Fine-tuning of Pre-trained Transformers for Hate, Ofensive, and Profane Content Detection in English and Marathi

Abstract

This paper describes neural models developed for the Hate Speech and Ofensive Content Identification in English and Indo-Aryan Languages Shared Task 2021. Our team called neuro-utmn-thales participated in two tasks on binary and fine-grained classification of English tweets that contain hate, ofensive, and profane content (English Subtasks A & B) and one task on identification of problematic content in Marathi (Marathi Subtask A). For English subtasks, we investigate the impact of additional corpora for hate speech detection to fine-tune transformer models. We also apply a one-vs-rest approach based on Twitter-RoBERTa to discrimination between hate, profane and ofensive posts. Our models ranked third in English Subtask A with the F1-score of 81.99% and ranked second in English Subtask B with the F1-score of 65.77%. For the Marathi tasks, we propose a system based on the Language-Agnostic BERT Sentence Embedding (LaBSE). This model achieved the second result in Marathi Subtask A obtaining an media safer. The Hate Speech and Ofensive Content Identification in English and Indo-Aryan

Keywords

• Hate speech
• ofensive language identification
• text classification
• transformer neural networks
• Twitter-

1. Introduction

Social media has a greater impact on our society. Social networks give us almost limitless freedom of speech and contribute to the rapid dissemination of information. However, these positive properties often lead to unhealthy usage of social media. Thus, hate speech spreading afects users’ psychological state, promotes violence, and reinforces hateful sentiments [ 1, 2 ]. This problem attracts many scholars to apply modern technologies in order to make social Languages Shared Task (HASOC) 2021 [ 3 ] aims to сompare and analyze existing approaches to identifying hate speech not only for English, but also for other languages. It focused on detecting hate, ofensive, and profane content in tweets, and ofering six subtasks. We participated in three of them: nEvelop-O (M. Glazkov) • English Subtask A: identifying hate, ofensive, and profane content from the post in

English [ 4 ]. • English Subtask B: discrimination between hate, profane, and ofensive posts in English. • Marathi Subtask A: identifying hate, ofensive, and profane content from the post in

Marathi [ 5 ].

The source code for our models is freely available1.

The paper is organized as follows. Section 2 contains a brief review of related works. Next, we describe our experiments on the binary and fine-grained classification of English tweets in Section 3. In Section 4, we present our model for hate, ofensive, and profane language identification in Marathi. We conclude this paper in Section 5. Finally, Section 6 contains acknowledgments.

2. Related Works

We briefly discuss works done related to harmful content detection in the past few years. Shared tasks related to hate speech and ofensive language detection from tweets was organized as a part of some workshops and conferences, such as FIRE [ 6, 7 ], SemEval, [ 8, 9 ], GermEval [ 10, 11 ], IberLEF [ 12 ], and OSACT [ 13 ]. The participants proposed a broad range of approaches from traditional machine learning techniques (for example, Support Vector Machines [ 14, 15 ], Random Forest [ 16 ]) to various neural architectures (Convolutional Neural Networks, CNN [17]; Long Short Term Memory, LSTM [18, 19]; Embeddings from Language Models, ELMo [20]; and Bidirectional Encoder Representations from Transformers, BERT [21, 22]). In most cases, BERT-based systems outperformed other approaches.

Most research on hate speech detection continues to be based on English corpora. Despite this, the harmful content is distributed in diferent languages. Therefore, there have been previous attempts at creating corpora and developing models for hate speech detection in common non-English languages, such as Arabic [ 13, 23 ], German [ 6, 7, 10, 11 ], Italian [24, 25], Spanish [ 9, 12 ], Hindi [ 6, 7 ], Tamil and Malayalam [ 7 ]. Several studies have focused on collecting hate speech corpora for Chinese [26], Portuguese [27], Polish [28], Turkish [29] and Russian [30] languages.

