File size: 82,145 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
{
    "paper_id": "2006",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T01:11:21.163412Z"
    },
    "title": "THE CHALLENGE OF ARABIC FOR NLP/MT",
    "authors": [
        {
            "first": "Mohamed",
            "middle": [],
            "last": "Maamouri",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Pennsylvania",
                "location": {
                    "country": "USA"
                }
            },
            "email": "maamouri@ldc.upenn.edu"
        },
        {
            "first": "Ann",
            "middle": [],
            "last": "Bies",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Pennsylvania",
                "location": {
                    "country": "USA"
                }
            },
            "email": "bies@ldc.upenn.edu"
        },
        {
            "first": "Seth",
            "middle": [],
            "last": "Kulick",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Pennsylvania",
                "location": {
                    "country": "USA"
                }
            },
            "email": "skulick@ldc.upenn.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Arabic diacritization (referred to sometimes as vocalization or vowelling), defined as the full or partial representation of short vowels, shadda (consonantal length or germination), tanween (nunation or definiteness), and hamza (the glottal stop and its support letters), is still largely understudied in the current NLP literature. In this paper, the lack of diacritics in standard Arabic texts is presented as a major challenge to most Arabic natural language processing tasks, including parsing. Recent studies (Messaoudi, et al. 2004; Vergyri & Kirchhoff 2004; Zitouni, et al. 2006 and Maamouri, et al. forthcoming) about the place and impact of diacritization in text-based NLP research are presented along with an analysis of the weight of the missing diacritics on Treebank morphological and syntactic analyses and the impact on parser development.",
    "pdf_parse": {
        "paper_id": "2006",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Arabic diacritization (referred to sometimes as vocalization or vowelling), defined as the full or partial representation of short vowels, shadda (consonantal length or germination), tanween (nunation or definiteness), and hamza (the glottal stop and its support letters), is still largely understudied in the current NLP literature. In this paper, the lack of diacritics in standard Arabic texts is presented as a major challenge to most Arabic natural language processing tasks, including parsing. Recent studies (Messaoudi, et al. 2004; Vergyri & Kirchhoff 2004; Zitouni, et al. 2006 and Maamouri, et al. forthcoming) about the place and impact of diacritization in text-based NLP research are presented along with an analysis of the weight of the missing diacritics on Treebank morphological and syntactic analyses and the impact on parser development.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Arabic NLP research, focusing mainly on Modern Standard Arabic (MSA) in this paper, faces two major challenges, not necessarily shared with many other natural languages: the first is its complex linguistic structure and the second, the specific features of its orthographic system. Arabic is a highly inflected language which has as a rich and complex morphological system. Any given Arabic lemma has usually more than one word form to represent it, which includes a root, its internal structure, prefixes, suffixes and clitics. The internal structure itself includes short vowels and vocalic length, which together carry the bulk of the morphological and morphosyntactic structures, and a consonantal skeleton, which, as in other Semitic languages, bears the weight of the lexical (semantic) structure. The Arabic orthographic system uses superscript and subscript diacritical marks (or diacritics) for the representation of the three short vowels (a, i, u) , and four letters \u202b\u0627(\u202c 'alif, \u202b\u0649\u202c 'imaala, \u202b\u0648\u202c waaw, and \u202b\u064a\u202c yaa') to mark vocalic length. The 'imaala, an undotted form of the letter yaa', is used idiosyncratically for certain words ending in the long vowel [-aa] while the yaa' and the waaw, in addition to being consonants in their own right, function as glides or semi-vowels and are used to represent long [-uw] and long [iy] . Many of the MSA grammatical functions, such as verb passive forms and irregular noun plural forms, also use short vowels. Finally, short vowels are also used to indicate mood, aspect and voice endings for verbs and case endings for nouns. Moreover, long vowels are mostly used in derivation and word formation: as in kataba 'to write' vs. kAtaba 'to correspond with'. The shadda (consonantal length or gemination) is another important diacritic which is used for the derivation of new words. The hamza is used just to mark the existence of the glottal stop. Its major issue is its complex graphemic support system and the fact that the hamza is frequently omitted and sometimes misused. Finally, it is to be noted that no further mention will be made in this paper of the sukun, which uses a small superscript zero-shaped grapheme and is in fact nothing more than the absence of a vowel. The sukun does not add anything to the written text and is only used for syllabic identification in speech.",
                "cite_spans": [
                    {
                        "start": 949,
                        "end": 958,
                        "text": "(a, i, u)",
                        "ref_id": null
                    },
                    {
                        "start": 1170,
                        "end": 1175,
                        "text": "[-aa]",
                        "ref_id": null
                    },
                    {
                        "start": 1322,
                        "end": 1327,
                        "text": "[-uw]",
                        "ref_id": null
                    },
                    {
                        "start": 1337,
                        "end": 1341,
                        "text": "[iy]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "INTRODUCTION",
                "sec_num": "1"
            },
            {
                "text": "From the comprehensive description above, we see that Arabic is a relatively complex and difficult language to analyze, not so much because of its difficult linguistic structure but mostly because of how that structure is impacted and made more complex by the orthographic issues of its written form. The view presented in this paper is that the Arabic NLP scientific community should become more aware of some of these Arabic script issues and more discriminative of their role and responsibility in Arabic NLP work. The present paper will focus on the role and impact of diacritization on Arabic Treebank annotation and Arabic parsing.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "INTRODUCTION",
                "sec_num": "1"
            },
            {
                "text": "The issue of diacritization in Arabic arises as the result of a mismatch between the orthographic conventions that have developed for written MSA and the Arabic language itself, including spoken MSA, with respect to the amount of linguistic information represented. MSA is generally written without diacritics, but the language itself, and also spoken MSA, of course includes all of the features that the diacritics would represent (short vowels, consonantal gemination, etc.). When working with MSA for NLP purposes, several choices must be made. Will a particular effort focus on written or spoken MSA (the issue of dialects being largely beyond the scope of this paper)? Spoken MSA must be transcribed to facilitate downstream processes, so in either case, we will be dealing with text of some kind. Will the text be vocalized/diacritized or not (and if so, how will this be accomplished)? Will the text be in the Arabic script or in a transliterated/Romanized from?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "REALITY OF ARABIC SPEECH AND TEXT",
