Offensive language detection in Hebrew: can other languages help?

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3715–3723 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3715 Offensive language detection in Hebrew: can other languages help? Natalia Vanetik1, Marina Litvak1, Chaya Liebeskind2, Omar Hmdia1, Rizek Abu Madeghem1 1Shamoon College of Engineering, Beer-Sheva, Israel, 2Jerusalem College of Technology, Jerusalem, Israel {natalyav, marinal}@ac.sce.ac.il, liebchaya@gmail.com, {omarhm, rezeqab}@ac.sce.ac.il Abstract Unfortunately, offensive language in social media is a common phenomenon nowadays. It harms many people and vulnerable groups. Therefore, automated detection of offensive language is in high demand and it is a serious challenge in multilingual domains. Various machine learning approaches combined with natural language techniques have been applied for this task lately. This paper contributes to this area in several aspects: (1) it introduces a new dataset of annotated Facebook comments in Hebrew; (2) it describes a case study with multiple supervised models and text representations for a task of offensive language detection in three languages, including two Semitic (Hebrew and Arabic) languages; (3) it reports evaluation results of cross-lingual and multilingual learning for detection of offensive content in Semitic languages; and (4) it discusses the limitations of these settings. Keywords: offensive language, Semitic languages, deep learning, BERT, cross-lingual learning, multilingual learn- ing 1. Introduction Multiple works on automated offensive language detec- tion show that contamination of social networks with offensive content is a new reality with serious outcomes affecting almost all of us. Moreover, it is an interna- tional phenomenon demanding multilingual solutions. Early works on offensive language detection used un- supervised lexicon-based approaches (Tulkens et al., 2016), while later more supervised approaches—with logistic regression, naive Bayes, decision trees, ran- dom forests, and support vector machines—were pro- posed (Davidson et al., 2017). Most of the re- cent papers report on the application of deep neural networks—long short-term memory networks, recur- rent neural networks, convolutional neural networks, gated recurrent units, transformers, and deep lan- guage models—frequently combined with word em- beddings, for separating offensive language from legit- imate texts (Zampieri et al., 2019c). In the last couple of years, transformer models like ELMo (Embeddings from Language Models) (Peters et al., 2018) and BERT (Bidirectional Encoder Representations from Trans- formers) (Devlin et al., 2018a; Reimers and Gurevych, 2019) have been most popular and successful for offen- sive language identification (Liu et al., 2019; Ranas- inghe et al., 2019). The clear majority of the offensive detection studies deal with English, partially because most available annotated datasets contain English data. For exam- ple, SemEval-2019 Task 6: Identifying and Categoriz- ing Offensive Language in Social Media (OffensEval) was based on the Offensive Language Identification Dataset (OLID), which contains over 14,000 English tweets. The main findings of this task can be found in (Zampieri et al., 2019c). Since social media became the most popular commu- nication tool worldwide, people from different coun- tries generate their content in various languages. The attention of international communities to this task emphasizes its “multilingual challenge” – many re- searchers contributed to this …

Content truncated to 500 words