File size: 65,560 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 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:06:24.738709Z"
},
"title": "Multi-Task Learning using AraBert for Offensive Language Detection",
"authors": [
{
"first": "Marc",
"middle": [],
"last": "Djandji",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "American University of Beirut",
"location": {}
},
"email": ""
},
{
"first": "Fady",
"middle": [],
"last": "Baly",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "American University of Beirut",
"location": {}
},
"email": ""
},
{
"first": "Wissam",
"middle": [],
"last": "Antoun",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "American University of Beirut",
"location": {}
},
"email": ""
},
{
"first": "Hazem",
"middle": [],
"last": "Hajj",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "American University of Beirut",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The use of social media platforms has become more prevalent, which has provided tremendous opportunities for people to connect but has also opened the door for misuse with the spread of hate speech and offensive language. This phenomenon has been driving more and more people to more extreme reactions and online aggression, sometimes causing physical harm to individuals or groups of people. There is a need to control and prevent such misuse of online social media through automatic detection of profane language. The shared task on Offensive Language Detection at the OSACT4 has aimed at achieving state of art profane language detection methods for Arabic social media. Our team \"BERTologists\" tackled this problem by leveraging state of the art pretrained Arabic language model, AraBERT, that we augment with the addition of Multi-task learning to enable our model to learn efficiently from little data. Our Multitask AraBERT approach achieved the second place in both subtasks A & B, which shows that the model performs consistently across different tasks.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "The use of social media platforms has become more prevalent, which has provided tremendous opportunities for people to connect but has also opened the door for misuse with the spread of hate speech and offensive language. This phenomenon has been driving more and more people to more extreme reactions and online aggression, sometimes causing physical harm to individuals or groups of people. There is a need to control and prevent such misuse of online social media through automatic detection of profane language. The shared task on Offensive Language Detection at the OSACT4 has aimed at achieving state of art profane language detection methods for Arabic social media. Our team \"BERTologists\" tackled this problem by leveraging state of the art pretrained Arabic language model, AraBERT, that we augment with the addition of Multi-task learning to enable our model to learn efficiently from little data. Our Multitask AraBERT approach achieved the second place in both subtasks A & B, which shows that the model performs consistently across different tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Offensive language, including hate speech, is a violent behavior that is becoming more and more pervasive across public social media platforms (Fosler-Lussier et al., 2012) . Hate speech was found to negatively impact the psychological well-being of individuals and to deteriorate inter-group relations on the societal level (Tynes et al., 2008) . As such, detection and prevention mechanisms should be setup to deal with such content. Machine learning algorithms can be employed to automatically detect these behaviors by relying on recent techniques in natural language processing that have shown propitious performance. A small number of works targeted the problem of simultaneously detecting both hate and offensive speech in Arabic. For example, Haddad et al. (2019) targeted the problem of hate and offensive speech detection for the Tunisian dialect using Support Vector Machine (SVM) and Naive Bayes classifiers trained on hand crafted features. Mulki et al. (2019) targeted the detection of profane language for the Levantine dialect using SVM and NB models trained on hand-crafted features. Although these works provided insights into the features that could be used for Arabic hate and offensive speech detection and introduced datasets for these specific dialects, they are limited to these specific dialects and do not target the problem of developing models that can learn efficiently with little data. In the 4th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT4) (Mubarak et al., 2020) the shared task on offensive language aimed at offensive and hate speech detection in Arabic tweets. The task is split up into two Subtasks: Subtask A) which aimed at detecting whether a tweet is offensive or not and Subtask B) which aimed at detecting whether a tweet is hate-speech or not. The organizers labeled a tweet as offensive if it contained explicit or implicit insults directed towards other people or inappropriate language. While a tweet labeled as hate speech contains targeted insults towards a group based on their nationality, ethnicity, gender, political or sport affiliation. Each subtask is evaluated independently with a macro-F1 score. The dataset had the following issues that also needed to be addressed: (i) The labeled tweets were written in dialectal Arabic which had inconsistent writing style and vocabulary (ii) The class labels were highly imbalanced especially in the hate speech case where only 5% of the data was labeled as hate speech. The models that we experimented with are all based on fine-tuning the Arabic Bidirectional Encoder Representation from Transformer (AraBERT) model (AUBMind-Lab, 2020) with different training classification schemes. To enable the model to learn from little data and not overfit to the dominant class, we train AraBERT in a multitask paradigm. Our contributions can be summarized as follows:",
