File size: 49,423 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
{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:06:26.045721Z"
    },
    "title": "ALT Submission for OSACT Shared Task on Offensive Language Detection",
    "authors": [
        {
            "first": "Sabit",
            "middle": [],
            "last": "Hassan",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "2\u00d6 zyegin University",
                "location": {}
            },
            "email": "sahassan2@hbku.edu.qa"
        },
        {
            "first": "Younes",
            "middle": [],
            "last": "Samih",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "2\u00d6 zyegin University",
                "location": {}
            },
            "email": "ysamih@hbku.edu.qa"
        },
        {
            "first": "Hamdy",
            "middle": [],
            "last": "Mubarak",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "2\u00d6 zyegin University",
                "location": {}
            },
            "email": "hmubarak@hbku.edu.qa"
        },
        {
            "first": "Ahmed",
            "middle": [],
            "last": "Abdelali",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "2\u00d6 zyegin University",
                "location": {}
            },
            "email": "aabdelali@hbku.edu.qa"
        },
        {
            "first": "Ammar",
            "middle": [],
            "last": "Rashed",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "2\u00d6 zyegin University",
                "location": {}
            },
            "email": "ammar.rasid@ozu.edu.tr"
        },
        {
            "first": "Shammur",
            "middle": [],
            "last": "Chowdhury",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "2\u00d6 zyegin University",
                "location": {}
            },
            "email": "shchowdhury@hbku.edu.qa"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In this paper, we describe our efforts at OSACT Shared Task on Offensive Language Detection. The shared task consists of two subtasks: offensive language detection (Subtask A) and hate speech detection (Subtask B). For offensive language detection, a system combination of Support Vector Machines (SVMs) and Deep Neural Networks (DNNs) achieved the best results on development set, which ranked 1st in the official results for Subtask A with F1-score of 90.51% on the test set. For hate speech detection, DNNs were less effective and a system combination of multiple SVMs with different parameters achieved the best results on development set, which ranked 4th in official results for Subtask B with F1-macro score of 80.63% on the test set.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In this paper, we describe our efforts at OSACT Shared Task on Offensive Language Detection. The shared task consists of two subtasks: offensive language detection (Subtask A) and hate speech detection (Subtask B). For offensive language detection, a system combination of Support Vector Machines (SVMs) and Deep Neural Networks (DNNs) achieved the best results on development set, which ranked 1st in the official results for Subtask A with F1-score of 90.51% on the test set. For hate speech detection, DNNs were less effective and a system combination of multiple SVMs with different parameters achieved the best results on development set, which ranked 4th in official results for Subtask B with F1-macro score of 80.63% on the test set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Detecting offensive language or hate speech on social media has gained a lot of interest recently. Use of offensive language and hate speech on social media can be an indication of hate crimes, toxic environment or level of antagonism against individuals or particular groups. Detecting offensive language and hate speech can also help in filtering out inappropriate content for users. Although there is a lot of research on detecting offensive language and hate speech (Agrawal and Awekar, 2018; Djuric et al., 2015; Davidson et al., 2017) , work on Arabic offensive language detection is still in its early stages with very few notable works (Mubarak and Darwish, 2019; Mubarak et al., 2017; Albadi et al., 2018; Alakrot et al., 2018) . Mubarak and Darwish (2019) report that only 1-2% of the tweets are offensive. The highly skewed distribution of data makes it extremely difficult to build useful datasets and effective systems. OS-ACT4 shared task (Mubarak et al., 2020) presents the problem of detecting offensive language and hate speech in Arabic tweets to the community. The shared task consists of 2 subtasks: offensive language detection (subtask A), and hate speech detection (subtask B). This paper describes the systems submitted for OSACT4 shared task on Offensive Language Detection by the team ALT. First, we experimented with classical machine learning classifiers such as Support Vector Machines (SVMs) that are trained on character and word-level features. Then, we experimented with Deep Neural Networks (DNNs) and Bidirectional Encoder Representations from Transformers (BERT). SVMs were seen to outperform the DNNs and BERT. Since we expect the different kinds of classifier to make different kinds of errors, we take the most promising and diverse individual classifiers and perform voting to decide the final output. Majority voting on SVMs, DNNs and BERT yielded better results than individual systems for subtask A on the development set. The best results on development set for subtask B were obtained by combining the output of different SVMs and considering an instance to be hate speech if any of the classifiers voted it to be hate speech.",
