Research And Application Of Prompt-based Fine-grained Category Generation Method

Category systems are crucial for building knowledge bases,and finegrained categories contain more semantic and relational information,providing irreplaceable value for downstream tasks such as entity set expansion.The emergence of new entities constantly increases the demand for updating and maintaining the category system.However,the updating of the category system is manually edited by knowledge editors,resulting in high costs and low efficiency.Therefore,automatically generating fine-grained categories is an important research problem.To address the above issues,this paper proposes a hint-based fine-grained category generation method and investigates the application of fine-grained categories in the entity set expansion task.The main contributions of the paper are as follows:(1)This paper proposes a two-stage framework for prompt-based finegrained category generation.In the category generation stage,a Seq2 Seq model is used to generate a large number of high-recall candidate categories.In the category selection stage,perplexity-based prompt ranking and rulebased ensemble ranking are designed to select categories,improving the accuracy of categories.Additionally,this paper constructs a fine-grained category generation dataset based on Wikipedia,including 23.8 million entities and contexts,1.1 million categories,and 53,882 fine-grained categories.Experimental results show that our method achieves state-of-theart performance in both ROUGE evaluation and text similarity assessment.(2)This paper proposes a Chinese MOOC entity set expansion method based on fine-grained category prompt.First,the Seq2 Seq model is used to generate fine-grained positive categories,and then negative categories are generated based on the masked language model.Entity completion templates are constructed using fine-grained positive category prompt to generate entities,and finally,entities are filtered by comparing their similarity with positive and negative categories and seed entity sets.This paper constructs an entity set expansion dataset based on the MOOCCube X dataset,containing45,000 entities and contexts as well as 476 courses.Experimental results show that our method achieves state-of-the-art performance in MAP evaluation.