Adversarial transfer network with bilinear attention for the detection of adverse drug reactions from social media

Abstract

The drug safety issues related to adverse drug reactions (ADRs) are gradually becoming more important to the public. With social media booming, increasingly more patients would like to share their reactions to get support from others. These published posts comprise a valuable resource for ADR identification because of their timeliness. However, available social media datasets are rare. Moreover, the informality of the social media text is also a challenge for ADR identification. PubMed and social media differ greatly in expression and syntax. Introducing the PubMed datasets, which are usually normative and numerous, may be helpful to ADR identification in social media. To this end, we propose an adversarial transfer framework for ADR identification that transfers the auxiliary features from PubMed to social media datasets to improve the generalization of the model and mitigate the noise caused by colloquial expression in social media. Additionally, we add dynamic weight to the loss function to offset the training slants caused by imbalanced training data. We experimentally evaluate the method we proposed on two social media datasets and two PubMed datasets. The results show that our proposed method can improve the performance of ADR identification from social media.