                "cite_spans": [
                    {
                        "start": 470,
                        "end": 496,
                        "text": "(Agrawal and Awekar, 2018;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 497,
                        "end": 517,
                        "text": "Djuric et al., 2015;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 518,
                        "end": 540,
                        "text": "Davidson et al., 2017)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 644,
                        "end": 671,
                        "text": "(Mubarak and Darwish, 2019;",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 672,
                        "end": 693,
                        "text": "Mubarak et al., 2017;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 694,
                        "end": 714,
                        "text": "Albadi et al., 2018;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 715,
                        "end": 736,
                        "text": "Alakrot et al., 2018)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 739,
                        "end": 765,
                        "text": "Mubarak and Darwish (2019)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 953,
                        "end": 975,
                        "text": "(Mubarak et al., 2020)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "In section 2, we describe the dataset and the tasks, in section 3, we describe our approach and compare results for subtask A, in section 4, we describe our approach and compare results for subtask B and in section 5, we provide conclusion of our work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1."
            },
            {
                "text": "In this section, we describe the dataset provided to the participants and the two subtasks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset and Task Description",
                "sec_num": "2."
            },
            {
                "text": "The dataset for OSACT Shared task consists of 10,000 Arabic tweets that are tagged for offensiveness and hate speech. The 10,000 tweets are split into training, development and testing sets as shown in Table 1 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 202,
                        "end": 209,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": "2.1."
            },
            {
                "text": "Dev Test 7,000 1,000 2,000 The organizers note that the data is highly skewed. Only 19% of the tweets are tagged as offensive and 5% of the tweets are tagged as hate speech. Table 2 shows examples from the training set. If a tweet has offensive language (insults or threats) targeting a group of people based on their origin (nationality, race, or ethnicity), their ideology (religion, political affiliation, etc.), gender or any other common characteristics, this is considered as hate speech, so all hate speech tweets are offensive according to this definition.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 174,
                        "end": 181,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Train",
                "sec_num": null
            },
            {
                "text": "The participants were required to produce labels indicating if a tweet is offensive (subtask A) or hate speech (subtask B). Each tweet took one of these labels for subtask A: \"OFF\" (offensive) or \"NOT OFF\" (not offensive), and \"HS\" (Hate Speech) or \"NOT HS\" (not Hate Speech) for subtask B. The full train and development data were made ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Task Description",
                "sec_num": "2.2."
            },
            {
                "text": "In this section, we discuss our approach in Subtask A. First, we describe the preprocessing step. Then, we describe the different models we experimented on and compare results of different models. Lastly, we perform an error analysis to understand the limitations of our models and the dataset.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Subtask A: Offensive Language Detection",
                "sec_num": "3."
            },
            {
                "text": "Preprocessing the tweets is an important step as the data from social media can be quite noisy as they contain a lot of emojis, text in mixed languages, excessive use of punctuation etc. It is important to note that some of our models (described in the next subsection) use pretrained word embeddings as feature. In order to reduce noise and be able to find more words in the embeddings, we perform the following steps for preprocessing the tweets.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing Tweets",
                "sec_num": "3.1."
            },
            {
                "text": "Step 1: Remove all words that contain non-Arabic characters.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing Tweets",
                "sec_num": "3.1."
            },
            {
                "text": "Step 2: Remove all diacritics.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing Tweets",
                "sec_num": "3.1."
            },
            {
                "text": "Step 3: Remove all punctuation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing Tweets",
                "sec_num": "3.1."
            },
            {
                "text": "Step 4: Replace repeated characters with only one.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing Tweets",
                "sec_num": "3.1."
