File size: 73,309 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
{
    "paper_id": "I05-1041",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:25:30.294827Z"
    },
    "title": "Improving Statistical Word Alignment with Ensemble Methods",
    "authors": [
        {
            "first": "Hua",
            "middle": [],
            "last": "Wu",
            "suffix": "",
            "affiliation": {},
            "email": "wuhua@rdc.toshiba.com.cn"
        },
        {
            "first": "Haifeng",
            "middle": [],
            "last": "Wang",
            "suffix": "",
            "affiliation": {},
            "email": "wanghaifeng@rdc.toshiba.com.cn"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This paper proposes an approach to improve statistical word alignment with ensemble methods. Two ensemble methods are investigated: bagging and cross-validation committees. On these two methods, both weighted voting and unweighted voting are compared under the word alignment task. In addition, we analyze the effect of different sizes of training sets on the bagging method. Experimental results indicate that both bagging and cross-validation committees improve the word alignment results regardless of weighted voting or unweighted voting. Weighted voting performs consistently better than unweighted voting on different sizes of training sets.",
    "pdf_parse": {
        "paper_id": "I05-1041",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This paper proposes an approach to improve statistical word alignment with ensemble methods. Two ensemble methods are investigated: bagging and cross-validation committees. On these two methods, both weighted voting and unweighted voting are compared under the word alignment task. In addition, we analyze the effect of different sizes of training sets on the bagging method. Experimental results indicate that both bagging and cross-validation committees improve the word alignment results regardless of weighted voting or unweighted voting. Weighted voting performs consistently better than unweighted voting on different sizes of training sets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Bilingual word alignment is first introduced as an intermediate result in statistical machine translation (SMT) [3] . Besides being used in SMT, it is also used in translation lexicon building [9] , transfer rule learning [10] , example-based machine translation [14] , etc. In previous alignment methods, some researchers employed statistical word alignment models to build alignment links [3] , [4] , [8] , [11] , [16] . Some researchers used similarity and association measures to build alignment links [1] , [15] .",
                "cite_spans": [
                    {
                        "start": 112,
                        "end": 115,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 193,
                        "end": 196,
                        "text": "[9]",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 222,
                        "end": 226,
                        "text": "[10]",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 263,
                        "end": 267,
                        "text": "[14]",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 391,
                        "end": 394,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 397,
                        "end": 400,
                        "text": "[4]",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 403,
                        "end": 406,
                        "text": "[8]",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 409,
                        "end": 413,
                        "text": "[11]",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 416,
                        "end": 420,
                        "text": "[16]",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 506,
                        "end": 509,
                        "text": "[1]",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 512,
                        "end": 516,
                        "text": "[15]",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "One issue about word alignment is how to improve the performance of a word aligner when the training data are fixed. One possible solution is to use ensemble methods [5] , [6] . The ensemble methods were proposed to improve the performance of classifiers. An ensemble of classifiers is a set of classifiers whose individual decisions are combined in some way (weighted or unweighted voting) to classify new examples. Many methods for constructing ensembles have been developed [5] . One kind of methods is to resample the training examples. These methods include bagging [2] , cross-validation committees [12] and boosting [7] . The two former methods generate the classifiers in parallel while boosting generates the classifiers sequentially. In addition, boosting changes the weights of the training instance that is provided as input to each inducer based on the previously built classifiers.",
                "cite_spans": [
                    {
                        "start": 166,
                        "end": 169,
                        "text": "[5]",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 172,
                        "end": 175,
                        "text": "[6]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 477,
                        "end": 480,
                        "text": "[5]",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 571,
                        "end": 574,
                        "text": "[2]",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 605,
                        "end": 609,
                        "text": "[12]",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 623,
                        "end": 626,
                        "text": "[7]",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we propose an approach to improve word alignment with ensemble methods. Although word alignment is not a classification problem, we can still build different word aligners by resampling the training data. If these aligners perform accurately and diversely on the corpus [6] , they can be employed to improve the word alignment results. Here, we investigate two ensemble methods: bagging and cross-validation committees. For both of the ensemble methods, we employ weighted and unweighted voting to build different ensembles. Experimental results indicate that both bagging and cross-validation committees improve the word alignment results. The weighted ensembles perform much better than the unweighted ensembles according to our word alignment results. In addition, we analyze the effect of different sizes of training data on the bagging algorithm. Experimental results also show that the weighted bagging ensembles perform consistently better than the unweighted bagging ensembles on different sizes of training sets.",
                "cite_spans": [
