Infusing Structured Medical Knowledge into Language Models: Graph Neural Prompting for Healthcare Question Answering

Current natural language processing models have made significant strides in understanding medical texts. Recent research leverages non-parametric external sources of relevant factual information. It taps into latent world knowledge to enhance performance and interpretability in the medical domain. However, a lack of in-depth understanding of structured medical knowledge limits its practical application in complex medical contexts. In this thesis, we present MedGNP, the first method for infusing structured medical knowledge into language models, a plug-and-play new approach applying a prompting framework by grounding knowledge to enhance models’ capabilities in addressing question-answering tasks within the medical domain. MedGNP incorporates various designs, including a graph attention network, cross-modal attention module, domain projector, and a bilinear model for self-supervised link prediction. We evaluated our method on three popular medical question-answering datasets: PubMedQA, BioASQ and MedQA. After knowledge grounding using MedGNP method, we get improvement across all datasets from 1.26% to 6.0% by leveraging 30M extra parameters, which takes only 12% of the original model. Furthermore, MedGNP allows a generalist model to achieve comparable results with state-of-the-art models that underwent a domain-specific pre-training extension while using up to 12x fewer trainable parameters, proving the effectiveness of this approach to successfully knowledge grounding language model. Finally, we discuss potential future work on this topic.