Abstract:
Designing effective automated techniques for proper identification and categorisation of hostile speech is essential, especially for low resource Indian languages without established datasets. A majority of Indians use Hindi for their interactions on social media. Multiple dialects of Hindi spoken by the users in different parts of the country further exacerbate the challenge of identifying hostile speech as they imply diverse patterns of expression of hostility. In this work, we experimented with a wide range of neural network models for hostility identification in Hindi - pre-trained word embeddings, stacked word embeddings, and fine-tuned XLM-RoBERTa model. We also analyzed the effectiveness of back translation as a data augmentation technique to assist in fine-tuning of XLM-RoBERTa model for hostility identification. The experiments are carried out on the dataset provided by Constraint 2021 shared task. Our team’s (Siva_Alfred on leader board) best neural network model achieves F1-weighted score of 0.90 for coarse-grained hostility detection and 0.54 F1-weighted score for fine-grained hostility identification.