File size: 77,587 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
{
    "paper_id": "D09-1024",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T16:39:26.140570Z"
    },
    "title": "Improved Word Alignment with Statistics and Linguistic Heuristics",
    "authors": [
        {
            "first": "Ulf",
            "middle": [],
            "last": "Hermjakob",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Southern California Information Sciences Institute",
                "location": {
                    "addrLine": "4676 Admiralty Way Marina del Rey",
                    "postCode": "90292",
                    "region": "CA",
                    "country": "USA"
                }
            },
            "email": "ulf@isi.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "We present a method to align words in a bitext that combines elements of a traditional statistical approach with linguistic knowledge. We demonstrate this approach for Arabic-English, using an alignment lexicon produced by a statistical word aligner, as well as linguistic resources ranging from an English parser to heuristic alignment rules for function words. These linguistic heuristics have been generalized from a development corpus of 100 parallel sentences. Our aligner, UALIGN, outperforms both the commonly used GIZA++ aligner and the state-of-theart LEAF aligner on F-measure and produces superior scores in end-to-end statistical machine translation, +1.3 BLEU points over GIZA++, and +0.7 over LEAF.",
    "pdf_parse": {
        "paper_id": "D09-1024",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "We present a method to align words in a bitext that combines elements of a traditional statistical approach with linguistic knowledge. We demonstrate this approach for Arabic-English, using an alignment lexicon produced by a statistical word aligner, as well as linguistic resources ranging from an English parser to heuristic alignment rules for function words. These linguistic heuristics have been generalized from a development corpus of 100 parallel sentences. Our aligner, UALIGN, outperforms both the commonly used GIZA++ aligner and the state-of-theart LEAF aligner on F-measure and produces superior scores in end-to-end statistical machine translation, +1.3 BLEU points over GIZA++, and +0.7 over LEAF.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Word alignment is a critical component in training statistical machine translation systems and has received a significant amount of research, for example, (Brown et al., 1993; Ittycheriah and Roukos, 2005; , including work leveraging syntactic parse trees, e.g., (Cherry and Lin, 2006; DeNero and Klein, 2007; Fossum et al., 2008) . Word alignment is also a required first step in other algorithms such as for learning sub-sentential phrase pairs (Lavie et al., 2008) or the generation of parallel treebanks (Zhechev and Way, 2002 ).",
                "cite_spans": [
                    {
                        "start": 155,
                        "end": 175,
                        "text": "(Brown et al., 1993;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 176,
                        "end": 205,
                        "text": "Ittycheriah and Roukos, 2005;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 263,
                        "end": 285,
                        "text": "(Cherry and Lin, 2006;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 286,
                        "end": 309,
                        "text": "DeNero and Klein, 2007;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 310,
                        "end": 330,
                        "text": "Fossum et al., 2008)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 447,
                        "end": 467,
                        "text": "(Lavie et al., 2008)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 508,
                        "end": 530,
                        "text": "(Zhechev and Way, 2002",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Yet word alignment precision remains surprisingly low, under 80% for state-of-the-art aligners on not closely related language pairs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Consider the following Arabic/English sentence pair with alignments built by the statistical word aligner LEAF:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Bitext Arabic: A A R K Q S P X @ P A K . Y K A K A J @ P A J R J @ , 6 -4 6 -4 Q J . J J JS R K A k . @ Q JS A @ 6 -4 6 -4 Q R P K . PP A K A A @ J K A Q J K Gloss: Won(1) Thai Paradorn Srichaphan(1) on/to(2) Australian Jason(2) Stoltenberg(3) 6(4) -4(5) and(3) 6(4) -4(5), and Czech Ji\u0159\u00ed(7) Van\u011bk(7) on/to German(6) Lars Burgsm\u00fcller 6(4) -4 and(3) 6(4) -4 Bitext English: Thailand 's(1) Baradorn Srichfan(1) beat(2) Australian Gayson(1) Stultenberg(3) 6(4) -6(4) 6(4) -4(5) , Czech player(1) Pierre(1) Vanic(7) beat(6) Germany(6) 's Lars Burgsmuller 6 -4 6 -4",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In the example above, words with the same index in the gloss for Arabic and the English are aligned to each other, alignment errors are underlined, translation errors are in italics. For example, the Arabic words for won and Srichaphan are aligned with the English words 's, Srichfan, Gayson, player and Pierre. As reflected in the example above, typical alignment problems include words that change sentence position between languages, such as verbs, which in Arabic are often sentence-initial (e.g. won/beat in the example above) function words without a clear and explicit equivalent in the other language (e.g. the Arabic /and in the example above) lack of robustness with respect to poor translations (e.g. Gayson Stultenberg instead of Jason Stoltenberg) or bad sentence alignment.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We believe we can overcome such problems with the increased use of linguistically based heuristics. We can model typical word order differences between English and Arabic using English parse trees and a few Arabic-specific phrase reordering heuristics. We can narrow the space of possible alignment candidates for function words using English parse trees and a few heuristics for each type of function word.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "These heuristics have been developed using a development corpus of 100 parallel sentences. The heuristics are generalizations based on patterns of misaligned words, misaligned with respect to a Gold Standard alignment for that development corpus.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The following sections describe how our word aligner works, first how relatively reliable content words are aligned, and then how function words and any remaining content words are aligned, with a brief discussion of an interesting issue relating to the Gold Standard we used. Finally we present evaluations on word alignment accuracy as well as the impact on end-to-end machine translation quality.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We divide the alignment process into two phases: first, we align relatively reliable content words, which in phase II we then use as a skeleton to align function words and remaining content words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase I: Content Words",
                "sec_num": "2"
            },
            {
