A Novel XLMR-DualCL Approach for Hate Speech Detection in Hindi-English Code-Mixed Data on Social Media

Abstract

Automated identification of hate speech in Hindi-English code-mixed content remains a significant open challenge in multilingual natural language processing, largely because users on Indian social media platforms routinely blend two languages within a single utterance, employ phonetic transliteration of Hindi in Roman script, and convey hostility through indirect or culturally situated expressions that resist straightforward lexical analysis. This study introduces XLMR-DualCL, a unified detection pipeline specifically designed to address these difficulties. The architecture integrates noise-aware text preprocessing, diversity-oriented data augmentation, deep multilingual encoding via XLM-RoBERTa, a dual contrastive learning objective that simultaneously exploits self-supervised and supervised signals, and focal loss to compensate for the class imbalance endemic to real-world hate speech corpora. Comprehensive evaluation is conducted across five publicly available benchmarks spanning HASOC, TRAC-1, ICHCL, THAR, and Kaggle Hinglish, supplemented by a self-curated corpus of Instagram and YouTube comments and a merged multi-source dataset. Across all evaluation settings the model delivers consistent and competitive results, recording F1-scores of 88.89, 89.83, 91.17, 88.49, 90.29, 91.62, and 87.92 on the seven corpora respectively, together with a peak classification accuracy of 91.67 on the self-collected data. Precision and recall values remain closely aligned throughout, confirming that the system avoids trading one error type against the other. Taken together, these outcomes demonstrate that coupling cross-lingual contextual encoding with dual contrastive representation learning substantially advances the state of hate speech detection in noisy, informal Hinglish social media environments.