File size: 64,566 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
{
    "paper_id": "2019",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:29:35.815919Z"
    },
    "title": "Towards Handling Verb Phrase Ellipsis in English-Hindi Machine Translation",
    "authors": [
        {
            "first": "Niyati",
            "middle": [],
            "last": "Bafna",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Ashoka University",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Dipti",
            "middle": [],
            "last": "Misra",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Indian Institute of Technology",
                "location": {
                    "settlement": "Hyderabad"
                }
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "English-Hindi machine translation systems have difficulty interpreting verb phrase ellipsis (VPE) in English, and commit errors in translating sentences with VPE. We present a solution and theoretical backing for the treatment of English VPE, with the specific scope of enabling English-Hindi MT, based on an understanding of the syntactical phenomenon of verb-stranding verb phrase ellipsis in Hindi (VVPE). We implement a rule-based system to perform the following sub-tasks: 1) Verb ellipsis identification in the English source sentence, 2) Elided verb phrase head identification 3) Identification of verb segment which needs to be induced at the site of ellipsis 4) Modify input sentence; i.e. resolving VPE and inducing the required verb segment. This system is tested in two parts. It obtains 94.83 percent precision and 83.04 percent recall on subtask (1), tested on 3900 sentences from the BNC corpus (Leech, 1992). This is competitive with state-of-the-art results. We measure accuracy of subtasks (2) and (3) together, and obtain a 91 percent accuracy on 200 sentences taken from the WSJ corpus(Paul and Baker, 1992). Finally, in order to indicate the relevance of ellipsis handling to MT, we carried out a manual analysis of the MT outputs of 100 sentences after passing it through our system. We set up a basic metric (1-5) for this evaluation, where 5 indicates drastic improvement, and obtained an average of 3.55.",
    "pdf_parse": {
        "paper_id": "2019",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "English-Hindi machine translation systems have difficulty interpreting verb phrase ellipsis (VPE) in English, and commit errors in translating sentences with VPE. We present a solution and theoretical backing for the treatment of English VPE, with the specific scope of enabling English-Hindi MT, based on an understanding of the syntactical phenomenon of verb-stranding verb phrase ellipsis in Hindi (VVPE). We implement a rule-based system to perform the following sub-tasks: 1) Verb ellipsis identification in the English source sentence, 2) Elided verb phrase head identification 3) Identification of verb segment which needs to be induced at the site of ellipsis 4) Modify input sentence; i.e. resolving VPE and inducing the required verb segment. This system is tested in two parts. It obtains 94.83 percent precision and 83.04 percent recall on subtask (1), tested on 3900 sentences from the BNC corpus (Leech, 1992). This is competitive with state-of-the-art results. We measure accuracy of subtasks (2) and (3) together, and obtain a 91 percent accuracy on 200 sentences taken from the WSJ corpus(Paul and Baker, 1992). Finally, in order to indicate the relevance of ellipsis handling to MT, we carried out a manual analysis of the MT outputs of 100 sentences after passing it through our system. We set up a basic metric (1-5) for this evaluation, where 5 indicates drastic improvement, and obtained an average of 3.55.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Verb phrase ellipsis is a particularly frequent form of ellipsis, both in speech and in text. English VPE is the elimination of a non-finite verb phrase, introduced by an auxiliary or the particle 'to' (Kenyon-Dean et al., 2016) .",
                "cite_spans": [
                    {
                        "start": 202,
                        "end": 228,
                        "text": "(Kenyon-Dean et al., 2016)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We observe that state-of-the-art MT systems often cannot correctly interpret and translate sentences with VPE. For example, (elliptical phrase and site of ellipsis are in bold and italics respectively):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Ram cooked the food quickly, but Shyam did not \u2192 In our initial analysis, we find that Google Translate could translate only 6/50 VPE sentences, taken from the WSJ corpus, correctly from English to Hindi. This motivated us to identify and resolve VPE in English and give this modified sentence to the machine translation system to check whether it improves the quality of translation. We used our MT system Anusaaraka for this purpose. 1 We present a rule-based approach that performs three sub-tasks in order to solve this problem: 1) ellipsis identification 2) identification of the antecedent head verb 3) addition of necessary auxiliaries and/or complements to the head verb. Finally, it modifies the input sentence by inducing the identified verb segment at the site of ellipsis. We tailor this algorithm for the purpose of English-Hindi MT, and therefore provide a special treatment to compound verbs (including phrasal verbs), serial verbs, and verb complements in English.",
                "cite_spans": [
                    {
                        "start": 436,
                        "end": 437,