            },
            {
                "text": "In our initial experiments, we noticed that the settings listed above produces the best results. Therefore, we keep the same preprocessing settings for all experiments. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preprocessing Tweets",
                "sec_num": "3.1."
            },
            {
                "text": "As the features for SVMs, we transform the tweets into bagof-n-grams vector weighted with logarithmic term frequencies (tf) multiplied with inverse document frequencies (idf). We created both character and word n-grams this way. We experimented with different ranges of character and word n-grams. We also experimented on using Mazajak embeddings (Abu Farha and Magdy, 2019) as features to SVM. Mazajak embeddings were trained on Twitter data, which matches the domain of data in our task. Therefore, we expect it to be more useful compared to other embeddings that are trained on different domain of data (BERT-Multilingual, for example). From table 3, we can see that the best results were obtained when character [1, 5] gram and word [1 \u2212 3] gram features were combined with pretrained Mazajak word embeddings.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "SVMs",
                "sec_num": "3.2.1."
            },
            {
                "text": "FastText is an efficient deep-learning based system for learning embeddings and performing text classification (Joulin et al., 2016) . Since the task of offensive language detection is a text classification problem, we experimented with FastText, but as we can see from Table 3 , FastText was outperformed by other systems.",
                "cite_spans": [
                    {
                        "start": 111,
                        "end": 132,
                        "text": "(Joulin et al., 2016)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 270,
                        "end": 277,
                        "text": "Table 3",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "FastText",
                "sec_num": "3.2.2."
            },
            {
                "text": "For the FFNN architecture, we use four hidden layers with a different number of units (1000, 500, 500, 100) and a sigmoid activation function in each layer, followed by an softmax output layer. To train the network, we use a batch size of 256, with maximum epochs of 50 and early stopping using a 10% of the training set with similar distribution of the labels. The models are then optimized using rmsprop optimization function. The parameter were initialized with small random numbers, sampled from a uniform distribution. Since other systems outperform FFNN, we have not tuned the architecture for different parameters such as number of hidden units or the learning rates, among others.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "FFNN",
                "sec_num": "3.2.3."
            },
            {
                "text": "We use two sets of features for this model. First, we have pretrained word embedding (Mazajak embeddings) features for each word. Second, we use CNN as characterlevel feature extractor. First layer of the network is used to project the input string to character embeddings, which is then passed through a convolutional layer and max-overtime pooling is applied to obtain fixed length representation of words. Character-level representations have been shown to capture morphology of words (Kim et al., 2015 ). If we were to use only word level features obtained from pretrained word embeddings, we would lose out information when a word does not appear in the vocabulary of the pre- ",
                "cite_spans": [
                    {
                        "start": 488,
                        "end": 505,
                        "text": "(Kim et al., 2015",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CNN-BiLSTM",
                "sec_num": "3.2.4."
            },
            {
                "text": "To have a better understanding of where our system is failing will help us improve our system in future and understand the limitation of the data. We examine 100 samples from the development data and attempt to identify where and why our best system (No. 11 from table 3) is failing. Examining the specific tweets provides us with some interesting insights. NOT HS ? Table 8 : Errors analysis of subtask B the table contains reference to Game of Thrones. This is offensive only if context of Game of Thrones is taken into account. We cannot expect the classifier to be correct on such instances. The second entry provides us with an example of error that we can tackle in the future. We see that the hashtag # is offensive toward immigrants. But it's out of vocabulary for our systems. In the future, we can attempt to parse the hashtags into its constituent words and see if it improves the performance. The third entry in the table is quite interesting. The sentence uses an offensive word, , which means \"stealing\", but the sentence itself is not offensive. It's likely our system picked up the offensive word but failed to take into account the context in which the word was used. In future, we can particularly target resolving issues of ambiguous use of words. Table 6 shows the confusion matrix of our best system. We can see that the classifier is much more likely to misclassify offensive instance as not offensive compared to misclassifying not offensive instances as offensive. This is to be expected as the data is highly skewed with only 19% of the instances being offensive.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 367,
                        "end": 374,
                        "text": "Table 8",
                        "ref_id": null
                    },
                    {
                        "start": 1267,
                        "end": 1274,
                        "text": "Table 6",
                        "ref_id": "TABREF8"
                    }
                ],
                "eq_spans": [],
                "section": "Error analysis",
                "sec_num": "3.3."
