Fine-Grained Category Generation for Sets of Entities

Category systems play an essential role in knowledge bases by groupings of semantically related entities. Category generation task aims to produce category suggestions which can help knowledge editors to expand a category system. Most past research has focused on solving coarse-grained problems, not fine-grained scenarios. In this paper, we propose a two-stage framework to generate fine-grained categories for sets of entities. In the category generation stage, we extract conceptual texts from the context of entities and then employ the Seq2Seq model to generate candidate categories. In the category selection stage, we cluster the entities and design discrete patterns using entity names for prompt ranking, which are further ensembled to preserve the final categories. We construct a new fine-grained category generation dataset based on Wikipedia. Experimental results demonstrate the effectiveness of the framework over the state-of-the-art abstractive summarization methods.