A relationship-first approach to joint entity relationship extraction

Entity relationship extraction, which aims to extract semantic relationships between two entities from unstructured text, is a key step in building knowledge graphs. Existing works using entity-first phased joint extraction methods, which has achieved good results on the task of entity overlapping triad extraction, but all suffer from a large number of redundant operations severe exposure bias phenomenon. In response, this paper proposes a relationship-first joint extraction method that prioritizes the extraction of relationships between entities to avoid redundant computations, then uses conditional layer normalization to fuse a priori information to mitigate exposure bias, and finally uses the generation of entity expression vectors that fuse relationship information for relationship extraction. In the experimental part, the model performance is evaluated and analyzed, and the experimental results show that the method in this paper achieves 91.4% and 92.2% F1 values on both NYT and WebNLG datasets, which is better than the existing models.