File size: 95,398 Bytes
6fa4bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:06:27.460141Z"
    },
    "title": "OSACT4 Shared Task on Offensive Language Detection: Intensive Preprocessing-Based Approach",
    "authors": [
        {
            "first": "Fatemah",
            "middle": [],
            "last": "Husain",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Kuwait University",
                "location": {
                    "country": "State of Kuwait"
                }
            },
            "email": "f.husain@ku.edu.kw"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "The preprocessing phase is one of the key phases within the text classification pipeline. This study aims at investigating the impact of the preprocessing phase on text classification, specifically on offensive language and hate speech classification for Arabic text. The Arabic language used in social media is informal and written using Arabic dialects, which makes the text classification task very complex. Preprocessing helps in dimensionality reduction and removing useless content. We apply intensive preprocessing techniques to the dataset before processing it further and feeding it into the classification model. An intensive preprocessing-based approach demonstrates its significant impact on offensive language detection and hate speech detection shared tasks of the fourth workshop on Open-Source Arabic Corpora and Corpora Processing Tools (OSACT). Our team wins the third place (3 rd) in the Sub-Task A Offensive Language Detection division and wins the first place (1 st) in the Sub-Task B Hate Speech Detection division, with an F1 score of 89% and 95%, respectively, by providing the state-of-the-art performance in terms of F1, accuracy, recall, and precision for Arabic hate speech detection.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "The preprocessing phase is one of the key phases within the text classification pipeline. This study aims at investigating the impact of the preprocessing phase on text classification, specifically on offensive language and hate speech classification for Arabic text. The Arabic language used in social media is informal and written using Arabic dialects, which makes the text classification task very complex. Preprocessing helps in dimensionality reduction and removing useless content. We apply intensive preprocessing techniques to the dataset before processing it further and feeding it into the classification model. An intensive preprocessing-based approach demonstrates its significant impact on offensive language detection and hate speech detection shared tasks of the fourth workshop on Open-Source Arabic Corpora and Corpora Processing Tools (OSACT). Our team wins the third place (3 rd) in the Sub-Task A Offensive Language Detection division and wins the first place (1 st) in the Sub-Task B Hate Speech Detection division, with an F1 score of 89% and 95%, respectively, by providing the state-of-the-art performance in terms of F1, accuracy, recall, and precision for Arabic hate speech detection.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Online offensive language detection is one of the most challenging text classification tasks to accomplish due to the ambiguity and informality of the language used in social media platforms. So far, online offensive language detection has been applied to various languages, such as English (Kwok and Wang, 2013; Davidson et al., 2017; Nobata et al., 2016; Pitsilis, Ramampiaro, Langseth, 2018) , German (Kent, 2018; Wiedemann et al., 2018) , Urdu (Mustafa et al., 2017) , Turkish (\u00d6zel et al., 2017) , Hindi (Bohra et al., 2018; Kapoor et al., 2018) , Danish (Derczynski, 2019) , and Arabic (Abozinadah, Mbaziira, and Jones, 2015; Mubarak, Darwish, and Magdy, 2017; Alakrot, Murray, and Nikolov, 2018; Mohaouchane, Mourhir, and Nikolov, 2019) . Regardless of the text language, a standard text classification pipeline consists of preprocessing, feature extraction, feature selection, and classification model. The preprocessing phase is the one that is the most distinguishable phase among the others based on the text language. Each language contains unique structures and rules, which need to be addressed using unique methods. We develop an intensive preprocessingbased classification model for Arabic offensive language detection. Among the participants of the fourth workshop on Open-Source Arabic Corpora and Corpora Processing Tools (OSACT) throughout the shared tasks, our team wins third place in Sub-Task A Offensive Language Detection division and wins first place in Sub-Task B Hate Speech Detection division.",
                "cite_spans": [
                    {
                        "start": 291,
                        "end": 312,
                        "text": "(Kwok and Wang, 2013;",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 313,
                        "end": 335,
                        "text": "Davidson et al., 2017;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 336,
                        "end": 356,
                        "text": "Nobata et al., 2016;",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 357,
                        "end": 394,
                        "text": "Pitsilis, Ramampiaro, Langseth, 2018)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 404,
                        "end": 416,
                        "text": "(Kent, 2018;",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 417,
                        "end": 440,
                        "text": "Wiedemann et al., 2018)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 448,
                        "end": 470,
                        "text": "(Mustafa et al., 2017)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 473,
                        "end": 500,
                        "text": "Turkish (\u00d6zel et al., 2017)",
                        "ref_id": null
                    },
                    {
                        "start": 509,
                        "end": 529,
                        "text": "(Bohra et al., 2018;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 530,
                        "end": 550,
                        "text": "Kapoor et al., 2018)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 560,
                        "end": 578,
                        "text": "(Derczynski, 2019)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 592,
                        "end": 631,
                        "text": "(Abozinadah, Mbaziira, and Jones, 2015;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 632,
                        "end": 666,
                        "text": "Mubarak, Darwish, and Magdy, 2017;",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 667,
                        "end": 702,
                        "text": "Alakrot, Murray, and Nikolov, 2018;",
                        "ref_id": null
                    },
                    {
                        "start": 703,
                        "end": 743,
                        "text": "Mohaouchane, Mourhir, and Nikolov, 2019)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "Text that contains some form of abusive behavior, exhibiting actions with the intention of harming others, is known as offensive language. This abusive behavior could lead to disturbances, disrespect, harm, insults, and anger, thus affecting the harmony of conversations. Wiedemann et al. (2018) describe offensive language as \"threats and discrimination against people, swear words or blunt insults\" (p.1). Hate speech, aggressive content, cyberbullying, and toxic comments are all different forms of offensive content (Schmidt & Wiegand, 2017) . Figure 1 shows an example of an offensive tweet. Hate speech is one of the most common forms of offensive language. A text that is targeted towards a group of peoplewith the intent to cause harm, violence, or social chaos is known as hate speech (Derczynski, 2019) . Davidson et al. (2017) define hate speech as \"a language that is used to express hatred towards a targeted group or is intended to be derogatory, to humiliate, or to insult the members of the group\" (p.1). Figure 2 illustrates an example of a hate speech tweet. Developing classification models, which can automatically detect offensive language and hate speech, is a very challenging task due to the following factors: a) the informality of the language used in posts from social media shows posts are usually written using symbols, short forms, and slang that are difficult to semantically process and understand by algorithms; b) the variation and the diversity of the Arabic language dialects and forms; and c) the small sample size of offensive samples; for example, in the dataset used in this study, only 5% of the tweets are labeled as hate speech and 19% of the tweets are labeled as offensive. We apply multiple preprocessing techniques to address the challenges of Arabic offensive language detection.",
                "cite_spans": [
                    {
