Hybrid Memory-Retrieval Model: Enhancing Trust in Medical Chatbots

ACL ARR 2025 May Submission4290 Authors

Abstract: Medical chatbots powered by large language models (LLMs) face two critical challenges: hallucination, where the model produces plausible but incorrect responses, and loss of context in multi-turn conversations. These issues undermine reliability and trust in healthcare settings. This paper introduces a hybrid memory-retrieval architecture designed to enhance factual grounding and conversational coherence. The system integrates a dual-retriever pipeline (BM25 and MedCPT) with long-term memory retrieval using ChromaDB. Retrieved documents and past interactions are fused via Reciprocal Rank Fusion and provided as input to a compact language model (Phi-2) for response generation. A fallback mechanism is employed when insufficient context is available to reduce hallucinated responses. Evaluation on the MedQuAD dataset demonstrates high semantic alignment (BERTScore F1 = 0.8644), improved fluency, and significantly faster response times compared to baseline retrieval-augmented models. These results support the effectiveness of combining structured memory with selective retrieval to develop more trustworthy medical dialogue systems.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Retrieval-Augmented Generation, Hallucination Mitigation, Memory Retrieval, Context Retention, Medical Chatbot

Languages Studied: english

Submission Number: 4290