"cite_spans": [
{
"start": 143,
"end": 172,
"text": "(Fosler-Lussier et al., 2012)",
"ref_id": "BIBREF10"
},
{
"start": 325,
"end": 345,
"text": "(Tynes et al., 2008)",
"ref_id": "BIBREF21"
},
{
"start": 751,
"end": 771,
"text": "Haddad et al. (2019)",
"ref_id": "BIBREF11"
},
{
"start": 954,
"end": 973,
"text": "Mulki et al. (2019)",
"ref_id": "BIBREF17"
},
{
"start": 1497,
"end": 1519,
"text": "(Mubarak et al., 2020)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "\u2022 Comprehensive evaluation including the impact of different sampling techniques and weighted loss functions that penalizes wrong predictions on the minority class in an attempt to balance the data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "\u2022 Propose a new model that combines AraBert and multi-task learning to achieve accurate predictions and address data imbalance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "\u2022 Propose a model that provides consistent performance on both hate and offensive speech detection with the presence of different Arabic dialects. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Hate Speech Detection An extensive overview of the different works on hate speech detection was done by (Al-Hassan and Al-Dossari, 2019), but very few works in the literature target the problem of Arabic hate speech detection. Albadi et al. (2018) introduced the first dataset containing 6.6K Arabic hate-speech tweets targeting religious groups. The authors compared a lexicon-based classifier, SVM classifier trained with character n-gram features, and a Deep Learning approach consisting of a GRU trained on AraVec embeddings (Soliman et al., 2017) . The GRU approach outperformed all other approaches with a 77% F1 score.",
"cite_spans": [
{
"start": 227,
"end": 247,
"text": "Albadi et al. (2018)",
"ref_id": "BIBREF3"
},
{
"start": 529,
"end": 551,
"text": "(Soliman et al., 2017)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Hate and Offensive Speech Detection",
"sec_num": "2.2."
},
{
"text": "Offensive Speech Detection For offensive speech detection in Arabic, different approaches can be found in the literature. Alakrot et al. (2018) , introduced a dataset for offensive speech in Arabic collected from 15K YouTube comments. For classifying the different comments, the data was preprocessed by removing stop words and diacritics, correcting misspelled words, then tokenization and stemming was performed in order to extract features that are used by a binary SVM classifier. Mohaouchane et al. 2019 ",
"cite_spans": [
{
"start": 122,
"end": 143,
"text": "Alakrot et al. (2018)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Hate and Offensive Speech Detection",
"sec_num": "2.2."
},
{
"text": "We based our approaches on the recently released AraBERT model. AraBERT is a Bidirectional representation of a text sequence, pretrained on a large Arabic corpus that achieved state of the art performance on multiple Arabic NLP tasks. Our best model is based on augmenting AraBERT with Multitask Learning, which solves the data imbalance problem by leveraging information from multiple tasks simultaneously. We also compare our best model with other approaches that are used to solve class imbalance issues such as balanced batch sampling and Multilabel classification. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed Models",
"sec_num": "3."
},
{
"text": "Multitask Learning is a learning paradigm that endows the developed models with the human-like abilities of transferring the important learned information between related tasks in what is called inductive transfer of knowledge under the assumption that commonalities exist between the learned tasks. Furthermore, the main advantages of MTL are that it reduces the requirements for large amounts of labeled data, improves the performance of a task with fewer data by leveraging the shared information from the related tasks with more data, and enables the model to be robust to missing observations for some tasks (Caruana, 1997; Qiu et al., 2017) . Given that little data is available for both hate and offensive classes, we use an MTL approach to augment the initial AraBERT model such that it can learn both tasks simultaneously, which reduces the overfitting effect induced by the dominant not offensive and not hate examples. Our MTL-Arabert model consists of two components as can be seen in Figure 1 : a part that gets trained by all the tasks' data in order to extract a general feature representation for all the tasks and a task-specific part that gets trained only by the task-specific examples to capture the task-specific characteristics.",
"cite_spans": [
{
"start": 613,
"end": 628,
"text": "(Caruana, 1997;",
"ref_id": "BIBREF6"
},
{
"start": 629,
"end": 646,
"text": "Qiu et al., 2017)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [
{
"start": 997,
"end": 1005,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Multitask Learning (MTL)",
"sec_num": "3.1."