                    {
                        "start": 285,
                        "end": 288,
                        "text": "[6]",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The remainder of the paper is organized as follows. Section 2 describes statistical word alignment. Section 3 describes the bagging algorithm. Section 4 describes the cross-validation committees. Section 5 describes how to calculate the weights used for voting. Section 6 presents the evaluation results. Section 7 discusses why the ensemble methods used in this paper are effective for the word alignment task. The last section concludes this paper and presents the future work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we use the IBM model 4 as our statistical word alignment model [3] . This model only allows word to word and multi-word to word alignments. Thus, some multi-word units cannot be correctly aligned. In order to tackle this problem, we perform word alignment in two directions (source to target and target to source) as described in [11] . In this paper, we call these two aligners bi-directional aligners. 1 Thus, for each sentence pair, we can get two alignment results. We use 1 S and 2 S to represent the bi-directional alignment sets. For alignment links in both sets, we use i for source words and j for target words.",
                "cite_spans": [
                    {
                        "start": 78,
                        "end": 81,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 345,
                        "end": 349,
                        "text": "[11]",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 419,
                        "end": 420,
                        "text": "1",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statistical Word Alignment",
                "sec_num": "2"
            },
            {
                "text": "}} 0 , | { | ) , {( 1 \u2265 = = = j j j j a a i i A j A S (1) }} 0 , | { | ) , {( 2 \u2265 = = = j j i i a i a j A A i S (2)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statistical Word Alignment",
                "sec_num": "2"
            },
            {
                "text": "Where, a j represents the index position of the source word aligned to the target word in position j. For example, if a target word in position j is connected to a source word in position i, then a j =i. If a target word in position j is connected to source words in positions i 1 and i 2 , then A j ={i 1 ,i 2 }. We name an element in the alignment set an alignment link. 2 ",
                "cite_spans": [
                    {
                        "start": 373,
                        "end": 374,
                        "text": "2",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statistical Word Alignment",
                "sec_num": "2"
            },
            {
                "text": "The bagging algorithm (derived from bootstrap aggregating) votes classifiers generated by different bootstrap replicates [2] . A bootstrap replicate is generated by uniformly sampling m instances from the training set with replacement. In general, T bootstrap replicates are built in the sampling process. And T different classifiers are built based on the bootstrap replicates. A final classifier is built from these T subclassifiers using weighted voting or unweighted voting. The original unweighted bagging algorithm is shown in Figure 1 . ",
                "cite_spans": [
                    {
                        "start": 121,
                        "end": 124,
                        "text": "[2]",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 533,
                        "end": 541,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Bagging",
                "sec_num": "3"
            },
            {
                "text": "Input: a training set }} ..., , 1 { ), , {( m i x y S i i \u2208 = an induction algorithm \u03a8 (1) For T j to 1 = { (2) j S =",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging",
                "sec_num": "3"
            },
            {
                "text": "\u2211 \u2208 = j j Y y y x C x C ) ), ( ( max arg ) ( * \u03b4 Where, 1 ) , ( = y x \u03b4 if y x = ; else 0 ) , ( = y x \u03b4",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging",
                "sec_num": "3"
            },
            {
                "text": ".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging",
                "sec_num": "3"
            },
            {
                "text": "Output: Classifier * C Fig. 1 . The Unweighted Bagging Algorithm",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 23,
                        "end": 29,
                        "text": "Fig. 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Bagging",
                "sec_num": "3"
            },
            {
                "text": "In this section, we apply the technique of bagging to word alignment, the detailed algorithm is shown in Figure 2 . In the algorithm, we first resample the training data to train the word aligners. We choose to resample the training set in the same way as the original bagging algorithm. With these different bootstrap replicates, we build the different word aligners. As described in Section 2, we perform word alignment in two directions to improve multiword alignment. Thus, on each bootstrap replicate, we train a word aligner in the source to target direction and another word aligner in the target to source direction, which is described in b) of step (1) .",
                "cite_spans": [
                    {
                        "start": 658,
                        "end": 661,
                        "text": "(1)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 105,
                        "end": 113,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Bagging the Statistical Word Aligner",
                "sec_num": "3.1"
            },
            {
                "text": "After building the different word aligners, we combine or aggregate the alignments generated by the individual alignment models to create the final alignments for each sentence pair. In this paper, the final alignment link for each word is chosen by performing a majority voting on the alignments provided by each instance of the model. The majority voting can be weighted or unweighted. For weighted voting, the weights of word alignment links produced by the bi-directional word aligners are trained from the training data, which will be further described in section 5. For unweighted voting, the best alignment link for a specific word or unit is voted by more than half of the word aligners in the ensemble. For those words that have no majority choice, the system simply does not align them. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging the Statistical Word Aligner",
                "sec_num": "3.1"
            },
            {
