Leveraging Large Language Models and Knowledge Graphs for Advanced Biomedical Question Answering Systems

This study introduces a Knowledge Base Question Answering (KBQA) system augmented with Large Language Models (LLMs) to address the complexities of biomedical data exploration and knowledge extraction. The proposed system leverages LLMs to translate natural language questions into structured graph queries, tailored for Knowledge Graph (KG) databases. This enables medical professionals to efficiently explore stored data with no need to master any graph query language syntax. The responses retrieved are then processed through the LLMs, which rephrase and refine this information to generate human-like answers. We evaluated our system using the BioASQ benchmark dataset, employing different knowledge sources, namely, PrimeKG and Hetionet. For the LLMs, LLAMA2-70B and GPT-4 were used. Cypher, adapted for Neo4j databases, served as the graph query language for our experiments. Initial results show that our system obtains good performance in the case of List-type questions. The presented comparative experiments indicate that the success of our proposed approach depends significantly on the quality of both the KGs and the LLMs used for generating queries. This underscores the necessity for more comprehensive KGs covering a wider range of biomedical knowledge.