GAR: Graph adversarial representation for adverse drug event detection on Twitter

Abstract

Adverse drug reaction events have become one of the main causes of patient death. Since traditional post-marketing surveillance systems based on spontaneous reports have a serious underreporting issue, in recent years research on the detection of adverse reaction events using social media such as Twitter as a data source has attracted increasing attention in recent year. Deep learning models usually rely on a large number of training samples. However, due to the characteristics of user-generated content and the time-consuming data annotation process, related research is faced with the problems caused by small-scale annotated datasets, which restricts deep learning models in achieving satisfactory results. Accordingly, we introduce two regularization methods are introduced at the representation level, i.e., graph embedding-based data augmentation and adversarial training, to improve the performance of detecting adverse events under such conditions. Besides, the applicable scope of these two methods is analyzed and discussed through experiments. Combined with the convolutional neural network, this paper proposes an adverse drug event detection framework that can make full use of the methods.