                "text": "Function words such as English a, ah, all, am, an, and, any, are, as, at, .. . are common words that often do not have an explicit equivalent word or words in the other side of the bitext. In our system, we use a list of 96 English and 110 Arabic function words with those characteristics. For the purposes of our algorithm, a word is a function word if and only if it is on the function word list for its language. A content word then is defined as a word that is neither a function word nor punctuation.",
                "cite_spans": [
                    {
                        "start": 23,
                        "end": 76,
                        "text": "English a, ah, all, am, an, and, any, are, as, at, ..",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase I: Content Words",
                "sec_num": "2"
            },
            {
                "text": "The approach for aligning content words in phase I is as follows: First, we score each combination of an Arabic content word and English content word in an aligned sentence and align those pairs that pass a threshold, typically generating too many alignments. Second, we compute a more comprehensive score that also takes into consideration matching alignments in the context around each alignment. Third, we eliminate inferior alignments that are incompatible with higherscoring alignments.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase I: Content Words",
                "sec_num": "2"
            },
            {
                "text": "The score in the first step is pointwise mutual information (PMI). The key resource to compute this PMI is an alignment lexicon generated beforehand by a statistical word alignment system from a large bitext. An alignment lexicon is a list of triples, each consisting of an English word, an Arabic word, and how often they have been aligned for a given bitext. Additional counts on how often each English and Arabic word occurs allow us use this alignment lexicon to compute PMI(e,f) = log pe;f pepf . We align those Arabic and English content words that have a PMI 0 and a minimum alignment lexicon count ( 10 initially). Using the alignment lexicon generated by a statistical word aligner to compute PMIs is the principal statistical component in our system. We explored alternative metrics such as the dicecoefficient that was used by other researchers in earlier alignment work, but found PMI to work better for our system.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase I: Content Words",
                "sec_num": "2"
            },
            {
                "text": "In a second step, we lay a window of size 5 around each aligned pair of Arabic and English words (counting only content words) and then add to the PMI score of the link itself the PMI scores of other links within that window, with a distance weight of 1 distance+1 . This yields a new score that takes into account whether a link is supported by context.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase I: Content Words",
                "sec_num": "2"
            },
            {
                "text": "In the third step, we check for overgenerated links, comparing links that share an Arabic or an English word. If a word on one side of the bitext is linked to multiple adjacent words on the other, we leave them alone, as one word in one language often corresponds to multiple words in the other. However, if a word on one side is linked to nonadjacent words on the other side, this flags an incompatibility, and we remove those links that have inferior context-sensitive scores. This removal is done one link at a time, with the lowest relative scores first.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase I: Content Words",
                "sec_num": "2"
            },
            {
                "text": "We boost the process we just described in a few ways. In the first alignment step, we also include as alignment candidates any content words that are string-identical on each side, such as ASCII numbers and ASCII words. We finally also include as alignment candidates those word pairs that are transliterations of each other to cover rare proper names (Hermjakob et al., 2008) , which is important for language pairs that don't share the same alphabet such as Arabic and English.",
                "cite_spans": [
                    {
                        "start": 352,
                        "end": 376,
                        "text": "(Hermjakob et al., 2008)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase I: Content Words",
                "sec_num": "2"
            },
            {
                "text": "We use a refined notion of context window that models word order differences between Arabic and English. Traversing a parse tree for English, we identify sub-trees for which the order in Arabic can be substantially different. In Arabic, for example, the verb is often sentence-initial. So for trees or subtrees identified by the parser as sentences, we generate an alternative reordering of its subtrees where the verb has been moved to the front. Similarly, in a noun phrase, we generate an alternative order where adjectives are moved to the right of the noun they modify.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reordering Using an English Parser",
                "sec_num": "2.1"
            },
            {
                "text": "For example, consider the sentence John bought a new car . We can reorder its parse tree both at the sentence level: (bought) (John) (a new car) (.) as well as at its object NP level: (a) (car) (new). If fully enumerated, this would yield these four reordering alternatives:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reordering Using an English Parser",
                "sec_num": "2.1"
            },
            {
                "text": "1. John bought a new car .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reordering Using an English Parser",
                "sec_num": "2.1"
            },
            {
                "text": "2. John bought a car new .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reordering Using an English Parser",
                "sec_num": "2.1"
            },
            {
                "text": "3. bought John a new car . 4. bought John a car new . We don't actually explicitly enumerate all variants but keep all reordering alternatives in a reordering forest, since the number of fully expanded reorderings grows exponentially with the number of phrases with reordering(s). At the beginning of Phase I, we compute from this reordering forest a minimum distance matrix, which, for specific instances of the words John and car would record a minimum distance of 1 (based on reordering 4, skipping the function word a).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reordering Using an English Parser",
                "sec_num": "2.1"
            },
            {
                "text": "For the example sentence at the beginning of the paper we would get reorderings including the following: Engl. orig.: thailand 's baradorn srichfan beat ... A reordering: beat thailand 's baradorn srichfan ... Arabic (gloss): won thai paradorn srichaphan ...",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reordering Using an English Parser",
                "sec_num": "2.1"
            },
            {
                "text": "In the above reordered English alternative, beat and thailand are next to each other, so their minimum distance is 1, which means that a link between English thailand and Arabic thai now strongly boosts the context-sensitive score between English beat and Arabic won.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Reordering Using an English Parser",