                "text": "Input: a training set }} ... 1 { ), , {( m i x y S i i \u2208 = a word alignment model M (1) For",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging the Statistical Word Aligner",
                "sec_num": "3.1"
            },
            {
                "text": "\u2211 + = j ts j st j j t t k s M t k s M t s W k s M )) ), , ( ( ) ), , ( ( ( * ) , ( max arg ) , ( * \u03b4 \u03b4 t is",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging the Statistical Word Aligner",
                "sec_num": "3.1"
            },
            {
                "text": "\u2211 = > + = T j ts j st j T t n t t k s M t k s M k s M 1 2 ) ( : * )) ), , ( ( ) ), , ( ( ( max arg ) , ( \u03b4 \u03b4 where, n(t)= \u2211 = + T j ts j st j t k s M t k s M 1 )) ), , ( ( ) ), , ( ( ( \u03b4 \u03b4",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging the Statistical Word Aligner",
                "sec_num": "3.1"
            },
            {
                "text": "Output: The final word alignment results",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging the Statistical Word Aligner",
                "sec_num": "3.1"
            },
            {
                "text": "The difference between bagging and cross-validation committees lies in the way to resample the training set. The cross-validation committees construct the training sets by leaving out disjoint subsets of the training data. For example, the training set can be randomly and evenly divided into N disjoint subsets. Then N overlapping training sets can be constructed by dropping out a different one of these N subsets. This procedure is the same as the one to construct training sets for N-fold cross-validation. Thus, ensembles constructed in this way are called cross-validation committees.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Fig. 2. The Bagging Algorithm for Word Alignment 4 Cross-Validation Committee",
                "sec_num": null
            },
            {
                "text": "For word alignment, we also divide the training set into N even parts and build N overlapping training sets. With the N sets, we build N alignment models as described above. Since the training sets are different, the word alignment results may be different for individual words. Using the same majority voting as described in Figure 2 , we get the final word alignment results.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 326,
                        "end": 334,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Fig. 2. The Bagging Algorithm for Word Alignment 4 Cross-Validation Committee",
                "sec_num": null
            },
            {
                "text": "In this paper, we compare both weighted voting and unweighted voting under our word alignment task. The algorithm in Figure 2 shows that the weights are related with the specific word alignment links and the specific word aligner. We calculate the weights based on the word alignment results on the training data.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 117,
                        "end": 125,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Weight Calculation",
                "sec_num": "5"
            },
            {
                "text": "As described in Section 3. (3) . This weight measures the association of the source part and the target part in an alignment link. This measure is like the Dice Coefficient. Smadja et al. [13] showed that the Dice Coefficient is a good indicator of translation association. ",
                "cite_spans": [
                    {
                        "start": 27,
                        "end": 30,
                        "text": "(3)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 188,
                        "end": 192,
                        "text": "[13]",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Weight Calculation",
                "sec_num": "5"
            },
            {
                "text": "+ = ' ' ) , ' ( ) ' , ( ) , ( * 2 ) , ( s t i t s count t s count t s count t s W (3)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u2211 \u2211",
                "sec_num": null
            },
            {
                "text": "We perform experiments on a sentence aligned English-Chinese bilingual corpus in general domain. There are about 320,000 bilingual sentence pairs in the corpus, from which, we randomly select 1,000 sentence pairs as testing data. The remainder is used as training data. In the sentence pairs, the average length of the English sentences is 13.6 words while the average length of the Chinese sentences is 14.2 words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training and Testing Set",
                "sec_num": "6.1"
            },
            {
                "text": "The Chinese sentences in both the training set and the testing set are automatically segmented into words. The segmentation errors in the testing set are post-corrected. The testing set is manually annotated. It has totally 8,651 alignment links. Among them, 866 alignment links include multiword units, which accounts for about 10% of the total links.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training and Testing Set",
                "sec_num": "6.1"
            },
            {
                "text": "We use the same evaluation metrics as in [17] . If we use G S to represent the set of alignment links identified by the proposed methods and R S to denote the reference alignment set, the methods to calculate the precision, recall, f-measure, and alignment error rate (AER) are shown in Equation 4, (5) , (6) , and (7). In addition, t-test is used for testing statistical significance. From the evaluation metrics, it can be seen that the higher the f-measure is, the lower the alignment error rate is. Thus, we will only show precision, recall and AER scores in the experimental results.",
                "cite_spans": [
                    {
                        "start": 41,
                        "end": 45,
                        "text": "[17]",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 299,
                        "end": 302,
                        "text": "(5)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 305,
                        "end": 308,
                        "text": "(6)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Metrics",
                "sec_num": "6.2"
            },
            {
                "text": "| S | | S S | G R G \u2229 = precision (4) | S | | S S | R R G \u2229 = recall (5) | | | | | | * 2 R G R G S S S S fmeasure + \u2229 = (6) fmeasure S S S S AER R G R G \u2212 = + \u2229 \u2212 = 1 | | | | | | * 2 1 (7)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Metrics",
                "sec_num": "6.2"
            },
            {
