A Method Based on Heterogeneous Graph and Contrastive Learning in Named Entity Recognition

Named entity recognition (NER) is a core natural language processing task. Traditional sequence labeling methods have trouble with nested entities because of single layer representations, and existing solutions are often complex or computationally costly. This paper presents a novel model, named HCG, using heterogeneous graph representation learning and contrastive optimization for NER. We introduce a gating mechanism, and jointly optimize CRF with contrast loss to optimize boundary prediction and effectively identify entities. Our model achieves 92.78% and 79.77% F1 on the CoNLL2003 and GENIA datasets, respectively, outperforming recent models. This work provides an efficient solution for complex NER.