3. English Subtasks A & B: Identification and Fine-grained

Classification of Hate, Ofensive, and Profane Tweets

The objective of English Subtasks A & B is to identify whether a tweet in English contains harmful content (Subtask A) and perform a fine-grained classification of posts into three categories, including: hate, ofensive, or profane (Subtask B). 3.1. Data The dataset provided to the participants of the shared task contains 4355 manually annotated social media posts divided into training (3074) and test (1281) sets. Table 1 presents the data description.

Further, we tested several data sampling techniques using diferent hate speech corpora as additional training data. Firstly, we evaluated the joint use of multilingual data provided by the organizers of HASOC 2021, including the English, the Hindi, and the Marathi training sets. Secondly, as the training sets were highly imbalanced, we applied the positive class random oversampling technique so that each training batch contained approximately the same number of samples. Besides, we experimented with the seq2seq-based data augmentation technique [31]. For this purpose, we fine-tuned the BART-base model for the denoising reconstruction task where 40% of tokens are masked and the goal of the decoder is to reconstruct the original sequence. Since the BART model [ 32 ] already contains the <mask> token, we use it to replace mask tokens. We generated one synthetic example for every tweet in the training set. Thus, the augmented data size is the same size as the size of the original training set. Finally, we evaluated the impact of additional training data, including: (a) the English dataset, used at HASOC 2020 [ 7 ]; (b) HatebaseTwitter, based on the hate speech lexicon from Hatebase2 [ 8 ]; (c) HatEval, a dataset presented at Semeval-2019 Task 5 [ 9 ] ; (d) Ofensive Language Identification Dataset (OLID), used in the SemEval-2019 Task 6 (OfensEval) [ 33 ]. All corpora except the HatebaseTwitter dataset contain non-intersective classes. Besides, all listed datasets are collected from Twitter. A representative sampling of additional data is shown in Table 2.

We preprocessed the datasets for Subtasks A & B in a similar manner. Inspired by [ 34 ], we used the following text preprocessing technique3: (a) removed all URLs; (b) replaced all user mentions with the $MENTION$ placeholder.

2https://hatebase.org/ 3https://pypi.org/project/tweet-preprocessor

3.2. Models

We conduct our experiments with neural models based on BERT [ 35 ] as they have achieved state-of-the-art results in harmful content detection. For example, BERT-based models proved eficient at previous HASOC shared tasks [ 7, 6 ] and SemEval [ 33, 36 ].

We used the following models: • BERT [ 35 ], a pre-trained model on BookCorpus [ 37 ] and English Wikipedia using a masked language modeling objective. • BERTweet [ 38 ], a pre-trained language model for English tweets. The corpus used to pre-train BERTweet consists of 850M English Tweets including 845M Tweets streamed from 01/2012 to 08/2019 and 5M Tweets related to the COVID-19 pandemic. • Twitter-RoBERTa for Hate Speech Detection [ 34 ], a RoBERTa [ 39 ] model trained on 58M tweets and fine-tuned for hate speech detection with the TweetEval benchmark. • LaBSE [ 40 ], a language-agnostic BERT sentence embedding model supporting 109 languages.

3.3. Experiments

For both Subtask A and Subtask B, we adopted pre-trained models from HuggingFace [ 41 ] and ifne-tuned them using PyTorch [ 42 ]. We fine-tuned each pre-trained language model for 3 epochs with the learning rate of 2e-5 using the AdamW optimizer [ 43 ]. We set batch size to 32 and maximum sequence size to 64. To validate our models during the development phase, we divided labelled data using the train and validation split in the ratio 80:20.

