Domain-Aware Healthcare Chatbot Incorporating BERT and RAG

Abstract

Existing healthcare chatbots suffer from diagnostic inaccuracy, poor context awareness, and reliance on static knowledge, leading to unsafe generic advice. We introduce a novel chatbot addressing these limitations. Our key contributions are: (1) A hybrid architecture combining BERT for query intent classification with a fine-tuned LLM for response generation; (2) Inclusion of Retrieval-Augmented Generation to dynamically access authoritative medical knowledge; and (3) A Gradio interface enabling user interaction. This approach significantly enhances response accuracy (alignment with medical facts), contextual relevance, and reliability (reducing model hallucinations). Evaluation shows a marked improvement over baselines: BLEU-4 increased from 14.8 to 27.5, and substantial gains in ROUGE-1 (19.7 to 25.6), ROUGE-2 (2.9 to 5.0), and ROUGE-L (9.3 to 16.1), demonstrating superior overall performance.