                "sec_num": "2"
            },
            {
                "text": "Each of these questions can be answered independently, and the answers will greatly affect the work to be done.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "REALITY OF ARABIC SPEECH AND TEXT",
                "sec_num": "2"
            },
            {
                "text": "It is obvious that when people speak MSA (or any dialectal Arabic) they must use the linguistic features of Arabic including its short vowels and other diacritical marks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Nature of Arabic Script-based Text",
                "sec_num": "2.1"
            },
            {
                "text": "Starting from recorded speech (as in MSA Broadcast News corpora), one has available and can therefore just transcribe all the existing phonetic data. The LDC CallHome Egyptian Colloquial Arabic (ECA) Corpus provides, in a Romanized form, an ASCII representation, which includes short vowels and other information relating to that which is transcribed with diacritics in the Arabic script. The ECA CallHome corpus has in fact all the information required for any NLP task, especially automatic speech recognition (ASR). As noted in Vergyri & Kirchhoff (2004) and , a Romanized transcription of Arabic could probably be an excellent way of providing full linguistic information to all Arabic research tasks if it weren't for its difficulty and the prohibitive training and annotation costs that it would entail. This excessive difficulty and the need for huge volume of data led to the inevitable use of what is readily available (namely, written Arabic script data) and made newswire Arabic the favourite and most used source data. Moreover, most acoustic material for Arabic ASR is in text also and the move to Arabic script based text followed Vergyri & Kirchhoff (2004) . A detailed account of the advantages and pitfalls of a two-prong Arabic orthography-based transcription is given in Maamouri et al. (2004b) , with the consonantal skeleton bare forms leading to a non-diacritized transcription (in the LDC Transcription using AMADAT's GREEN layer) and a fully diacritized one (in the same transcription tool's YELLOW layer). This last move by the Arabic NLP scientific community led to a closer look at diacritization.",
                "cite_spans": [
                    {
                        "start": 531,
                        "end": 557,
                        "text": "Vergyri & Kirchhoff (2004)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 1145,
                        "end": 1171,
                        "text": "Vergyri & Kirchhoff (2004)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 1290,
                        "end": 1313,
                        "text": "Maamouri et al. (2004b)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Nature of Arabic Script-based Text",
                "sec_num": "2.1"
            },
            {
                "text": "Though the use of diacritics is extremely important in setting up grammatical functions leading to acceptable text understanding and correct reading or analysis, diacritical markings are rarely present in real-world/life situations. It is true that they are rarely visible in out-of-school written documents and they do not, as a rule, appear in most printed materials in the Arab region. This predominant/generalized practice of graphemic under-representation of linguistic information concerns not only the short vowels but also the shadda (consonantal length) and the hamza (glottal stop). The use of diacritics is usually restricted to the early years of formal education and the sacred Koranic text and seems to be limited to whatever length of time is considered sufficient for the learner to be initiated to reading without the missing information (Maamouri 1998) . This generally amounts to about six years and most times more. Nowadays, vocalized Arabic text seems to be only used in pure deference to the needs of young and inexperienced learners. It is to be noted that diacritized MSA text does exist outside of the Koran in numerous sources, such as the rich and important heritage Arabic literature books. However, this source of diacritized data is not used by the NLP community, usually because neither the language nor the domains are of acceptable currency. Anywhere else, in Arabic newswire and in most other Arabic-script based transcriptions, the norm and 'real-world' data is Arabic non-diacritized text, whether MSA or dialectal.",
                "cite_spans": [
                    {
                        "start": 855,
                        "end": 870,
                        "text": "(Maamouri 1998)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Importance of Diacritics",
                "sec_num": "2.2"
            },
            {
                "text": "The loss of the internal diacritics (such as short vowels or shadda) leads to the following types of ambiguity, as exemplified in a given MSA lemma: \u202b\ufecb\ufee0\ufee2\u202c Elm. The situation of this specific bare graphemic form is as follows: (a) An ambiguity within 'core' POS tags, which distinguishes between the different lexical senses of the same 'core' POS tag. Example: The bare form \u202b\ufecb\ufee0\ufee2\u202c Elm can be diacritized as \u202b\u0650\ufee0\ufee2\u202c \u202b\ufecb\u202c Eilm (a noun meaning 'science, learning') or \u202b\u064e\ufee2\u202c \u202b\u064e\ufee0\u202c \u202b\ufecb\u202c Ealam, another noun meaning 'flag'. (b) A second type of 'core' POS tag ambiguity distinguishes between different lexical senses leading to different core POS tags. The same bare form \u202b\ufecb\ufee0\ufee2\u202c Elm, which was diacritized as two different nouns above, can additionally be diacritized as three different verb forms, all lexically and semantically connected. Example: \u064e \u202b\u0650\ufee2\u202c \u202b\u064e\ufee0\u202c \u202b\ufecb\u202c Ealima for 3 rd Person Masculine, Singular, Perfective Verb (MSA Verb Form I) meaning 'he learned/knew'; \u064e \u202b\u0650\ufee2\u202c \u202b\u064f\ufee0\u202c \u202b\ufecb\u202cEulima for 3 rd Person Singular, Passive Verb (MSA Verb Form I) meaning 'it/he was learned' and \u064e \u202b\ufc60\ufee2\u202c \u202b\u064e\ufee0\u202c \u202b\ufecb\u202c Eal~ama for the Intensifying, Causative, Denominative Verb (MSA Verb Form II) meaning 'he taught.' Ambiguities (a) and (b) are lexical/morphological ambiguities.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Diacritics and Ambiguity",
                "sec_num": "2.3"
            },
            {
                "text": "(c) Finally, a huge amount of ambiguity occurs at the structural/grammatical level, where the use of short vowels is correlated with case (nominal) and mood/aspect (verbal) information. This information is rendered by the use of one of six possible diacritics. Using the above example, we have the following:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Diacritics and Ambiguity",
                "sec_num": "2.3"
            },
            {
                "text": "\u064f \u202b\u0650\ufee0\ufee2\u202c \u202b\ufecb\u202c / \u064c \u202b\u0650\ufee0\ufee2\u202c \u202b\ufecb\u202c",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Diacritics and Ambiguity",
                "sec_num": "2.3"
            },
            {