                        "start": 272,
                        "end": 295,
                        "text": "Wiedemann et al. (2018)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 520,
                        "end": 545,
                        "text": "(Schmidt & Wiegand, 2017)",
                        "ref_id": null
                    },
                    {
                        "start": 794,
                        "end": 812,
                        "text": "(Derczynski, 2019)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 815,
                        "end": 837,
                        "text": "Davidson et al. (2017)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 548,
                        "end": 556,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    },
                    {
                        "start": 1021,
                        "end": 1029,
                        "text": "Figure 2",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "The preprocessing techniques include conversion of emoticons and emoji, conversion of hashtags, normalizing different forms of Arabic letters, normalizing Arabic dialects to Modern Standard Arabic (MSA), normalizing words by categorization, and basic cleaning processes. These intensive preprocessing techniques report valuable influence on the system performance. We train a Support Vector Machine (SVM)based classifier using character-based count vectorizer (2-5 characters). Results report an F1 score of 89.82% for the offensive language detection model and 95.16% for the hate speech detection model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "In the rest of this paper, we organize the content as follows: Section 2 discusses related work of Arabic offensive language detection on social media, Section 3 introduces data description, details of preprocessing, and the methodology of our models, and experimental results are discussed in Section 4. We also present the conclusion of our work at the end of the paper.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "There are multiple studies investigating offensive Arabic tweets to identify abusive Twitter accounts (Abozinadah, Mbaziira, & Jones, 2015; Abozinadah & Jones, 2017; Abozinadah, 2017) . Abozinadah, Mbaziira, and Jones (2015) construct an initial dataset starting with 500 Twitter accounts based on a set of Arabic swear words. Then, they check the most recent 50 tweets, profile pictures, and hashtags for each of these 500 Twitter accounts in order to reach a dataset of 350,000 Twitter accounts and 1,300,000 tweets with balanced classes; half were labelled abusive while the other half were labelled as non-abusive. Next, they use three types of features, including profile-based features, tweet-based features, and social graph features to train three classifiers: Naive Bayes (NB), SVM, and Decision Tree (J48). The results show that the NB-based classifier outperforms the other classifiers when used with 100 features and 10 tweets for each account with an accuracy score of 85% (Abozinadah, Mbaziira, & Jones, 2015; Abozinadah, 2017) .",
                "cite_spans": [
                    {
                        "start": 102,
                        "end": 139,
                        "text": "(Abozinadah, Mbaziira, & Jones, 2015;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 140,
                        "end": 165,
                        "text": "Abozinadah & Jones, 2017;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 166,
                        "end": 183,
                        "text": "Abozinadah, 2017)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 186,
                        "end": 224,
                        "text": "Abozinadah, Mbaziira, and Jones (2015)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 986,
                        "end": 1023,
                        "text": "(Abozinadah, Mbaziira, & Jones, 2015;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 1024,
                        "end": 1041,
                        "text": "Abozinadah, 2017)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2."
            },
            {
                "text": "Another approach for detecting offensive language was adopted by Mubarak, Darwish, and Magdy (2017) for the purpose of detecting vulgar and pornographic obscene speech in Arabic social media by applying a simple obscene phrases list-based approach. They use a Twitter dataset consisting of 175 million tweets to extract a list of seed words for obscene phrases through manual assessment. Then, they utilize the list to construct 3 sublists of obscene words, phrases, and hashtags using multiple measurements, such as the Log Odds Ratio (LOR) for unigrams and bigrams. Intrinsic and extrinsic evaluations were used to evaluate these lists. The intrinsic evaluation consists of manual coding for a list of 100 words that are randomly selected from each list to be marked as either obscene or not. The extrinsic evaluation consists of recall, precision, and F1 measures using a dataset that they developed for the purpose of this evaluation. They then select 10 Egyptian Twitter users from the top controversial users. After that, they extract 100 tweets with at least 10 replies for each user. The final dataset has 100 original tweets and 1,000 replies tweets. Each tweet along with its replies were submitted to CrowdFlower to become coded by three annotators from Egypt using three classes: obscene, offensive, and clean; the inter-annotator agreement is 84%. As a result, they develop a linear match model for each labeled tweet; for example, if a match with a phrase from the list occurs, then, it will then predict a label of obscenity to that tweet. The results among all lists show a highest F1 score of 60%, which demonstrate that a list-based approach is very limited and not a good choice for an obscene detection system. Arabic language has also been studied also by Nikolov (2018a, 2018b) for automatic detection of offensive language. They construct a dataset from YouTube comments based on selecting channels that have controversial videos about celebrities. Their final dataset includes 167,549 comments posted by 84,354 users, and 87,388 replies posted by 24,039 users from 150 YouTube videos (Alakrot, Murray, & Nikolov, 2018a) . Two labels were used for the classes: positive to label offensive comments and negative to label ones that are not offensive (Alakrot, Murray, & Nikolov, 2018a) . They train an SVM-based classifier using two features: character ngram (n= 1-5) and word-level features. The results show the best performance when using the SVM-based classifier with 10-fold cross validation and word-level features with 90.05% of an accuracy score (Alakrot, Murray, & Nikolov, 2018b ).",
                "cite_spans": [
                    {
                        "start": 65,
                        "end": 99,
                        "text": "Mubarak, Darwish, and Magdy (2017)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 1777,
                        "end": 1799,
                        "text": "Nikolov (2018a, 2018b)",
                        "ref_id": null
                    },
                    {
                        "start": 2108,
                        "end": 2143,
                        "text": "(Alakrot, Murray, & Nikolov, 2018a)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 2271,
                        "end": 2306,
                        "text": "(Alakrot, Murray, & Nikolov, 2018a)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 2575,
                        "end": 2609,
                        "text": "(Alakrot, Murray, & Nikolov, 2018b",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2."
            },
            {
                "text": "Mohaouchane, Mourhir, and Nikolov (2019) explore multiple deep learning models to classify offensive Arabic language for YouTube comments using the same dataset developed by Alakrot, Murray, and Nikolov (2018a) . They create 300-dimension word embedding using AraVec, which is an Arabic word embedding tool, trained on Twitter dataset and skip-gram model. Four deep learning models were evaluated for classifying offensive comments, including Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Bi-LSTM with an attention mechanism, and a combined model of CNN and LSTM. The results demonstrate an overall better performance for CNN with a highest accuracy score of 87.84%, a precision score of 86.10%, and an F1 score of 84.05%, while the combined CNN-LSTM model shows a better recall score of 83.46% (Mohaouchane, Mourhir, & Nikolov, 2019) .",
                "cite_spans": [
                    {
                        "start": 174,
                        "end": 210,
                        "text": "Alakrot, Murray, and Nikolov (2018a)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 833,
                        "end": 872,
                        "text": "(Mohaouchane, Mourhir, & Nikolov, 2019)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2."