                "text": "For the bagging method, we use ten word aligners trained on five different bootstrap replicates. Among them, five aligners are trained in the source to target direction. The other five aligners are trained in the target to source direction. The bagging method will be compared with a baseline method using the entire training data. For this baseline method, we also train bi-directional models. Based on the alignment results on the entire training data, we calculate the alignment weights for the two word aligners as described in Section 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Results for Bagging",
                "sec_num": "6.3"
            },
            {
                "text": "The results using weighted voting are shown in Table 1 . The number in brackets of the first column describes the number of word aligners used in the ensembles. For example, in the ensemble \"bagging (4)\", two word aligners are trained in the source to target direction and the other two are trained in the target to source direction.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 47,
                        "end": 54,
                        "text": "Table 1",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation Results for Bagging",
                "sec_num": "6.3"
            },
            {
                "text": "From the results, it can be seen that the bagging methods obtain significantly better results than the baseline. The best ensemble achieves an error rate reduction of 7.34% as compared with the baseline. The results show that increasing the number of word aligner does not greatly reduce the word alignment error rate. The reduction is even smaller when the number increases from 8 to 10. In order to further analyze the effect of the weights on the word alignment results, we also use unweighted voting in the ensembles. The results are shown in Table 2 . The baseline method also trains bi-directional aligners with the entire training data. The final word alignment results are obtained by taking an unweighted voting on the two alignment results produced by the bi-directional aligners. That is the same as that by taking the intersection of the two word alignment results. Increasing the number of word aligners in the ensembles, the unweighted bagging method does not greatly reduce AER. However, the ensembles obtain much lower error rate as compared with the baseline. The best ensemble achieves a relative error rate reduction of 10.64%, indicating a significant improvement. From the experimental results, we find that there are no multiword alignment links selected in the ensembles. This is because unweighted voting in this paper requires more than half of the word aligners in the ensembles to vote for the same link. Thus, there should be bidirectional word aligners voting for the target alignment link. The intersection of bidirectional word alignment results produced by the IBM models only creates single word alignments. It can also be seen from the Equations (1) and (2) in Section 2.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 547,
                        "end": 554,
                        "text": "Table 2",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation Results for Bagging",
                "sec_num": "6.3"
            },
            {
                "text": "Comparing the results obtained using weighted voting in Table 1 and those obtained using unweighted voting in Table 2 , we find that (1) the weighted bagging methods are much better than the unweighted bagging methods; (2) the ensembles using unweighted voting obtain higher precision but lower recall than those using weighted voting. For example, the weighted voting \"bagging (10)\" achieves a relative error rate reduction of 20.59% as compared with the corresponding unweighted voting. This indicates that the method used to calculate voting weights described in section 5 is very effective.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 56,
                        "end": 63,
                        "text": "Table 1",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 110,
                        "end": 117,
                        "text": "Table 2",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation Results for Bagging",
                "sec_num": "6.3"
            },
            {
                "text": "For the cross-validation committees, we divide the entire training data into five disjoint subsets. For each bootstrap replicate, we leave one out. Thus, each replicate includes 80% sentence pairs of the full training data. For each replicate, we train bidirectional word alignment models. Thus, we totally obtain ten individual word aligners. The baseline is the same as shown in Table 1 . The results obtained using weighted voting are shown in Table 3 . The number in the brackets of the first column describes the number of word aligners used in the ensembles. From the results, it can be seen that the cross-validation committees perform better than the baseline. The best ensemble \"validation (10)\" achieves an error rate reduction of 7.39% as compared with the baseline, indicating a significant improvement. The results also show that increasing the number of word aligner does not greatly reduce the word alignment error rate.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 381,
                        "end": 388,
                        "text": "Table 1",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 447,
                        "end": 454,
                        "text": "Table 3",
                        "ref_id": "TABREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation Results for Cross-Validation Committees",
                "sec_num": "6.4"
            },
            {
                "text": "As described in section 6.3, we also use unweighted voting for the cross-validation committees. The results are shown in Table 4 . The baseline is the same as described in Table 2 . From the results, it can be seen that increasing the number of word aligners in the ensembles, the alignment error rate is reduced. The best ensemble achieves a relative error rate reduction of 8.04% as compared with the baseline, indicating a significant improvement. Comparing the results in Table 3 and Table 4 , we find that the weighted methods are also much better than the unweighted ones. For example, the weighted method \"Validation (10)\" achieves an error rate reduction of 22.87% as compared with the corresponding unweighted method.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 121,
                        "end": 128,
                        "text": "Table 4",
                        "ref_id": "TABREF5"
                    },
                    {
                        "start": 172,
                        "end": 179,
                        "text": "Table 2",
                        "ref_id": "TABREF3"
                    },
                    {
                        "start": 476,
                        "end": 483,
                        "text": "Table 3",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 488,
                        "end": 495,