                "text": "Eilmu/EilmN (NOM Noun + Definite and Indefinite), \u064e \u202b\u0650\ufee0\ufee2\u202c \u202b\ufecb\u202c / \u064b \u202b\u0650\ufee0\ufee4\ufe8e\u202c \u202b\ufecb\u202c Eilma/EilmAF (ACCU Noun + Definite and Indefinite) and \u0650 \u202b\u0650\ufee0\ufee2\u202c \u202b\ufecb\u202c / \u064d \u202b\u0650\ufee0\ufee2\u202c \u202b\ufecb\u202c Eilmi/EilmK (GEN Noun + Definite and Indefinite).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Diacritics and Ambiguity",
                "sec_num": "2.3"
            },
            {
                "text": "The loss, even partial, of diacritics frequently leads to a significant increase of linguistic ambiguity (both structural and lexical), which can only be resolved by contextual information and an adequate knowledge of the language. In order to be able to read/transcribe, annotators have to provide their own grammatical interpretations and bring to task considerable additional knowledge of syntax, vocabulary, and sometimes contextual interpretation in order to obtain correct and meaningful vocalizations which will allow them to reach acceptable word recognition and sense disambiguation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Diacritics and Ambiguity",
                "sec_num": "2.3"
            },
            {
                "text": "The considerable lexical ambiguity consequent to loss of diacritics is observed in Debili, et al. 2002, who calculate that an Arabic non-diacritized dictionary word form had 2.9 possible diacritized forms on average and that an Arabic text containing 23K word forms showed an average ratio of 1:11.6 (quoted in Vergyri & Kirchhoff 2004 ).",
                "cite_spans": [
                    {
                        "start": 311,
                        "end": 335,
                        "text": "Vergyri & Kirchhoff 2004",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Diacritics and Ambiguity",
                "sec_num": "2.3"
            },
            {
                "text": "When we look at the availability of Arabic text data, the situation boils down to the following:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "'Real-World' Arabic Text Data",
                "sec_num": "2.4"
            },
            {
                "text": "(a) Unvocalized/non-diacritized Arabic text for MSA (and even for newly written dialectal Arabic) seems to be the most available material for the speech research community and the main data source for all other NLP research needs (mostly in newswire form).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "'Real-World' Arabic Text Data",
                "sec_num": "2.4"
            },
            {
                "text": "(b) Since non-diacritized text prevails, the Arabic NLP community seems to have accepted using it as the de facto 'real world' information material without feeling an obligation to question its choice/use, even espousing the idea sometimes that the robustness of software algorithms can deal with the problem and reduce the negative effect of the missing information on their research.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "'Real-World' Arabic Text Data",
                "sec_num": "2.4"
            },
            {
                "text": "(c) While it is possible, as noted in Vergyri & Kirchhoff (2004) to collect available MSA (or even dialectal) public data and to transcribe it manually with full or partial restoration of the missing diacritics, obtaining thus an acceptably diacritized form, this practise has not been continued. The prohibitive cost and the usually unequal and questionable quality of human/manual diacritization have led the scientific Arabic NLP community and its sponsors to focus more on volume of unvoweled data so far.",
                "cite_spans": [
                    {
                        "start": 38,
                        "end": 64,
                        "text": "Vergyri & Kirchhoff (2004)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "'Real-World' Arabic Text Data",
                "sec_num": "2.4"
            },
            {
                "text": "(d) Most NLP Arabic research -even research dealing with diacritization -makes use of text-based information only and makes little use of diacritics even when they exist. No significant use is made of diacritics in the acoustic data -even when work starts from a speech source.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "'Real-World' Arabic Text Data",
                "sec_num": "2.4"
            },
            {
                "text": "The NLP scientific community is slowly becoming aware of the vital importance of diacritization in Arabic text-based research. The problem posed to the Arabic NLP community is how to diacritize text data in order to reach full linguistic information and better research results. In the remaining sections, a brief presentation of the place and impact of diacritization in text-based NLP research will be reviewed, and an analysis of the weight of the missing diacritics on Treebank morphological and syntactic analyses and the impact on parser development will be given.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "DIACRITIZATION RESEARCH REVIEWED",
                "sec_num": "3"
            },
            {
                "text": "A look at current research shows that it is possible to restore/recover and provide much diacritization information automatically or semi-automatically, if it is not included in the transcription text -via manual annotation, from the annotator's 'virtual knowledge' or from available acoustic information (as heard in MSA Broadcast News). If the source is text-based information, diacritization could be determined from morphological and contextual knowledge. Whether one is working from the perspective of Acoustic Modelling -mainly ASR systems -or from that of Language Modelling, the available knowledge sources that could be used for most appropriate diacritization of a scriptbased Arabic text forms are the following:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "DIACRITIZATION RESEARCH REVIEWED",
                "sec_num": "3"
            },
            {
                "text": "\u2022 analysis of the morphological structure: segmentation of words/lemmas into stems, roots and patterns \u2022 consideration of the syntactic context in which the word/lemma form occurs \u2022 knowledge added from the context of speech/acoustic data accompanying the transcription.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "DIACRITIZATION RESEARCH REVIEWED",
                "sec_num": "3"
            },
            {
                "text": "A fully automatic approach to diacritization (reported in Vergyri & Kirchhoff 2004) was presented in Gal (2002) . Gal used an HMM-based bigram model, which was used for decoding diacritized sentences from non-diacritized sentences. Gal applied this technique to the Koranic text, achieving 14% word error. Kirchhoff et al. (2002) and Lamel (2003) : Since the LDC CallHome ECA corpus was distributed with both Romanized and script-based transcriptions, the above work compared error rates of recognizers trained on both transcriptions. The results show that the loss of information due to training on script forms is significantly worse with a relative increase of 10% in Word Error Rate (WER). Kirchhoff et al. (2002) show that the MSA WER ranged from 9% to 28%, depending on whether or not case endings were counted.",
                "cite_spans": [
                    {
                        "start": 58,
                        "end": 83,
                        "text": "Vergyri & Kirchhoff 2004)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 101,
                        "end": 111,
                        "text": "Gal (2002)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 306,
