Bilingual Mental Health Chatbot Using DeBERTa-Based Intent Classification and Neural Translation

The creation of smart and culturally diverse mental health support systems is a burning issue, particularly in the case of low-resource language communities. This paper has proposed a bilingual conversational agent that used the deep learning model and helped two different users with English and Telugu. The drive is to propel the provision of a scalable language intelligent mental health support by means of natural dialogue systems. Our solution consists of combining a transformer-based architecture based on DeBertas, which processes intent recognition, but also layers of Bidirectional Long Short-Term Memory or BiLSTM to recognize it. Likewise, we combine MarianMT-based neural machine translation models to communicate in both directions between a hybrid and web-based translation between English and Telugu. A validation accuracy of 84.58 percent was accrued, and the chatbot also improved greatly after 50 epochs with an improved intent dataset with emotional and support-seeking patterns. The standard metrics (BLEU, CHRF, TER and ROUGE-L) were used to evaluate translation systems, where the English->Telugu model performed with a BLEU of 58.65, CHRF 80.98, TER 27.64, ROUGE-L 0.09 and Telugu->English model with BLEU 51.45, CHRF 71.13, TER 36.05 and ROUGE-L 0.75. Lastly, a real-time GUI containing chatbot and translation engines was connected to a quality check system based on the BLEU quality assessment, allowing to make mental health conversations across languages in their natural form, with a high level of faithfulness.