Offensive Language Detection in Social Media Using Transformers and Importance of Pre-training

Abstract

 Abstract —Being exposed to offensive language on social media platforms is relatively higher because of anonymity and distant self-expression compared to real communication. Billions of contents are shared daily on these platforms, making it impossible to detect offensive posts with manual editorial processes. This situation arises the need for automatic detection of offensive language in social media posts to provide users' online safety. In this paper, we applied different Machine Learning (ML) models on over manually annotated 36,000 Turkish tweets to detect the use of offensive language messages automatically. According to the results, the most successful model for predicting offensive language is pre-trained transformer-based ELECTRA model with 0.8216 F-1 score. We also obtained the highest F-1 score with 0.8342 in this dataset up to now by combining transformer-based ELECTRA and BERT models in an ensemble model.