PLP-NER: Point-Line-Plane Context Fusion for Named Entity Recognition

Shuning Mao, Wu Yuan, ZiXiang Deng, Wen Yuan, Yidian Huang

Keywords: NER;POINT-LINE-PLANE;Dynamic Linear Chain CRF

Abstract: Current state-of-the-art Named Entity Recognition systems commonly leverage an architecture that integrates BERT with Conditional Random Fields. Nevertheless, BERT is inherently constrained in capturing comprehensive global contextual semantics due to its Masked Language Modeling pre-training objective. To address this limitation, A novel “point–line–plane” contextual fusion framework is proposed. Within this paradigem, the [CLS] token functions as a “plane” that provides a compressed global representation, while the attention weights between the [CLS] token and individual tokens form a “line”, which captures semantic topological relationships. These multi-grained features are subsequently incorporated into token representations via a Graph Neural Network, considerably enriching their contextual expressiveness. Furthermore, we introduce a Dynamic Linear-Chain CRF that adaptively models label transitions using attention-mechanized probability estimates, thereby overcoming the inflexibility of conventional CRFs. Extensive experiments on multiple benchmark datasets demonstrate that our approach consistently and significantly surpasses competitive baselines, achieving a notable 3.91 point gain in F1-score.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 9452