Multi-perspective context aggregation for document-level relation extraction

The task of document-level relation extraction (RE) involves integration of information within and across multiple sentences of a document and extraction of complex semantic relations between multiple named entities. However, effective aggregation of local and nonlocal contexts information in the document continues to be a challenging research question. This study proposes a novel document-level RE model, called the multi-perspective context aggregation (MPCA), that aggregates document context information from multi-perspective at different layers. Specifically, this aggregated context information not only comes from the pre-training stage but is also reflected during node construction and classification. Experimental results show that our model achieves desirable performance on two public datasets for document-level RE and is particularly effective in extracting relations between entities with multiple mentions.