                "sec_num": "2.1"
            },
            {
                "text": "Another challenge to content word alignment is morphological variation which can create data sparsity in the alignment lexicon. For example, in a given bitext sentence, the Arabic word AlAw-DAE might be translated as situational, for which there might be no support in the alignment lexicon. However the PMI between AlAwDAE and situation might be sufficiently high. Additionally, there is another Arabic word, AlHAlAt, which often translates as both situation and situational.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Variation",
                "sec_num": "2.2"
            },
            {
                "text": "To take advantage of such constellations, we built morphological variation lists for both Arabic and English, lists that for a given head word such as situational lists variants such as situation, and situations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Variation",
                "sec_num": "2.2"
            },
            {
                "text": "We built these lists in a one-time process by identifying superficially similar words, i.e. those that vary only with respect to an ending or a prefix, and then semantically validating such candidates using a pivot word in the other language such as AlHAlAt that has sufficiently strong alignment lexicon co-alignment counts with both situation and situational. The alignment lexicon co-alignment count of an Arabic word w ar and an English word w en is considered strong enough, if it is at least 2.0 and at least 0.001 times as high as the highest co-alignment count of w ar with any English word; words shorter than four letters are excluded from consideration. So because situation and situational are superficially similar and they are both have a strong alignment count with AlHAlAt in the alignment lexicon, situation is added to the English morphological variation list as a variant of situational and vice versa.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Variation",
                "sec_num": "2.2"
            },
            {
                "text": "Exploring whether we can align situational and AlAwDAE in the bitext, we find that situational is a morphological variant of situation (based on our morphological variation list for English); next we find that based on the alignment lexicon, there is a positive PMI between situation and AlAw-DAE, which completes the chain between situational and AlAwDAE, so we include them as an alignment candidate after all. The PMI of such a morphological-variation-based candidate is weighted by a 'penalty' factor of 0.5 when compared with the PMI of any competing alignment candidate without such morphological-variation step.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Variation",
                "sec_num": "2.2"
            },
            {
                "text": "Similarly, the English pivot word situations can be used to semantically validate the similarity between Arabic AlAwDAE and AwDAE for our Arabic morphological variation list. The resulting Arabic morphological variation list has entries for 193,263 Arabic words with an average of 4.2 variants each; our English morphological variation list has 57,846 entries with 2.8 variants each.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Variation",
                "sec_num": "2.2"
            },
            {
                "text": "At the end of phase I, most content words will be aligned with relatively high precision. Since function words often do not have an explicit equivalent word or words in the other side of a bitext, they can not be aligned as reliably as content words based on bilingual PMI. 1 Note that due to data sparsity, some content words will remained unaligned in phase I and will subsequently be aligned in phase II as explained in section 3.3.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Morphological Variation",
                "sec_num": "2.2"
            },
            {
                "text": "In Phase II, we align function words, punctuation, and some remaining content words. Function words can be classified into three categories: monovalent, divalent and independent. Monovalent function words modify one head; they include articles (which modify nouns), possessive pronouns, demonstrative adjectives and auxiliary verbs. Divalent function words connect two words or phrases; they include conjunctions and prepositions. Independent function words include nonpossessive pronouns and copula (e.g. is as a main verb). Each of these types of function words is aligned according to its own heuristics.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase II: Function Words",
                "sec_num": "3"
            },
            {
                "text": "In this section we present three representative examples, one for articles (monovalent), one for prepositions (divalent), as well as a structural heuristic.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phase II: Function Words",
                "sec_num": "3"
            },
            {
                "text": "Monovalent function words have the simplest heuristics. Recall that Arabic does not have articles (only a definite prefix Al-added to one or more words in a definite noun phrase), so there is usually no explicit equivalent of the English article on the Arabic side.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Articles",
                "sec_num": "3.1"
            },
            {
                "text": "For an English article, our system identifies the English head word that it modifies based on the English parse tree, and then aligns it with the same Arabic word(s) which that head word is aligned with.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Articles",
                "sec_num": "3.1"
            },
            {
                "text": "Divalent function words are much more interesting. In many cases, an English preposition corresponds to an explicit Arabic preposition in basi-1 It is this lack of reliability that is the defining characteristic of our function words, differentiating them from the concept of marker words used in EBMT chunking (Way and Gough, 2002) . cally the same position. Alignment in that case is straightforward. However, some Arabic prepositions and even more English prepositions do not have an explicit counterpart on the other side. We call such prepositions orphan prepositions. The English preposition of is almost always orphaned in this way.",
                "cite_spans": [
                    {
                        "start": 311,
                        "end": 332,
                        "text": "(Way and Gough, 2002)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Prepositions",
                "sec_num": "3.2"
            },
            {
                "text": "The decision how to align such an orphan preposition is not trivial. Consider the bitext island of Basilan/jzyrp bAsylAn, a typical (NP1 (P NP2)) construction on the English side. Should we co-align the preposition of with the head of NP1 or the head of NP2? In English syntax, the preposition is grouped with NP2, but a preposition is often better \"motivated\" by NP1. We therefore decided to use the English parse tree to identify the heads of both NP1 and NP2, identify the Arabic words aligned to these heads as candidates, and then align the preposition to the Arabic candidate word with which it has the highest bilingual PMI. It turns out that in most cases this will be the candidate on the \"left\". For the example at the top of this paragraph, of will be aligned with jzyrp (\"island\"), which is actually desirable for MT, as it facilitates subsequent rule extraction of type \"island of X/jzyrp X\". We refer to this orphan preposition alignment style as MT-style.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Prepositions",
