Document-level relation extraction via graph transformer networks and temporal convolutional networks

Abstract

• We propose a novel document-level relation extraction framework DocRE-HGNN. • Temporal convolutional networks is used to encode document with non-local dependency. • Heterogeneous GNN is adopted to obtain entity-entity representation automatically. • DocRE-HGNN can obtain both long inter-sentence paths and short intra-sentence paths. • Experiments on document-level biomedical datasets demonstrate model’s effectiveness. Relation Extraction (RE) aims at extracting meaningful relation facts between entities in texts. It is an important semantic processing task in the field of natural language processing (NLP) and has many applications. Traditional RE focuses on extracting entity relationships from a single input sentence. Recently, the research scope has been extended from sentence level to document level. However, compared with sentence-level RE, document-level RE, which needs to identify the inter-sentence relations from entities scattered in different sentences, is more complex and still lacks of solutions. To solve this problem, we propose a novel document-level RE method based on Heterogeneous Graph Neural Networks in this paper. Concretely, to obtain token embeddings containing long-distance dependency signals well, we encode the document with Temporal Convolutional Networks, whose dilated convolution and residual structure allow the effective and efficient preservation of historical information. To better describe the interaction between different elements, we construct the input documents as heterogeneous graphs with different node and edge types and utilize Graph Transformer Networks to generate semantic paths. Numerical experiments on two document-level biomedical datasets demonstrate the effectiveness of the proposed method.