A Novel Knowledge Enhanced Large Language Model Augmented Framework for Medical Question Answering

Leveraging domain-specific knowledge from pre-trained large language models and knowledge graphs for reasoning in the medical question answering task has emerged as a prominent research field. However, the accuracy of the inference results is restricted by multiple factors, including the quality of analyzed topic entities, the selected inference path in the knowledge graph, and the absence of mutual updating for embedding representations from large language models and knowledge graphs. In this paper, we propose a novel medical question answering framework based on the domain-specific large language model, aiming to enhance the quality of topic entities by implementing the retrieval augmentation technique. Inspired by the concept of chain-of-thought reasoning, we introduce a joint reasoning approach based on analyzed topic entities to facilitate the generation of accurate inference paths. Additionally, we design a unified embedding mechanism that combines representations from both the large language model and graph neural networks, incorporating a pooling operation for predicting answers to input questions. To the best of our knowledge, this work signifies the pioneering efforts in implementing the retrieval augmentation technique and the joint reasoning approach within the context of the medical question answering task. Experimental results on public benchmark datasets demonstrate that the introduced method outperforms state-of-the-art baseline approaches.