                        "end": 329,
                        "text": "Kirchhoff et al. (2002)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 334,
                        "end": 346,
                        "text": "Lamel (2003)",
                        "ref_id": null
                    },
                    {
                        "start": 694,
                        "end": 717,
                        "text": "Kirchhoff et al. (2002)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Automatic Diacritization of Training Data /Arabic Text Research",
                "sec_num": "3.1"
            },
            {
                "text": "Vergyri & Kirchhoff (2004) indicate that lower diacritization error rates were produced when more linguistic information is included (morphological and syntactic context) in combination with acoustics. Nonetheless, the best word error rate reported for diacritizing Egyptian Arabic was 41.6%. The authors mention that they intend to apply knowledge-poor diacritization procedures to dialects of Arabic for which morphological analyzers do not exist in their future work. Safadi, et al. (2006) report on an unsupervised method of diacritizing that builds on a combination of automatic tagging and manually written rules. They report an error rate of 10-20%, as determined by expert evaluation of their output. Nelken and Shieber (2005) use a weighted finite-state transducers for diacritic restoration, including also the separation of clitics. They report a diacriticization error rate of 12.79% and word error rate of 23.61% when including the case endings. Disregarding the case endings for restoration and evaluation improves the scores to 6.35% and 7.33%. Zitouni, et al. (2006) report on a maximum entropy approach to restoring a comprehensive list of diacritics, treating the problem as one of sequence classification.",
                "cite_spans": [
                    {
                        "start": 8,
                        "end": 26,
                        "text": "& Kirchhoff (2004)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 471,
                        "end": 492,
                        "text": "Safadi, et al. (2006)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 709,
                        "end": 734,
                        "text": "Nelken and Shieber (2005)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 1060,
                        "end": 1082,
                        "text": "Zitouni, et al. (2006)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Automatic Diacritization of Training Data /Arabic Text Research",
                "sec_num": "3.1"
            },
            {
                "text": "They report a diacritic error rate of 5.1% (and a word error rate of 17.3%) in a fully diacritized setting. They also find a significant improvement in error rates when case endings are disregarded, with the scores improving to 2.2% and 7.2%. This study did not make any mention of the hamza, especially on whether its inconsistent and sometimes lacking representation had any appreciable impact on the final results. One notable aspect of this work is that they used a part of speech tagger to generate tags used as features in the model, and they report that this improves the score. Given the connection between some diacritic restoration and part of speech tags, as discussed in Section 2.3, this is not surprising. However, they do not discuss what POS tag set is used, making it difficult to understand precisely how their approach captures this relationship. Messaoudi, et al. (2004) found that it was useful to transcribe short vowels in order to build acoustic models for Arabic broadcast news. They report a 10% improvement in Arabic WER with the inclusion of diacritics.",
                "cite_spans": [
                    {
                        "start": 866,
                        "end": 890,
                        "text": "Messaoudi, et al. (2004)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Automatic Diacritization of Training Data /Arabic Text Research",
                "sec_num": "3.1"
            },
            {
                "text": "In our Arabic Treebank effort at LDC, MSA text data has so far been morphologically and syntactically been annotated using the following combination of semi-automatic and manual/human annotation:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semi-Automatic Diacritization of Arabic Text",
                "sec_num": "3.2"
            },
            {
                "text": "\u2022 Morphological knowledge provided by Buckwalter analyzer (stemmer) -1.5% of unknown word forms (typos, named entities, 'not in lexicon,' etc.) (Buckwalter 2002) \u2022 Probabilistic contextual knowledge provided by tagger (still in development)",
                "cite_spans": [
                    {
                        "start": 144,
                        "end": 161,
                        "text": "(Buckwalter 2002)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semi-Automatic Diacritization of Arabic Text",
                "sec_num": "3.2"
            },
            {
                "text": "\u2022 Word forms with full case endings provided by human annotation \u2022 Syntactic (tree structure) output provided by the Bikel implementation of the Collins parser (publicly available at http://www.cis.upenn.edu/~dbikel/), which is being enhanced by a University of Pennsylvania/BBN team of researchers (Kulick, et al. forthcoming) \u2022 Corrections, final trees and final word forms provided by human annotators",
                "cite_spans": [
                    {
                        "start": 299,
                        "end": 327,
                        "text": "(Kulick, et al. forthcoming)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semi-Automatic Diacritization of Arabic Text",
                "sec_num": "3.2"
            },
            {
                "text": "We did not initially have all the tools to address the problem of diacritization when the LDC Arabic Treebank project began at the end of 2001. We decided to annotate our first Arabic Treebank segment, ATB: Part 1 (also known as the AFP Corpus), by having annotators supply word-internal lexical identity vocalization only, because that is how people normally read Arabic -taking/assuming the normal risks taken by all Arabic readers, with the assumption that any interpretation of the case or mood chosen would be the acceptable interpretation of an educated native speaker/annotator.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semi-Automatic Diacritization of Arabic Text",
                "sec_num": "3.2"
            },
            {
                "text": "In our second Arabic Treebank segment, ATB: Part 2 (also known as the UMAAH Corpus), we decided that it would improve annotation and the overall usefulness of the corpus to complete the vocalization of the morphologically analyzed texts by adding the necessary case and mood endings at the Part-of-speech + Gloss (MPG) level of annotation. Up to six case endings (-a, -u, -i for definite and N, AF, K for indefinite nominals) were automatically added in Tim Buckwalter's Morphological Analyzer (BAMA) output alternatives. Our annotators had to select not only the correct/appropriate word-internal lexical identity vocalization for each targeted token from the lexicon-based set of output analyses provided by BAMA, but they also had to make sure that the selection incorporated the correct case and mood endings.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semi-Automatic Diacritization of Arabic Text",
                "sec_num": "3.2"
            },
            {
