Domain-Adaptive Multitask BERT with Graph Context Modeling for Code-Mixed Hinglish Sentiment Classification

Hinglish, a widely used Hindi–English code-mixed language on social media, presents unique challenges for sentiment analysis due to transliteration variability, script mixing, and inconsistent grammar. To address these issues, this study proposes the Cross-Lingual Domain-Adaptive Multitask Graph-Enhanced Bidirectional Encoder Representations from Transformers (CDMG-BERT), tailored for code-mixed sentiment classification. The model integrates four key components: English-to-Hinglish embedding alignment, domain-adversarial training using 92k English samples, multitask learning with auxiliary POS tagging, and graph-based token relational modeling for long-range contextual refinement. Experimental results on a 12,000-sample IIT Bombay Hinglish dataset show that CDMG-BERT outperforms mBERT, XLM-R, and deep learning baselines, with an F1-score of 84.8%. Ablation analysis suggests that each architectural module works well, with cross-lingual alignment and domain adaptation providing the largest gains. These results demonstrate that the model is robust enough to handle spelling differences, code-mixing intensity, and inconsistencies between the Roman and Devanagari scripts. This makes CDMG-BERT a good choice for sentiment analysis in a multilingual environment with limited resources.