Few-Shot Relation Classification Method Based on Semantic Feature Fusion and Relation Network

Current few-shot relation classification models based on prototype networks ignore the dependencies between query instances and support set instance representation vectors, and ignore local features of text leading to bias in the calculation of prototype centers, which affects the classification performance of the model. To solve the problem, a few-shot relation classification method based on semantic feature fusion and relation networks is proposed to enrich the feature vector representation by learning the dependencies between instances in the character layer and semantic layer, combining the relation networks to obtain the relation score, adjusting the calculation of the prototype center to enhance the support instances that are more closely related to the query instances as well as the categories. Tested on the FewRell.0 dataset, the accuracy of the method is higher than that of other few-shot learning models, which proves that the method can effectively enhance the prototype network and significantly increase the accuracy of few-shot relation classification task.