                "text": "For our third treebank segment, ATB: Part 3 (also known as the ANNAHAR Corpus), we decided to fully vocalize the text, adding the final missing piece, mood and voice endings for verbs in all of the alternatives presented by the BAMA output to our POS annotators.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semi-Automatic Diacritization of Arabic Text",
                "sec_num": "3.2"
            },
            {
                "text": "The LDC experience shows that the Arabic Treebank research team had to adjust its initial MSA Treebank annotation in order to include a 'diacritization' which related mostly to the addition of inflectional endings on top of the full short vowel representation (mostly provided by BAMA) in its morphological layer of analysis . The ensuing fully diacritized POS output was used by syntactic annotators and provided them with a single interpretation (which is the result of word-disambiguation) that they had to accept/confirm or contest for the syntactic layer. The availability of a fully diacritized text, always present in the syntactic annotation tool, made the syntactic annotation task easier and decreased the annotation responsibility load, leading hopefully to more annotation consistency, less time on the job and less annotation stress in general. It is our belief that the results of the morphological and POS annotation and word disambiguation used in all consequent segments of the Arabic Treebank led to a scientifically sound methodology for diacritizing bare MSA text. Although we did not originally set out with this goal, as a result of annotation necessity and through the morphological annotation process, we have now produced nearly a million words of diacritized MSA newswire text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semi-Automatic Diacritization of Arabic Text",
                "sec_num": "3.2"
            },
            {
                "text": "The first experiments on parsing the Arabic Treebank were reported in Bikel (2004) , for a small section of the early work on the ATB from the AFP section, consisting of 149K tokens for training and 11K for testing. The results were a recall/precision of 75.4/76.0 for sentences of length <=40, and 72.5/73.4 for all sentences.",
                "cite_spans": [
                    {
                        "start": 70,
                        "end": 82,
                        "text": "Bikel (2004)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "This work used a combination of the bare data, together with the morphological Parts of Speech tags resulting from the POS annotation step. The large size of the POS tagset was problematic for the parser, since it fragmented the data in a way that the current configuration of the parser was not well suited for. In addition, it was always possible for \"new\" tags that the parser had not seen in the training data to be created by a new combination of affixes. Therefore, the tagset was mapped down to a reduced number of tags by Ann Bies.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "The reason given by Bikel for using the bare data is directly relevant to the concerns of this paper: \"We only used the unvoweled data, because that would ultimately be necessary for any real-world [defined here as an NLP situation in which the text is all too often bare] parser. However, because the purpose was for bootstrapping treebank annotation, we used the gold-standard, Bies-mapped part-of-speech tags.\" (p. 80, Bikel 2004) There are two points that should be made about this:",
                "cite_spans": [
                    {
                        "start": 422,
                        "end": 433,
                        "text": "Bikel 2004)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "(1) It is not necessarily true that the bare data would be the input to the parser for a real-world parser. There could be a preprocessing step reconstructing the diacritics, or at least some of them. Also, the bare data and corresponding tags are already the result of a certain amount of morphological analysis and tokenization, since various clitics have been separated from their tokens as they originally appeared in the \"real-world\" data, such as conjunction prefixes and pronoun suffixes.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "(2) Even though the \"Bies-mapped\" tags are being used by the parser, the gold POS tags are still taken as the input before that mapping. The gold tags, however, correspond to the diacritized data. It is in fact reasonable to say that the bare data together with the gold POS tags carries the same information as the diacritized data. This therefore seems to us to be an inconsistency in the original parsing experiments. (Due to the lack of a part-of-speech tagger at the time, however, there was no choice but to use the gold part-of-speech tags.)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "In general, the role of diacritics in a NLP pipeline that includes parsing is very much an open question. It seems fair to say that the end process of an NLP pipeline that aims to extract some sort of syntactic structure and semantic meaning from the initial text should have sufficient \"understanding\" that it can restore the diacritics to the text by the end of the pipeline. This however does not speak to the question of at what point in the pipeline diacritics should be restored, and the possibility that different diacritics perhaps should be restored at different times. The steps taken for diacritic restoration for manual annotation may not be the appropriate ones for an NLP pipeline.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "There are two aspects to the problem of how the parser might utilize diacritic information.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "One question concerns what diacritic information might be useful for the parser. While the earlier work, as mentioned above, used the bare text, there has been very little work examining whether a parser can make use of vocalized text. The Arabic Treebank now gives us a corpus to carry out such experiments, as we discuss in the following subsection.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "However, in addition to exploring which diacritic information is useful for the parser, we must also be concerned with what might be available to the parser outside the context of these experiments and outside the context of Treebank research. As discussed in Section 3, there is a growing body of work on diacritic restoration. However, Zitouni, et al. (2006) and Nelken & Shieber (2005) both report that it is much harder to restore the diacritics representing case information. This is not surprising, since as Nelken & Shieber write, \"including case information naturally yields proportionally worse accuracy. Since case markings encode higher-order grammatical information, they would require a more powerful grammatical model than offered by finite state methods\".",
                "cite_spans": [
                    {
                        "start": 338,
                        "end": 360,
                        "text": "Zitouni, et al. (2006)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 365,
                        "end": 388,
                        "text": "Nelken & Shieber (2005)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "While noting these concerns and issues, for the experiments reported here we continue to train and test the data as it is tokenized following the stage of POS annotation, and we continue to use the gold tags, while varying the level of vocalization in the words. We recognize that in any actual use of the parser on bare text, there would need to be a level of preprocessing for tokenization and perhaps some level of diacritic restoration. For the current work, however, our goal is to utilize the diacritization that is present in the ATB for understanding how different amounts of vocalization affect the parser.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "PARSER DEVELOPMENT: HOW DOES DIACRITIZATION IMPACT PARSING?",