                "sec_num": "3.2"
            },
            {
                "text": "According to the gold standard alignment guidelines used for the LDC Gold Standard however, an orphan preposition should always be aligned to the \"right\", to bAsylAn in the example above. We therefore implemented an alternative GS-style (for \"Gold Standard\") to be able to later evaluate the impact of these alternatives alignment styles.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Prepositions",
                "sec_num": "3.2"
            },
            {
                "text": "The question whether GIZA or LEAF alignments will indeed give meaningful scores to support the MT-style attachments will be answered by the MT experiments described in section 4.3.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Prepositions",
                "sec_num": "3.2"
            },
            {
                "text": "Here is a more complex example with Arabic (A), its gloss (G) and English (E):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Prepositions",
                "sec_num": "3.2"
            },
            {
                "text": "Arabic: P @ k . J J QJ A @ H @ Q K A @ HP A @ Yk A @ Gloss: sunday attacked aircraft american on/to area jiwar Engl.: on sunday american aircraft attacked the area of jiwar For the Arabic orphan preposition /ElY (\"on/to\"), our system identifies two candidates based on the English parse tree: attacked and area. Based on a higher mutual information, our system then aligns Arabic ElY (\"on/to\") with English attacked, which results in the English word attacked now being aligned to both Arabic attacked and the Arabic on/to, even though they are not adjacent. In the Gold Standard, Arabic on/to is aligned with English area, and LEAF aligns it with English on (yes, the one preceding Sunday). This is apparently very tempting as Arabic on/to is often translated as English on, but here it is incorrect, and our system avoids this tempting alignment because it is ruled out linguistically.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Prepositions",
                "sec_num": "3.2"
            },
            {
                "text": "Note that in some cases, such as sentence-initial prepositional phrases, there is only one candidate; occasionally, when relevant content words remain unaligned, no candidate can be identified, in which case the orphan preposition remains unaligned as well.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Prepositions",
                "sec_num": "3.2"
            },
            {
                "text": "It is not uncommon that content words that we would like to be aligned are not supported by the alignment lexicon, due to general data sparsity or maybe a somewhat unorthodox translation. In those cases we can use structure and word order knowledge to make reasonable alignments anyway.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Adjectives",
                "sec_num": "3.3"
            },
            {
                "text": "Consider an English noun phase ADJ-E NOUN-E and the corresponding Arabic NOUN-A ADJ-A. If the nouns are already aligned, but the adjectives are not yet aligned, we can use the English parse tree to identify ADJ-E as a modifier to NOUN-E, and, aware that adjectives in Arabic post-modify their nouns, identify the corresponding Arabic word based on structure and word order alone. This can be done the other way around as well (link nouns based on already aligned adjectives) and other elements of other phrases as well.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Adjectives",
                "sec_num": "3.3"
            },
            {
                "text": "As more and more function words and remaining content words get aligned, heuristics that weren't applicable before may now apply to the remaining unaligned words, so we perform four passes through a sentence pair to align unaligned words using heuristics. We found that an additional fifth pass did not yield any further improvements.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Example: Adjectives",
                "sec_num": "3.3"
            },
            {
                "text": "We evaluated our word aligner in terms of both alignment accuracy and its impact on an end-toend machine translation system.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "We evaluated our word aligner against a Gold Standard distributed by LDC. The human align-ments of the sentences in this Gold Standard are based on the 2006 GALE Guidelines for Arabic Word Alignment Annotation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Alignment Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "Both the 100-sentence development set and the separate 837-sentence test set are Arabic newswire sentences from LDC2006E86. The test set includes only sentences for which our English parser (Soricut and Marcu, 2003) could produce a parse tree, which effectively excluded a few very long sentences.",
                "cite_spans": [
                    {
                        "start": 190,
                        "end": 215,
                        "text": "(Soricut and Marcu, 2003)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Alignment Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "In the first set of experiments, we compare two settings of our UALIGN system with other aligners, GIZA++ (Union) (Och and Ney, 2003) and LEAF (with 2 iterations) . The GIZA++ aligner is based on IBM Model 4 (Brown et al., 1993) . We chose GIZA Union for our comparison, because it led to a higher BLEU score for our overall MT system than other GIZA variants such as GIZA Intersect and Grow-Diag. The two settings of our system vary in the style on how to align orphan prepositions. Besides precision, recall and (balanced) F-measure, we also include an F-measure variant strongly biased towards recall ( =0.1), which found to be best to tune their LEAF aligner for maximum MT accuracy. GIZA++ and LEAF alignments are based on a parallel training corpus of 6.6 million sentence pairs, incl. the LDC2006E86 set mentioned above.",
                "cite_spans": [
                    {
                        "start": 114,
                        "end": 133,
                        "text": "(Och and Ney, 2003)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 208,
                        "end": 228,
                        "text": "(Brown et al., 1993)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Alignment Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "Prec. Recall F-0.5 F-0. with UALIGN using LEAF alignment lexicon.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Aligner",
                "sec_num": null
            },
            {
                "text": "Our aligner outperforms both GIZA and LEAF on all metrics. Not surprisingly, the GS-style alignments, which align \"orphan\" prepositions according to Gold Standard guidelines, yield higher scores than MT-style alignments. And interestingly by a remarkably high margin.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Aligner",
                "sec_num": null
            },
            {
