An Active Learning Approach for Identifying Adverse Drug Reaction-Related Text from Social Media Using Various Document Representations

Adverse drug reaction (ADR) is a major health concern. Identifying text that mentions ADRs from a large volume of social media data discussing other topics is a key preliminary but nontrivial task for drug-ADR pair detection. This task suffers from severe imbalance issue. Moreover, prior studies have overlooked the simultaneous use of high-level abstract information contained in data and the domain-specific information embedded in knowledge bases. Therefore, we propose a novel multi-view active learning approach, in which a selection strategy is tailored to the imbalanced dataset and various document representations are regarded as multi views. We capture data-driven and domain-specific information by resorting to deep learning methods and handcrafted feature engineering, respectively. Experimental results demonstrate the effectiveness of our proposed approach.