics",
                "volume": "29",
                "issue": "4",
                "pages": "589--637",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Collins. 2003. Head-driven statistical models for natural language parsing. Computational Linguis- tics, 29(4):589-637.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "A unified architecture for natural language processing: deep neural networks with multitask learning",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Collobert",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Weston",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of ICML",
                "volume": "",
                "issue": "",
                "pages": "160--167",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Collobert and J. Weston. 2008. A unified archi- tecture for natural language processing: deep neural networks with multitask learning. In Proceedings of ICML, pages 160-167.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Towards incremental parsing of natural language using recursive neural networks",
                "authors": [
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Costa",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Frasconi",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Lombardo",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Soda",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "F. Costa, P. Frasconi, V. Lombardo, and G. Soda. 2003. Towards incremental parsing of natural language us- ing recursive neural networks. Applied Intelligence.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Adaptive subgradient methods for online learning and stochastic optimization",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Duchi",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Hazan",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Singer",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "JMLR",
                "volume": "12",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. Duchi, E. Hazan, and Y. Singer. 2011. Adaptive sub- gradient methods for online learning and stochastic optimization. JMLR, 12, July.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Distributed representations, simple recurrent networks, and grammatical structure",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "L"
                        ],
                        "last": "Elman",
                        "suffix": ""
                    }
                ],
                "year": 1991,
                "venue": "Machine Learning",
                "volume": "7",
                "issue": "",
                "pages": "195--225",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. L. Elman. 1991. Distributed representations, sim- ple recurrent networks, and grammatical structure. Machine Learning, 7(2-3):195-225.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Efficient, feature-based, conditional random field parsing",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "R"
                        ],
                        "last": "Finkel",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Kleeman",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "959--967",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. R. Finkel, A. Kleeman, and C. D. Manning. 2008. Efficient, feature-based, conditional random field parsing. In Proceedings of ACL, pages 959-967.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The necessity of parsing for predicate argument recognition",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Gildea",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "239--246",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Gildea and M. Palmer. 2002. The necessity of pars- ing for predicate argument recognition. In Proceed- ings of ACL, pages 239-246.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Learning taskdependent distributed representations by backpropagation through structure",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Goller",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "K\u00fcchler",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Proceedings of the International Conference on Neural Networks",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. Goller and A. K\u00fcchler. 1996. Learning task- dependent distributed representations by backprop- agation through structure. In Proceedings of the In- ternational Conference on Neural Networks.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Parsing Inside-Out",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Goodman",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. Goodman. 1998. Parsing Inside-Out. Ph.D. thesis, MIT.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Training factored pcfgs with expectation propagation",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Hall",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Hall and D. Klein. 2012. Training factored pcfgs with expectation propagation. In EMNLP.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Neural network probability estimation for broad coverage parsing",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Henderson",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of EACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. Henderson. 2003. Neural network probability esti- mation for broad coverage parsing. In Proceedings of EACL.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Discriminative training of a neural network statistical parser",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Henderson",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. Henderson. 2004. Discriminative training of a neu- ral network statistical parser. In ACL.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Better k-best parsing",
                "authors": [
                    {
                        "first": "Liang",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Chiang",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the 9th International Workshop on Parsing Technologies (IWPT 2005",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Liang Huang and David Chiang. 2005. Better k-best parsing. In Proceedings of the 9th International Workshop on Parsing Technologies (IWPT 2005).",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Improving Word Representations via Global Context and Multiple Word Prototypes",
                "authors": [
                    {
                        "first": "E",
                        "middle": [
                            "H"
                        ],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "Y"
                        ],
