MPTTF-BERT: Multi-Prefix-Tuning Template Fusion by BERT for Zero-shot English Text Relation Extraction Model

Relation extraction is the core task of information extraction. How to extract important relation features quickly and accurately from massive English texts has become a difficult task in English text relation extraction. Meanwhile, Prefix-tuning is widely used in zero-shot natural language processing tasks, but the existing zeroshot relationship extraction model based on Prefix-tuning is difficult to construct answer space mapping and depends on manual template selection, which cannot achieve good results. To solve the above problems, this paper proposes a novel zero-shot English text relation extraction model based on multi-Prefix-tuning template fusion via BERT. Firstly, the task of extracting zero-shot relation is defined as the task of mask language model, and the construction of answer space mapping is abandoned. The words output from template and the relation description text are compared in the word vector space to judge the relation category. Secondly, the part of speech of the description text of the relation class to be extracted is introduced as the feature, and the weight between the feature and each template is learned. Finally, this weight is used to fuse the results of multiple template outputs to reduce the performance penalty caused by manually selected Prefix-tuning templates. The experimental results show that the F1 values obtained by MPTTF-BERT model on DuIE, COAE-2016-Task3, FinRE data sets with different features are 93.73%, 91.49% and 49.46%, respectively, which is significantly better than the comparison models. In addition, the ablation experiment and fixed-length selection experiments are used to further verify that the MPTTF-BERT model can effectively improve the effect of English text relation extraction, indicating the feasibility and effectiveness of the new method.