            },
            {
                "text": "For subtask B, The preprocessing is the same as section 3. In this section, we describe the different models we experimented on for subtask B, present the results for the different models and discuss errors made by the models.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Subtask B: Hate Speech Detection",
                "sec_num": "4."
            },
            {
                "text": "Since same tweets are used for both the subtasks, for subtask B, we focused only on those systems that were promising in subtask A. These systems include SVMs, CNN-BiLSTM and M-BERT. The accuracy, precision, recall and F1 score are reported in Table 7 . The precision, recall and F1 scores are macro averaged.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 244,
                        "end": 251,
                        "text": "Table 7",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "System Descriptions",
                "sec_num": "4.1."
            },
            {
                "text": "As the features for SVMs, we use the same features as subtask A. In addition, we also experiment with bagged SVM (5 estimators). The SVMs outperformed CNN-BiLSTM and M-BERT in this subtask as well.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "SVMs",
                "sec_num": "4.1.1."
            },
            {
                "text": "We keep the same structure and settings of the CNN-BiLSTM used in subtask A.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "CNN-BiLSTM",
                "sec_num": "4.1.2."
            },
            {
                "text": "Once again, we follow the same methodology as subtask A. The only difference is that the M-BERT is now fine-tuned on hate speech detection.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "M-BERT",
                "sec_num": "4.1.3."
            },
            {
                "text": "For subtask B, we opt for a slightly different ensemble method. The best ensemble on the development set was obtained by only considering the three SVMs (Nos. 1,2,3 from table 7). We also change the voting mechanism such that it's no longer majority voting. We consider an instance to be hate speech if any of the three SVMs vote it to be hate speech. This voting scheme was outperforming majority voting scheme on the development set. The official result on this ensemble is shown in ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ensemble Method",
                "sec_num": "4.1.4."
            },
            {
                "text": "We attempt to perform error analysis similar to subtask A by collecting 100 samples from development set and examining them. Unfortunately, we could not identify any particular reasons for the system to fail. We speculate this to be because of the data for subtask B being extremely skewed (with only 5% of the data being hate speech). Table 8 contains examples of errors made by the best system (no. 6 from table 7). Table 10 shows the confusion matrix of our best system on subtask 2. As expected, because of the extreme imbalance of the data, we can see that the classifier is prone to misclassifying hate speech instances as non hate speech instances.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 336,
                        "end": 343,
                        "text": "Table 8",
                        "ref_id": null
                    },
                    {
                        "start": 418,
                        "end": 426,
                        "text": "Table 10",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Error analysis",
                "sec_num": "4.2."
            },
            {
                "text": "To conclude, we experimented heavily with classical machine learning and deep learning approaches to detect if a tweet is offensive or contains hate speech. We achieve state of the art results for offensive language detection by combination of SVMs, CNN-BiLSTM and Multilingual BERT. We achieve competitive results on hate speech detection with a system combination of SVMs. Our error analysis indicates certain types of errors for offensive language detection that can be addressed in future. In future, we aim to explore augmentation of hate speech data to build better systems.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion and Future Work",
                "sec_num": "5."
