Multi-relation Word Pair Tag Space for Joint Entity and Relation Extraction

Abstract

Joint entity and relation extraction from unstructured texts is a crucial task in natural language processing and knowledge graph construction. Recent approaches still suffer from error propagation and exposure bias because most models decompose joint entity and relation extraction into several separate modules for cooperation. In addition, the mode of multi-module cooperation to complete the joint extraction task ignores the information interaction between entities and relations. Most modeling methods are based on the pattern of token pairs, which leads to ambiguous information about entities to a certain extent. To address these issues, in this work, we creatively propose a method to transform the extraction task of complex triples under multiple relations into a fine-grained classification problem based on word pairs. Specifically, to fully utilize entity information and facilitate decoding, the proposed model uses a tag strategy specific to the feature of the entity itself. Extensive experiments show that the performance achieved by the proposed model outperforms public benchmarks and delivers consistent gain on complex scenarios of overlapping triples.