                "text": "In a second set of experiments, we measure the impact of using different input alignment lexicon used by our aligner on alignment accuracy. In one case UALIGN uses as input the alignment lexicon produced by LEAF, in the other the alignment lexicon produced by GIZA. All experiments in table 2 are for UALIGN. Table 2 : Alignment precision, recall, F-measure ( =0.5), F-measure( =0.1), all of UALIGN, for different alignment styles, different input alignment lexicons.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 309,
                        "end": 316,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Aligner",
                "sec_num": null
            },
            {
                "text": "As LEAF clearly outperforms GIZA on F-0.1 (79.0 vs. 69.5, see table 1), the alignment lexicon based on LEAF is better, so it is not surprising that when we use an alignment lexicon based on GIZA, all metrics degrade, and consistently so for both alignment styles. However the drop in F-0.1 of about 1 point (80.2 ! 79.4 and 83.0 ! 81.9) is much smaller than the differences between the underlying aligners themselves. Our aligner therefore degrades quite gracefully for a worse alignment lexicon. Table 3 shows how much LEAF and UALIGN differ in the percentage of Arabic and English words aligned (correctly or incorrectly). LEAF is much more aggressive in making alignments, aligning almost every Arabic word. Our aligner still leaves some 8% of all words in a sentence unaligned (an opportunity for further improvements). For comparison, in the Gold Standard, 4-5% of all words in our test corpus are left unaligned.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 497,
                        "end": 504,
                        "text": "Table 3",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Aligner",
                "sec_num": null
            },
            {
                "text": "To better understand the impact of several alignment system sub-components, we ran a number of experiments disabling individual sub-components and then comparing the resulting alignment scores with those of the full system. We also measured alignment scores running Phase II with 0 to 5 passes. The test set was the same as in section 4.1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Impact of Sub-Components",
                "sec_num": "4.2"
            },
            {
                "text": "Prec. Recall F-0. Table 4 : Impact of sub-components on alignment precision, recall, F-measure, with GS-style attachments, based on the LEAF alignment lexicon.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 18,
                        "end": 25,
                        "text": "Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "System",
                "sec_num": null
            },
            {
                "text": "Special sub-components of Phase I include adding link candidates for ASCII-string-identical words and transliterated names (see last paragraph before section 2.1), reordering using an English parser (section 2.1) and morphological variation (section 2.2). Each of these sub-components provides a small boost to F-0.1, ranging from +0.1 to +0.5. The second part of the table shows alignment scores before and after each pass of Phase II. Our full system includes 4 passes; an additional 5th pass did not yield any further improvements. Note that during Phase II, precision drops. This is a reflection of (1) our strategy to first align relatively reliable content words in Phase I, followed by less reliable function words and remaining content words, and (2) the challenges of building reliable Gold Standard alignments for function words and non-literal translations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System",
                "sec_num": null
            },
            {
                "text": "The ultimate test for a word aligner is to measure its impact on an end-to-end machine translation system. For this we aligned 170,863 pairs of Arabic/English newswire sentences from LDC, trained a state-of-the-art syntax-based statistical machine translation system (Galley et al., 2006) on these sentences and alignments, and measured BLEU scores (Papineni et al., 2002) on a separate set of 1298 newswire test sentences. Besides swapping in a new set of alignments for the same set of training sentences, and automatically retuning the parameters of the translation system for each set of alignments, no other changes or adjustments were made to the existing MT system.",
                "cite_spans": [
                    {
                        "start": 267,
                        "end": 288,
                        "text": "(Galley et al., 2006)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 349,
                        "end": 372,
                        "text": "(Papineni et al., 2002)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "MT Experiments",
                "sec_num": "4.3"
            },
            {
                "text": "In the first set of experiments, we compare two settings of our UALIGN system with other aligners, again GIZA++ (Union) and LEAF (with 2 iterations). The two settings vary in the alignment lexicon that the UALIGN aligner uses as input.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "MT Experiments",
                "sec_num": "4.3"
            },
            {
                "text": "Aligner BLEU GIZA 47.4 LEAF 48.0 UALIGN using GIZA alignment-lexicon 48.4 UALIGN using LEAF alignment-lexicon 48.7 Table 5 : BLEU scores in end-to-end statistical MT system based on different aligners. Both UALIGN variants use MT-style alignments.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 115,
                        "end": 122,
                        "text": "Table 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "MT Experiments",
                "sec_num": "4.3"
            },
            {
                "text": "With a BLEU score of 48.7, UALIGN using a LEAF alignment-lexicon is significantly better than both GIZA (+1.3) and LEAF (+0.7). This and other significance assertions in this paper are based on paired bootstrap resampling tests with 95% confidence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "MT Experiments",
                "sec_num": "4.3"
            },
            {
                "text": "UALIGN using a GIZA alignment-lexicon significantly outperforms GIZA itself (+1.0).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "MT Experiments",
                "sec_num": "4.3"
            },
            {
                "text": "In a second experiment, we measured the impact of the two alignment styles on BLEU. Recall that for GS-style alignments, orphan prepositions are always co-aligned to the right, following Gold Standard annotation guidelines, whereas for MT-style alignments, mutual information is used to decide whether to align orphan prepositions to the left or to the right.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "MT Experiments",
                "sec_num": "4.3"
            },
            {
                "text": "48.0 UALIGN with GS-style alignments 48.0 UALIGN with MT-style alignments 48.7 Table 6 : BLEU scores in end-to-end statistical MT system based on different alignment styles for orphan prepositions. Both UALIGN variants use a LEAF alignment lexicon.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 79,
                        "end": 86,
                        "text": "Table 6",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "BLEU LEAF",
                "sec_num": null
            },
            {
