A Domain Knowledge Graph Construction Method Based on Joint Extraction of GlobalPointer

In order to reduce the effect of ternary overlap on the joint extraction, the GlobalPointer global pointer joint extraction strategy is used to construct a vertical domain knowledge map. The method maps the inputs into embedding vectors through the BERT pre-training model and transmits them to the Transformer bi-directional encoder to obtain dynamic word vectors that can characterize the polysemy of words. Then the idea of GlobalPointer global normalization is applied to consider the first and last information of the entity, and the first and last information of the recognized entity is introduced as a whole to be predicted, and the ternary extraction is transformed into quintuple extraction to realize the joint decoding of the information. Finally, the test is carried out on the DuIE public dataset, and the results show that compared with the mainstream TPLinker, CasRel and other baseline models, the proposed model performs the best in terms of F1 scores, and the comprehensive F1 value reaches 82.62%.