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Mazajak: An online Arabic sentiment analyser",
                "authors": [
                    {
                        "first": "Abu",
                        "middle": [],
                        "last": "Farha",
                        "suffix": ""
                    },
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Magdy",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the Fourth Arabic Natural Language Processing Workshop",
                "volume": "",
                "issue": "",
                "pages": "192--198",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Abu Farha, I. and Magdy, W. (2019). Mazajak: An online Arabic sentiment analyser. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 192-198, Florence, Italy, August. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Deep learning for detecting cyberbullying across multiple social media platforms",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Agrawal",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Awekar",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Agrawal, S. and Awekar, A. (2018). Deep learning for de- tecting cyberbullying across multiple social media plat- forms. CoRR, abs/1801.06482.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "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. (2018). To- wards accurate detection of offensive language in online communication in arabic. Procedia Computer Science, 142:315 -320. Arabic Computational Linguistics.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "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": "2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM)",
                "volume": "",
                "issue": "",
                "pages": "69--76",
                "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. In 2018 IEEE/ACM Interna- tional Conference on Advances in Social Networks Anal- ysis and Mining (ASONAM), pages 69-76, Aug.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "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": [
                            "W"
                        ],
                        "last": "Macy",
                        "suffix": ""
                    },
                    {
                        "first": "I",
                        "middle": [],
                        "last": "Weber",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Davidson, T., Warmsley, D., Macy, M. W., and Weber, I. (2017). Automated hate speech detection and the prob- lem of offensive language. CoRR, abs/1703.04009.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
                "authors": [
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "M.-W",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "4171--4186",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2019). BERT: Pre-training of deep bidirectional trans- formers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Pa- pers), pages 4171-4186, Minneapolis, Minnesota, June. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Hate speech detection with comment embeddings",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Djuric",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Morris",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Grbovic",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Radosavljevic",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Bhamidipati",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 24th International Conference on World Wide Web, WWW '15 Companion",
                "volume": "",
                "issue": "",
                "pages": "29--30",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Djuric, N., Zhou, J., Morris, R., Grbovic, M., Radosavlje- vic, V., and Bhamidipati, N. (2015). Hate speech de- tection with comment embeddings. In Proceedings of the 24th International Conference on World Wide Web, WWW '15 Companion, page 29-30, New York, NY, USA. Association for Computing Machinery.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Fasttext.zip: Compressing text classification models",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Joulin",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Grave",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Bojanowski",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Douze",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "J\u00e9gou",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Mikolov",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Joulin, A., Grave, E., Bojanowski, P., Douze, M., J\u00e9gou, H., and Mikolov, T. (2016). Fasttext.zip: Compressing text classification models. CoRR, abs/1612.03651.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Character-aware neural language models",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Jernite",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [
                            "A"
                        ],
                        "last": "Sontag",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "M"
                        ],
                        "last": "Rush",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kim, Y., Jernite, Y., Sontag, D. A., and Rush, A. M. (2015). Character-aware neural language models. CoRR, abs/1508.06615.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Arabic offensive language classification on twitter",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Mubarak",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Darwish",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "International Conference on Social Informatics",
                "volume": "",
                "issue": "",
                "pages": "269--276",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mubarak, H. and Darwish, K. (2019). Arabic offensive language classification on twitter. In International Con- ference on Social Informatics, pages 269-276. Springer.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "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). Abu- sive language detection on Arabic social media. In Pro- ceedings of the First Workshop on Abusive Language On- line, pages 52-56, Vancouver, BC, Canada, August. As- sociation for Computational Linguistics.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Overview of osact4 arabic offensive language detection shared task",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Mubarak",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Darwish",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [],
                        "last": "Magdy",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Elsayed",
                        "suffix": ""
                    },
                    {
                        "first": "Al-Khalifa",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "4",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mubarak, H., Darwish, K., Magdy, W., Elsayed, T., and Al- Khalifa, H. (2020). Overview of osact4 arabic offensive language detection shared task. 4.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF0": {
                "num": null,
                "html": null,
                "content": "<table/>",
                "type_str": "table",
                "text": ""
            },
            "TABREF2": {
                "num": null,
                "html": null,
                "content": "<table><tr><td>: Examples from the dataset</td></tr></table>",
                "type_str": "table",
                "text": ""
            },
            "TABREF5": {
                "num": null,
                "html": null,
                "content": "<table><tr><td>Tweet True</td><td>Majority Reason</td></tr><tr><td>Label</td><td>Label</td></tr></table>",
                "type_str": "table",
                "text": "Comparison of different systems submitted to subtask A"
            },
            "TABREF6": {
                "num": null,
                "html": null,
                "content": "<table><tr><td>: Error analysis of subtask A</td></tr></table>",
                "type_str": "table",
                "text": "Wikipedia text, is quite different from the Twitter data used in the shared task. Articles on Wikipedia are typically written in a formal way and follow the structure and rules of grammar. Text on social media platforms such as Twitter, on the other hand, can be very informal and chaotic."
