Heterogeneous Knowledge Grounding for Medical Question Answering with Retrieval Augmented Large Language Model

The Large Language Model (LLM) is renowned for its ability to encode a vast amount of general domain knowledge, enabling it to excel in question-answering, dialogue systems, and summarization tasks. However, the medical domain presents a unique challenge to LLM due to the distribution of medical knowledge, which follows a long-tail pattern. Existing approaches address this challenge by injecting medical knowledge into LLM through single sources such as medical textbooks or medical knowledge bases. However, medical knowledge is distributed across multiple heterogeneous information sources. A medical question-answering system can enhance answer coverage and confidence by considering these diverse knowledge sources together. To bridge this gap, we propose a novel approach called Heterogeneous Knowledge Retrieval-Augmented LLM for medical domain question answering. Our experiments, conducted on the MedQA-USMLE dataset, demonstrate promising performance improvements. These results underscore the importance of harnessing heterogeneous knowledge sources in the medical domain.