Full-attention Based Drug Drug Interaction Extraction Exploiting User-generated Content

Abstract

When a patient takes multiple medications simultaneously under treatment, it is vital for the doctor to comprehend all interactions between drugs in the prescription entirely. Drug drug interaction (DDI) extraction aims to obtain interactions between drugs from biomedical literature automatically. Nowadays, researchers apply artificial intelligence and natural language processing techniques to perform DDI extraction task. Existing DDI extraction methods have utilized some kinds of external resources such as biomedical databases or ontologies to offer more knowledge and improve the performance. However, these kinds of external resources are delayed because of the hardship of updating. User-generated content (UGC) is another sort of external biomedical resource which is up-to-date and can be updated rapidly. We attempt to utilize UGC resource in our deep learning DDI extraction method to provide more fresh information. We propose a DDI extraction method that merges UGC information and contextual information together by a new attention mechanism called full-attention. We conduct a series of experiments on the DDI 2013 Evaluation dataset to evaluate our method. UGC-DDI outperforms the other state-of-the-art methods and achieves a competitive F-score of 0.712.