},
{
"text": "1. Shared Part: Contains the pretrained AraBert model that gets tuned by the combined loss of both tasks in order to learn a shared set of information between both tasks 2. Task-specific layers: These consist of a task-specific dense layer that are dedicated to extracting the unique information per task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Multitask Learning (MTL)",
"sec_num": "3.1."
},
{
"text": "Multilabel Classification Multilabel classification is the task of classifying a single instance with multiple labels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Other Approaches",
"sec_num": "3.2."
},
{
"text": "We considered using this approach for two main reasons. Firstly, the subtasks are very coherent as they both try to solve problems that behaviorally fall under the same general idea, detecting violent behaviors. Secondly, considering that subtask B has very little hate speech labeled data and that all hate speech data is also labeled as offensive, we assumed that a multilabel classifier would help leverage and provide a better understanding of the hate speech instances as they are being trained simultaneously with the offensive instances. We also explored oversampling the Task B instances and made sure that each training batch included samples of hate speech data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Other Approaches",
"sec_num": "3.2."
},
{
"text": "Weighted Cross-Entropy loss Cross-entropy loss is useful in classification tasks, since the loss increases as the predicted probability diverges from the actual label. The Weighted version, penalizes each class differently, according to the given weight. The weighted cross-entropy loss of a class i with weight W i is shown in 1, the weight vector is given in 2",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Other Approaches",
"sec_num": "3.2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(x i ) = \u2212W i log exp(x i ) j exp(x j )",
"eq_num": "(1)"
}
],
"section": "Other Approaches",
"sec_num": "3.2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "W i = N o Samples N o Classes \u00d7 Count(i)",
"eq_num": "(2)"
}
],
"section": "Other Approaches",
"sec_num": "3.2."
},
{
"text": "Balanced batch sampling We re-sample the dataset in such a way that we under-sample the majority class and over-sample the minority class at the same time. Which reduces information loss due to under-sampling, and minimizes overfitting due to over-sampling, since the over/under-sampling is done to a lesser extent compared to independently implementing over/under-sampling.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Other Approaches",
"sec_num": "3.2."
},
{
"text": "The dataset for both tasks is the same containing 10K tweets that were annotated for offensiveness with labels (OFF or NOT OFF) and hate speech with labels (HS or NOT HS). The data was split by the competition organizers into 70% training set, 10% development set, and 20% test set. Table 1 shows the data distribution among the different labels and splits. By examining Table 1 , it can be seen that the data is very imbalanced having only 5% of the examples labeled as hate speech and 20% of the examples labeled as offensive in the training dataset, which makes the tasks much harder and calls for methods that can learn efficiently from little data. ",
"cite_spans": [],
"ref_spans": [
{
"start": 283,
"end": 290,
"text": "Table 1",
"ref_id": "TABREF2"
},
{
"start": 371,
"end": 378,
"text": "Table 1",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Data Description",
"sec_num": "4.1."
},
{
"text": "For preprocessing the data, we tokenized Arabic words with the Farasa Arabic segmenter (Abdelali et al., 2016) so that the input would be compatible with the AraBERT input. For example, \" -Almadrasa\" becomes \" -Al+ madras +T\". We also removed all mentions of the user tokens \"USER\", retweet mentions \"RT USER:\", URL tokens, the \"<LF>\" tokens, diacritics, and emojis. As for hashtags, we replaced the underscore within a hashtag \" \" with a white space to regain separate understandable tokens, and we pad the hashtag with a whitespace as well. For instance, \" \" turns into \" #\".",
"cite_spans": [
{
"start": 87,
"end": 110,
"text": "(Abdelali et al., 2016)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "4.2."
