Consumer Medical Question Answering: Challenges and Approaches

Abstract

Consumer Medical Question Answering (CQA) requires continual community support and innovation in the fields of information retrieval and natural language processing. In this thesis, we highlight research directions and investigate solutions to improve consumer medical question answering: (1) Given the limited resources, medical question answering systems typically rely on pre-trained open-domain models as part of the question answering pipeline. These pre-trained models are not designed for use in the biomedical domain, resulting in worse representations of biomedical language. We propose two strategies for improving biomedical representation learning for consumer question answering by modifying the models' architecture and vocabulary, (2) consumers often have trouble formulating their medical questions for their information needs correctly and therefore, find it difficult to locate trustworthy peer-reviewed biomedical information regarding their medical concerns. We propose a method that combines statistical keyword modelling alongside deep learning techniques to allow consumers to search for reputable information that an expert would find given the same information need, and (3) medical question answering datasets often have simplifying assumptions about the consumer's question to make the task more feasible. These datasets often assume a shorter question from the consumer, free of grammatical errors and lexical gaps. We curate a dataset that uses more realistic consumer questions paired with a verified medical expert's response. Given the difficulty of using consumer questions, we conduct a feasibility study using three QA pipelines. We conclude the thesis with answers to our research questions and highlight future directions for medical question answering for consumers.