                "text": "While the GS-style alignments yielded a 2.8 point higher F-0.1 score (83.0 vs. 80.2), the MTstyle alignments result in a significantly better BLEU score (48.7 vs. 48.0) . This shows that (1) a seemingly small difference in alignment styles can have a remarkably high impact on both BLEU scores and alignment accuracy as measured against a Gold Standard, and that (2) optimizing alignment accuracy against an alignment Gold Standard does not necessarily optimize BLEU in end-to-end MT. The latter has been observed by other researchers before, but these results additionally suggest that the gold-standard annotation style might itself have to shoulder part of the blame.",
                "cite_spans": [
                    {
                        "start": 153,
                        "end": 168,
                        "text": "(48.7 vs. 48.0)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "BLEU LEAF",
                "sec_num": null
            },
            {
                "text": "In a small random \"sanity check\" sample from the 170,863 training sentences for the MT experiment, we found cases where the sentence in one language contained much more material than the sentence in the other language. Consider, for example the following sentence pair (with spurious material underlined): Arabic:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Noise Robustness",
                "sec_num": "4.4"
            },
            {
                "text": ", V Y J @ A R JK @ X @ K@ [ JK Q k @ Y JK .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Noise Robustness",
                "sec_num": "4.4"
            },
            {
                "text": "A J A \\K @ Gloss: but also there-is clause another stipulates on/to that if not established the-hotel , English: but , also there is another clause that stipulates that if the hotel is not established , then the government shall be compensated .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Noise Robustness",
                "sec_num": "4.4"
            },
            {
                "text": "Both LEAF and UALIGN correctly align the English \"but , also ... not established ,\" with the Arabic side. LEAF further aligns all words in the spurious English \"then the government shall be compensated .\" with seemingly random material on the Arabic side, whereas UALIGN leaves these spurious words completely unaligned. It would be reasonable to speculate that this behavior, observed in several cases, may be contributing to the good BLEU scores.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus Noise Robustness",
                "sec_num": "4.4"
            },
            {
                "text": "Building on existing statistical aligners, our new word aligner significantly outperforms the best word aligner to date in both alignment error rate and BLEU score.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5"
            },
            {
                "text": "We have developed an approach to word alignment that combines a statistical component with linguistic heuristics. It is novel in that it goes beyond generic resources such as parsers, adding heuristics to explicitly model word order differences and function word alignment.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5"
            },
            {
                "text": "The approach has numerous benefits. Our system produces superior results both on alignment accuracy and end-to-end machine translation quality. Alignments have a high precision. The system is fast (about 0.7 seconds per sentence), and sentences are aligned individually so that a large corpus can easily be aligned on several computers in parallel. All alignment links are tagged with additional information, such as which phase and/or heuristic created them, yielding extensive explanatory power to the developer for easy understanding on how the system arrived at a given alignment. Our approach needs and uses a parser for only one side (English) and not for the other (Arabic).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5"
            },
            {
                "text": "On the other hand, some of the components of this aligner are language-specific, such as word order heuristics, the list of specific function words, and morphological variation lists. While these parts of the system need to be adapted for new languages, the overall architecture and types of heuristics and function words are languageindependent. Chinese for example has different specific types of function words such as aspect markers and measure words. But these fall into the existing category of monovalent function words and will be treated according the same principles as other monovalent function words (section 3.1). Similarly, Japanese postpositions would be treated like other divalent function words (such as Arabic or English prepositions). The author and developer has a basic knowledge of Arabic in general, and an intermediate knowledge of Arabic grammar, which means that no intimate knowledge of Arabic was required to develop the languagespecific components. This same author and developer recently started to adapt UALIGN to Chinese-English word alignment.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5"
            },
            {
                "text": "The alignment rate is still somewhat low. We plan to increase it by enlarging our development set beyond 100 sentences and adding further heuristics, as well as generalizing the output word alignment structure to allow alignments of words to larger constituents in a tree, and to explicitly assert that some words are not covered by the other side of a bitext to model poor translations and poor sentence alignments.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion",
                "sec_num": "5"
            }
        ],
        "back_matter": [
            {
                "text": "This research was supported under DARPA Contract No. HR0011-06-C-0022. The author would like to thank Kevin Knight and the anonymous reviewers for their helpful suggestions, and Steve DeNeefe for running the end-to-end MT evaluations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgment",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "The Mathematics of Statistical Machine Translation: Parameter Estimation",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [
                            "E"
                        ],
                        "last": "Brown",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "Della"
                        ],
                        "last": "Vincent",
                        "suffix": ""
                    },
                    {
                        "first": "Stephen",
                        "middle": [
                            "A Della"
                        ],
                        "last": "Pietra",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [
                            "L"
                        ],
                        "last": "Pietra",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Mercer",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "Computational Linguistics",
                "volume": "19",
                "issue": "",
                "pages": "263--311",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Peter E. Brown, Vincent J. Della Pietra, Stephen A. Della Pietra and Robert L. Mercer. 1993. The Math- ematics of Statistical Machine Translation: Parame- ter Estimation. In Computational Linguistics Vol. 19(2), pages 263-311.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Soft Syntactic Constraints for Word Alignment Through Discriminative Training",
