Optimizing boundary dynamics for nested named entity recognition via semantic refinement and trimming.

Abstract

Nested Named Entity Recognition (Nested NER) addresses the complex task of identifying and classifying entity spans embedded within other entities in textual data. Despite advances, existing span-based methods primarily rely on exhaustive span enumeration without adequately accounting for subtle semantic differences at entity boundaries, leading to boundary ambiguity and inaccurate entity delineation. This limitation significantly affects the precision of nested entity detection, particularly in densely nested contexts. To overcome these challenges, we propose a novel approach named Semantic Refinement and Trimming (SRT). Specifically, SRT employs a biaffine attention mechanism to construct detailed semantic representations for candidate spans, which are subsequently refined through a Boundary-aware Semantic Refinement Module (BSRM). This module leverages a learnable convolutional kernel to explicitly capture fine-grained semantic differences between overlapping spans, addressing boundary ambiguity at the semantic level. Moreover, we introduce a Boundary Trimming Module (BTM), which effectively reduces irrelevant span noise through a dual-pathway architecture, simultaneously facilitating top-down semantic refinement and bottom-up semantic restoration. Extensive experiments conducted on established nested (ACE04, ACE05, GENIA) and flat (CoNLL03) NER benchmarks demonstrate that our proposed SRT method achieves state-of-the-art performance.