                "sec_num": "4"
            },
            {
                "text": "We have shifted the parser experiments to use ATB3 (Annahar) instead of the data Bikel was working with. The ATB3 corpus is larger than ATB1, and was done after more experience was gained with the annotation process. We did a balanced 80/10/10 split on the corpus for training/development/test, so that eight out of every 10 sentences was used for training, and so on.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "The input to the parser was varied in four ways:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "Bare: is using the bare data (ATB3 Annahar) Full: is using the fully diacritized form of the words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "No Case: is using a modified form of the diacritized data, in which the case endings for nouns and adjectives have been deleted. This was based on the hypothesis that the extra case endings were causing data fragmentation. For example, suppose the bare word is \u064b \u202b\u0652\u0622\ufef4\ufe8e\u202c \u202b\u0623\ufee3\ufef4\ufeae\u202c \">myrokyAF\", with the diacritized form \u064b \u202b\u0651\ufe8e\u202c \u202b\u0650\ufef4\u202c \u202b\u0652\u0622\u202c \u202b\u0650\ufef4\ufeae\u202c \u202b\u064e\ufee3\u202c \u202b\u0623\u202c \">amiyrokiy~+AF\" (American+ [acc.indef] As discussed in Bikel (2004) , the parser considers all words occurring below some threshold (here, 6) to be \"unknown\" words. This is somewhat of a misnomer, as they can be considered rare words, for which their lexical information is not stored the same way as with more common words. The idea is that this will help the handling of words during parsing that were not seen during training, which will also be treated as \"unknown\" words. Table 1 shows the # of words and instances (i.e., word types and word tokens) encountered during training, along with the % of unknown words and % unknown instances. To compare these numbers for the ATB to the Penn Treebank, we also trained the parser on a training size of the Penn Treebank roughly the same as our training size for the ATB.",
                "cite_spans": [
                    {
                        "start": 376,
                        "end": 387,
                        "text": "[acc.indef]",
                        "ref_id": null
                    },
                    {
                        "start": 404,
                        "end": 416,
                        "text": "Bikel (2004)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 826,
                        "end": 833,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "One thing to note is the interestingly high percentage of unknown words, even for the Wall Street Jounal (WSJ). This is no doubt because of the large number of nouns that occur only one or two times, for the \"Bare\" Arabic run, the % of unknown words is roughly the same, while the % of unknown instances is significantly higher than for the WSJ, and indeed the number of words is significantly higher as well, although the number of instances is roughly the same. This indicates a somewhat flatter distribution of words in the ATB. Switching to the \"Full\" run, the situation worsens, as expected.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "Since the words now include the full tokenization, including the case endings, a single unvocalized entry can now be shattered into many different fully vocalized forms. The # of words for the same # of instances increases from 32204 to 43141, with an increase to 21.24% of the unknown instances.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "However, using the \"No Case\" version of the diacriticized data, the data sparseness is actually less than with the bare form, and the same is even more true when the mood suffixes are eliminated as well. Our suspicion is that this is because the diacritized form, in addition to adding the diacritics, also normalizes the orthographic variations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "For example, the lemma [<ihAnap_1] (insult/contempt) occurs 4 times with a determiner prefix and feminine singular ending, but with three different spellings: Al<hAnp (1 time), AlAhAnp (2 times), and Al<hAn_p (1 time). The diacritized versions of these are all the same, however: Al+<ihAn+ap (leaving aside the case endings). So there are likely many cases in which a word that appears in different orthographic forms, each of which is \"unknown\", can be unified as one entry in the diacriticized form and thus not treated as an \"unknown word\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "We also performed parsing runs for each of the four models, to see what sort of effect the different training and test input had on the overall score. In general, there is a significant difference in the score for sentences <=40 and all sentences, more so than in the initial Bikel results, perhaps because the sentences in ATB3 have more conjunction at the S level than in the ATB1, although we have not confirmed that. We include testing results only for sentences <=40, because it is then much faster to run the parser. Our concern is looking at differences in parsing results, with the assumption that the scores for including all sentences will continue to lag behind.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "For the experiments, as mentioned above, we are continuing to use the gold POS tags in the \"Bies-reduced\" form, while recognizing the inherent limits of this approach, as discussed above. The results of the four runs are shown in Table 2 . As has been generally recognized, the scores for ATB are much lower than for the WSJ. The results for our baseline with the \"Bare\" data are somewhat different from those reported in Bikel (75.4/76.0), for two reasons: (1) Different corpora are being used -ATB3 instead of ATB1, and (2) We are taking advantage of various improvements to the parser that are discussed in a separate paper.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 230,
                        "end": 237,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Parser Experiments",
                "sec_num": "4.1"
            },
            {
                "text": "As can be seen in Table 2 above, there is little difference in the parsing results. Given all of the open questions around the status of parsing Arabic, it is not entirely clear how to evaluate these results. Still, there is little change in the overall scores, and there is no correlation between the improvement in unknown word frequency and parsing improvement.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 18,
                        "end": 25,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Discussion of Parser Experiments",
                "sec_num": "4.2"
            },
            {
                "text": "It is clear that the parser needs a substantial overhaul as to its handling of morphological information. For example, nouns with and without a determiner Al-prefix are treated as two separate words, with no relation to each other. This is clearly wrong. But even that aside, there are several ways to expand on this work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion of Parser Experiments",
                "sec_num": "4.2"
            },
            {