            },
            {
                "text": "To our knowledge, the only Arabic hate speech detection studies are the studies of Mishra (2018, 2019) and the study of Chowdhury et al. (2019) , which investigate religious hate speech in Arabic language for Twitter data, both using the same dataset. Albadi, Kurdi, and Mishra (2018) develop a logistic regression-based model and an SVM-based model using a character n-gram feature (n= 1 to 4), and a Gated Recurrent Unit (GRU) based on the Recurrent Neural Network (RNN) with the Twitter Continuous Bag-of-Word (CBOW) 300-dimension embedding model provided by AraVec, batches of size 32, and Adam as the optimizer. Their findings have indicated that for some religious minorities in the Middle East -Jews, Atheists, and Shia-almost half of the tweets that were mentioning these minorities were referring to them within a hate speech content (Albadi, Kurdi, & Mishra, 2018) . Furthermore, results report best performance when using the GRU-based model with an F1 score of 77% (Albadi, Kurdi, & Mishra, 2018) . Albadi, Kurdi, and Mishra (2019) enhance the same GRU-based model with additional temporal, users, and content features in another study, then report the state-of-the-art performance in terms of a recall score of 84%. Chowdhury et al (2019) extend previous studies done by Mishra (2018, 2019) to investigate the effects of community interactions and social representations in detecting religious hate speech in the Arabic Twitter sphere. They use multiple features, including word embedding, node embedding, sentence representation, and character n-gram features (n = 1 -4). Several classification models were explored using multiple combinations of features, such as GRU, logistic regression, SVM, LSTM, Bi-LSTM, CNN and Bi-GRU in addition to combining multiple models, using self-attention mechanisms, and using Node2Vec criteria. Therefore, results have shown a high accuracy score of 81% that was obtained by the model containing a combination of Bi-GRU, CNN, and NODE2VEC. While displaying the best F1 score, recall and precision were recorded by the model that combined LSTM, CNN and NODE2VEC: 89%, 78%, and 86% respectively. Previous studies focus on features extraction and classification model, while we focus more on text preprocessing. The preprocessing steps we follow in this study are not identical to any of the preprocessing steps of the previous studies.",
                "cite_spans": [
                    {
                        "start": 83,
                        "end": 102,
                        "text": "Mishra (2018, 2019)",
                        "ref_id": null
                    },
                    {
                        "start": 120,
                        "end": 143,
                        "text": "Chowdhury et al. (2019)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 843,
                        "end": 874,
                        "text": "(Albadi, Kurdi, & Mishra, 2018)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 977,
                        "end": 1008,
                        "text": "(Albadi, Kurdi, & Mishra, 2018)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 1011,
                        "end": 1043,
                        "text": "Albadi, Kurdi, and Mishra (2019)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 1284,
                        "end": 1303,
                        "text": "Mishra (2018, 2019)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2."
            },
            {
                "text": "The shared task of the fourth workshop on Open-Source Arabic Corpora and Corpora Processing Tools (OSACT) in Language Resources and Evaluation Conference (LREC) 2020 provides Twitter dataset for offensive language detection in Arabic language. The main goal of this shared task is to identify and categorize Arabic offensive language in Twitter. The organizers collect tweets through Twitter API and annotate them hierarchically regarding offensive language and offense type. The task is divided into two subtasks: a) detecting if a post is offensive or not offensive; and b) identifying offensive content type of an offensive post as hate speech or not hate speech.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset Description",
                "sec_num": "3.1"
            },
            {
                "text": "The provider of the dataset performs some preprocessing to ensure the privacy of users. Twitter user mentions were substituted by \"@USER\", URLs had simply been substituted by \"URL\", and empty lines were replaced by \"<LF>\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset Description",
                "sec_num": "3.1"
            },
            {
                "text": "The shared task issues the dataset in three different parts, training dataset, development dataset and testing dataset. The summary of datasets distribution is presented in Table   1 . Training dataset and development dataset are provided with their actual labeled, while the testing dataset consists of 2,000 unlabeled tweets for competition evaluation purposes. The training dataset consists of 6,839 tweets with 1,371 offensive tweets and 350 hate speech tweets. The development dataset consists of 1,000 tweets with 179 offensive tweets and 44 hate speech tweets. We explore multiple characteristics of the dataset including most frequent emoji, emoticons, and words. The most frequent emoticons and emoji are very similar in both classes. Table 2 and Table 3 show the top 5 most frequent emoji and emoticons for both not offensive and offensive classes, respectively. Most frequent words are also very similar in both classes. Figure 3 and Figure 4 show the most frequent words for both not offensive class and offensive class, respectively. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 173,
                        "end": 182,
                        "text": "Table   1",
                        "ref_id": "TABREF1"
                    },
                    {
                        "start": 744,
                        "end": 763,
                        "text": "Table 2 and Table 3",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 932,
                        "end": 940,
                        "text": "Figure 3",
                        "ref_id": "FIGREF2"
                    },
                    {
                        "start": 945,
                        "end": 953,
                        "text": "Figure 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Dataset Description",
                "sec_num": "3.1"
            },
            {
                "text": "Emoji % Emoticon % 18% :< 44% \u2764 12% :) 23% 5% :( 16% 5% ;D 6% \u2665 4% :D 6%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Labels",
                "sec_num": null
            },
            {
                "text": "Emoji and emoticons are often used to convey feelings and attitude, which is very valuable to the task of offensive detection. Knowing some contextual information about the author of the post can provide insight into the textual content of the post. Moreover, a previous study converts an emoji to a textual label, and uses it to provide sentimental features to train a classifier for aggression detection, thus reporting better performance than the other model that does not consider emoji conversion (Orasan, 2018) . Thus, we consider converting emoji and emoticons to an Arabic textual label that describe the content of them. To have a more robust system that can scale and cover more emoji and emoticons, we then extract the entire set of emoji defined by Unicode.org (Unicode Organization, n.d.). Beautifulsoup4 4.8.2, a Python package for parsing HTML and XML documents, was used to scrape the emoji list available at the Unicode Organization website. We extract the Unicode and name of each emoji using Beutifulsoup4. The emoji list contains 1,374 emoji. The Unicode Organization website provides textual descriptions for each emoji written in English. We extract these textual descriptions, then, we use Translate 1.0.7 python package to translate them into Arabic language. Thus, during preprocessing of the data, each emoji is converted to its Arabic description. For example, is replaced by \" \u202b\u0648\ufe9f\ufbab\u202c \u202b\ufedf\ufeb4\ufe8e\u0646\u202c \u202b\ufee3\ufeca\u202c \u202b\"\ufbfe\ufed0\ufee4\ufeb0\u202c . In addition, we manually construct an emoticons list that includes a total of 140 emoticons with their textual descriptions in Arabic language. For example, \":-X\" is replaced by \" \u202b\u0627\ufedf\ufee0\ufeb4\ufe8e\u0646\u202c \u202b\ufee3\ufecc\ufed8\ufeee\u062f\u202c \". Next, we analyze the description phrase as a regular textual phrase in tweets so that it could maintain its semantic meaning after removing the original emoji and emoticon.",