Table 3 shows the performance of our models on the validation subset for Subtask A in terms of macro-averaging F1-score (F1), precision (P), and recall (R). As can be seen from the table, BERT, BERTweet, and LaBSE show very close results during validation. Despite this, LaBSE jointly fine-tuned on three mixed multilingual datasets shows the highest precision score. The use of Twitter-RoBERTa increases the F1-score by 1.5-2.5% compared to other classification models. Based on this, we chose Twitter-RoBERTa for further experiments. We found out that neither the random oversampling technique nor the use of the augmented and additional data shows a performance improvement, except the joint use of the original dataset and the HatebaseTwitter dataset that gives an F1-score growth of 0.09% and a precision growth of 0.28% compared to basic Twitter-RoBERTa.

For our oficial submission for Subtask A, we designed a soft-voting ensemble of five TwitterRoBERTa jointly fine-tuned on the original training set and the HatebaseTwitter dataset (see Table 4). For Subtask B, we used the following one-vs-rest approach to discrimination between hate, profane, and ofensive posts.

• First, we applied our Subtask A binary models to identify non hate-ofensive examples. • Second, we fine-tuned three Twitter-RoBERTa binary models to delimit examples of hate-vs-profane, hate-vs-ofensive, and ofensive-vs-profane classes. The training dataset was extended with the HatebaseTwitter dataset. • Finally, we compared the results of binary models. If the result was defined uniquely, we used it as a predicted label. Otherwise, we chose the label in proportion to the number of examples in the training set.

This can be illustrated briefly by the following examples.

– Let the models show the following results: ∗ hate-vs-profane→hate; ∗ hate-vs-ofensive→hate; ∗ ofensive-vs-profane→ofensive.

Thus, classes have the following votes: hate – 2, ofensive - 1, profane – 0. Then we predict the HATE label. – If the results are: ∗ hate-vs-profane→profane; ∗ hate-vs-ofensive→hate; ∗ ofensive-vs-profane→ofensive, we have the class votes, such as hate – 1, ofensive - 1, profane – 1. Then we choose the PRFN label as the most common label in the training set.

4. Marathi Subtask A: Identifying Hate, Ofensive, and Profane

Content from the Post

4.1. Data For the Marathi task, we used the original training and test sets provided by the organizers of the HASOC 2021. The whole dataset contains 2499 tweets, including: 1874 training and 625 test examples. The training set consists of 1205 texts of the NOT class and 669 texts of the HOF class. We used raw data as an input for our models. Following [ 44, 45 ], we experimented with the combination of the English, the Hindi, and the Marathi training sets provided by the organizers.

4.2. Models

We evaluated the following models: • XLM-RoBERTa [ 46 ], a transformer-based multilingual masked language model supporting 100 languages. • LaBSE [ 40 ], a language-agnostic BERT sentence embedding model pre-trained on texts in 109 languages.

4.3. Experiments

We experimented with the above-mentioned language models fine-tuned on monolingual and multilingual data. For model evaluation during the development phase, we used the random train and validation split in the ratio 80:20 with a fixed seed. We set the same model parameters as for English tasks.

Table 5 illustrates the results. It can be seen that LaBSE outperforms XLM-RoBERTa in all cases. Moreover, the F1-score of LaBSE fine-tuned only on the Marathi dataset are higher than the results of LaBSE fine-tuned on multilingual data. XLM-RoBERTa, on the other hand, mostly benefits from multilingual fine-tuning.

For our final submission, we used a soft-voting ensemble of five LaBSE fine-tuned on the oficial Marathi dataset provided by the organizers of the competition. The results of this model on the test set are shown in Table 6.

5. Conclusion

In this paper, we have presented the details about our participation in the HASOC Shared Task 2021. We have explored an application of domain-specific monolingual and multilingual BERT-based models to the tasks of binary and fine-grained classification of Twitter posts. We also proposed a one-vs-rest approach to discrimination between hate, ofensive, and profane tweets. Further research can focus on analyzing the efectiveness of various text preprocessing techniques for harmful content detection and exploring how diferent transfer learning approaches can afect classification performance.