                        "text": "1",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Our solution is based on an understanding of verb phrase ellipsis in Hindi. Hindi does not exhibit VPE in the same manner as English; however, it exhibits a phenomenon called verb-stranding verb phrase ellipsis (VVPE) (Manetta, 2018) . This means that a verbal phrase, including objects and other arguments, may be completed elided, stranding the head verb, which then appears at the site of ellipsis.",
                "cite_spans": [
                    {
                        "start": 218,
                        "end": 233,
                        "text": "(Manetta, 2018)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "We propose that English VPE can be transferred into Hindi VVPE in order to improve MT results on the modified sentence. For this, we claim that it is enough to identify the head verb antecedent in the English sentence, and perform an analysis to support the claim.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Finally, we want to see that MT systems indeed perform better on statements with ellipsis after our treatment. Since we want to test the performance of the MT on a very specific facet: i.e. the elliptical clause, we do not utilize standard evaluation metrics but set up and define our own scale, and perform a manual analysis of 100 sentences. This evaluation dataset is meant to be merely indicative of the benefits of ellipsis handling for MT. The results are explained below.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "There have been several previous works addressing the problem of antecedent head resolution for VPE in English. Cheung et al adapt the Margin-Infused-Relaxed Algorithm (MIRA) for target detection and antecedent resolution and obtain an accuracy of 65 percent (Kenyon-Dean et al., 2016 (Nielsen, 2005 ) (Hardt, 1992) . Liu et al, 2016 experiment with various joint modelling techniques, and obtain a recall of 83.46 percent for boundary identification (Liu et al., 2016) . Earlier works include Daniel Hardt's linguistically motivated rule-based system, that eliminates impossible antecedents by looking at be-do conflicts, contained antecedents and assigned scores based on co-reference of the noun subjects and clausal relationships, such as 'as' constructions. (Hardt, 1992) The necessity for tools to deal with VPE is widely recognized in literature, for the purposes of information extraction, finding event co-occurrence, etc (Kenyon-Dean et al., 2016) . In the context of MT, a possible solution is to identify the antecedent, or the source verb phrase, and induce it at the site of ellipsis to gain a legitimate, simplified sentence. This solution functions by reiterating the antecedent at the site of ellipsis. By breaking the link between the ellipsis and the antecedent, we give the MT two independent clauses, which it can translate without error.",
                "cite_spans": [
                    {
                        "start": 259,
                        "end": 284,
                        "text": "(Kenyon-Dean et al., 2016",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 285,
                        "end": 299,
                        "text": "(Nielsen, 2005",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 302,
                        "end": 315,
                        "text": "(Hardt, 1992)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 451,
                        "end": 469,
                        "text": "(Liu et al., 2016)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 763,
                        "end": 776,
                        "text": "(Hardt, 1992)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 931,
                        "end": 957,
                        "text": "(Kenyon-Dean et al., 2016)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background and Our Contribution",
                "sec_num": "2"
            },
            {
                "text": "Indeed, the previous works listed above are aimed at identifying the antecedent verb phrase, which includes compulsorily a head verb, and optionally its arguments and adverbials. While it is important to pick up arguments and adverbials of the head verb for the purpose of comprehension, there are some problems that it introduces: namely, it must often disambiguate by context and therefore can be a great source of error, and in the context of MT, it might make the output sentence clumsy and unnatural.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background and Our Contribution",
                "sec_num": "2"
            },
            {
                "text": "Since we are looking at antecedent head resolution from a particular angle i.e. transforming the English sentence containing VPE in order to align with Hindi VVPE and therefore help the MT, our problem does not require us to find the antecedent boundary of the head verb, while we do provide an additional treatment of certain verb constructions in English according to the manner in which they would be translated into Hindi. As far as we are aware, this is the first attempt to tailor English VPE resolution in the context of English-Hindi MT.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background and Our Contribution",
                "sec_num": "2"
            },
            {
                "text": "Verb phrase ellipsis in English is introduced by an auxiliary. We consider five classes of el-lipsis depending upon the auxiliary at the site of ellipsis: 1) to_be, 2) to_have, 3) to_do, 4) modals, and 5) to_particle ellipsis. Cheung et al include a sixth class: the do-so anaphora, while acknowledging that modals and Do-X anaphora are not technically auxiliaries (Kenyon-Dean et al., 2016) . We have chosen to identify ellipsis introduced by modals, however, because 1. the behaviour of the former is identical with ellipsis by true auxiliaries, 2. Likewise, we observe that it poses a problem to state-of-the-art MT English-Hindi systems. We do not identify Do-X anaphora in this system, however, since this is simply a pronominalization of the antecedent rather than eliding. These have a different treatment than VPE -for example, they may be directly pronominalized in Hindi as well, rather than transferred into VVPE. Indeed, MT systems are able to do this:",