                "text": "One aspect that needs to be investigated more is the relation between the diacritics and the Part of Speech tags. As discussed in Section 2, diacritics can distinguish between different 'core' POS tags (and so different tags in the reduced tag set used by the parser.) It is likely that the gold Part of Speech tags, even in their reduced form, are masking some of the benefits of using the diacritized data. For example, \"ktb\" occurs 56 times in the corpus, 17 as NOUN, 6 as PV_PASS (passive perfective verb), and 33 as PV (perfective verb). As a NOUN, it has the vowels \"kutub\", as a PV_PASS it has \"kutib\", and as PV it has \"katab\". Therefore, in this case the vowelization does not add much beyond the Part of Speech tags. It is possible therefore that the full benefits of vocalization can only be seen in the context of a wider NLP pipeline than just the parser, including in particular the Part-of-Speech tagger. Related to this, Habash (2005) reports a drop in ambiguity when considering tokens only within the same Part of Speech tag. Also notable in the connection is the work of Habash & Rambow (2005) , who describe an integrated approach to tokenization, Part of Speech tagging and morphological disambiguation. Of particular interest is the close connection between POS tagging and morphological disambiguation. While this connection is clearly related to the concerns expressed here, they do not include a step of diacritazation.",
                "cite_spans": [
                    {
                        "start": 937,
                        "end": 950,
                        "text": "Habash (2005)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 1090,
                        "end": 1112,
                        "text": "Habash & Rambow (2005)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion of Parser Experiments",
                "sec_num": "4.2"
            },
            {
                "text": "In our view, the categorization of the ambiguities resolved by diacritic restoration discussed in Section 2 deserves detailed empirical analysis based on the data the parser is using. As just discussed, there is a close connection between POS tags and some diacritic restoration (ambiguity (b) in Section 2.3). Another type of ambiguity is that within 'core' POS tags, distinguishing for example between two different nouns (ambiguity (a) in Section 2.3). Since so many nouns occur infrequently enough that they are categorized as \"unknown\" by the parser, this will probably make less of a difference. However, this may account for some of the flatness of the distribution of words in the ATB, as discussed above. The utility of the case and mood/aspect markers (ambiguity (c) in Section 2.3) for the parser is even more of an open question, and it seems reasonable that such restoration should take place as a byproduct of the parsing process, rather than as a preprocessing step.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion of Parser Experiments",
                "sec_num": "4.2"
            },
            {
                "text": "It seems worthwhile to explore this point more, to determine not just what the ambiguity of unvocalized forms is, but also what kinds of ambiguity matter the most to our research.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Discussion of Parser Experiments",
                "sec_num": "4.2"
            },
            {
                "text": "The role of diacritization in the annotation process for the Arabic Treebank is now firmly established, and this data has been available and quite useful to the scientific community. In general, however, the correct way to utilize diacritization in various Natural Language Processing tasks is more of an open question. In Section 4 we have described some initial experiments exploring the role of diacritics in parsing. In our view, one of the primary tasks for this line of investigation is a more systematic investigation of the ambiguities that different diacritics resolve, and their interaction with the Part of Speech tags.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CONCLUSION",
                "sec_num": "5"
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "On the Parameter Space of Lexicalized Statistical Parsing Models",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Bikel",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Bikel. (2004). On the Parameter Space of Lexicalized Statistical Parsing Models. Ph.D. Dissertation. University of Pennsylvania.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Arabic Morphological Analyzer Version 1.0. Linguistic Data Consortium catalog number LDC2002L49",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Buckwalter",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "T. Buckwalter. (2002). Arabic Morphological Analyzer Version 1.0. Linguistic Data Consortium catalog number LDC2002L49, ISBN 1-58563-257-0.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "An HMM Approach to Vowel Restoration in Arabic and Hebrew",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "ACL-02 Workshop on Computational Approaches to Semitic Languages",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Y. Gal. (2002). 'An HMM Approach to Vowel Restoration in Arabic and Hebrew'. In ACL-02 Workshop on Computational Approaches to Semitic Languages.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Introduction to Arabic Natural Language Processing",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Habash",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "N. Habash. (2005). 'Introduction to Arabic Natural Language Processing'. ACL Tutorial.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Habash",
                        "suffix": ""
                    },
                    {
                        "first": "O",
                        "middle": [],
                        "last": "Rambow",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "N. Habash and O. Rambow. (2005). 'Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop'. ACL 2005.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Novel Approaches to Arabic Speech Recognition -Final report of the JHU Summer workshop",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Kirchhoff",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "K. Kirchhoff, et al. (2002). Novel Approaches to Arabic Speech Recognition -Final report of the JHU Summer workshop. John Hopkins University.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "forthcoming). 'Parsing the Arabic Treebank: Analysis and Improvements",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Kulick",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Gabbard",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Marcus",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Kulick, R. Gabbard and M. Marcus. (2006, forthcoming). 'Parsing the Arabic Treebank: Analysis and Improvements'. University of Pennsylvania.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Arabic Diglossia and its Impact on the quality of education in the Arab region'. Discussion paper prepared for World Bank at \"The Mediterranean Development Forum",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Maamouri",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "World Bank",
                "volume": "",
                "issue": "",