6. Acknowledgments

The work on multi-label text classification was carried out by Anna Glazkova and supported by the grant of the President of the Russian Federation no. MK-637.2020.9. hate speech and ofensive content identification in Indo-European languages, in: FIRE (Working Notes), 2020, pp. 168–174. [17] A. Ribeiro, N. Silva, Inf-hateval at semeval-2019 task 5: Convolutional neural networks for hate speech detection against women and immigrants on Twitter, in: Proceedings of the 13th International Workshop on Semantic Evaluation, 2019, pp. 420–425. [18] A. K. Mishraa, S. Saumyab, A. Kumara, IIIT_DWD@ HASOC 2020: Identifying ofensive content in Indo-European languages (2020). [19] A. Montejo-Ráez, S. M. Jiménez-Zafra, M. A. García-Cumbreras, M. C. Díaz-Galiano, SINAIDL at SemEval-2019 task 5: Recurrent networks and data augmentation by paraphrasing, in: Proceedings of the 13th International Workshop on Semantic Evaluation, 2019, pp. 480–483. [20] M. Bojkovsky, M. Pikuliak, STUFIIT at SemEval-2019 task 5: Multilingual hate speech detection on twitter with MUSE and ELMo embeddings, in: Proceedings of the 13th International Workshop on Semantic Evaluation, 2019, pp. 464–468. [21] J. Risch, A. Stoll, M. Ziegele, R. Krestel, hpiDEDIS at GermEval 2019: Ofensive language identification using a German BERT model., in: KONVENS, 2019. [22] P. Liu, W. Li, L. Zou, NULI at SemEval-2019 task 6: Transfer learning for ofensive language detection using bidirectional transformers, in: Proceedings of the 13th international workshop on semantic evaluation, 2019, pp. 87–91. [23] N. Albadi, M. Kurdi, S. Mishra, Are they our brothers? analysis and detection of religious hate speech in the Arabic twittersphere, in: 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), IEEE, 2018, pp. 69–76. [24] C. Bosco, D. Felice, F. Poletto, M. Sanguinetti, T. Maurizio, Overview of the EVALITA 2018 hate speech detection task, in: EVALITA 2018-Sixth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian, volume 2263, CEUR, 2018, pp. 1–9. [25] M. Sanguinetti, F. Poletto, C. Bosco, V. Patti, M. Stranisci, An Italian Twitter corpus of hate speech against immigrants, in: Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), 2018. [26] X. Tang, X. Shen, Y. Wang, Y. Yang, Categorizing ofensive language in social networks: A Chinese corpus, systems and an explanation tool, in: China National Conference on Chinese Computational Linguistics, Springer, 2020, pp. 300–315. [27] R. P. de Pelle, V. P. Moreira, Ofensive comments in the Brazilian web: a dataset and baseline results, in: Anais do VI Brazilian Workshop on Social Network Analysis and Mining, SBC, 2017. [28] M. Ptaszynski, A. Pieciukiewicz, P. Dybała, Results of the PolEval 2019 shared task 6: First dataset and open shared task for automatic cyberbullying detection in Polish Twitter (2019). [29] Ç. Çöltekin, A corpus of Turkish ofensive language on social media, in: Proceedings of the 12th Language Resources and Evaluation Conference, 2020, pp. 6174–6184. [30] L. Komalova, A. Glazkova, D. Morozov, R. Epifanov, L. Motovskikh, E. Mayorova, Automated classification of potentially insulting speech acts on social network sites, in: International Conference on Digital Transformation and Global Society, Springer, 2021. [31] V. Kumar, A. Choudhary, E. Cho, Data augmentation using pre-trained transformer models, in: Proceedings of the 2nd Workshop on Life-long Learning for Spoken Language Systems, M. Ott, L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual representation learning at scale, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 8440–8451.

References

[1] L. E. Beausoleil , Free, hateful, and posted: rethinking first amendment protection of hate speech in a social media world , BCL Rev . 60 ( 2019 ) 2101 .