},
{
"text": "We should also mention that these preprocessing steps are precisely applied to all the experiments conducted for both subtasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "4.2."
},
{
"text": "Both tasks were evaluated using the unweighted-average F1 of all classes, which is the macro-F1 score. Given the high imbalance in the dataset and that the macro-F1 score is penalized by the minority class, achieving a high macro-F1 score is challenging. Table 2 and 3 provide the results of our models on the development and test set, respectively. All three models were trained on the whole training set for five epochs with a batch size of 32 and a sequence length of 256 in a GPU-accelerated environment. The epoch-model that achieved the highest macro-F1 score on the dev-set is reported in Table 2 . We only show the results of our best MTL model on the test data in Table 3 as provided by the competition organizers. Our Multitask approach shows consistent performance on both the dev and test sets across both tasks. The results show that training both tasks jointly in a Multitask setting improves the model generalizability with the presence of little data for each task. The results for the hate speech task are not as good as the offensive language task due to the minimal number of hate speech training examples, which constitute 5% of the training data. Although when combined, balanced batch sampling and weighted loss achieved the second best results on task A. When used separately, both approaches performed worse than the baseline model. This might be due to the overfitting effect of oversampling the minority class. While examining the false predictions of our MTL model on the dev set, we noticed that the model was classifying tweets with a negative sentiment as offensive tweets. While it is intuitive for offensive tweets to have a negative sentiment by nature, our model did not capture the fact that not all tweets with negative sentiment are offensive. On another note, the use of words that are offensive in a nonoffensive context was found to confuse the model. For example, the words \" \" and \" \" in the following tweets (720, 828), respectively, were not used with an offensive intent and made the model classify both tweets as offensive. We also found that the model has learned that a tweet cannot be hate-speech unless it is offensive, which would be ideal in case the offensive prediction was perfect. However, in our case, this also made the model falsely predict three tweets as hate-speech after they were falsely predicted as offensive. Furthermore, tweets 785, 881 in the dev set were found to be mislabeled as hate speech, and the model was able to detect this error showing a good understanding of what characterizes hate speech in a tweet. Finally, we found our model to falsely predict tweets that mostly contain mockery, sarcasm, or quoting other offensive/hateful statements. Future work should explore the use of data augmentation techniques such as adversarial examples and learning from little data approaches such as meta-learning in order to enable state-of-the-art Natural Language Understanding (NLU) models such as AraBERT to be trained efficiently with little data.",
"cite_spans": [],
"ref_spans": [
{
"start": 255,
"end": 262,
"text": "Table 2",
"ref_id": "TABREF3"
},
{
"start": 596,
"end": 603,
"text": "Table 2",
"ref_id": "TABREF3"
},
{
"start": 673,
"end": 680,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "4.3."
},
{
"text": "The presence of hate speech and offensive language on Arabic social platforms is a major issue affecting the social lives of many individuals in the Arab world. The lack of annotated data and the presence of different dialects constitutes major challenges for automated Arabic offensive and hate speech detection systems. In this paper, we proposed the use of pre-trained Arabic BERT for accurate classification of the different tweets. We further augment the AraBERT model using Multitask Learning to enable the model to jointly learn both tasks efficiently with the presence of little labeled data per-task. Our results show the superiority of our proposed Multitask AraBERT model over single-task and Multilabel AraBERT. We explore different methods in order to cope with the presence of imbalanced training classes such as the use a weighted loss function and data re-sampling techniques, but found these methods to not introduce any improvements. Our method achieved the second place on both tasks in the OSACT4 competition.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5."
