Joint Relational Triple Extraction Based on Multicore Convolutional Attention and Entity Topic Enhancement

Abstract

Relation extraction is a fundamental task in natural language processing, usually tied to named entity recognition. While existing relational triple extraction methods can improve performance to some extent, these models tend to treat the identified entities as meaningless categorical labels, ignoring the thematic attributes embedded in the entities in a particular context. As a result, we propose a relationship extraction model called MCATE. The model is dedicated to mining the topic semantics of entities and assigning appropriate attention weights to entity vectors and full-text information. Specifically, we constructed two modules sequentially between the subtasks of Named Entity Recognition(NER) and Relationship Extraction(RE), named as Subject Topic Filter(STF) and Multicore Convolutional Semantic Fusion(MCSF). STF deeply refines the thematic information of the extracted entity vectors on the basis of NER, which will play an important role in entity-relationship matching. MCSF combines the local information where the entities are located with the full text content to further enrich the semantic features of the text. Extensive experiments on both NYT and WebNLG datasets show that our model indeed achieves an excellent performance.