                        "text": "Table 4",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation Results for Cross-Validation Committees",
                "sec_num": "6.4"
            },
            {
                "text": "According to the evaluation results, bagging and cross-validation committees achieve comparable results. In order to further compare bagging and cross-validation committees, we classify the alignment links in the weighted ensembles into two classes: single word alignment links (SWA) and multiword alignment links (MWA). SWA links only include one-to-one alignments. MWA links refer to those including multiword units in the source language or/and in the target language. The SWA and MWA for the bagging ensembles are shown in Table 5 and Table 6 . The SWA and MWA for the cross-validation committees are shown in Table 7 and Table 8 . The AERs of the baselines for SWA and MWA are 0.1531 and 0.8469, respectively. From the results, it can be seen that the single word alignment results are much better than the multiword alignment results for both of the two methods. This indicates that it is more difficult to align the multiword units than to align single words.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 527,
                        "end": 546,
                        "text": "Table 5 and Table 6",
                        "ref_id": null
                    },
                    {
                        "start": 614,
                        "end": 621,
                        "text": "Table 7",
                        "ref_id": null
                    },
                    {
                        "start": 626,
                        "end": 633,
                        "text": "Table 8",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Bagging vs. Cross-Validation Committees",
                "sec_num": "6.5"
            },
            {
                "text": "Comparing the bagging methods and validation committees, we find that these two methods obtain comparable results on both the single word alignment links and multiword alignment links. This indicates that the different resampling methods in these two ensemble methods do not much affect the results on our word alignment task.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Bagging vs. Cross-Validation Committees",
                "sec_num": "6.5"
            },
            {
                "text": "In this section, we investigate the effect of the size of training data on the ensemble methods. Since the difference between bagging and cross-validation committees is very small, we only investigate the effect on the bagging ensembles.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Different Sizes of Training Data",
                "sec_num": "6.6"
            },
            {
                "text": "We randomly select training data from the original training set described in Section 6.1 to construct different training sets. We construct three training sets, which include 1/4, 1/2 and 3/4 of sentence pairs of the original training set, respectively.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Different Sizes of Training Data",
                "sec_num": "6.6"
            },
            {
                "text": "For each of the training set, we obtain five bootstrap replicates and train ten word aligners. The results of ensembles consisting of ten word aligners are shown in Table  9 and Table 10 . Table 9 and Table 10 show the weighted and unweighted bagging results, respectively. The methods to construct the baselines for different training sets in Table 9 and Table 10 are the same as those in Table 1 and Table 2 , respectively. For convenience, we also list the results using the original training set in the tables. The first column describes the size of the training sets used for the ensembles. The last column presents the relative error rate reduction (RERR) of the ensembles as compared with the corresponding baselines. From the results, it can be seen that both weighted and unweighted bagging ensembles are effective to improve word alignment results. The weighted ensembles perform consistently better than the unweigted ensembles on different sizes of training sets. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 165,
                        "end": 173,
                        "text": "Table  9",
                        "ref_id": null
                    },
                    {
                        "start": 178,
                        "end": 186,
                        "text": "Table 10",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 189,
                        "end": 209,
                        "text": "Table 9 and Table 10",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 344,
                        "end": 364,
                        "text": "Table 9 and Table 10",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 390,
                        "end": 409,
                        "text": "Table 1 and Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Different Sizes of Training Data",
                "sec_num": "6.6"
            },
            {
                "text": "Both bagging and cross-validation committees utilize multiple classifiers to make different assumptions about the learning system. Bagging requires that the learning system should not be stable, so that small changes to the training set would lead to different classifiers. Breiman [2] also noted that poor predicators could be transformed into worse ones by bagging. In this paper, the learning system is the word alignment model described in Section 2. The classifiers refer to the different word aligners trained on different bootstrap replicates. In our experiments, although word alignment models do not belong to unstable learning systems, bagging obtains better results on all of the datasets. This is because the training data is insufficient or subject to data sparseness problem. Thus, changing the training data or resampling the training data causes the alternation of the trained parameters of the alignment model. The word aligners trained on a different bootstrap replicate produce different word alignment links for individual words. Using majority voting, the ensembles can improve the alignment precision and recall, resulting in lower alignment error rates.",
                "cite_spans": [
                    {
                        "start": 282,
                        "end": 285,
                        "text": "[2]",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "7"
            },
            {
                "text": "The experiments also show that weighted voting is better than unweighted voting. The advantage of weighted voting is that it can select the good word alignment link even if only one aligner votes for it in the ensembles. This is because the selected alignment link gets much higher weight than the other links.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "7"
            },
            {
