Multilingual Offensive Language Identification for Low-resource Languages

Offensive content is pervasive in social media and a reason for concern to companies government organizations. Several studies have been recently published investigating methods detect the various forms of such (e.g., hate speech, cyberbullying, cyberaggression). The clear majority these deal with English partially because most annotated datasets available contain data. In this article, we take advantage by applying cross-lingual contextual word embeddings transfer learning make predictions low-resource languages. We project on comparable data Arabic, Bengali, Danish, Greek, Hindi, Spanish, Turkish. report results 0.8415 F1 macro Bengali TRAC-2 shared task [23], 0.8532 Danish 0.8701 Greek OffensEval 2020 [58], 0.8568 Hindi HASOC 2019 [27], 0.7513 Spanish SemEval-2019 Task 5 (HatEval) [7], showing that our approach compares favorably best systems submitted recent tasks three Additionally, competitive performance Arabic Turkish using training development sets task. all languages confirm robustness