A Few-Shot Relation Extraction Method for Enhancing Entity Attention

Abstract

The aim of the few-shot relation extraction (FSRE) method is to study the relation classification problem with fewer samples. An effective few-shot relation extraction model EnAttConceptFERE is proposed to effectively classify relationships through externally introduced entity concept information and the greater use of internal information. First, we introduce an entity-level vector representation, which uses selects appropriate entity concepts by comparing the similarity between the semantics of entity pairs in a sentence and the semantics of the concepts corresponding to the entities. In addition, the access to external resources is often limited and the introduction of noise cannot be avoided. Therefore, this paper is based on fully mining the effective information of the sample itself, and by introducing the entity self-attention module, the model can pay greater attention to the information of entity pairs that have an impact on relationship extraction. In order to verify the performance of EnAttConceptFERE, experiments are conducted on the FSRE benchmark dataset FewRel. Under the few-shot task setting of 5 way1shot (N=5,K=1) and 10way1shot (N=10,K=1), the accuracy rate is improved by 2.53% and 1.06%, and under the task setting of 5way5shot(N=5,K=5), the accuracy was improved by 1.31% compared with the TD-Proto model, demonstrating the effectiveness and superiority of the EnAttConceptFERE model.