                "text": "Two ensemble methods are employed in this paper to improve word alignment results: bagging and cross-validation committees. Both of these two methods obtain better results than the original word aligner without increasing any training data. In this paper, we use two different voting methods: weighted voting and unweighted voting. Experimental results show that the weighted bagging method and weighted cross-validation committees achieve an error rate reduction of 7.34% and 7.39% respectively, as compared with the original word aligner. Results also show that weighted voting is much better than unweighted voting on the word alignment task. Unweighted voting obtains higher precision but lower recall than weighted voting. In addition, the weighted voting used in this paper obtains multiword alignment links while the unweighted voting cannot.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "8"
            },
            {
                "text": "We also compare the two ensemble methods on the same training data and testing data. Bagging and cross-validation committees obtain comparable results on both single word alignment links and multiword alignment links. This indicates that the different resampling methods in these two ensemble methods do not much affect the results under our word alignment task.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "8"
            },
            {
                "text": "We also investigate the bagging method on different sizes of training sets. The results show that both weighted voting and unweighted voting are effective to improve word alignment results. Weighted voting performs consistently better than unweigted voting on different sizes of training sets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "8"
            },
            {
                "text": "In future work, we will investigate more ensemble methods on the word alignment task such as the boosting algorithm. In addition, we will do more research on the weighting schemes in voting.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "8"
            },
            {
                "text": "The GIZA++ toolkit is used to perform statistical alignment. It is located at http://www.fjoch.com/GIZA++.html.2 Our definition of alignment link is different from that in[11]. In[11], alignment links are classified into possible links and sure links. In our paper, both one-to-one and non one-to-one links are taken as sure links.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "A Simple Hybrid Aligner for Generating Lexical Correspondences in Parallel Texts",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Ahrenberg",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Merkel",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Andersson",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proc. of the 36 th Annual Meeting of the Association for Computational Linguistics and the 17 th Int. Conf. on Computational Linguistics (ACL/COLING-1998)",
                "volume": "",
                "issue": "",
                "pages": "29--35",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ahrenberg, L., Merkel, M., Andersson, M.: A Simple Hybrid Aligner for Generating Lexi- cal Correspondences in Parallel Texts. In Proc. of the 36 th Annual Meeting of the Associa- tion for Computational Linguistics and the 17 th Int. Conf. on Computational Linguistics (ACL/COLING-1998), 29-35",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Bagging Predicators",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Breiman",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Machine Learning",
                "volume": "24",
                "issue": "1",
                "pages": "123--140",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Breiman, L.: Bagging Predicators. Machine Learning (1996), 24(1): 123-140",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "The Mathematics of Statistical Machine Translation: Parameter Estimation. Computational Linguistics",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "F"
                        ],
                        "last": "Brown",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [
                            "D"
                        ],
                        "last": "Pietra",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [
                            "D"
                        ],
                        "last": "Pietra",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Mercer",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "",
                "volume": "19",
                "issue": "",
                "pages": "263--311",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Brown, P. F., Pietra, S. D., Pietra, V. D., Mercer, R.: The Mathematics of Statistical Ma- chine Translation: Parameter Estimation. Computational Linguistics (1993), 19(2): 263-311",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "A Probability Model to Improve Word Alignment",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Cherry",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proc. of the 41 st Annual Meeting of the Association for Computational Linguistics (ACL-2003)",
                "volume": "",
                "issue": "",
                "pages": "88--95",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Cherry, C., Lin, D.: A Probability Model to Improve Word Alignment. In Proc. of the 41 st Annual Meeting of the Association for Computational Linguistics (ACL-2003), pp. 88-95",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Machine Learning Research: Four Current Directions. AI Magazine (1997)",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Dietterich",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "18",
                "issue": "",
                "pages": "97--136",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dietterich, T.: Machine Learning Research: Four Current Directions. AI Magazine (1997), 18 (4): 97-136",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Ensemble Methods in Machine Learning",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Dietterich",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proc. of the First Int. Workshop on Multiple Classifier Systems",
                "volume": "",
                "issue": "",
                "pages": "1--15",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dietterich, T.: Ensemble Methods in Machine Learning. In Proc. of the First Int. Work- shop on Multiple Classifier Systems (2000), 1-15",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Experiments with a new boosting algorithm",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Freund",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Schapire",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Machine Learning: Proc. of the Thirteenth International Conference",
                "volume": "",
                "issue": "",
