AGCN: Attention-based graph convolutional networks for drug-drug interaction extraction

Abstract

Extracting drug-drug interaction (DDI) relations is one of the most typical tasks in the field of biomedical relation extraction. Automatic DDI extraction from the biomedical corpus is central to the mining of knowledge hidden in the biomedical literature. Existing approaches for DDI extraction primarily focus on either the contextual or the structural information of the sentence, despite their complementary role. Also, previous studies do not even exploit the entire knowledge of the input sentence, which could lead to a loss of crucial clues. In this paper, we propose an Attention-based Graph Convolutional Networks (AGCN) to address these issues. In contrast to the existing DDI extraction methods, the AGCN is designed to leverage contextual and structural knowledge together, where GCN is employed in combination with encoders based on recurrent networks. Additionally, we apply a novel attention-based pruning strategy to optimallyuse syntactic information while ignoring irrelevant information, in contrast to previous rule-based pruning methods. Therefore, AGCN can take advantage of the context and structure of the input sentence as efficiently as possible. We evaluate our model using a dominant DDI extraction corpus. The experimental results demonstrate the effectiveness of our model, which outperforms existing approaches.