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Farasa: A fast and furious segmenter for arabic",
"authors": [
{
"first": "A",
"middle": [],
"last": "Abdelali",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Darwish",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Durrani",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Mubarak",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations",
"volume": "",
"issue": "",
"pages": "11--16",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Abdelali, A., Darwish, K., Durrani, N., and Mubarak, H. (2016). Farasa: A fast and furious segmenter for ara- bic. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pages 11-16.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Detection of hate speech in social networks: a survey on multilingual corpus",
"authors": [
{
"first": "A",
"middle": [],
"last": "Al-Hassan",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Al-Dossari",
"suffix": ""
}
],
"year": 2019,
"venue": "6th International Conference on Computer Science and Information Technology",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Al-Hassan, A. and Al-Dossari, H. (2019). Detection of hate speech in social networks: a survey on multilin- gual corpus. In 6th International Conference on Com- puter Science and Information Technology.",
"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.",
"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. IEEE. AUBMind-Lab. (2020). https://github.com/aub- mind/arabert.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Multilingual detection of hate speech against immigrants and women in twitter",
"authors": [],
"year": null,
"venue": "Proceedings of the 13th International Workshop on Semantic Evaluation",
"volume": "",
"issue": "",
"pages": "54--63",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Semeval-2019 task 5: Multilingual detection of hate speech against immigrants and women in twitter. In Pro- ceedings of the 13th International Workshop on Semantic Evaluation, pages 54-63.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Multitask learning. Machine learning",
"authors": [
{
"first": "R",
"middle": [],
"last": "Caruana",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "28",
"issue": "",
"pages": "41--75",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Caruana, R. (1997). Multitask learning. Machine learn- ing, 28(1):41-75.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Universal sentence encoder",
"authors": [
{
"first": "D",
"middle": [],
"last": "Cer",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "S.-Y",
"middle": [],
"last": "Kong",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Hua",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Limtiaco",
"suffix": ""
},
{
"first": "R",
"middle": [
"S"
],
"last": "John",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Constant",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Guajardo-Cespedes",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Yuan",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Tar",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1803.11175"
]
},
"num": null,
"urls": [],
"raw_text": "Cer, D., Yang, Y., Kong, S.-y., Hua, N., Limtiaco, N., John, R. S., Constant, N., Guajardo-Cespedes, M., Yuan, S., Tar, C., et al. (2018). Universal sentence encoder. arXiv preprint arXiv:1803.11175.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"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",
"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 prob- lem of offensive language. In Eleventh international aaai conference on web and social media.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"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": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1810.04805"
]
},
"num": null,
"urls": [],
"raw_text": "Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2018). Bert: Pre-training of deep bidirectional trans- formers for language understanding. arXiv preprint arXiv:1810.04805.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Proceedings of the 2012 conference of the north american chapter of the association for computational linguistics: Human language technologies",
"authors": [
{
"first": "E",
"middle": [],
"last": "Fosler-Lussier",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Riloff",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Bangalore",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fosler-Lussier, E., Riloff, E., and Bangalore, S. (2012). Proceedings of the 2012 conference of the north ameri- can chapter of the association for computational linguis- tics: Human language technologies. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "T-hsab: A tunisian hate speech and abusive dataset",
"authors": [
{
"first": "H",
"middle": [],
"last": "Haddad",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Mulki",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Oueslati",
"suffix": ""
}
],
"year": 2019,
"venue": "International Conference on Arabic Language Processing",
"volume": "",
"issue": "",
