An improved two-stage zero-shot relation triplet extraction model with hybrid cross-entropy loss and discriminative reranking

Diyou Li, Lijuan Zhang, Juncheng Zhou, Jie Huang, Neal N. Xiong, Lei Zhang, Jian Wan

Abstract: Zero-shot relation triplet extraction (ZeroRTE) aims to extract relation triplets from unstructured text under zero-shot conditions, where the relation sets in the training and testing stages are disjoint. However, most of the existing ZeroRTE models do not fully utilize the generation ability of the model and lack of precise measurement for the sorting of triplets. To this end, we propose a novel ZeroRTE model based on two training stages in this paper, including a generative training stage and a discriminative training stage. In the generative training stage, our model designs the hybrid cross-entropy loss function, which combines the forward cross-entropy loss function with the reverse cross-entropy loss function to improve the generation quality of relation triplets. In the discriminative training stage, our model integrates a reranking task to enhance the sorting accuracy of our model for candidate triplets. We evaluate the proposed model on two ZeroRTE datasets (FewRel and Wiki-ZSL), and relevant experimental results fully demonstrate the effectiveness of our method.