                "pages": "3--6",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Maamouri. (1998). 'Arabic Diglossia and its Impact on the quality of education in the Arab region'. Discussion paper prepared for World Bank at \"The Mediterranean Development Forum.\" World Bank, Marrakech, 3-6 September 1998. http://www.worldbank.org/wbi/mdf/mdf2/papers/humandev/education/maamour i.pdf",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Developing an Arabic Treebank: Methods, Guidelines, Procedures, and Tools",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Maamouri",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Bies",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of the 20th International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Maamouri and A. Bies. (2004). 'Developing an Arabic Treebank: Methods, Guidelines, Procedures, and Tools'. In Proceedings of the 20th International Conference on Computational Linguistics (COLING 2004).",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "The Penn Arabic Treebank: Building a Large-Scale Annotated Arabic Corpus",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Maamouri",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Bies",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Buckwalter",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Mekki",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of NEMLAR",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Maamouri, A. Bies, T. Buckwalter and W. Mekki. (2004). 'The Penn Arabic Treebank: Building a Large-Scale Annotated Arabic Corpus'. In Proceedings of NEMLAR 2004.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "forthcoming). 'Diacritization in Arabic Treebank Annotation and Parsing",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Maamouri",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Bies",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Kulick",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Maamouri, A. Bies and S. Kulick. (2006, forthcoming). 'Diacritization in Arabic Treebank Annotation and Parsing'. University of Pennsylvania.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Dialectal Arabic Orthographybased Transcription & CTS Levantine Arabic Collection",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Maamouri",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Graff",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Jin",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Cieri",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "EARS PI Meeting and RT-04 Workshop, IBM Executive Conference Center",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M. Maamouri, D. Graff, H. Jin and C. Cieri. (2004b). 'Dialectal Arabic Orthography- based Transcription & CTS Levantine Arabic Collection'. EARS PI Meeting and RT-04 Workshop, IBM Executive Conference Center, Palisades, NY, USA. November 7-11, 2004. www.sainc.com/richtrans2004/",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The LIMSI RT-04 BN Arabic System",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Messaoudi",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Lamel",
                        "suffix": ""
                    },
                    {
                        "first": "J-L",
                        "middle": [],
                        "last": "Gauvain",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of the EARS RT-04 Workshop",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A. Messaoudi, L. Lamel and J-L. Gauvain. (2004). 'The LIMSI RT-04 BN Arabic System'. Proceedings of the EARS RT-04 Workshop, 29.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Arabic diacritization using weighed finite-state transducers'",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Nelken",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [
                            "M"
                        ],
                        "last": "Shieber",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the ACL Workshop on Computational Approaches to Semitic Languages",
                "volume": "",
                "issue": "",
                "pages": "79--86",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R. Nelken and S. M. Shieber. (2005). 'Arabic diacritization using weighed finite-state transducers'. In Proceedings of the ACL Workshop on Computational Approaches to Semitic Languages. Association for Computational Linguistics. Ann Arbor, Michigan. URL {http://www.aclweb.org/anthology/W/W05/W05-0711}: 79-86.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Computational Methods to Vocalize Arabic Texts",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Safadi",
                        "suffix": ""
                    },
                    {
                        "first": "O",
                        "middle": [],
                        "last": "Dakkak",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Ghneim",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Second Workshop on Internationalizing SSML",
                "volume": "",
                "issue": "",
                "pages": "30--31",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "H. Safadi, O. Dakkak and N. Ghneim. (2006). 'Computational Methods to Vocalize Arabic Texts'. Second Workshop on Internationalizing SSML, Crete, 30-31 May 2006.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Automatic diacritization of Arabic for Acoustic Modeling in Speech Recognition",
                "authors": [
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Vergyri",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Kirchhoff",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Proceedings of the Workshop on Computational Approaches to Arabic Script-based Languages. COLING",
                "volume": "",
                "issue": "",
                "pages": "66--73",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. Vergyri and K. Kirchhoff. (2004). 'Automatic diacritization of Arabic for Acoustic Modeling in Speech Recognition'. In Proceedings of the Workshop on Computational Approaches to Arabic Script-based Languages. COLING 2004, Geneva: 66-73.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Maximum Entropy Based Restoration of Arabic Diacritics",
                "authors": [
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Zitouni",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "S"
                        ],
                        "last": "Sorensen",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sarikaya",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "577--584",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "I. Zitouni, J. S. Sorensen and R. Sarikaya. (2006). 'Maximum Entropy Based Restoration of Arabic Diacritics'. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics. July 2006 Workshop on Computational Approaches to Semitic Languages. Association for Computational Linguistics. Sydney, Australia. URL {http://www.aclweb.org/anthology/P/P06/P06-1073}: 577-584.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF0": {
                "type_str": "table",
                "text": ") and tag ADJ+CASE_INDEF_ACC. For the No Case run, this would be treated as \u0651 \u202b\u0650\ufef2\u202c \u202b\u0652\u0622\u202c \u202b\u0650\ufef4\ufeae\u202c \u202b\u064e\ufee3\u202c \u202b\u0623\u202c \">amiyrokiy~\".",
                "content": "<table><tr><td colspan=\"5\">No Case, Mood: is a more extreme form of the No Case run, in which the</td></tr><tr><td colspan=\"3\">mood suffixes have been deleted as well.</td><td/><td/></tr><tr><td/><td>#Words</td><td>#Instances</td><td>%Unknown</td><td>% Unknown</td></tr><tr><td/><td/><td/><td>words</td><td>instances</td></tr><tr><td>WSJ English</td><td>23333</td><td>296872</td><td>79.43%</td><td>10.61%</td></tr><tr><td>Bare Annahar</td><td>32204</td><td>296748</td><td>80.37%</td><td>15.42%</td></tr><tr><td>MSA</td><td/><td/><td/><td/></tr><tr><td>Full</td><td>43141</td><td>296748</td><td>84.95%</td><td>21.24%</td></tr><tr><td>No Case</td><td>31524</td><td>296748</td><td>80.28%</td><td>14.94%</td></tr><tr><td>No Case,</td><td>30454</td><td>296748</td><td>79.31%</td><td>14.45%</td></tr><tr><td>Mood</td><td/><td/><td/><td/></tr></table>",
                "html": null,
                "num": null
            },
            "TABREF1": {
                "type_str": "table",
                "text": "Unknown word frequencies.",
                "content": "<table/>",
                "html": null,
                "num": null
            }
        }
    }
}