Drug-Drug Interaction Extraction via Attentive Capsule Network with an Improved Sliding-Margin Loss

Abstract

Relation extraction (RE) is an important task in information extraction. Drug-drug interaction (DDI) extraction is a subtask of RE in the biomedical field. Existing DDI extraction methods are usually based on recurrent neural network (RNN) or convolution neural network (CNN) which have finite feature extraction capability. Therefore, we propose a new approach for addressing the task of DDI extraction with consideration of sequence features and dependency characteristics. A sequence feature extractor is used to collect features between words, and a dependency feature extractor is designed to mine knowledge from the dependency graph of sentence. Moreover, we use an attention-based capsule network for DDI relation classification, and an improved sliding-margin loss is proposed to well learn relations. Experiments demonstrate that incorporating capsule network and improved sliding-margin loss can effectively improve the performance of DDI extraction.