                "cite_spans": [
                    {
                        "start": 365,
                        "end": 391,
                        "text": "(Kenyon-Dean et al., 2016)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ellipsis in English",
                "sec_num": "3"
            },
            {
                "text": "Although Mr. Azoff won't produce films at first, it is possible that he could do so later, the sources said \u2192 ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ellipsis in English",
                "sec_num": "3"
            },
            {
                "text": "We know that state-of-the-art machines cannot interpret VPE. (Voita et al., 2019) . Translation to Hindi from English requires a prediction of the elided English verb. An elided VP consists of a head verb, optionally along with its object arguments and adverbials. For example, it is clear that the adverbial is interpreted as part of the VP in the following sentence:",
                "cite_spans": [
                    {
                        "start": 61,
                        "end": 81,
                        "text": "(Voita et al., 2019)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "I could walk quickly, at that age, across traffic-filled roads, but now I cannot. (created example) One solution to eliminate errors due to VPE, is to identify the antecedent and induce it at the site of ellipsis entirely, as a preliminary step before translation, thus eliminating the ellipsis entirely. However, this naive approach has the following issues:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "1. Making the decision of whether to import a particular adjunct is complicated, and might be governed by semantic context, and (when verbal), by emphasis.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "2. Reiterating the entire verb phrase at the site of ellipsis may sound clumsy and unnatural, and make for a worse translation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "Identifying the boundaries of the elided verb phrase is a much harder task than identifying the antecedent head verb. We note, for example, the consistent drop in Exact Match accuracy in Neilson's study of Hardt's algorithm over different corpora (Nielsen, 2005) -sometimes dipping as much as 30 percent lower. The matter is not as simple as picking up the entire verb sub-tree -always importing all arguments and adjuncts is not permissible, since the correct interpretation often depends on surrounding knowledge. For example, in Ram would have liked to eat out with you on Sunday afternoon, but he can't. (created example)",
                "cite_spans": [
                    {
                        "start": 247,
                        "end": 262,
                        "text": "(Nielsen, 2005)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "There are more than one interpretations of what Ram can't do: eat out, eat out with you, or eat out with you on Sunday afternoon. A native speaker selects one of these in context. If we are seeking to eliminate ellipsis with our system, we must select one of these interpretations to paste at the site of ellipsis, or the whole verb phrase. We may also not adopt intermediate policies such as importing noun objects but not adjuncts, because this risks placing undue emphasis on certain arguments, and may result in an incorrect semantics.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "The second problem with this strategy is that it may render sentences, after pasting, clumsy, unnatural or nonsensical, as illustrated, respectively, by the following example, taken from the WSJ corpus (Bos and Spenader, 2011 ).",
                "cite_spans": [
                    {
                        "start": 202,
                        "end": 225,
                        "text": "(Bos and Spenader, 2011",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "The Volokhs were afraid that they'd end up like a friend of theirs who'd applied for a visa and waited for 10 years, having been demoted from his profession of theoretical mathematician to shipping clerk. They didn't (end up like\u2026shipping clerk)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Resolution in the Context of MT",
                "sec_num": "4"
            },
            {
                "text": "Verb-stranding verb phrase ellipsis is the phenomenon wherein a verb is stranded at a site of ellipsis, and its internal arguments are elided. Raj also read (the paper twice). (Manetta, 2018) Here, 'paRha' provides access to the internal arguments i.e. the direct object and the adverb via VVPE. Manetta supports her analysis theoretically and by a survey across native speakers. This provides us an intuition for a solution for our larger problem: that is, instead of disambiguating the antecedent boundaries to resolve English VPE as simplification for English-Hindi MT, we may provide the Hindi clause containing ellipsis only the head verb, which has access to the internal arguments of the antecedent. If we induce the head verb at the site of ellipsis in the English before translating, now, then we create a valid, syntactical Hindi sentence with all the interpretations of the original sentence intact. We illustrate with an example: Resolving the ellipsis in (b) we get an English sentence with 'eat' at the previous site of ellipsis, that means that Ram is incapable of eating. However, (c) in Hindi, that similarly has 'eat' at previous site of English VPE, is a perfectly acceptable translation of the original sentence, exhibiting VVPE. In (c) the stranded verb 'kha' (eat) has access to the internal arguments 'aapke saath', (with you), 'Sunday dopahar' (Sunday afternoon), etc., and therefore it carries all the interpretations of the original. This is the core idea of the system.",
