Named Entity Recognition (NER) is commonly treated as a sequence labeling task, which encounters significant challenges when named entities appear in nested forms. Span-based methods, which enumerate all possible spans in a sentence as candidate entities and then classify them, provide a straightforward solution for nested NER. However, this approach has difficulty effectively capturing semantic information within spans as well as the interactions between semantic representations. In addition, several challenges remain, including the need for precise annotation of nested entity boundaries, high computational complexity, and substantial computational costs. To address these challenges, this paper proposes a nested entity recognition method based on a Multi-head Tri-Affine Attention mechanism (MTAA). First, MTAA is employed to generate span representations that integrate both boundary features and internal span information, enabling better modeling of interactions for longer spans (entities). Next, regularization techniques are applied to smooth annotated entity boundaries and alleviate model overfitting. Finally, entity recognition is performed using span scores. Experimental results demonstrate that the proposed method outperforms traditional span-based approaches, achieving state-of-the-art F1 scores on nested NER datasets ACE2004, ACE2005, and GENIA, as well as on the flat dataset RESUME. Moreover, the proposed method exhibits advantages in both space complexity and time complexity.
Zhang et al. (Thu,) studied this question.