Fine-grained entity typing based on hyperbolic representation and label-context interaction

Abstract

Fine-grained entity typing (FET) is a crucial task in natural language processing (NLP), which aims to assign detailed type labels to entities based on context. Accurate entity typing is essential for many downstream applications, such as knowledge graph construction, information retrieval, and question answering. However, existing FET methods face significant challenges in capturing the hierarchical structure of entity types and effectively leveraging contextual information. Many prior approaches either rely on label co-occurrence statistics, which may introduce noise, or utilize hyperbolic space, which performs well for ultra-fine entities but struggles with coarse-grained entity types. Furthermore, the lack of effective label-context interaction limits the model's ability to filter out irrelevant type labels, leading to suboptimal entity typing performance. To address these issues, we propose a novel FET framework that integrates hyperbolic representation and label-context interaction. First, we map the hierarchical structure of entity labels into hyperbolic space, allowing for a more effective representation of type relationships. A graph convolutional network (GCN) is then employed to model label dependencies while filtering out noisy co-occurrence information. Additionally, we introduce a label-context interaction module using attention mechanism to refine type selection by modeling semantic correlations between context and labels. This mechanism dynamically enhances the relevance of selected type labels while mitigating noise. Experiments on multiple public datasets demonstrate the effectiveness of combining hyperbolic representation with label-context interaction for FET.