Research on relation extraction model of overlapping entity based on attention mechanism

Abstract

Relation extraction refers to get the triple structure composed of semantic relation entity pairs from unstructured text, which is an important part of tasks such as knowledge graphs. At present, the joint extraction model is in common used to avoid the impact of overlapping entities, but there are the following problems. First, the dependencies between text words are not fully considered, and the recognition performance of entities with long spans is low. Insufficient utilization of information makes it difficult to fully extract implicit relationships. In order to address these issues, this text proposes an improved joint learning model, which builds text semantic representation through BERT pre-training, obtains relation type representation as an additional mapping through a multi-label classification method, and sequentially uses multi-layer BiLSTM combined with highway network to obtain semantic information, and combine The attention mechanism obtains the entity location score, and the pointer network is used to obtain the entity location. The experiments of this method on the common dataset of relation extraction task is effective.