Adverse Drug Events Detection, Extraction and Normalization from Online Comments of Chinese Patent Medicines

Chinese Patent Medicines (CPMs) are welcomed by many people around the world, but the lack of information about their Adverse Drug Reactions (ADRs) is a big issue for drug safety. To get this information, we need to analyze from a number of real-world Adverse Drug Events (ADEs). However, current surveillance systems can only capture a small portion of them and there is a significant time lag in processing the reported data. With the rapid growth of E-commerce in recent years, quantities of patient-oriented user comments are posted on social media in real-time, making it of great value to automatically discover ADEs. To this end, we build a dataset containing 17K patient-oriented user posts about CPMs and further propose a new model that jointly performs ADE detection, extraction and normalization. Different from most previous works dealing with these tasks independently, we show how multi-task learning helps tasks to facilitate each other. To better deal with colloquial expressions and confusing statements in user comments, we leverage standard ADR-terms as prior knowledge as well as finding clues from other related comments. Experimental results show that our model outperforms previous work by a substantial margin.