Coreference Resolution for Structured Drug Product Labels

Abstract

FDA drug package inserts provide comprehensive and authoritative information about drugs. DailyMed database is a repository of structured product labels extracted from these package inserts. Most salient information about drugs remains in free text portions of these labels. Extracting information from these portions can improve the safety and quality of drug prescription. In this paper, we present a study that focuses on resolution of coreferential information from drug labels contained in DailyMed. We generalized and expanded an existing rule-based coreference resolution module for this purpose. Enhancements include resolution of set/instance anaphora, recognition of appositive constructions and wider use of UMLS semantic knowledge. We obtained an improvement of 40% over the baseline with unweighted average F1-measure using B-CUBED, MUC, and CEAF metrics. The results underscore the importance of set/instance anaphora and appositive constructions in this type of text and point out the shortcomings in coreference annotation in the dataset.