                "cite_spans": [
                    {
                        "start": 502,
                        "end": 516,
                        "text": "(Orasan, 2018)",
                        "ref_id": "BIBREF26"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Emoji and Emoticon Conversion",
                "sec_num": "3.2.1"
            },
            {
                "text": "The Arabic dialects have various forms, based on geography and social class (Habash, 2010) . Arabic dialects are the Arabic languages that have often been used in usergenerated contents such as Twitter. Habash (2010) categorizes the Arabic dialects into seven dialects: Gulf, Egyptian, Iraqi, North African, Yemenite, Levantine, and Maltese Arabic, which is not always considered one of the Arabic dialects. It is very crucial to consider the variations among different dialects when detecting offensive language, as some words that have exactly the same pronunciation and spelling might have different meanings. For example, the word \u202b\ufecb\ufe8e\ufed3\ufbff\ufe94\"\u202c -Afiah\" means health in Gulf, Egyptian, Iraqi, and Levantine dialects ; however, it also means fire in Moroccan Arabic. We try to solve the variation in dialects by a dimensionality reduction approach. We reduce the dimensionality of the data by normalizing the variation of dialects for a set of nouns to be converted from dialectal Arabic to MSA. For example, the variations of the word boy, \u202b,\"\u0631\ufe9f\ufede\"\u202c \u202b,\"\u0632\ufedf\ufee4\ufe94\"\u202c and \u202b\"\u0632\u0648\u0644\"\u202c are converted to \u202b.\"\u0648\ufedf\ufeaa\"\u202c The set of nouns is manually constructed based on manual insepction for a sample of the dataset and based on our own experience as native Arabic speakers.",
                "cite_spans": [
                    {
                        "start": 76,
                        "end": 90,
                        "text": "(Habash, 2010)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 203,
                        "end": 216,
                        "text": "Habash (2010)",
                        "ref_id": "BIBREF17"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Arabic Dialects Normalization",
                "sec_num": "3.2.2"
            },
            {
                "text": "Normalizing and reducing the dimensionality can improve the performance of the model. From our manual inspection for a sample of tweets from the dataset, we notice that it is very common to mention name of animals among the hate speech tweets. Thus, we manually create a list of the most common animal names used in different Arabic dialects, such as \u202b\ufedb\ufee0\ufe90\"\u202c -dog\", \u202b\ufea7\ufee8\ufeb0\ufbfe\ufeae\"\u202c -pig\", \u202b\ufea3\ufbff\ufe94\"\u202c -snake\". The list considers the variation in dialects in animal names, such as the word cat in MSA is \"Qetta/ \u202b\ufed7\ufec4\ufe94\u202c \", while in Egyptian it is \"Otta/ \u202b,\"\u0623\u0637\ufe94\u202c in Levantine it is \"Bisse/\u202b,\"\ufe91\ufeb4\ufe94\u202c in Gulf it is \"Qatwa/\u202b,\"\ufed7\ufec4\ufeee\u0629\u202c in Moroccan it is \"Qetta/ \u202b\ufed7\ufec4\ufe94\u202c \", in Iraqi it is \"Bazzuna/\u202b,\"\ufe91\ufeb0\u0648\ufee7\ufe94\u202c and in Yamani it is \"demah/\u202b.\"\u062f\ufee3\ufe94\u202c In addition to the variation in dialects, the list includes variations of the same animal names, such as the female name, male name, plural name. Thus, for the word cat, we include singular female name \"Qetta/ \u202b\ufed7\ufec4\ufe94\u202c \", singular male name \"Qett/ \u202b\ufed7\ufec2\u202c \", two female name \"Qettatan/ \u202b\ufed7\ufec4\u202c \u202b\ufe98\ufe8e\u0646\u202c \", two male name \"Qettan/ \u202b\ufed7\ufec4\u202c \u202b\ufe8e\u0646\u202c \", and plural name \"Qettat/ \u202b\ufed7\ufec4\u202c \u202b\ufec2\u202c \". The list contains a total of 335 lexicons. All animal names that are listed and occurred in the dataset were converted to the word animal in Arabic \u202b.\"\ufea3\ufbff\ufeee\u0627\u0646\"\u202c Accordingly, all animal names that are included in tweets are reduced to only one word.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Words Categorization",
                "sec_num": "3.2.3"
            },
            {
                "text": "Arabic letters can be written in various format depending on the location of the letter within the word. We normalize Alif \u202b\u0625\u060c\u0622\u060c\u0623(\u202c to \u202b\u0627\u202c ), Alif Maqsura \u202b\u0626\u060c\u064a(\u202c to \u202b,)\u0649\u202c and Ta Marbouta \u202b\u0629(\u202c to \u202b.)\u0647\u202c Letters that were repeated more than two times within a word were reduced to two times only.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Letters Normalization",
                "sec_num": "3.2.4"
            },
            {
                "text": "Hashtags are commonly used in Twitter to highlight important phrases within the tweet. Thus, it is very important to consider hashtags during the preprocessing phase to convert hashtags into a meaningful format. We remove the \"#\" symbol and replace \"_\" by a space. For example, the hashtag \u202b\"\u0627\ufedf\ufbad\ufefc\u0644_\u0627\ufedf\ufe98\ufecc\ufe8e\u0648\u0646#\"\u202c is converted to \u202b\u0627\ufedf\ufbad\ufefc\u0644\"\u202c \u202b\u0627\u202c \u202b\ufedf\ufe98\ufecc\ufe8e\u0648\u0646\u202c \", which is easier for the system to understand and process.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Hashtag Segmentation",
                "sec_num": "3.2.5"
            },
            {
                "text": "Tweets were filtered to remove numbers, kashida, HTML tags, more than one space, three or more repetitions of any character, and some symbols or terms (e.g., \" _\", '\"', \"\\\"\", \"'s\", \"...\", \"!\", \"?\", \"I\", \"@USER\", \"USER\", \"URL\", \".\", \";\", \":\", \"/\", \"\\\\\", \", \", \"#\", \"@\", \"$\", \"&\", \")\", \"(\", \"\\\"). We borrow the list of Arabic stopwords defined by Alrefaie",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Miscellaneous",
                "sec_num": "3.2.6"
            },
            {
                "text": "5% 10% 15% 20% \u202b\ufbfe\ufe8e\u202c \u202b\ufdf2\u202c \u202b\u0627\ufee7\ufe96\u202c \u202b\u0627\ufedf\ufee0\ufef2\u202c \u202b\u0627\ufe91\ufee6\u202c (2016), which contains 750 stopwords. Furthermore, we remove diacritics that were used in tweets containing text from the holy Qur'an or poetry.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "0%",
                "sec_num": null
            },
            {
                "text": "As can be noticed from Table 1 , the classes are highly imbalanced, with only 5% hate speech tweets and 19% offensive tweets. we try to solve the problem with the imbalanced distribution of classes by using up-sampling. The original training dataset has 7,000 tweets with a first label hierarchy of 1,410 offensive tweets and 5,590 not offensive tweets. After up-sampling, the number of tweets for each class, offensive tweets and not offensive tweets, becomes 5,590. The second label hierarchy is sharply imbalanced, the original training dataset consists of 361 hate speech tweets and 6639 not hate speech tweet. After up-sampling, the number of tweets for each class, hate speech tweets and not hate speech tweets, becomes 6,639. We use resample function from Python scikit-learn library to implement upsampling. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 23,
                        "end": 30,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Upsampling",
                "sec_num": "3.2.7"
            },
            {
                "text": "We use two features in training the models, which consists of character-based count vectorizer and TF-IDF vectorizer, both with 2 to 5 characters. Previous studies highlight the importance of using character-based features over wordbased features for offensive language detection because it is a language independent feature that can work with misspelling errors or obfuscating offensive words, which are commonly practiced on social media posts (Bohra et al., 2018; Nobata et al., 2016) . Both features are implemented using Python scikit-learn library.",
                "cite_spans": [
                    {
                        "start": 446,
                        "end": 466,
                        "text": "(Bohra et al., 2018;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 467,
                        "end": 487,
                        "text": "Nobata et al., 2016)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Feature Extraction",
                "sec_num": "3.3"
            },