                "cite_spans": [
                    {
                        "start": 176,
                        "end": 191,
                        "text": "(Manetta, 2018)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hindi VVPE",
                "sec_num": "5"
            },
            {
                "text": "a (Original",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hindi VVPE",
                "sec_num": "5"
            },
            {
                "text": "While we do not require to identify the boundaries of the head verb, we do need to identify all the components of the verb. This may be required in several cases, such as phrasal verbs, idioms, or serial verbs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dealing with Multi-Word Verbs in English",
                "sec_num": "6"
            },
            {
                "text": "These can be categorized into phrasal verbs and prepositional verbs. The former is a class of verbs that consists of a head verb and a preposition, like \"take over\", \"get around\" or \"sink in\". Since phrasal verbs are opaque, we require to pick up its preposition to maintain the correct sense in which it appears. For example, in She hasn't got over her old failures as yet, although she should. (created example)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Compound Verbs",
                "sec_num": "6.1"
            },
            {
                "text": "We want the stranded segment to be \"get over\", not simply \"get\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Compound Verbs",
                "sec_num": "6.1"
            },
            {
                "text": "Prepositional verbs are also verbs followed by a preposition e.g \"stare at\", \"care for\", but they are different from phrasal verbs in certain ways. Their meaning is derived primarily from the head verb. We can see that in the first case, \"stare at\", it would admissible to strand simply \"stare\", although in the subsequent case, the sense of 'care' changes without the preposition. It is necessary, therefore, to pick up the prepositional verb as a unit.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Compound Verbs",
                "sec_num": "6.1"
            },
            {
                "text": "In general, when the head verb has verb complement arguments, it is not necessary to pick them up, because the stranded head verb in Hindi will have access to them. For example, (b) is a legitimate translation of (a), with the stranded verb \"wants\". However, this treatment assumes that the sequence of complements in the head verb lexically translate into a sequence of simple verbs in Hindi. This might not be the case. The reason we cannot strand is because 'padega' (has: obligation aspect) is an auxiliary of 'jaana' (to go) in the Hindi sentence, not the head verb of the compound verb \"jaana padega\" (has to go). The head here is \"jaana\", and therefore it must be induced in the site of ellipsis. Therefore, with the English verb \"has\", among others, we must also pick up its complements to induce at the site of ellipsis. To generalize, we require the complement cE of a head verb hE, when hE + cE results in a VV or verb-light verb complex in Hindi.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Verbal Complements",
                "sec_num": "6.2"
            },
            {
                "text": "Examples of serial verbs in English are a I'll go see if she's okay b Why don't you run get a taxi?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Serial Verbs",
                "sec_num": "6.3"
            },
            {
                "text": "These can be treated as verb-complement series, as the Stanford Universal Dependency Framework treats them. Hindi also treats the verb complex as a head verb followed by complements. Taking the first example, it is acceptable to strand \"jaati\" (go) from this sentence, in: ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Serial Verbs",
                "sec_num": "6.3"
            },
            {
                "text": "We are using a dependency tree of the input sentence for all tasks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Algorithm to Identify Ellipsis and Head Verb of Antecedent",
                "sec_num": "7"
            },
            {
                "text": "We assume that each word w in the input sentence that belongs to our five classes, is a site of ellipsis. Then we go through a process of elimination. We have a different set of rules for each class.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Identification: Rules and Results",
                "sec_num": "7.1"
            },
            {
                "text": "Some of the basic criteria for elimination include:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Identification: Rules and Results",
                "sec_num": "7.1"
            },
            {
                "text": "1. w is a copula (for to_be) 2. w is an auxiliary child of a verb (for to_be, to_do, to_have, modals)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Identification: Rules and Results",
                "sec_num": "7.1"
            },
            {