[2] M. Bilewicz , W. Soral, Hate speech epidemic. the dynamic efects of derogatory language on intergroup relations and political radicalization , Political Psychology 41 ( 2020 ) 3 - 33 .

[3] S. Modha , T. Mandl , G. K. Shahi , H. Madhu , S. Satapara , T. Ranasinghe , M. Zampieri, Overview of the HASOC subtrack at FIRE 2021: Hate speech and ofensive content identiifcation in English and Indo-Aryan languages and conversational hate speech , in: FIRE 2021 : Forum for Information Retrieval Evaluation, Virtual Event , 13th -17th December 2021 , ACM, 2021 .

[4] T. Mandl , S. Modha , G. K. Shahi , H. Madhu , S. Satapara , P. Majumder , J. Schäfer , T. Ranasinghe , M. Zampieri , D. Nandini , A. K. Jaiswal , Overview of the HASOC subtrack at FIRE 2021 : Hate speech and ofensive content identification in English and Indo-Aryan languages , in: Working Notes of FIRE 2021 - Forum for Information Retrieval Evaluation , CEUR , 2021 . URL: http://ceur-ws.org/.

[5] S. Gaikwad , T. Ranasinghe , M. Zampieri , C. M. Homan , Cross-lingual ofensive language identification for low resource languages: The case of Marathi , in: Proceedings of RANLP , 2021 .

[6] T. Mandl , S. Modha , P. Majumder , D. Patel , M. Dave , C. Mandlia , A. Patel , Overview of the HASOC track at FIRE 2019 : Hate speech and ofensive content identification in Indo-European languages , in: Proceedings of the 11th forum for information retrieval evaluation , 2019 , pp. 14 - 17 .

[7] T. Mandl , S. Modha , A. Kumar M , B. R. Chakravarthi , Overview of the HASOC track at FIRE 2020 : Hate speech and ofensive language identification in Tamil, Malayalam, Hindi, English and German , in: Forum for Information Retrieval Evaluation , 2020 , pp. 29 - 32 .

[8] T. Davidson , D. Warmsley , M. Macy , I. Weber , Automated hate speech detection and the problem of ofensive language , in: Proceedings of the International AAAI Conference on Web and Social Media , volume 11 , 2017 .

[9] V. Basile , C. Bosco , E. Fersini , N. Debora , V. Patti , F. M. R. Pardo , P. Rosso , M. Sanguinetti , et al., Semeval-2019 task 5: Multilingual detection of hate speech against immigrants and women in Twitter , in: 13th International Workshop on Semantic Evaluation , 2019 , pp. 54 - 63 .

[10] M. Wiegand , M. Siegel , J. Ruppenhofer , Overview of the germeval 2018 shared task on the identification of ofensive language , in: 14th Conference on Natural Language Processing KONVENS 2018 , 2018 .

[11] J. M. Struß , M. Siegel , J. Ruppenhofer , M. Wiegand , M. Klenner , et al., Overview of GermEval task 2 , 2019 shared task on the identification of ofensive language ( 2019 ).

[12] M. Taulé , A. Ariza , M. Nofre , E. Amigó, P. Rosso , Overview of detoxis at IberLEF 2021: Detection of toxicity in comments in Spanish , Procesamiento del Lenguaje Natural 67 ( 2021 ) 209 - 221 .

[13] H. Mubarak , K. Darwish , W. Magdy , T. Elsayed , H. Al-Khalifa , Overview of OSACT4 Arabic ofensive language detection shared task , in: Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Ofensive Language Detection , 2020 , pp. 48 - 52 .

[14] F. Schmid , J. Thielemann , A. Mantwill , J. Xi , D. Labudde , M. Spranger , Fosil-ofensive language classification of German tweets combining SVMs and deep learning techniques , in: KONVENS , 2019 .

[15] S. Hassan , Y. Samih , H. Mubarak , A. Abdelali , A. Rashed , S. A. Chowdhury , Alt submission for osact shared task on ofensive language detection , in: Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Ofensive Language Detection , 2020 , pp. 61 - 65 .