"pages": "251--263",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Haddad, H., Mulki, H., and Oueslati, A. (2019). T-hsab: A tunisian hate speech and abusive dataset. In Inter- national Conference on Arabic Language Processing, pages 251-263. Springer.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Overview of the hasoc track at fire 2019: Hate speech and offensive content identification in indo-european languages",
"authors": [
{
"first": "T",
"middle": [],
"last": "Mandl",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Modha",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Majumder",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Patel",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Dave",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Mandlia",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Patel",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 11th Forum for Information Retrieval Evaluation",
"volume": "",
"issue": "",
"pages": "14--17",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mandl, T., Modha, S., Majumder, P., Patel, D., Dave, M., Mandlia, C., and Patel, A. (2019). Overview of the hasoc track at fire 2019: Hate speech and offensive con- tent identification in indo-european languages. In Pro- ceedings of the 11th Forum for Information Retrieval Evaluation, pages 14-17.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Detecting offensive language on arabic social media using deep learning",
"authors": [],
"year": 2019,
"venue": "Sixth International Conference on Social Networks Analysis, Management and Security (SNAMS)",
"volume": "",
"issue": "",
"pages": "466--471",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Detecting offensive language on arabic social media us- ing deep learning. In 2019 Sixth International Confer- ence on Social Networks Analysis, Management and Se- curity (SNAMS), pages 466-471. IEEE.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"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
},
"BIBREF16": {
"ref_id": "b16",
"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
},
"BIBREF17": {
"ref_id": "b17",
"title": "L-hsab: A levantine twitter dataset for hate speech and abusive language",
"authors": [
{
"first": "H",
"middle": [],
"last": "Mulki",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Haddad",
"suffix": ""
},
{
"first": "C",
"middle": [
"B"
],
"last": "Ali",
"suffix": ""
},
{
"first": "Alshabani",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the Third Workshop on Abusive Language Online",
"volume": "",
"issue": "",
"pages": "111--118",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mulki, H., Haddad, H., Ali, C. B., and Alshabani, H. (2019). L-hsab: A levantine twitter dataset for hate speech and abusive language. In Proceedings of the Third Workshop on Abusive Language Online, pages 111-118.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "A short-term rainfall prediction model using multi-task convolutional neural networks",
"authors": [
{
"first": "M",
"middle": [],
"last": "Qiu",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Shi",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Chu",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2017,
"venue": "2017 IEEE International Conference on Data Mining (ICDM)",
"volume": "",
"issue": "",
"pages": "395--404",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Qiu, M., Zhao, P., Zhang, K., Huang, J., Shi, X., Wang, X., and Chu, W. (2017). A short-term rainfall prediction model using multi-task convolutional neural networks. In 2017 IEEE International Conference on Data Mining (ICDM), pages 395-404. IEEE.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A survey on hate speech detection using natural language processing",
"authors": [
{
"first": "A",
"middle": [],
"last": "Schmidt",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Wiegand",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media",
"volume": "",
"issue": "",
"pages": "1--10",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Schmidt, A. and Wiegand, M. (2017). A survey on hate speech detection using natural language processing. In Proceedings of the Fifth International Workshop on Nat- ural Language Processing for Social Media, pages 1-10.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Aravec: A set of arabic word embedding models for use in arabic nlp",
"authors": [
{
"first": "A",
"middle": [
"B"
],
"last": "Soliman",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Eissa",
"suffix": ""
},
{
"first": "S",
"middle": [
"R"
],
"last": "El-Beltagy",
"suffix": ""
}
],
"year": 2017,
"venue": "Procedia Computer Science",
"volume": "117",
"issue": "",
"pages": "256--265",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Soliman, A. B., Eissa, K., and El-Beltagy, S. R. (2017). Aravec: A set of arabic word embedding models for use in arabic nlp. Procedia Computer Science, 117:256-265.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Online racial discrimination and psychological adjustment among adolescents",
"authors": [
{
"first": "B",
"middle": [
"M"
],
"last": "Tynes",
"suffix": ""
},
{
"first": "M",
"middle": [
"T"
],
"last": "Giang",
"suffix": ""
},
{
"first": "D",
"middle": [
"R"
],
"last": "Williams",
"suffix": ""
},
{
"first": "G",
"middle": [
"N"
],
"last": "Thompson",