                "authors": [
                    {
                        "first": "Colin",
                        "middle": [],
                        "last": "Cherry",
                        "suffix": ""
                    },
                    {
                        "first": "Dekang",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the 44th Annual Meeting on Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "105--112",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Colin Cherry and Dekang Lin. 2006. Soft Syntactic Constraints for Word Alignment Through Discrimi- native Training. In Proceedings of the 44th Annual Meeting on Association for Computational Linguis- tics, Sydney, Australia, pages 105-112.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Tailoring Word Alignments to Syntactic Machine Translation",
                "authors": [
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Denero",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of the 45th Annual Meeting on Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "17--24",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "John DeNero and Dan Klein. 2007. Tailoring Word Alignments to Syntactic Machine Translation. In Proceedings of the 45th Annual Meeting on Associ- ation for Computational Linguistics, Prague, Czech Republic, pages 17-24.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Using Syntax to Improve Word Alignment Precision for Syntax-Based Machine Translation",
                "authors": [
                    {
                        "first": "Victoria",
                        "middle": [],
                        "last": "Fossum",
                        "suffix": ""
                    },
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Knight",
                        "suffix": ""
                    },
                    {
                        "first": "Steven",
                        "middle": [],
                        "last": "Abney",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the ACL Workshop on Statistical Machine Translation",
                "volume": "",
                "issue": "",
                "pages": "44--52",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Victoria Fossum, Kevin Knight and Steven Abney. 2008. Using Syntax to Improve Word Alignment Precision for Syntax-Based Machine Translation. In Proceedings of the ACL Workshop on Statistical Ma- chine Translation, Columbus, Ohio, pages 44-52.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Getting the Structure Right for Word Alignment: LEAF",
                "authors": [
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Fraser",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Marcu",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of Conference for Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "51--60",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alexander Fraser and Daniel Marcu. 2007. Getting the Structure Right for Word Alignment: LEAF. In Proceedings of Conference for Empirical Methods in Natural Language Processing (EMNLP), Prague, Czech Republic, pages 51-60.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Measuring Word Alignment Quality for Statistical Machine Translation",
                "authors": [
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Fraser",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Marcu",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Computational Linguistics",
                "volume": "33",
                "issue": "",
                "pages": "293--303",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alexander Fraser and Daniel Marcu. 2007. Mea- suring Word Alignment Quality for Statistical Ma- chine Translation. In Computational Linguistics Vol. 33(3), pages 293-303.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Scalable Inference and Training of Context-Rich Syntactic Translation Models",
                "authors": [
                    {
                        "first": "Michel",
                        "middle": [],
                        "last": "Galley",
                        "suffix": ""
                    },
                    {
                        "first": "Jonathan",
                        "middle": [],
                        "last": "Graehl",
                        "suffix": ""
                    },
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Knight",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Marcu",
                        "suffix": ""
                    },
                    {
                        "first": "Steve",
                        "middle": [],
                        "last": "Deneefe",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Ignacio",
                        "middle": [],
                        "last": "Thayer",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the 44th Annual Meeting on Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "961--968",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer 2006. Scalable Inference and Training of Context-Rich Syntactic Translation Models. In Pro- ceedings of the 44th Annual Meeting on Association for Computational Linguistics, Sydney, Australia, pages 961-968.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Name Translation in Statistical Machine Translation: Learning When to Transliterate",
                "authors": [
                    {
                        "first": "Ulf",
                        "middle": [],
                        "last": "Hermjakob",
                        "suffix": ""
                    },
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Knight",
                        "suffix": ""
                    },
                    {
                        "first": "Hal",
                        "middle": [],
                        "last": "Daum\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Iii",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the 46th Annual Meeting on Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "389--397",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ulf Hermjakob, Kevin Knight, and Hal Daum\u00e9 III 2008. Name Translation in Statistical Machine Translation: Learning When to Transliterate. In Proceedings of the 46th Annual Meeting on Asso- ciation for Computational Linguistics, Columbus, Ohio, pages 389-397.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "A Maximum Entropy Word Aligner for Arabic-English Machine Translation",
                "authors": [
                    {
                        "first": "Abraham",
                        "middle": [],
                        "last": "Ittycheriah",
                        "suffix": ""
                    },
                    {
                        "first": "Salim",
                        "middle": [],
                        "last": "Roukos",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of Joint Conference of Human Language Technology and Empirical Methods in Natural Language Processing (HLT/EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "89--96",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Abraham Ittycheriah and Salim Roukos. 2005. A Maximum Entropy Word Aligner for Arabic- English Machine Translation. In Proceed- ings of Joint Conference of Human Language Technology and Empirical Methods in Natural Language Processing (HLT/EMNLP), Vancouver, British Columbia, Canada, pages 89-96.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Syntax-Driven Learning of Sub-Sentential Translation Equivalents and Translation Rules from Parsed Parallel Corpora",
