Prompt-based Few-shot Learning for Table-based Fact Verification

Natural language processing has been a hot topic of research, but existing research is mainly limited to unstructured information such as natural language sentences and documents, and less research has been done on structured information such as tables. The main object of this paper is the table-based fact verification task, under which there is only one TABFACT dataset. Most of the existing methods on this dataset are based on pre-trained models and need to be fine-tuned again if a new dataset appears. And some previous work on natural language sentences has shown that prompt approach can achieve good performance with few samples. Therefore, in this paper, we adopt the prompt approach for experiments on the table fact detection task by manually designing templates for hinting the pre-trained model. Meanwhile, to enhance the generalization of the model, we introduce a multi-pair mapping relationship in the Answer Engineering phase. Experiments on the TABFACT dataset show that using the prompt method for table-based fact verification task in the case of few samples can be effective, providing a new way for optimizing table-related tasks in the case of few samples.