Team UKNLP at TREC 2017 Precision Medicine Track: A Knowledge-Based IR System with Tuned Query-Time Boosting

This paper describes the system architecture of the University of Kentucky Natural Language Processing (UKNLP) team’s entry for the TREC 2017 Precision Medicine Track. The goal of the challenge is to retrieve useful precision medicine-related information (abstracts, clinical trials) for the given synthetic cancer patient cases, each of which consists of a neoplastic condition, genetic variants, demographic details, and any additional information (e.g., comorbidities). We explored query expansion techniques using well-known broad knowledge sources such as the Unified Medical Language System (UMLS) and the Medical Subject Headings (MeSH) for each abstract, and additional specialized sources such as the Catalogue Of Somatic Mutations In Cancer (COSMIC) database, which allowed us to construct boosted queries. We conducted several experiments with model averaging techniques and our final system architecture placed 6th (in terms of infNDCG and R-prec) among 29 teams that submitted runs to the scientific abstract retrieval task.