A Joint Method of Entity Relation Extraction in the Field of Electric Power Public Opinion

To improve the response capability of public opinion on sudden power events and to solve the problem of inaccurate recognition of long entity boundaries due to the high overlap of power equipment names, a joint extraction method of entity relationships for power public opinion is proposed. First, the semantic distance between spans is increased by fusing the semantic features at the Chinese radical level, then the span unit features are updated by constructing graph neural networks using the predicted potential relationship information and span context information. Finally, the remaining entities are predicted by combining span units and potential relationships to complete the triple. The method is validated on the power opinion dataset and the public dataset, and the results of the experiment demonstrate that it has a good recognition effect on the highly overlapping entities and an adequate extraction effect on the public dataset.