Exploiting adversarial transfer learning for adverse drug reaction detection from texts.

Abstract

Adverse Drug Reactions (ADRs) are extremely hazardous to patients. ADR Detection aims to automatically determine whether a sentence is related to an ADR, which is a fundamental study for public health monitoring tasks, particularly for pharmacovigilance. Benchmark corpora are mostly sampled from biomedical literature or social media, but most of them are on small scales. Correspondingly, existing ADR detection models are either trained with additional corpora that are annotated manually or jointly trained with the ADR detection and the entity mention extraction task. However, directly training a method with additional corpora sampled from different sources may introduce noises and impact the performance of neural networks. Besides, jointly training a method with different tasks requires the annotation for other tasks, which still increases the annotation workload. To address the above issues, we formulate ADR detection as a text classification task and introduce an adversarial transfer learning framework into ADR detection. Our method focuses on exploiting a source corpus to improve the performance on small target corpora which only contain hundreds of training instances. Also, adversarial learning is applied to prevent corpus-specific features from being introduced into shared space so that corpora from different sources can be leveraged with minimum extra noises. Experimental results on three different benchmark corpora show that our proposed method consistently outperforms other state-of-the-art methods, especially on small corpora.