Prompt-Driven Cross-Lingual Dialogue Generation for Low-Resource Indian Languages: A Case Study on Bengali and Hindi

The proposed work introduces an improved prompt-driven system for dialog generation in Bengali and Hindi, two low-resource Indian languages. To achieve semantic consistency and cultural integrity, the system is equipped with an Adaptive Prompt Controller (APC) that dynamically selects between role-based, few-shot, and zero-shot instructional prompts based on dialogue context. The research employs both automatic scores and human-oriented measures to evaluate two leading models, Claude Sonnet-4 and GPT-4-Turbo, in addition to four multilingual large language models (lLMs). The results show that GPT-4-Turbo and Claude Sonnet-4 outperform earlier models, achieving a higher semantic alignment and a conversation success rate of as high as 91.2%. The architecture establishes a scalable, resource-light, and culturally appropriate model for cross-lingual conversation systems in Indic languages and is fully replicable.