A Span Information Fusion-Based End-to-End Relation Extraction Model for Power Knowledge Graph

Abstract

Entity relation extraction is a crucial step towards constructing knowledge graph. Most of the current researches on entity relationship extraction focus on joint extraction in terms of entities and relationships. However, due to that the context information respectively from entities and relationships is very different, existing joint extraction approaches often suffer from the information noise caused by both them, which may significantly affect the performance of the whole model and bring about a lower efficacy in extracting relationships. In this paper, we proposed an end-to-end relation extraction method, which uses a text representation enhanced pretraining model and fuses span information for entity relation extraction. We use the BERT model to pre-train the datasets from power project management. In order that the output of BERT contains as much contextual knowledge as possible, the external knowledge is embedded into a vector representation and further spliced into the token embeddings in the BERT input for downstream tasks. Then, we respectively use the Span-level NER method to extract all possible fragment arrangements, and SoftMax to judge the entity type of each Span. At last, when performing relation extraction, the entity boundary and types are taken as identifiers and further added into the front and end of entity span, which is used as the input of the relational model for predicting the possible relationship between the pair of spans. Experiments were made on the power project management dataset, and the results show that entity types can provide very important information for relation extraction, and the performance of the proposed method is also competitive against the state-of-the-art entity-relationship joint extraction method.