Contextualized Relevance Feedback for Precision Medicine

摘要

Precision Medicine (PM) is viewed as an information retrieval (IR) task, in which biomedical articles containing treatment information about specific diseases or genetic variants are retrieved in response to patient record, aiming at providing medical evidence to the point-of-care. Previous PM approaches are mostly based on unigram matching of individual query terms, or concepts, to the target articles to produce the ranking list, while ignoring the context of the matched query terms of concepts. To this end, this paper presents a preliminary investigation of utilizing contextualized representation of text for pseudo relevance feedback (PRF) to enhance PM search effectiveness. By considering the multi-aspect word relations, we propose a BERTNPRF model to integrate PRF with the fine-tuned BERT model for contextualized interaction of document-document pairs. Experimental results on the standard Text REtrieval Conference (TREC) PM track benchmark show that our proposed method with interpolation can improve the performance in PM.