            },
            "TABREF7": {
                "num": null,
                "html": null,
                "content": "<table><tr><td/><td colspan=\"2\">: Official results on subtask A test set</td></tr><tr><td/><td>NOT OFF</td><td>OFF</td></tr><tr><td colspan=\"2\">NOT OFF 789</td><td>32</td></tr><tr><td>OFF</td><td>25</td><td>154</td></tr></table>",
                "type_str": "table",
                "text": ""
            },
            "TABREF8": {
                "num": null,
                "html": null,
                "content": "<table/>",
                "type_str": "table",
                "text": "Confusion matrix on subtask A dev set"
            },
            "TABREF9": {
                "num": null,
                "html": null,
                "content": "<table><tr><td colspan=\"2\">No. Model</td><td>Features</td><td>Acc.</td><td>Prec.</td><td>Recall</td><td>F1</td></tr><tr><td>1</td><td>SVM</td><td>Mazajak SG-250 + char[1-5] + word[1-3]</td><td>96.2</td><td>72</td><td>78.2</td><td>74.7</td></tr><tr><td>2</td><td/><td>Char [2-6] Gram</td><td>97.1</td><td>72.5</td><td>88.8</td><td>78.2</td></tr><tr><td>3</td><td>Bagged SVM</td><td>Mazajak SG-250 + char[1-5] + word[1-3]</td><td>96.4</td><td>68.9</td><td>81.3</td><td>73.4</td></tr><tr><td>4</td><td>CNN-BiLSTM</td><td colspan=\"2\">Mazajak SG-100 + CNN Feature Extractor 95.9</td><td>76.1</td><td>67.5</td><td>70.9</td></tr><tr><td>5</td><td>M-BERT</td><td/><td>95.7</td><td>72.4</td><td>74.9</td><td>73</td></tr><tr><td>11</td><td>Ensemble(1+2+3)</td><td/><td>96.6</td><td>80</td><td>78.7</td><td>79.3</td></tr><tr><td/><td colspan=\"4\">Table 7: Comparison of different systems submitted to subtask B</td><td/><td/></tr><tr><td/><td/><td colspan=\"2\">Tweet True</td><td/><td colspan=\"2\">Majority Reason</td></tr><tr><td/><td/><td/><td>Label</td><td/><td>Label</td><td/></tr><tr><td/><td/><td/><td>HS</td><td/><td>NOT</td><td>?</td></tr><tr><td/><td/><td/><td>HS</td><td/><td>NOT</td><td>?</td></tr></table>",
                "type_str": "table",
                "text": ""
            },
            "TABREF10": {
                "num": null,
                "html": null,
                "content": "<table><tr><td>Official</td><td>Acc.</td><td>Prec.</td><td>Recall</td><td>F1</td></tr><tr><td>Rank</td><td/><td/><td/><td/></tr><tr><td>4th</td><td>96.6%</td><td>83.8%</td><td>78.1%</td><td>80.6%</td></tr></table>",
                "type_str": "table",
                "text": ""
            },
            "TABREF11": {
                "num": null,
                "html": null,
                "content": "<table><tr><td/><td>NOT HS</td><td>HS</td></tr><tr><td colspan=\"2\">NOT HS 940</td><td>16</td></tr><tr><td>HS</td><td>18</td><td>26</td></tr></table>",
                "type_str": "table",
                "text": "Official results on subtask B test set"
            },
            "TABREF12": {
                "num": null,
                "html": null,
                "content": "<table/>",
                "type_str": "table",
                "text": "Confusion matrix on subtask B dev set"
            }
        }
    }
}