Consumer Health Question Answering Using Off-the-Shelf Components

Abstract

In this paper, we address the task of open-domain health question answering (QA). The quality of existing QA systems heavily depends on the annotated data that is often difficult to obtain, especially in the medical domain. To tackle this issue, we opt for PubMed and Wikipedia as trustworthy document collections to retrieve evidence. The questions and retrieved passages are passed to off-the-shelf question answering models, whose predictions are then aggregated into a final score. Thus, our proposed approach is highly data-efficient. Evaluation on 113 health-related yes/no question and answer pairs demonstrates good performance achieving AUC of 0.82.