Abstract:
The swift escalation in hostile content on the web and specifically on Online Social Media (OSM) has lately become a matter of concern that we must tackle. The situation worsens to a whole different level with recent events such as COVID-19 pandemic, BLM, and #MeToo movements. Even though the existing systems address the problem of hostile post detection in one or more dimensions, e.g., hate, fake, etc., there has not been sufficient studies that address multiple hostile dimensions in a unified system. Moreover, a significant majority of the existing systems devour the English language, and research in regional languages (e.g., Hindi, Bengali, etc.) do not get adequate attention. To this end, in this paper, we tackle the hostile post detection in Hindi for four dimensions – fake, hate, offensive, and defamation. We propose HostileNet, a novel deep learning framework that leverages the HindiBERT-based contextual representations and hand-crafted lexicon features for the hostile post classification. Moreover, we also propose a novel mechanism to further fine-tune the attention vectors w.r.t. each hostile dimension. We evaluate HostileNet on the CONSTRAINT-2021’s multi-label Hindi shared task dataset in both coarse-grained (hostile vs. non-hostile) and fine-grained (fake vs. hate vs. offensive vs. defamation) setups. Our evaluation shows that HostileNet outperforms various existing systems including the best performing system as reported in the CONSTRAINT-2021 shared task for both the setups. Furthermore, we provide a thorough analyses of the obtained results in forms of ablation study, error analysis, attention heapmap analysis, lexicon feature analysis, etc. We make the code and the curated multi-label hostile lexicon available for research use at https://github.com/LCS2-IIITD/HostileNet. git.