                "pages": "148--156",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Freund, Y., Schapire, R.: Experiments with a new boosting algorithm. In Machine Learn- ing: Proc. of the Thirteenth International Conference (1996), 148-156",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Symmetric Word Alignments for Statistical Machine Translation",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Matusov",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Zens",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Ney",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proc. of the 20 th Int. Conf. on Computational Linguistics (COLING-2004)",
                "volume": "",
                "issue": "",
                "pages": "219--225",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matusov, E., Zens, R., Ney H.: Symmetric Word Alignments for Statistical Machine Translation. In Proc. of the 20 th Int. Conf. on Computational Linguistics (COLING-2004), 219-225",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Automatic Construction of Clean Broad-Coverage Translation Lexicons",
                "authors": [
                    {
                        "first": "I",
                        "middle": [
                            "D"
                        ],
                        "last": "Melamed",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proc. of the 2 nd Conf. of the Association for Machine Translation in the Americas (AMTA-1996)",
                "volume": "",
                "issue": "",
                "pages": "125--134",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Melamed, I. D.: Automatic Construction of Clean Broad-Coverage Translation Lexicons. In Proc. of the 2 nd Conf. of the Association for Machine Translation in the Americas (AMTA-1996), 125-134",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "A Best-first Alignment Algorithm for Automatic Extraction of Transfer Mappings from Bilingual Corpora",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Menezes",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [
                            "D"
                        ],
                        "last": "Richardson",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "Proc. of the ACL 2001 Workshop on Data-Driven Methods in Machine Translation",
                "volume": "",
                "issue": "",
                "pages": "39--46",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Menezes, A., Richardson, S.D.: A Best-first Alignment Algorithm for Automatic Extrac- tion of Transfer Mappings from Bilingual Corpora. In Proc. of the ACL 2001 Workshop on Data-Driven Methods in Machine Translation (2001), 39-46",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Improved Statistical Alignment Models",
                "authors": [
                    {
                        "first": "F",
                        "middle": [
                            "J"
                        ],
                        "last": "Och",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Ney",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proc. of the 38 th Annual Meeting of the Association for Computational Linguistics (ACL-2000)",
                "volume": "",
                "issue": "",
                "pages": "440--447",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Och, F. J., Ney, H.: Improved Statistical Alignment Models. In Proc. of the 38 th Annual Meeting of the Association for Computational Linguistics (ACL-2000), 440-447",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Improving Committee Diagnosis with Resampling Techniques",
                "authors": [
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Parmanto",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Munro",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Doyle",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "8",
                "issue": "",
                "pages": "882--888",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Parmanto, B., Munro, P., Doyle, H.: Improving Committee Diagnosis with Resampling Techniques. In Touretzky, D., Mozer, M., Hasselmo, M. (Ed..): Advances in Neural In- formation Processing Systems (1996), Vol. 8, 882-888",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Translating Collocations for Bilingual Lexicons: a Statistical Approach. Computational Linguistics",
                "authors": [
                    {
                        "first": "F",
                        "middle": [
                            "A"
                        ],
                        "last": "Smadja",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [
                            "R"
                        ],
                        "last": "Mckeown",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Hatzivassiloglou",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "",
                "volume": "22",
                "issue": "",
                "pages": "1--38",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Smadja, F. A., McKeown, K. R., Hatzivassiloglou, V.: Translating Collocations for Bilingual Lexicons: a Statistical Approach. Computational Linguistics (1996), 22 (1):1-38",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Review Article: Example-Based Machine Translation",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Somers",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Machine Translation",
                "volume": "14",
                "issue": "",
                "pages": "113--157",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Somers, H.: Review Article: Example-Based Machine Translation. Machine Translation (1999), 14: 113-157",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Lexical Token Alignment: Experiments, Results and Application",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Tufis",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Barbu",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Proc. of the 3 rd Int. Conf. on Language Resources and Evaluation (LREC-2002",
                "volume": "",
                "issue": "",
                "pages": "458--465",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tufis, D., Barbu, M.: Lexical Token Alignment: Experiments, Results and Application. In Proc. of the 3 rd Int. Conf. on Language Resources and Evaluation (LREC-2002), 458-465",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Stochastic Inversion Transduction Grammars and Bilingual Parsing of Parallel Corpora",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Computational Linguistics",
                "volume": "23",
                "issue": "3",
                "pages": "377--403",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wu, D.: Stochastic Inversion Transduction Grammars and Bilingual Parsing of Parallel Corpora. Computational Linguistics (1997), 23(3): 377-403",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Improving Domain-Specific Word Alignment with a General Bilingual Corpus",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Machine Translation: From Real Users to Research: 6 th Conf. of the Association for Machine Translation in the Americas",