[16] B. Ray , A. Garain , JU at HASOC 2020: Deep learning with RoBERTa and random forest for 2020 , pp. 18 - 26 .

[32] M. Lewis , Y. Liu , N. Goyal , M. Ghazvininejad , A. Mohamed , O. Levy , V. Stoyanov , L. Zettlemoyer, Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension , in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , 2020 , pp. 7871 - 7880 .

[33] M. Zampieri , S. Malmasi , P. Nakov , S. Rosenthal , N. Farra , R. Kumar , Semeval -2019 task 6: Identifying and categorizing ofensive language in social media (OfensEval) , in: Proceedings of the 13th International Workshop on Semantic Evaluation , 2019 , pp. 75 - 86 .

[34] F. Barbieri , J. Camacho-Collados , L. E. Anke , L. Neves, TweetEval: Unified benchmark and comparative evaluation for tweet classification , in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings , 2020 , pp. 1644 - 1650 .

[35] J. Devlin , M.- W. Chang , K. Lee , K. Toutanova , Bert: Pre-training of deep bidirectional transformers for language understanding , arXiv preprint arXiv: 1810 . 04805 ( 2018 ).

[36] J. Pavlopoulos , J. Sorensen , L. Laugier , I. Androutsopoulos , SemEval -2021 task 5: Toxic spans detection , in: Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021) , 2021 , pp. 59 - 69 .

[37] Y. Zhu , R. Kiros , R. Zemel , R. Salakhutdinov , R. Urtasun , A. Torralba , S. Fidler , Aligning books and movies: Towards story-like visual explanations by watching movies and reading books , in: Proceedings of the IEEE international conference on computer vision , 2015 , pp. 19 - 27 .

[38] D. Q. Nguyen , T. Vu , A. T. Nguyen, BERTweet: A pre-trained language model for English tweets , in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , 2020 , pp. 9 - 14 .

[39] Y. Liu , M. Ott , N. Goyal , J. Du , M. Joshi , D. Chen , O. Levy , M. Lewis , L. Zettlemoyer , V. Stoyanov , Roberta: A robustly optimized bert pretraining approach , arXiv preprint arXiv: 1907 . 11692 ( 2019 ).

[40] F. Feng , Y. Yang , D. Cer , N. Arivazhagan , W. Wang , Language-agnostic BERT sentence embedding , arXiv preprint arXiv: 2007 . 01852 ( 2020 ).

[41] T. Wolf , J. Chaumond , L. Debut , V. Sanh , C. Delangue , A. Moi , P. Cistac , M. Funtowicz , J. Davison , S. Shleifer , et al., Transformers: State-of-the-art natural language processing , in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , 2020 , pp. 38 - 45 .

[42] A. Paszke , S. Gross , F. Massa , A. Lerer , J. Bradbury , G. Chanan, T. Killeen , Z. Lin , N. Gimelshein , L. Antiga , et al., Pytorch: An imperative style, high-performance deep learning library , Advances in neural information processing systems 32 ( 2019 ) 8026 - 8037 .

[43] I. Loshchilov , F. Hutter , Decoupled weight decay regularization , in: International Conference on Learning Representations , 2018 .

[44] S. Mishra , S. Prasad , S. Mishra , Multilingual joint fine-tuning of transformer models for identifying trolling , aggression and cyberbullying at TRAC 2020 , in: Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying , 2020 , pp. 120 - 125 .

[45] P. Singh , P. Bhattacharyya , CFILT IIT Bombay at HASOC 2020: Joint multitask learning of multilingual hate speech and ofensive content detection system ., in: FIRE (Working Notes) , 2020 , pp. 325 - 330 .

[46] A. Conneau , K. Khandelwal , N. Goyal , V. Chaudhary , G. Wenzek , F. Guzmán , É. Grave,