A Joint Entity Relation Extraction Method Based on Type Prediction

Entity relation joint extraction is to extract entity and relation facts from unstructured text at the same time, which is a key step in constructing knowledge graph. Most of the existing joint extraction models adopt the method of extracting entities first and then classifying the relation. Although the performance is improved, the relation classifier is affected by redundant entities. Therefore, we propose a joint extraction model based on Relational Type Prediction (RTP). The type prediction attention mechanism is used to predict the relationship types of sentences, and a new sentence representation based on specific relationship types is obtained. Then each word in the sentence is mapped to the label space, the entity associated with the relation type is selected using softmax function, the entity location is marked using BIOES labeling strategy, and the triples are extracted. Experimental results on two public data sets show that this model can alleviate the problem of entity pair redundancy in overlapping entity relationship extraction, improve the interaction between entity relationships, and extract triples effectively.