            {
                "text": "We explore the effect of each preprocessing technique on the performance of the model before applying them to the final models. Previous studies on offensive language detection report high performance when using an SVMbased classifier (Abozinadah and Jones, 2017; Schmidt and Wiegand, 2017; Albadi, Kurdi, and Mishra, 2018) . The SVM classifier focuses on maximizing the margin, the distance of the closest points to the hyperplane that separate between instances of classes using a linear function (Goodfellow, Bengio, and Courville, 2016) . Thus, we decide to be in-line with the findings from earlier studies and use the linear SVM-based classifier in this step trained using the first label hierarchy of the dataset; offensive or not offensive, for the purpose of this exploration task. We did not use the testing dataset during this step. The training dataset used to train the models and the development dataset used to evaluate the models. During this preliminary modeling phase, two main goals were targeted. The first goal is to identify if count vectorizer outperforms TF-IDF vectorizer or not. To accomplish this goal, we trained two SVM-based models using the raw dataset without performing any preprocessing technique, one model applies the count vectorizer and the other one applies the TF-IDF vectorizer. Results are shown in Table 5 . The TF-IDF vectorizer is 2% better than the count vectorizer in precision score. However, having a comprehensive measurement that consider multiple factors in evaluating the model is very important. The dataset is highly imbalanced, so we consider F1 score in evaluating the performance rather than accuracy. The accuracy score is often gives misleading results in similar situation with imbalanced dataset. F1 score is the harmonic mean for recall and precision. The count vectorizer outperforms the TF-IDF vectorizer by 2% in F1 score. This finding is consistent with Wiedeman et al. 2018finding, which reports the unsuitability of TF-IDF vectorizer over twitter datasets because tweets are very short, making it not an optimal source for IDF estimations. Thus, we consider adopting the count vectorizer as the feature for the rest of the models. The second goal of this preliminary modeling phase is to evaluate each preprocessing technique separately and measure their effects on the performance. Table 6 presents the results from the preliminary preprocessing exploration models. Only two preprocessing techniques were merged; hashtags segmentation and miscellaneous cleaning processes due to their relatedness and similarity in term of platform specific attributes. Preliminary results illustrate the contribution of each preprocessing technique to the performance of the model; however, we also expect to have different results when used on larger dataset and when used for hate speech detection. The results report the highest contribution in term of F1 score to letters normalization, followed by dialect normalization and word categorization, then, emoji and emoticon conversion, miscellaneous cleaning and hashtag segmentation, and finally, upsampling. The results demonstrate some issues with the upsampling technique, it reduces F1 score from 79%, as it shown in Table 5 , to 71%. Consequently, we decide to apply all the preprocessing techniques except upsampling for the main models. ",
                "cite_spans": [
                    {
                        "start": 235,
                        "end": 263,
                        "text": "(Abozinadah and Jones, 2017;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 264,
                        "end": 290,
                        "text": "Schmidt and Wiegand, 2017;",
                        "ref_id": null
                    },
                    {
                        "start": 291,
                        "end": 323,
                        "text": "Albadi, Kurdi, and Mishra, 2018)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 499,
                        "end": 540,
                        "text": "(Goodfellow, Bengio, and Courville, 2016)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 1341,
                        "end": 1348,
                        "text": "Table 5",
                        "ref_id": "TABREF7"
                    },
                    {
                        "start": 2352,
                        "end": 2359,
                        "text": "Table 6",
                        "ref_id": "TABREF9"
                    },
                    {
                        "start": 3227,
                        "end": 3234,
                        "text": "Table 5",
                        "ref_id": "TABREF7"
                    }
                ],
                "eq_spans": [],
                "section": "Preliminary Models",
                "sec_num": "3.4.1"
            },
            {
                "text": "The baseline models include two linear SVM-based models, both trained on dataset without any sort of preprocessing technique and using a count vectorizer with 2 to 5 characters. The first baseline model classifies tweets to either Offensive (OFF) or Not Offensive (NOT_OFF), while the second one classifies tweets to either Hate Speech (HS) or Not Hate Speech (NOT_HS). To assess our goal in investigating the effect of preprocessing on offensive detection and hate speech detection, we use the same feature and classifier for baseline models and main models. All Models are implemented using Python scikit-learn library.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Baseline Models",
                "sec_num": "3.4.2"
            },
            {
                "text": "The main model for offensive language detection, which classifies tweets to either offensive (OFF) or not offensive (NOT_OFF), is a Linear SVM-based classifier. The model is trained on full preprocessed tweets that have been preprocessed using all techniques mentioned earlier; conversion of emoticons and emoji, conversion of hashtags, normalizing different forms of Arabic letters, normalizing Arabic dialects to MSA, normalizing words by categorization, and other miscellaneous cleaning processes. As we mentioned earlier, the feature used in training the model is the character-based count vectorizer with 2 to 5 characters. The model is implemented using Python scikitlearn library.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sub-Task A Model : Offensive Language Detection",
                "sec_num": "3.4.3"
            },
            {
                "text": "The same exact model of Sub-Task A is used for Sub-Task B, including the same preprocessing and feature extraction techniques. However, the model classifies tweets to either hate speech (HS) or not hate speech (NOT_HS). Table 7 shows the results of performance evaluation for the baseline model and the main model for offensive language detection task, and Table 8 shows the results for hate speech detection task. The baseline models were evaluated using the development dataset while the main models were evaluated using the testing dataset through the shared task competition evaluators. The dataset is highly imbalanced, thus, the accuracy score might not be very informative to evaluate the performance. The F1 score increased from 79% for the baseline model before preprocessing to 89% after preprocessing for Sub-Task A for offensive detection, which was ranked the 3rd place. For Sub-Task B for hate speech detection, F1 score increased sharply from 67% to 95%, which was ranked the 1st place.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 220,
                        "end": 227,
                        "text": "Table 7",
                        "ref_id": "TABREF11"
                    },
                    {
                        "start": 357,
                        "end": 364,
                        "text": "Table 8",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Sub-Task B Model : Hate Speech Detection",
                "sec_num": "3.4.4"
            },
            {
                "text": "The most noticeable point from the tables is that in Sub-Task B results are better than in Sub-Task A, given the fact that the class distribution is more skewed than that of Sub-Task A and the number of training instance is much smaller than Sub-Task A. This interesting point demonstrates how data preprocessing adds value to the performance, even for hate speech class that is very rare. Table 8 : Performance evaluation of hate speech detection task",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 390,
                        "end": 397,
                        "text": "Table 8",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Experiment Results",
                "sec_num": "4."
