Enhanced Biomedical Factoid Question Answering through Biomedical Knowledge Integration

Factoid question answering requires short, factual responses. In the biomedical domain, this involves extracting specific answers from articles. Despite recent progress using pre-trained language models (LMs) for passage retrieval and text comprehension, challenges persist due to limited biomedical datasets, affecting LM accuracy. This thesis introduces the Biomedical Knowledge-enhanced Question Answering Framework (BK-QAF), incorporating knowledge from the Unified Medical Language System (UMLS) to enhance comprehension and reasoning. The framework uses a graph attention network (GAT) to prioritize UMLS concepts by relevance, facilitating precise answers.

BK-QAF’s effectiveness was assessed using three distinct GAT architectures, demonstrating robustness across different configurations. Empirical evaluations with the BioASQ datasets show that BK-QAF significantly outperforms state-of-the-art baselines in Strict Accuracy (SAcc) and Mean Reciprocal Rank (MRR). This thesis highlights BK-QAF’s potential to improve performance in biomedical question answering.