KSRE-CNER: A Knowledge and Semantic Relation Enhancement Framework for Chinese NER

Chinese named entity recognition (CNER) constitutes a pivotal undertaking entailing the identification and classification of named entities present within Chinese text. Traditional approaches based on CNN and BiLSTM have been effective for sequence labeling tasks. Additionally, graph neural networks (GNNs) have shown promising results in improving Chinese NER performance by incorporating lexical knowledge. However, these methods may still face challenges in handling ambiguity and inaccurate boundary recognition in Chinese NER. To tackle these challenges, we propose a knowledge and semantic relation enhancement framework. This framework integrates N-gram information and lexical knowledge into a gated graph neural network (GGNN) to capture Chinese lexical information and reduce ambiguity. Moreover, we leverage the Transformer model to update the weight information of each node, aiming to eliminate the influence of incorrect matching lexicons and augment the model’s capability to recognize entity boundaries. Comprehensive experiments conducted on diverse datasets, including Resume, CCKS2017, MSRA, and a self-constructed History dataset, substantiate that our proposed model attains comparable results.