Entity-level classification of adverse drug reactions: a comparison of neural network models

This paper presents our experimental work on neural network models for entity-level adverse drug reaction (ADR) classification. Aspect-level sentiment classification, which aims to determine the sentimental class of a specific aspect conveyed in user opinions, have been actively studied for more than 10 years. In the past few years, several neural network models have been proposed to address this problem. While these models have a lot in common, there are some architecture components that distinguish them from each other. We investigate the applicability of neural network models for ADR classification. We conduct extensive experiments on various pharmacovigilance text sources including biomedical literature, clinical narratives, and social media and compare the performance of five state-of-the-art models as well as a feature-rich SVM in terms of the accuracy of ADR classification.