Mention-centered Graph Neural Network for Document-level Relation Extraction

TL;DR

This paper introduces a mention-centered graph neural network that infers cross-sentence dependencies for document-level relation extraction, improving relation modeling by reasoning over mention relations and addressing NA instance issues.

Contribution

It proposes a novel mention convolution approach with an improved ranking loss to better model inter-sentence relations and handle incomplete annotations in document-level relation extraction.

Findings

Cross-sentence mention dependencies are crucial for better relation extraction.

The proposed method outperforms existing approaches on benchmark datasets.

Reasoning over mention relations enhances the extraction of higher-level compositional relations.

Abstract

Document-level relation extraction aims to discover relations between entities across a whole document. How to build the dependency of entities from different sentences in a document remains to be a great challenge. Current approaches either leverage syntactic trees to construct document-level graphs or aggregate inference information from different sentences. In this paper, we build cross-sentence dependencies by inferring compositional relations between inter-sentence mentions. Adopting aggressive linking strategy, intermediate relations are reasoned on the document-level graphs by mention convolution. We further notice the generalization problem of NA instances, which is caused by incomplete annotation and worsened by fully-connected mention pairs. An improved ranking loss is proposed to attend this problem. Experiments show the connections between different mentions are crucial to document-level relation extraction, which enables the model to extract more meaningful higher-level compositional relations.