                "text": "3. w has a direct object noun child (for to_be, to_do, to_have, modals)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Identification: Rules and Results",
                "sec_num": "7.1"
            },
            {
                "text": "We perform part 1 of our two-part testing at this stage. These rules, tested on 3900 sentences from the BNC corpus (Leech, 1992), give us a precision of 94.83 percent and a recall of 83.04 percent.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "w is not an xcomp child (for to_particle)",
                "sec_num": "4."
            },
            {
                "text": "If we find an ellipsis, then we perform 2 subtasks:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Head Resolution: Algorithm",
                "sec_num": "7.2"
            },
            {
                "text": "1. Find the head verb 2. Supplement the head verb",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Antecedent Head Resolution: Algorithm",
                "sec_num": "7.2"
            },
            {
                "text": "We collect all the verbs in the input, eliminating according to the constraints on to_be/non-to_be ellipsis discussed in Section 3. We use a score-based approach, as does Hardt (Hardt, 1992) .",
                "cite_spans": [
                    {
                        "start": 177,
                        "end": 190,
                        "text": "(Hardt, 1992)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Finding the Head Verb",
                "sec_num": "7.2.1"
            },
            {
                "text": "Here are the features that we look at, for verb v as a candidate for ellipsis e: 1. Positive scores for noun subjects matching in number, negative score for noun subjects not matching in 'passivity', positive score if both noun subjects are proper, positive score for identical noun subjects.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Finding the Head Verb",
                "sec_num": "7.2.1"
            },
            {
                "text": "2. Negative score if e belongs to the complement clause child of v. For example, in: He told me that he had passed the exam, and then he told me that John hadn't. \"told\" gets ruled out as the clause \"John hadn't\" is a complement child of \"told\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Finding the Head Verb",
                "sec_num": "7.2.1"
            },
            {
                "text": "3. Negative score if v belongs to the complement clause child of e. 4. Positive score if v has an auxiliary in the same class as e. For example: Sita could walk while texting, but now she can't.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Finding the Head Verb",
                "sec_num": "7.2.1"
            },
            {
                "text": "5. Finally, we assign a positive score to the first verb that we obtain by backtracking up the dependency tree from e, if it is contained in our candidate verbs. This gives us the correct antecedent head several times, even when it is far away. It captures clausal relationships between antecedent verb and e, such as as\u2026as, \u2026than, which Hardt also mentions in their scoring algorithm.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Finding the Head Verb",
                "sec_num": "7.2.1"
            },
            {
                "text": "6. We then evaluate the scores. If there is a clash, we choose the closest verb, advantaging forward ellipsis to backward ellipsis, as the latter is widely acknowledged to be much rarer than the former.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Finding the Head Verb",
                "sec_num": "7.2.1"
            },
            {
                "text": "As we said earlier, the head verb might need to be supplemented before it is 'stranded'. For example, in:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "Both banks have been battered, as other Arizona banks have, by falling real estate prices.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "The above algorithm will identify \"battered\" as the head verb; however, we do need to supplement it with the auxiliary \"been\", to create a grammatical verb after inducing.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "We perform three sub-tasks in supplementing the main verb v:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "1. Add auxiliaries: we add any auxiliaries of v, after skipping the first if any that belongs to the same class as e. For example, we skip the auxiliary \"have\" in the above example.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "2. Add particles: here, we check whether v is part of a phrasal verb/prepositional verb, and add the preposition(s) if so. The dependency tree marks particle dependants of the verb: however, since it doesn't always do so, we maintain and import a list of common phrasal verbs/prepositional verbs for reference.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "3. Add verb complements. Similarly, we maintain and import a list of verbs of which verb complements, if any, we need to pick up, since they result in nonstrandable verbs in isolation in Hindi: \"let\", \"have\". These also include verbs which may not always give a correct lexical translation in isolation, e.g. \"feel\", \"seem\", for which it is safer to also induce the complements.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "The lists above can always be augmented, of course. Currently, our system does not deal with idioms, but one solution that we suggest is to maintain a list of frequently occurring idioms and induce them as a whole.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Supplement Main Verb",
