Multi-category hierarchical pointers network with lexical interaction for Chinese nested entity extraction
Abstract
Pointer networks are a common approach for addressing nested entity recognition. However, most existing studies employ linear labeling schemes for entity prediction, which fail to adequately represent multi-class entities and thus struggle to effectively handle overlapping structures. Moreover, these methods often overlook contextual lexical relationships, resulting in blurred entity boundaries. To address these limitations, this paper proposes a lexicon-enhanced pointer network. First, a multi-layer pointer mechanism is organized by entity categories, and a multi-label cross-entropy loss is introduced to optimize the labeling framework for better handling of overlapping entities. Subsequently, the LEBERT module is integrated into the BERT backbone to incorporate lexical information, thereby enhancing the modeling of word-pair relationships. Experimental results on three Chinese datasets—Resume, Weibo, and CMeEE-V2 show that the model achieves F1 scores of 97.23%, 72.73%, and 74.66%, respectively, demonstrating robust performance across both nested and non-nested entity recognition tasks. These findings validate that the proposed model improves feature integration, enhances entity recognition capability, and exhibits stronger scalability.