                        "last": "Ng",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "E. H. Huang, R. Socher, C. D. Manning, and A. Y. Ng. 2012. Improving Word Representations via Global Context and Multiple Word Prototypes. In ACL.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "A unified sentence space for categorical distributionalcompositional semantics: Theory and experiments",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Kartsaklis",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Sadrzadeh",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Pulman",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Proceedings of 24th International Conference on Computational Linguistics (COLING): Posters",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Kartsaklis, M. Sadrzadeh, and S. Pulman. 2012. A unified sentence space for categorical distributional- compositional semantics: Theory and experiments. Proceedings of 24th International Conference on Computational Linguistics (COLING): Posters.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Accurate unlexicalized parsing",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "423--430",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Klein and C. D. Manning. 2003a. Accurate un- lexicalized parsing. In Proceedings of ACL, pages 423-430.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Fast exact inference with a factored model for natural language parsing",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "NIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Klein and C.D. Manning. 2003b. Fast exact in- ference with a factored model for natural language parsing. In NIPS.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Parser showdown at the wall street corral: An empirical investigation of error types in parser output",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "K"
                        ],
                        "last": "Kummerfeld",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Hall",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "R"
                        ],
                        "last": "Curran",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. K. Kummerfeld, D. Hall, J. R. Curran, and D. Klein. 2012. Parser showdown at the wall street corral: An empirical investigation of error types in parser out- put. In EMNLP.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Tiled convolutional neural networks",
                "authors": [
                    {
                        "first": "Q",
                        "middle": [
                            "V"
                        ],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Ngiam",
                        "suffix": ""
                    },
                    {
                        "first": "Z",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Chia",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [
                            "W"
                        ],
                        "last": "Koh",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "Y"
                        ],
                        "last": "Ng",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "NIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Q. V. Le, J. Ngiam, Z. Chen, D. Chia, P. W. Koh, and A. Y. Ng. 2010. Tiled convolutional neural net- works. In NIPS.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Probabilistic cfg with latent annotations",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Matsuzaki",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Miyao",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Tsujii",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "T. Matsuzaki, Y. Miyao, and J. Tsujii. 2005. Proba- bilistic cfg with latent annotations. In ACL.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Effective self-training for parsing",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Mcclosky",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Charniak",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Johnson",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "NAACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. McClosky, E. Charniak, and M. Johnson. 2006. Ef- fective self-training for parsing. In NAACL.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Wide coverage natural language processing using kernel methods and neural networks for structured data",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Menchetti",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Costa",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Frasconi",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Pontil",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Pattern Recognition Letters",
                "volume": "26",
                "issue": "12",
                "pages": "1896--1906",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Menchetti, F. Costa, P. Frasconi, and M. Pon- til. 2005. Wide coverage natural language pro- cessing using kernel methods and neural networks for structured data. Pattern Recognition Letters, 26(12):1896-1906.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Linguistic regularities in continuous spaceword representations",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Yih",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Zweig",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "HLT-NAACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "T. Mikolov, W. Yih, and G. Zweig. 2013. Linguis- tic regularities in continuous spaceword representa- tions. In HLT-NAACL.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Improved inference for unlexicalized parsing",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Petrov",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "NAACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Petrov and D. Klein. 2007. Improved inference for unlexicalized parsing. In NAACL.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Learning accurate, compact, and interpretable tree annotation",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Petrov",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Barrett",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Thibaux",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "433--440",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Petrov, L. Barrett, R. Thibaux, and D. Klein. 2006. Learning accurate, compact, and interpretable tree annotation. In Proceedings of ACL, pages 433-440.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Online) subgradient methods for structured prediction",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Ratliff",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "A"
                        ],