"suffix": ""
}
],
"year": 2008,
"venue": "Journal of adolescent health",
"volume": "43",
"issue": "6",
"pages": "565--569",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tynes, B. M., Giang, M. T., Williams, D. R., and Thomp- son, G. N. (2008). Online racial discrimination and psy- chological adjustment among adolescents. Journal of adolescent health, 43(6):565-569.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Hateful symbols or hateful people? predictive features for hate speech detection on twitter",
"authors": [
{
"first": "Z",
"middle": [],
"last": "Waseem",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Hovy",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the NAACL student research workshop",
"volume": "",
"issue": "",
"pages": "88--93",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Waseem, Z. and Hovy, D. (2016). Hateful symbols or hate- ful people? predictive features for hate speech detec- tion on twitter. In Proceedings of the NAACL student research workshop, pages 88-93.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Semeval-2019 task 6: Identifying and categorizing offensive language in social media (offenseval)",
"authors": [
{
"first": "M",
"middle": [],
"last": "Zampieri",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Malmasi",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Nakov",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Rosenthal",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Farra",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Kumar",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1903.08983"
]
},
"num": null,
"urls": [],
"raw_text": "Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N., and Kumar, R. (2019). Semeval-2019 task 6: Identi- fying and categorizing offensive language in social me- dia (offenseval). arXiv preprint arXiv:1903.08983.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "The trained Multitask Learning model given an input offensive tweet"
},
"TABREF1": {
"num": null,
"type_str": "table",
"html": null,
"text": "The most popular classifiers in the literature are SVM, NB, LSTM, CNN, GRU. The best performing systems employed a CNN model and AraVec embeddings for offensive speech detection and a GRU model on AraVec embeddings for hatespeech detection. Very little work can be found in the literature for Arabic hate and offensive speech detection. The current work does not address the multiple dialects and little data challenges for these tasks.",
"content": "<table><tr><td>, ex-</td></tr><tr><td>plored the use of different Deep Learning architectures for</td></tr><tr><td>offensive language detection. AraVec embeddings of each</td></tr><tr><td>comment were used to train several models: CNN-LSTM,</td></tr><tr><td>CNN-BiLSTM with attention, Bi-LSTM, and CNN model</td></tr><tr><td>on the dataset proposed in (Alakrot et al., 2018) where the</td></tr><tr><td>CNN model was found to provide the best F1 score. In</td></tr><tr><td>Mubarak and Darwish (2019) 36 million tweets were col-</td></tr><tr><td>lected and used it to train a FastText deep learning model</td></tr><tr><td>and SVM classifier on character n-gram features where it</td></tr><tr><td>was found that the Arabic FastText DL model provided the</td></tr><tr><td>best results.</td></tr></table>"
},
"TABREF2": {
"num": null,
"type_str": "table",
"html": null,
"text": "The data distribution for both tasks. The first two rows show the class distribution of task A. The second two rows show the class distribution of task B",
"content": "<table><tr><td>Class</td><td colspan=\"2\">Training Developement</td></tr><tr><td>NOT OFF</td><td>5468</td><td>821</td></tr><tr><td>OFF</td><td>1371</td><td>179</td></tr><tr><td>NOT HS</td><td>6489</td><td>956</td></tr><tr><td>HS</td><td>350</td><td>44</td></tr></table>"
},
"TABREF3": {
"num": null,
"type_str": "table",
"html": null,
"text": "The performance of the different approaches on the development set for both tasks using the Macro-F1 score metric. It can be seen that the Multitask approach outperforms all other approaches",
"content": "<table><tr><td>Model</td><td>Macro-F1 Offensive Language</td><td>Hate Speech</td></tr><tr><td>AraBERT</td><td>89.56</td><td>80.60</td></tr><tr><td>AraBERT-S*</td><td>87.24</td><td>79.42</td></tr><tr><td>AraBERT-W**</td><td>88.17</td><td>79.85</td></tr><tr><td>AraBERT-SW***</td><td>90.02</td><td>78.13</td></tr><tr><td>Multilable AraBERT</td><td>89.41</td><td>79.83</td></tr><tr><td>Multilable AraBERT*</td><td>89.55</td><td>80.81</td></tr><tr><td>Multitask AraBERT</td><td>90.15</td><td>83.41</td></tr><tr><td colspan=\"2\">* AraBERT with balanced batch sampling</td><td/></tr><tr><td colspan=\"2\">** AraBERT with weighted loss</td><td/></tr><tr><td colspan=\"3\">*** AraBERT with both balanced batch sampling and weighted loss</td></tr></table>"
},
"TABREF4": {
"num": null,
"type_str": "table",
"html": null,
"text": "The performance of the Multitask Learning (MTL) model on the test set for both tasks using the Macro-F1 score metric.",
"content": "<table><tr><td>Model</td><td colspan=\"2\">Task A: Macro-F1 Task B: Macro-F1</td></tr><tr><td>Multitask AraBERT</td><td>90</td><td>82.28</td></tr></table>"
}
}
}
} |