Query Refinement: Negation Detection and Proximity Learning Georgetown at TREC 2014 Clinical Decision Support Track.

摘要

In this paper we describe our efforts for TREC Clinical Decision Support Track 2014. Our system takes medical case narratives as input and returns relevant biomedical articles to answer clinical questions to determine the patient's diagnosis, the tests the patient should receive, and how the patient should be treated. We model each topic as highly representative keyword-based structured queries. Since both the topics and returned documents are written in highly technical language, we address the traditional vocabulary gap present within medical information retrieval, while also focusing on employing methodologies to refine our queries by detecting negation and applying query proximity learning. We hypothesized terms with high frequency among the topics, which are likely to create noise and impair the return of highly relevant documents. Our top two runs utilizing negation detection perform above the median for P@10, R-Prec, and infAP, and our other three runs that utilized proximity learning performed approximately consistent with the median. More research is required to explore the potential benefits of proximity learning over a more robust set of topics.