Joint Extraction of Entities and Relations in the News Domain

Abstract

AbstractExtracting entities and relationships between entities from news text information is the core task of building news knowledge graphs. In recent years, with the rise of knowledge graphs, the joint extraction of entity relationships has become a research hotspot in the field of natural language processing. Aiming at the problem that there are many entities in news text data and overlapping relationships between entities are common, this paper first proposes a labeling strategy around the central entity, which transforms the extraction of entities and relationships into sequence labeling problems. After that, this paper also proposes a joint extraction model, which is based on pre-trained language and combined with the improved Bi-directional Long Short-Term Memory (BiLSTM) and Conditional Random Field (CRF) model to achieve entity and relationship extraction. The experimental results on two public news datasets show that our proposed joint extraction model has different degrees of improvement in accuracy and recall compared with other popular joint extraction models. The F1 value on NYT and DuIE both achieved the highest values, reaching 71.6% and 81.4%, which proves that the method proposed in this paper is effective. KeywordsJoint extractionDeep learningRelation overlap