                "authors": [
                    {
                        "first": "Alon",
                        "middle": [],
                        "last": "Lavie",
                        "suffix": ""
                    },
                    {
                        "first": "Alok",
                        "middle": [],
                        "last": "Parlikar",
                        "suffix": ""
                    },
                    {
                        "first": "Vamshi",
                        "middle": [],
                        "last": "Ambati",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the ACL/HLT Second Workshop on Syntax and Structure in Statistical Translation (SSST-2)",
                "volume": "",
                "issue": "",
                "pages": "87--95",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alon Lavie, Alok Parlikar and Vamshi Ambati. 2008. Syntax-Driven Learning of Sub-Sentential Transla- tion Equivalents and Translation Rules from Parsed Parallel Corpora. In Proceedings of the ACL/HLT Second Workshop on Syntax and Structure in Statis- tical Translation (SSST-2), Columbus, Ohio, pages 87-95.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Models of translational equivalence among words",
                "authors": [
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Melamed",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Computational Linguistics",
                "volume": "26",
                "issue": "",
                "pages": "221--249",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dan Melamed. 2000. Models of translational equiv- alence among words. In Computational Linguistics Vol. 26(2), pages 221-249.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "A Systematic Comparison of Various Statistical Alignment Models",
                "authors": [
                    {
                        "first": "Josef",
                        "middle": [],
                        "last": "Franz",
                        "suffix": ""
                    },
                    {
                        "first": "Hermann",
                        "middle": [],
                        "last": "Och",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ney",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Computational Linguistics",
                "volume": "29",
                "issue": "",
                "pages": "19--51",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Franz Josef Och and Hermann Ney. 2003. A System- atic Comparison of Various Statistical Alignment Models. In Computational Linguistics Vol. 29(1), pages 19-51.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The Alignment Template Approach to Statistical Machine Translation",
                "authors": [
                    {
                        "first": "Josef",
                        "middle": [],
                        "last": "Franz",
                        "suffix": ""
                    },
                    {
                        "first": "Hermann",
                        "middle": [],
                        "last": "Och",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ney",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Computational Linguistics",
                "volume": "30",
                "issue": "",
                "pages": "417--449",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Franz Josef Och and Hermann Ney. 2004. The Alignment Template Approach to Statistical Ma- chine Translation. In Computational Linguistics Vol. 30(4), pages 417-449.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "BLEU: a Method for Automatic Evaluation of Machine Translation",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Papineni",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Roukos",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Ward",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [
                            "J"
                        ],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of the 40th Annual Meeting on Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "311--318",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "K. Papineni, S. Roukos, T. Ward, and W. J. Zhu. 2002. BLEU: a Method for Automatic Evaluation of Ma- chine Translation. In Proceedings of the 40th An- nual Meeting on Association for Computational Lin- guistics, Philadelphia, PA, pages 311-318.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Sentence Level Discourse Parsing Using Syntactic and Lexical Information",
                "authors": [
                    {
                        "first": "Radu",
                        "middle": [],
                        "last": "Soricut",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Marcu",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology",
                "volume": "",
                "issue": "",
                "pages": "149--156",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Radu Soricut and Daniel Marcu. 2003. Sentence Level Discourse Parsing Using Syntactic and Lexical In- formation. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology, Edmonton, Canada, pages 149-156.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "wEBMT: developing and validating an example-based machine translation system using the world wide web",
                "authors": [
                    {
                        "first": "Andy",
                        "middle": [],
                        "last": "Way",
                        "suffix": ""
                    },
                    {
                        "first": "Nano",
                        "middle": [],
                        "last": "Gough",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "In Computational Linguistics",
                "volume": "29",
                "issue": "3",
                "pages": "421--457",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andy Way, Nano Gough. 2003. wEBMT: develop- ing and validating an example-based machine trans- lation system using the world wide web In Compu- tational Linguistics Vol. 29(3), pages 421-457.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Automatic Generation of Parallel Treebanks",
                "authors": [
                    {
                        "first": "Ventsislav",
                        "middle": [],
                        "last": "Zhechev",
                        "suffix": ""
                    },
                    {
                        "first": "Andy",
                        "middle": [],
                        "last": "Way",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of 22nd International Conference on Computational Linguistics (COLING)",
                "volume": "",
                "issue": "",
                "pages": "1105--1112",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ventsislav Zhechev, Andy Way. 2008. Automatic Generation of Parallel Treebanks. In Proceed- ings of 22nd International Conference on Compu- tational Linguistics (COLING), Manchester, UK, pages 1105-1112.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF2": {
                "type_str": "table",
                "content": "<table/>",
                "num": null,
                "text": "Percentages of Arabic and English words aligned",
                "html": null
            }
        }
    }
}