                "sec_num": "7.2.2"
            },
            {
                "text": "In part 2 of our testing, we tested this system on 200 sentences from the WSJ corpus and got an accuracy of 91 percent on antecedent head resolution. Here are some errors, and their analysis:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "Mr. Dinkins also has failed to allay Jewish voters' fears about his association with the Rev. Jesse Jackson, despite the fact that few local non-Jewish politicians have been as vocal for Jewish causes in the past 20 years as Mr. Dinkins has.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "Here, the algorithm identifies \"failed\" instead of \"been\". It awards \"failed\" for common noun subject, common auxiliary and being the first verb upon backtracking, whereas \"been\" is only awarded for common auxiliary.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "But Sony also says in its filing that the Warner contract \"doesn't require that Guber and Peters take any affirmative steps to produce motion pictures; it simply rewards them when they do and prohibits them from producing for another entertainment company.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "Here the algorithm identifies \"rewards\" as the source verb instead of \"take\". It awards \"take\" for noun subject number, \"require\" for common auxiliary, and \"rewards\" by backtracking. Finally, it resolves the tie by choosing \"rewards\" which is the closest. Now they know who you mean and you know who you mean -but no one else does.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "The algorithm gives \"mean\" instead of \"know\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "\"mean\" gets awarded for noun subject number, and \"know\" is awarded for backtracking; however, \"mean\" wins the tie since it is the closest.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "There are errors introduced by the POS tagger; for example, in:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "A good half-hour into breakfast at the Palmer House, Mr. O'Brien looks up from his plate after Mr. Straszheim says something about people who believe interest rates are about to nosedive -\"I'm one of them who hope they will, with 6 billion in debt on the books. \"nosedive\" is not recognized as a verb, similar to, in another examples, \"program-trade\", and \"move\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "There are errors introduced by the parser:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "The text by Patrick O'Connor is a tough read, but the pictures make her magnetism clear and help explain why Ernest Hemmingway called Baker, \"The most sensational women anybody ever saw -or ever will.\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "Here, the algorithm wrongly identifies \"called\" as the antecedent main verb, instead of \"saw\". The clause \"or ever will\" is labelled a conjugate dependent of the noun Baker, instead of a conjunct of the verb \"saw\". If this had been so, \"saw\" would have got scores for a common subject (\"anybody\") and being the verb obtained upon backtracking.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "\"A lot of people think I will give away the store, but I can assure you I will not,\" he says.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "The algorithm identifies \"assure\". This would be avoided if the dependency marked \"I will not\" as a complement clause of \"assure\" -however the dependency misses this relation. If \"assure\" was given a penalty on this grounds, the next highest candidate is indeed the correct one: \"give\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results and Error Analysis for Antecedent Head Resolution",
                "sec_num": "8"
            },
            {
                "text": "We now show that inducing the head elliptical verb makes the input sentence easier for the MT system to comprehend. We perform a manual analysis of 100 sentences with ellipsis, taken from the WSJ corpus: we create a \"before\" and \"after\" translation pair for each, and compare to identify improvements. This was done by two fluent speakers of Hindi and English. We define a scale (1-5) to quantify this improvement:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "5: Improvement from incoherent to perfect translation of ellipsis, and surroundings 4: The meaning is fairly clearer than it was in the original 3: The translation is as good or as bad as it originally was 2: The meaning is fairly more obscure than it originally was 1: The sentence is rendered completely incoherent from an original good translation",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "We add some flags to further nuance this scale: we also mark the translations for fluidity, i.e. if the translation while rendered better is still not fluent (f) or if the translation while not making the meaning clearer is rendered more fluent (F), and for the overall meaning of the entire sentence -i.e. beyond the clauses of the antecedent and the ellipsis. These markers, however, are only for exceptional cases, since most of the translations we got, both \"before\" and \"after\" were not fluent.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "The average score over 100 sentences was found to be 3.55, with 18 cases of correct nonfluent sentences, and 5 cases of incorrect sentences with especial improvement in fluency.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "These are the large-scale sources of lack of improvement that we found: 1. Sentences that the MT cannot translate overall due to other complexities such as nested clauses, etc., for which its output is close to gibberish, show little to no improvement with addition of the elliptical verb. These sentences passed through a system that can handle them, sometimes Google Translate, almost always show high improvement with the addition of the verb. We had about 35 such sentences, all marked 3, sometimes marked for improvement or deterioration in fluency.