Comparison of Pretrained Embeddings to Identify Hate Speech in Indian Code-Mixed Text
Shubhanker Banerjee, Bharathi Raja Chakravarthi, John P. McCrae
Abstract
Two or more languages used in the same sentence is known as the code-mixed text. The phenomenon is abundant in social media due to multilingualism. It poses a considerable challenge for classic NLP tools trained on monolingual corpora. Automatic hate speech detection in code-mixed text becomes even more challenging due to non-standard variations in the spelling, grammar and writing in foreign scripts. Pre-trained models provide word embedding trained on massive monolingual corpora, which are now ubiquitous forms of word representation to classifying text. In this paper, we compare pretrained models and create an ensemble model for code-mixed data of hate speech classification task on Hindi-English data. We have also experimented with using word embedding for CNN networks and showed that XLNet performs better for hate speech detection in code-mixed text.