            },
            {
                "text": "The preprocessing phase is one of the key phases within the text classification pipeline, including offensive language and hate speech classification. The ambiguity and informality of social media language increase the complexity of achieving high classification performance, particularly for Arabic text that has multiple dialects. Our study shows the competitive results obtained for offensive language detection and hate speech detection by using intensive preprocessing techniques to filter and clean the dataset before feeding it into the rest of the phases for the text classification pipeline. Moreover, results report the significant impacts of preprocessing on a very challenging task, such as the hate speech classification, with very small sample size of 5% from the overall dataset.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "5."
            },
            {
                "text": "In the future, we hope to enhance the available studies of offensive language detection and hate speech detection by investigating our preprocessing techniques using some deep learning models. Another future direction is to use more advanced features in training the model, such as word embedding.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "5."
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Detection of Abusive Accounts with Arabic Tweets",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Abozinadah",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Mbaziira",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "International Journal of Knowledge Engineering",
                "volume": "1",
                "issue": "2",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.7763/IJKE.2015.V1.19"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Abozinadah, E. , Mbaziira, A., and Jones, J. (2015). Detection of Abusive Accounts with Arabic Tweets. International Journal of Knowledge Engineering, Vol. 1, No. 2. DOI: 10.7763/IJKE.2015.V1.19",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Detecting Abusive Arabic Language Twitter Accounts Using a Multidimensional Analysis Model",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Abozinadah",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Abozinadah, E. (2017). Detecting Abusive Arabic Language Twitter Accounts Using a Multidimensional Analysis Model.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "A Statistical Learning Approach to Detect Abusive Twitter Accounts",
                "authors": [
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Abozinadah",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the International Conference on Compute and Data Analysis",
                "volume": "130280",
                "issue": "",
                "pages": "6--13",
                "other_ids": {
                    "DOI": [
                        "10.1145/3093241.3093281"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Abozinadah, E., and Jones, J. (2017). A Statistical Learning Approach to Detect Abusive Twitter Accounts. Proceedings of the International Conference on Compute and Data Analysis, 130280, 6-13. https://doi.org/10.1145/3093241.3093281",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Dataset Construction for the Detection of Anti-Social Behaviour in Online Communication in Arabic",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Alakrot",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Murray",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [
                            "S"
                        ],
                        "last": "Nikolov",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Procedia Computer Science",
                "volume": "142",
                "issue": "",
                "pages": "174--181",
                "other_ids": {
                    "DOI": [
                        "10.1016/j.procs.2018.10.473"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Alakrot, A., Murray, L., and Nikolov, N. S. (2018a). Dataset Construction for the Detection of Anti- Social Behaviour in Online Communication in Arabic. Procedia Computer Science, 142, 174-181. https://doi.org/https://doi.org/10.1016/j.procs.2018. 10.473",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Towards Accurate Detection of Offensive Language in Online Communication in Arabic",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Alakrot",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Murray",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [
                            "S"
                        ],
                        "last": "Nikolov",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Procedia Computer Science",
                "volume": "142",
                "issue": "",
                "pages": "315--320",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alakrot, A., Murray, L., and Nikolov, N. S. (2018b). Towards Accurate Detection of Offensive Language in Online Communication in Arabic. Procedia Computer Science, 142, 315-320.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Are they Our Brothers? Analysis and Detection of Religious Hate Speech in the Arabic Twittersphere",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Albadi",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Kurdi",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Mishra",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Albadi, N., Kurdi, M., and Mishra, S. (2018). Are they Our Brothers? Analysis and Detection of Religious Hate Speech in the Arabic Twittersphere. 2018",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM)",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "69--76",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), 69-76.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Investigating the effect of combining GRU neural networks with handcrafted features for religious hatred detection on Arabic Twitter space",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Albadi",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Kurdi",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Mishra",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Soc. Netw. Anal. Min",
                "volume": "9",
                "issue": "1",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Albadi, N., Kurdi, M., and Mishra, S. (2019). Investigating the effect of combining GRU neural networks with handcrafted features for religious hatred detection on Arabic Twitter space. Soc. Netw. Anal. Min., vol. 9, no. 1, p. 41.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Arabic-stop-words",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Alrefaie",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alrefaie, M. (2016). Arabic-stop-words [Github Repository]. Retrieved on January 24, 2020 from https://github.com/mohataher/arabic-stop-words",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "A Dataset of Hindi-English Code-Mixed Social Media Text for Hate Speech Detection",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Bohra",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Vijay",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Singh",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [
                            "S"
                        ],
                        "last": "Akhtar",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Shrivastava",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W18-1105"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Bohra, A., Vijay, D., Singh, V., Akhtar, S. S., & Shrivastava, M. (2018). A Dataset of Hindi-English Code-Mixed Social Media Text for Hate Speech Detection. https://doi.org/10.18653/v1/W18-1105",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "ARHNet -Leveraging Community Interaction for Detection of Religious Hate Speech in Arabic",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Chowdhury",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Didolkar",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Sawhney",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Shah",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop",
                "volume": "",
                "issue": "",
                "pages": "273--280",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chowdhury, A., Didolkar, A., Sawhney, R., and Shah, R. (2019). ARHNet -Leveraging Community Interaction for Detection of Religious Hate Speech in Arabic. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop. Association for Computational Linguistics, Florence, Italy, 273-280.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Automated Hate Speech Detection and the Problem of Offensive Language",