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "2. Most sentences of the type \"as did\", as in 'X ate apples as did Y' fail to show any improvement and often show a deterioration in fluency. This is because the processed sentence is rendered ungrammatical and perhaps more incomprehensible.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "Again, Google Translate often shows improvement on these sentences from \"before\" to \"after\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "An example of the first instance is:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "(1) American Enterprise Institute scholar Norman Ornstein in the Oct. 21 TV Guide on \"What TV News Doesn't Report About Congressand Should\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "The system induces the elliptical verb complex \"report\". Both Anusaaraka and Google output incorrect translations for the original sentence, although Google shows errors only due to the ellipsis. Therefore, it is able to improve its translation after we induce the elliptical verb:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "(2) amerikee (Score: 3f, for deterioration in fluency)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "The original gives a similarly incorrect output, though not quite as ungrammatical. This output, even after the elliptical verb has been added, conveys little to no meaning in Hindi.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "Here are some micro-level sources of nonimprovement:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "1. When an antecedent verb is being used idiomatically, the MT system may interpret it literally when stranded in the elliptical clause, even if it catches the correct sense in the antecedent clause, possible because the former construction is more unusual.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "2. In general, the VVPE construction fails if the system makes two different lexical (Score: 4)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "We note here that these figures are dependent upon how well the base translation system can translate the original. We performed the same analysis on samples from this dataset with Google Translation and got consistently better results per each batch of 10, and an average of 3.7. This indicative exercise is intended to give an idea of why targeted VPE handling for specific language pairs holds significance in bettering MT results.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation of Effect on MT outputs",
                "sec_num": "9"
            },
            {
                "text": "The concept and the system that we have introduced above, for handling ellipsis in a targeted manner to improve English-Hindi MT, are still in their nascent stages. There are three primary entry points for future work: firstly, the conceptual negotiation of the phenomenon in English and Hindi. We have decided, as we explain, only to induce the main verb. While this gives satisfactory results most of the times, it also fails in some cases: it might help, for example, to make a decision to induce object arguments in these cases. Secondly, the identification of VPE in English, and the antecedent resolution. We are already dealing with complex verbs, verbal complements etc. in a certain manner, but this treatment invites further and more rigorous work, both in terms in nuance with the treatment, and how exhaustive we are with the lists that we have drawn up, introducing, for example, treatment of idioms. Thirdly, the application of our system as over the input, before it is passed through the MT. There are certain problems that may be solved by pipelining this ellipsis handling differently into the MT system: we leave this to future investigation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Future Work",
                "sec_num": "10"
            },
            {
                "text": "https://ltrc.iiit.ac.in/Anusaaraka/anu_ home.html",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "interpretations in the antecedent and elliptical clause. This may be for different reasons: e.g. on transitive verbs, since they appear in the elliptical verb without their objects, which is unnatural. The MT system will possibly attempt to catch an intransitive sense of the verb in such a situation. However, this is only in certain few cases of such verbs.An example of both of the above is:During the takeover, Mr. Hahn said he would put his account up for review if WPP's bid were successful, but he didn't.The system induces the elliptical verb complex \"put up\". Both Anusaaraka and Google output incorrect translations for the original sentence, although Google interprets the idiomatic meaning of \"put up\" correctly. Therefore, it is able to improve its translation after we induce the elliptical verb:Anusaaraka makes error type 1 (literal interpretation of verb): 4 (Score: 3) Google makes error type 2 (different lexical interpretation). In this case, it is minor, because the meaning of the sentence is restored from the original. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "annex",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "An annotated corpus for the analysis of vp ellipsis",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Bos",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Spenader",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Language Resources and Evaluation",
                "volume": "45",