                "volume": "",
                "issue": "",
                "pages": "262--271",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wu, H., Wang, H.: Improving Domain-Specific Word Alignment with a General Bilingual Corpus. In Frederking R., Taylor, K. (Eds.): Machine Translation: From Real Users to Re- search: 6 th Conf. of the Association for Machine Translation in the Americas (AMTA- 2004), 262-271",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "num": null,
                "text": "replicate of S by sampling m items from S with replacement b) Train the bi-directional alignment models the number of sentence pairs) For each word s : a) For weighted voting",
                "uris": null
            },
            "FIGREF1": {
                "type_str": "figure",
                "num": null,
                "text": "the word or phrase in the target sentence;",
                "uris": null
            },
            "TABREF1": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td colspan=\"4\">st j M in the source to target direction and a word aligner ts j M in the target to source</td></tr><tr><td colspan=\"4\">direction. That is to say, we obtain two different word alignment sets st j S and ts j S for</td></tr><tr><td>each of the bootstrap replicate. For each word alignment link</td><td>, ( t s</td><td>)</td><td>produced by st j M</td></tr><tr><td>or ts</td><td/><td/><td/></tr></table>",
                "text": "1, on each bootstrap replicate j, we train a word aligner"
            },
            "TABREF2": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Bagging (4)</td><td>0.8035</td><td>0.7898</td><td>0.2034</td></tr><tr><td>Bagging (6)</td><td>0.8048</td><td>0.7922</td><td>0.2015</td></tr><tr><td>Bagging (8)</td><td>0.8061</td><td>0.7948</td><td>0.1996</td></tr><tr><td>Bagging (10)</td><td>0.8064</td><td>0.7948</td><td>0.1994</td></tr><tr><td>Baseline</td><td>0.7870</td><td>0.7826</td><td>0.2152</td></tr></table>",
                "text": "Weighted Bagging Results"
            },
            "TABREF3": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Bagging (4)</td><td>0.9230</td><td>0.6073</td><td>0.2674</td></tr><tr><td>Bagging (6)</td><td>0.9181</td><td>0.6200</td><td>0.2598</td></tr><tr><td>Bagging (8)</td><td>0.9167</td><td>0.6307</td><td>0.2527</td></tr><tr><td>Bagging (10)</td><td>0.9132</td><td>0.6347</td><td>0.2511</td></tr><tr><td>Baseline</td><td>0.9294</td><td>0.5756</td><td>0.2810</td></tr></table>",
                "text": "Unweighted Bagging Results"
            },
            "TABREF4": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Validation (4)</td><td>0.8059</td><td>0.7913</td><td>0.2015</td></tr><tr><td>Validation (6)</td><td>0.8070</td><td>0.7928</td><td>0.2002</td></tr><tr><td>Validation (8)</td><td>0.8063</td><td>0.7933</td><td>0.2002</td></tr><tr><td>Validation (10)</td><td>0.8068</td><td>0.7947</td><td>0.1993</td></tr><tr><td>Baseline</td><td>0.7870</td><td>0.7826</td><td>0.2152</td></tr></table>",
                "text": "Evaluation Results for Weighted Cross-Validation Committees"
            },
            "TABREF5": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Validation (4)</td><td>0.9199</td><td>0.5943</td><td>0.2779</td></tr><tr><td>Validation (6)</td><td>0.9174</td><td>0.6124</td><td>0.2655</td></tr><tr><td>Validation (8)</td><td>0.9154</td><td>0.6196</td><td>0.2610</td></tr><tr><td>Validation (10)</td><td>0.9127</td><td>0.6245</td><td>0.2584</td></tr><tr><td>Baseline</td><td>0.9294</td><td>0.5756</td><td>0.2810</td></tr></table>",
                "text": "Evaluation Results for Unweighted Cross-Validation Committees"
            },
            "TABREF6": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Bagging (4)</td><td>0.8263</td><td>0.8829</td><td>0.1463</td></tr><tr><td>Bagging (6)</td><td>0.8270</td><td>0.8845</td><td>0.1452</td></tr><tr><td>Bagging (8)</td><td>0.8270</td><td>0.8877</td><td>0.1437</td></tr><tr><td>Bagging (10)</td><td>0.8265</td><td>0.8876</td><td>0.1440</td></tr><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Bagging (4)</td><td>0.4278</td><td>0.1815</td><td>0.7451</td></tr><tr><td>Bagging (6)</td><td>0.4432</td><td>0.1896</td><td>0.7344</td></tr><tr><td>Bagging (8)</td><td>0.4540</td><td>0.1884</td><td>0.7336</td></tr><tr><td>Bagging (10)</td><td>0.4620</td><td>0.1896</td><td>0.7311</td></tr><tr><td colspan=\"4\">Table 7. Single Word Alignment Results for Weighted Cross-Validation Committees</td></tr><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Validation (4)</td><td>0.8282</td><td>0.8833</td><td>0.1452</td></tr><tr><td>Validation (6)</td><td>0.8285</td><td>0.8847</td><td>0.1443</td></tr><tr><td>Validation (8)</td><td>0.8275</td><td>0.8851</td><td>0.1447</td></tr><tr><td>Validation (10)</td><td>0.8277</td><td>0.8867</td><td>0.1438</td></tr><tr><td colspan=\"4\">Table 8. Multiword Alignment Results for Weighted Cross-Validation Committees</td></tr><tr><td>Method</td><td>Precision</td><td>Recall</td><td>AER</td></tr><tr><td>Validation (4)</td><td>0.4447</td><td>0.1908</td><td>0.7330</td></tr><tr><td>Validation (6)</td><td>0.4538</td><td>0.1931</td><td>0.7291</td></tr><tr><td>Validation (8)</td><td>0.4578</td><td>0.1942</td><td>0.7273</td></tr><tr><td>Validation (10)</td><td>0.4603</td><td>0.1942</td><td>0.7268</td></tr></table>",
                "text": "Single Word Alignment Results for the Weighted Bagging Methods Multiword Alignment Results for the Weighted Bagging Methods"
            },
            "TABREF7": {
                "html": null,
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Data</td><td>Precision</td><td>Recall</td><td>AER</td><td>Baseline (AER)</td><td>RERR</td></tr><tr><td>1/4</td><td>0.7684</td><td>0.7517</td><td>0.2316</td><td>0.2464</td><td>6.00%</td></tr><tr><td>1/2</td><td>0.7977</td><td>0.7775</td><td>0.2125</td><td>0.2293</td><td>7.33%</td></tr><tr><td>3/4</td><td>0.8023</td><td>0.7869</td><td>0.2055</td><td>0.2184</td><td>5.89%</td></tr><tr><td>All</td><td>0.8064</td><td>0.7948</td><td>0.1994</td><td>0.2152</td><td>7.34%</td></tr><tr><td>Data</td><td>Precision</td><td>Recall</td><td>AER</td><td>Baseline (AER)</td><td>RERR</td></tr><tr><td>1/4</td><td>0.8960</td><td>0.6033</td><td>0.2789</td><td>0.3310</td><td>15.72%</td></tr><tr><td>1/2</td><td>0.9077</td><td>0.6158</td><td>0.2662</td><td>0.3050</td><td>12.72%</td></tr><tr><td>3/4</td><td>0.9140</td><td>0.6270</td><td>0.2562</td><td>0.2943</td><td>12.95%</td></tr><tr><td>All</td><td>0.9132</td><td>0.6347</td><td>0.2511</td><td>0.2810</td><td>10.64%</td></tr></table>",
                "text": "Weighted Bagging Results on Different Sizes of Training Sets Unweighted Bagging Results on Different Sizes of Training Sets"
            }
        }
    }
}