Abstract:
Detecting and eliminating offensive and hate speech in social media content is an important concern as hate
and offensive speech can have serious consequences in society ranging from ill-education among youth to hate
crimes. Offensive speech identification in countries like India poses several additional challenges due to the
usage of code-mixed and romanized variants of multiple languages by the users in their posts on social media.
HASOC-Dravidian-CodeMix - FIRE 2020 extended the task of offensive speech identification to Dravidian
languages. In this paper, we describe our approach in HASOC Dravidian Code-mixed 2020, which topped two
out of three tasks(F1-weighted scores - 0.95 and 0.90) and stood second in the third task lagging the top model
only by 0.01 points((F1-weighted score - 0.77). We propose a novel and flexible approach of selective translation
and transliteration to be able to reap better results out of fine-tuning and ensembling multilingual transformer
networks like XLM-RoBERTa and mBERT. Further, we implemented pre-trained, fine-tuned and ensembled
versions of XLM-RoBERTa for offensive speech classification. We open source our work to facilitate further
experimentation.