                "authors": [
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Davidson",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Warmsley",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Macy",
                        "suffix": ""
                    },
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Weber",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Eleventh International AAAI Conference on Web and Social Media About offensive content. arXiv.org",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Davidson, T., Warmsley, D., Macy, M., and Weber, I. (2017). Automated Hate Speech Detection and the Problem of Offensive Language. AAAI Publications, Eleventh International AAAI Conference on Web and Social Media About offensive content. arXiv.org. Retrieved from http://search.proquest.com/docview/2074118430/",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Offensive Language and Hate Speech Detection for Danish",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Derczynski",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Derczynski, L. (2019). Offensive Language and Hate Speech Detection for Danish. arXiv.org. Retrieved from http://search.proquest.com/docview/2273023012/",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Deep Learning",
                "authors": [
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Goodfellow",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Courville",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Goodfellow, I., Bengio, B., Courville, A. (2016). Deep Learning. MIT press. Retrieved from http://www.deeplearningbook.org/contents/ml.html",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Introduction to Arabic natural language processing",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Habash",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "",
                "volume": "3",
                "issue": "",
                "pages": "1--185",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Habash, N. (2010). Introduction to Arabic natural language processing (Vol. 3, pp. 1-185).",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Mind Your Language: Abuse and Offense Detection for Code-Switched Languages",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Kapoor",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Kumar",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Rajput",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Shah",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Kumaraguru",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Zimmermann",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kapoor, R., Kumar, Y., Rajput, K., Shah, R., Kumaraguru, P., & Zimmermann, R. (2018). Mind Your Language: Abuse and Offense Detection for Code- Switched Languages. arXiv.org. Retrieved from http://search.proquest.com/docview/2111970667/",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "German Hate Speech Detection on Twitter",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Kent",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "14th Conference on Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kent, S. (2018). German Hate Speech Detection on Twitter. Proceedings of GermEval 2018, 14th Conference on Natural Language Processing (KONVENS 2018) Vienna, Austria -September 21.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Locate the Hate: Detecting Tweets Against Blacks",
                "authors": [
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Kwok",
                        "suffix": ""
                    },
                    {
                        "first": "Wang",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the Twenty-Seventh AAAI Conference on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "1621--1622",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kwok, I., and Wang, Y. (2013). Locate the Hate: Detecting Tweets Against Blacks. Proceedings of the Twenty- Seventh AAAI Conference on Artificial Intelligence, pp.1621-1622",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Detecting Offensive Language on Arabic Social Media using Deep Learning",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Mohaouchane",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Mourhir",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [
                            "S"
                        ],
                        "last": "Nikolov",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mohaouchane, H., Mourhir, A., and Nikolov, N. S. (2019). Detecting Offensive Language on Arabic Social Media using Deep Learning, (December).",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Abusive Language Detection on Arabic Social Media",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Mubarak",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Darwish",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Magdy",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the First Workshop on Abusive Language Online",
                "volume": "",
                "issue": "",
                "pages": "52--56",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mubarak, H., Darwish, K., and Magdy, W. (2017). Abusive Language Detection on Arabic Social Media. Proceedings of the First Workshop on Abusive Language Online. Association for Computational Linguistics (ACL), 2017. pp. 52-56.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Early Detection of Controversial Urdu Speeches from Social Media. Data Science and Pattern Recognition, Ubiquitous International, Volume1, Number2",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Mustafa",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Nawaz",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Ferzund",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Lali",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Shahzad",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Viger",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mustafa, R., Nawaz, M., Ferzund, J., Lali, M., Shahzad, B., and ournier-Viger, P. (2017). Early Detection of Controversial Urdu Speeches from Social Media. Data Science and Pattern Recognition, Ubiquitous International, Volume1, Number2. ISSN 2520-4165.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Abusive language detection in online user content",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Nobata",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Tetreault",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Thomas",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Mehdad",
                        "suffix": ""
                    },
                    {
                        "first": "Chang",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 25th international conference on world wide web",
                "volume": "",
                "issue": "",
                "pages": "145--153",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nobata, C., Tetreault, J., Thomas, A., Mehdad, Y., and Chang, Y. (2016). Abusive language detection in online user content. In Proceedings of the 25th international conference on world wide web, pp. 145-153. International World Wide Web Conferences Steering Committee, 2016",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Aggressive language identification using word embeddings and sentiment features",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Orasan",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Orasan, C. (2018). Aggressive language identification using word embeddings and sentiment features. Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying. Association for Computational Linguistics, http://hdl.handle.net/2436/621749",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Detection of cyberbullying on social media messages in Turkish",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Ozel",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Sarac",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Akdemir",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Aksu",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Computer Science and Engineering",
                "volume": "",
                "issue": "",
                "pages": "366--370",
                "other_ids": {