                "issue": "4",
                "pages": "463--494",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J. Bos and J. Spenader. An annotated corpus for the analysis of vp ellipsis. Language Resources and Evaluation, 45(4):463-494, 2011.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "An algorithm for vp ellipsis",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Hardt",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Proceedings of the 30th annual meeting on Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "9--14",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Hardt. An algorithm for vp ellipsis. In Proceed- ings of the 30th annual meeting on Association for Computational Linguistics, pages 9-14. As- sociation for Computational Linguistics, 1992.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Verb phrase ellipsis resolution using discriminative and margin-infused algorithms",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Kenyon-Dean",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "C K"
                        ],
                        "last": "Cheung",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Precup",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1734--1743",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "K. Kenyon-Dean, J. C. K. Cheung, and D. Pre- cup. Verb phrase ellipsis resolution using dis- criminative and margin-infused algorithms. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1734-1743, 2016.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Leech. 100 million words of english: the british national corpus (bnc)",
                "authors": [
                    {
                        "first": "G",
                        "middle": [
                            "N"
                        ],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "G. N. Leech. 100 million words of english: the british national corpus (bnc). 1992.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Exploring the steps of verb phrase ellipsis",
                "authors": [
                    {
                        "first": "Z",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [
                            "G"
                        ],
                        "last": "Pellicer",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Gillick",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the Workshop on Coreference Resolution Beyond OntoNotes (CORBON 2016)",
                "volume": "",
                "issue": "",
                "pages": "32--40",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Z. Liu, E. G. Pellicer, and D. Gillick. Exploring the steps of verb phrase ellipsis. In Proceedings of the Workshop on Coreference Resolution Be- yond OntoNotes (CORBON 2016), pages 32-40, 2016.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Verb-phrase ellipsis and complex predicates in hindi-urdu",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Manetta",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Natural Language & Linguistic Theory",
                "volume": "",
                "issue": "",
                "pages": "1--39",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "E. Manetta. Verb-phrase ellipsis and complex predicates in hindi-urdu. Natural Language & Linguistic Theory, pages 1-39, 2018.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "A corpus-based study of Verb Phrase Ellipsis Identification and Resolution. PhD thesis",
                "authors": [
                    {
                        "first": "L",
                        "middle": [
                            "A"
                        ],
                        "last": "Nielsen",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "L. A. Nielsen. A corpus-based study of Verb Phrase Ellipsis Identification and Resolution. PhD the- sis, King's College London, 2005.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "The design for the wall street journal-based csr corpus",
                "authors": [
                    {
                        "first": "D",
                        "middle": [
                            "B"
                        ],
                        "last": "Paul",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "M"
                        ],
                        "last": "Baker",
                        "suffix": ""
                    }
                ],
                "year": 1992,
                "venue": "Proceedings of the workshop on Speech and Natural Language",
                "volume": "",
                "issue": "",
                "pages": "357--362",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. B. Paul and J. M. Baker. The design for the wall street journal-based csr corpus. In Proceed- ings of the workshop on Speech and Natural Lan- guage, pages 357-362. Association for Computa- tional Linguistics, 1992.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "When a good translation is wrong in context: Contextaware machine translation improves on deixis, ellipsis, and lexical cohesion",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Voita",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sennrich",
                        "suffix": ""
                    },
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Titov",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1905.05979"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "E. Voita, R. Sennrich, and I. Titov. When a good translation is wrong in context: Context- aware machine translation improves on deixis, ellipsis, and lexical cohesion. arXiv preprint arXiv:1905.05979, 2019.",
                "links": null
            }
        },
        "ref_entries": {}
    }
}