                        "last": "Bagnell",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Zinkevich",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Eleventh International Conference on Artificial Intelligence and Statistics (AIStats)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "N. Ratliff, J. A. Bagnell, and M. Zinkevich. 2007. (On- line) subgradient methods for structured prediction. In Eleventh International Conference on Artificial Intelligence and Statistics (AIStats).",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Learning continuous phrase representations and syntactic parsing with recursive neural networks",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "Y"
                        ],
                        "last": "Ng",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the NIPS-2010 Deep Learning and Unsupervised Feature Learning Workshop",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Socher, C. D. Manning, and A. Y. Ng. 2010. Learn- ing continuous phrase representations and syntactic parsing with recursive neural networks. In Proceed- ings of the NIPS-2010 Deep Learning and Unsuper- vised Feature Learning Workshop.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Dynamic Pooling and Unfolding Recursive Autoencoders for Paraphrase Detection",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [
                            "H"
                        ],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Pennington",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "Y"
                        ],
                        "last": "Ng",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "NIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Socher, E. H. Huang, J. Pennington, A. Y. Ng, and C. D. Manning. 2011a. Dynamic Pooling and Un- folding Recursive Autoencoders for Paraphrase De- tection. In NIPS. MIT Press.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Parsing Natural Scenes and Natural Language with Recursive Neural Networks",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "Y"
                        ],
                        "last": "Ng",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "ICML",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Socher, C. Lin, A. Y. Ng, and C.D. Manning. 2011b. Parsing Natural Scenes and Natural Language with Recursive Neural Networks. In ICML.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Semantic Compositionality Through Recursive Matrix-Vector Spaces",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Huval",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "Y"
                        ],
                        "last": "Ng",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Socher, B. Huval, C. D. Manning, and A. Y. Ng. 2012. Semantic Compositionality Through Recur- sive Matrix-Vector Spaces. In EMNLP.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Max-margin parsing",
                "authors": [
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Taskar",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Collins",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Koller",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "1--8",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "B. Taskar, D. Klein, M. Collins, D. Koller, and C. Man- ning. 2004. Max-margin parsing. In Proceedings of EMNLP, pages 1-8.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "Porting statistical parsers with data-defined kernels",
                "authors": [
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Titov",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Henderson",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "I. Titov and J. Henderson. 2006. Porting statistical parsers with data-defined kernels. In CoNLL-X.",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Constituent parsing with incremental sigmoid belief networks",
                "authors": [
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Titov",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Henderson",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "I. Titov and J. Henderson. 2007. Constituent parsing with incremental sigmoid belief networks. In ACL.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "Word representations: a simple and general method for semisupervised learning",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Turian",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Ratinov",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "384--394",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. Turian, L. Ratinov, and Y. Bengio. 2010. Word rep- resentations: a simple and general method for semi- supervised learning. In Proceedings of ACL, pages 384-394.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "From frequency to meaning: Vector space models of semantics",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "D"
                        ],
                        "last": "Turney",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Pantel",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Journal of Artificial Intelligence Research",
                "volume": "37",
                "issue": "",
                "pages": "141--188",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P. D. Turney and P. Pantel. 2010. From frequency to meaning: Vector space models of semantics. Jour- nal of Artificial Intelligence Research, 37:141-188.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF1": {
                "type_str": "figure",
                "num": null,
                "text": "Test sentences of semantic transfer for PP attachments. The CVG was able to transfer semantic word knowledge from two related training sentences. In contrast, the Stanford parser could not distinguish the PP attachments based on the word semantics. Three binary composition matrices",
                "uris": null
            },
            "TABREF1": {
                "num": null,
                "text": "Comparison of parsers with richer state representations on the WSJ. The last line is the self-trained re-ranked Charniak parser.",
                "content": "<table/>",
                "html": null,
                "type_str": "table"
            },
            "TABREF3": {
                "num": null,
                "text": "Detailed comparison of different parsers.",
                "content": "<table><tr><td>the largest sources of improved performance over</td></tr><tr><td>the original Stanford factored parser is in the cor-</td></tr><tr><td>rect placement of PP phrases. When measuring</td></tr><tr><td>only the F1 of parse nodes that include at least one</td></tr><tr><td>PP child, the CVG improves the Stanford parser</td></tr><tr><td>by 6.2% to an F1 of 77.54%. This is a 0.23 re-</td></tr><tr><td>duction in the average number of bracket errors</td></tr><tr><td>per sentence. The 'Other' category includes VP,</td></tr><tr><td>PRN and other attachments, appositives and inter-</td></tr><tr><td>nal structures of modifiers and QPs.</td></tr></table>",
                "html": null,
                "type_str": "table"
            }
        }
    }
}