CGCL: A Novel Collaborative Graph Contrastive Learning Network for Chinese NER

Jicheng Yu,Zefeng Feng,Jiawei Li,Yixiu Qin,Yuncheng Jiang

DOI: https://doi.org/10.1007/978-981-97-5501-1_13

2024-01-01

Abstract:As an essential task in Natural Language Processing (NLP), Named Entity Recognition (NER) aims to extract key entities in text. Most existing methods utilize dictionaries and embeddings to solve the Chinese NER task, which is effective. However, the methods of using lexical embedding are limited by the fixed content of the dictionary, and a large number of self-matched words will blur text boundaries to interfere with the prediction. Due to the multiple meanings of Chinese characters, these methods cannot cope with confusing text boundaries and emerging Chinese entities. To address these challenges, we propose the Collaborative Graph Contrastive Learning (CGCL) network for Chinese NER. Specifically, we propose two innovative modules for Chinese NER. Firstly, we extract the grammatical information of the sentence to solve the interference problem of vocabulary knowledge and the fuzzy problem of entity boundaries. Secondly, we combine five collaborative graphs to construct three sample graph groups, and we construct a contrastive learning network to improve recognition performance. Finally, experimental results on the benchmark Chinese NER datasets show that CGCL is effective and. CGCL outperforms the state-of-the-art (SOTA) results.