                    "DOI": [
                        "10.1109/UBMK.2017.8093411"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ozel, S., Sarac, E., Akdemir, S., and Aksu, H. (2017). Detection of cyberbullying on social media messages in Turkish. 2017 International Conference on Computer Science and Engineering (UBMK), 366- 370. https://doi.org/10.1109/UBMK.2017.8093411",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Effective hate-speech detection in Twitter data using recurrent neural networks",
                "authors": [
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Pitsilis",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Ramampiaro",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Langseth",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Applied Intelligence",
                "volume": "48",
                "issue": "12",
                "pages": "4730--4742",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Pitsilis, G., Ramampiaro, H., and Langseth, H. (2018). Effective hate-speech detection in Twitter data using recurrent neural networks. Applied Intelligence, 48(12), pp. 4730-4742.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Transfer Learning from LDA to BiLSTM-CNN for Offensive Language Detection in Twitter",
                "authors": [
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Wiedemann",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Ruppert",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Jindal",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Biemann",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wiedemann, G., Ruppert, E., Jindal, R., and Biemann, C. (2018). Transfer Learning from LDA to BiLSTM- CNN for Offensive Language Detection in Twitter.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "text": "Example for an offensive tweet.",
                "type_str": "figure",
                "uris": null
            },
            "FIGREF1": {
                "num": null,
                "text": "Example for a hate speech tweet.",
                "type_str": "figure",
                "uris": null
            },
            "FIGREF2": {
                "num": null,
                "text": "Most frequent 5 words in not offensive tweets",
                "type_str": "figure",
                "uris": null
            },
            "TABREF1": {
                "html": null,
                "text": "",
                "num": null,
                "type_str": "table",
                "content": "<table/>"
            },
            "TABREF2": {
                "html": null,
                "text": "The top 5 most frequent emoji and emoticons for not offensive tweets",
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Emoji</td><td>%</td><td>Emoticon</td><td>%</td></tr><tr><td/><td>38.87%</td><td>;D</td><td>40%</td></tr><tr><td/><td>8.52%</td><td>:&lt;</td><td>33.33%</td></tr><tr><td>\u2764</td><td>3.83%</td><td>:(</td><td>13.33%</td></tr><tr><td/><td>3.74%</td><td>:)</td><td>6.67%</td></tr><tr><td/><td>2.43%</td><td>d:</td><td>6.67%</td></tr></table>"
            },
            "TABREF3": {
                "html": null,
                "text": "The top 5 most frequent emoji and emoticons for offensive tweets",
                "num": null,
                "type_str": "table",
                "content": "<table/>"
            },
            "TABREF4": {
                "html": null,
                "text": "illustrates example of tweets before preprocessing and after preprocessing.",
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Tweet Before Cleaning</td><td/><td colspan=\"2\">Tweet After Cleaning</td></tr><tr><td colspan=\"2\">\u202b\ufe9b\ufe8e\ufbfe\ufeae\u202c \u202b\ufeb7\ufecc\ufeae\u202c \u202b\ufe91\ufeee\u202c \u202b\ufbfe\ufe8e\u202c \u202b\ufecb\ufee4\ufeae\u0631\u0631\u0631\u0631\u0631\u0631\u0631\u064a\u202c \u202b\ufbfe\ufe8e\u202c</td><td colspan=\"2\">\u202b\u0627\ufedf\ufe92\ufedc\ufe8e\u0621\u202c \u202b\u0648\ufe9f\ufbab\u202c</td><td>\u202b\ufe9b\ufe8e\u066e\ufeae\u202c \u202b\ufeb7\ufecc\ufeae\u202c \u202b\ufe91\ufeee\u202c \u202b\u066e\ufe8e\u202c \u202b\ufecb\ufee4\ufeae\u0631\u0649\u202c \u202b\u066e\ufe8e\u202c</td></tr><tr><td/><td/><td colspan=\"2\">\u202b\ufe91\ufebc\ufeee\u062a\u202c \u202b\u0627\ufedf\ufe92\ufedc\ufe8e\u0621\u202c \u202b\u0648\ufe9f\ufbab\u202c</td><td>\u202b\ufecb\ufe8e\u0644\u202c \u202b\ufe91\ufebc\ufeee\u062a\u202c</td></tr><tr><td/><td/><td>\u202b\u0648\ufe9f\ufbab\u202c</td><td>\u202b\ufecb\ufe8e\u0644\u202c \u202b\ufe91\ufebc\ufeee\u062a\u202c \u202b\u0627\ufedf\ufe92\ufedc\ufe8e\u0621\u202c \u202b\u0648\ufe9f\ufbab\u202c</td><td>\u202b\ufecb\ufe8e\u0644\u202c</td></tr><tr><td/><td/><td>\u202b\ufed7\ufee0\ufe90\u202c</td><td>\u202b\u0627\ufebb\ufed4\ufeae\u202c \u202b\ufed7\ufee0\ufe90\u202c</td><td>\u202b\ufecb\ufe8e\u0644\u202c \u202b\ufe91\ufebc\ufeee\u062a\u202c \u202b\u0627\ufedf\ufe92\ufedc\ufe8e\u0621\u202c</td></tr><tr><td/><td/><td/><td>\u202b\u0627\ufebb\ufed4\ufeae\u202c \u202b\ufed7\ufee0\ufe90\u202c</td><td>\u202b\u0627\ufebb\ufed4\ufeae\u202c \u202b\ufed7\ufee0\ufe90\u202c</td><td>\u202b\u0627\ufebb\ufed4\ufeae\u202c</td></tr><tr><td colspan=\"2\">@USER: \u202b\ufbfe\ufe8e\u202c \u202b\ufeb7\ufecc\ufe90\u202c \u202b\ufbfe\ufe8e\u202c \u202b\u0627\ufee7\ufe8e\ufee7\ufbff\ufbab\u202c \u202b\ufe91\ufec4\ufee0\ufeee\u202c</td><td colspan=\"2\">\u202b\u0648\u066e\ufefc\u202c \u202b\u0627\ufee7\ufebc\ufeae\u0627\ufed3\ufef0\u202c \u202b\u066e\ufe8e\u202c \u202b\ufeb7\ufecc\ufe90\u202c \u202b\u066e\ufe8e\u202c \u202b\u0627\ufee7\ufe8e\ufee7\u066e\ufbab\u202c \u202b\ufe91\ufec4\ufee0\ufeee\u202c</td></tr><tr><td colspan=\"2\">\u202b\ufecb\ufbff\ufeb8\ufedc\ufee2\u202c \u202b\ufebb\ufed4\ufeee\u0641\u202c \u202b\ufecb\ufee0\ufef0\u202c \u202b\u0648\ufbfe\ufefc\u202c \u202b\u0627\ufee7\ufebc\ufeae\u0627\ufed3\ufef2\u202c</td><td/><td>\u202b\u0648\ufe91\ufee8\ufeb0\ufbfe\ufee8\ufedc\ufee2\u202c</td><td>\u202b\ufecb\u066e\ufeb8\ufedc\ufee2\u202c \u202b\ufebb\ufed4\ufeee\u0641\u202c</td></tr><tr><td>\u202b\ufee8\ufeb0\ufbfe\ufee8\ufedc\ufee2\u202c</td><td>\u202b\u0648\ufe91\u202c</td><td/></tr><tr><td colspan=\"2\">RT @USER: \u202b\ufe97\ufe92\ufeaa\u0623\u202c \u202b\ufee3\ufee4\ufedc\ufee6\u202c \u202b\ufedb\ufe8e\u0646\u202c \u202b\ufee3\ufe8e\u202c</td><td colspan=\"2\">\u202b\ufea3\u066e\ufeee\u0627\u0646\u202c \u202b\u066e\ufe8e\u202c \u202b\ufe97\ufe92\ufeaa\u066e\ufede\u202c \u202b\ufe97\ufea8\ufeb4\ufeae\u202c \u202b\ufe91\ufeaa\u0644\u202c \u202b\ufe91\u066e\ufbab\u202c \u202b\ufe97\ufe92\ufeaa\u0627\u202c \u202b\ufee3\ufee4\ufedc\ufee6\u202c</td></tr><tr><td colspan=\"2\">\u202b\ufbfe\ufe8e\ufe91\ufee6\u202c \u202b\ufea3\ufee4\ufe8e\u0631\u202c \u202b\ufbfe\ufe8e\u202c \u202b\ufe97\ufe92\ufeaa\ufbfe\ufede\u202c \u202b\ufe97\ufea8\ufeb4\ufeae\u202c \u202b\ufee3\ufe8e\u202c \u202b\ufe91\ufeaa\u0644\u202c \u202b\ufe91\ufbff\ufbab\u202c</td><td/><td>\u202b\ufecf\ufe92\u066e\ufef0\u202c \u202b\u066e\ufe8e\u202c \u202b\ufea3\u066e\ufeee\u0627\u0646\u202c \u202b\u066e\ufe8e\ufe91\ufee6\u202c</td></tr><tr><td colspan=\"2\">\u202b\ufecf\ufe92\ufbff\ufbff\ufbff\ufbff\ufef2\u202c \u202b\ufbfe\ufe8e\u202c \u202b\u0627\ufedf\ufea4\ufee4\ufe8e\u0631\u202c</td><td/></tr></table>"
            },
            "TABREF5": {
                "html": null,
                "text": "Examples of tweets before preprocessing and after preprocessing",
                "num": null,
                "type_str": "table",
                "content": "<table/>"
            },
            "TABREF7": {
                "html": null,
                "text": "",
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>: Performance results for preliminary models</td></tr><tr><td>based on features</td></tr></table>"
            },
            "TABREF9": {
                "html": null,
                "text": "Preliminary performance evaluation results from each preprocessing technique on offensive detection",
                "num": null,
                "type_str": "table",
                "content": "<table/>"
            },
            "TABREF11": {
                "html": null,
                "text": "Performance evaluation of offensive language detection task",
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td>Model</td><td>Precision</td><td>Recall</td><td>F1</td><td>Accuracy</td></tr><tr><td>Baseline Model</td><td>74%</td><td>63%</td><td>67%</td><td>96%</td></tr><tr><td>Main Model</td><td>95%</td><td>95%</td><td>95%</td><td>95%</td></tr></table>"
            }
        }
    }
}