Unsupervised Domain Adaptation with Contrastive Learning for Cross-domain Chinese NER

Anonymous

Abstract: Understanding and recognizing the entities of Chinese articles highly relies on the fully supervised learning based on the domain-specific annotation corpus. However, this paradigm fails to generalize over other unlabeled domain data which consists of different entities semantics and domain knowledge. To address this domain shift issue, we propose the framework of unsupervised Domain Adaptation with Contrastive learning for Chinese NER (DAC-NER). We follow Domain Separation Network (DSN) framework to leverage private-share pattern to capture domain-specific and domain-invariant knowledge. Specifically, we enhance the Chinese word by injecting external lexical knowledge base into the context-aware word embeddings, and then combine with sentence-level semantics to represent the domain knowledge. To learn the domain-invariant knowledge, we replace the conventional adversarial method with novel contrastive regularization to further improve the generalization abilities. Extensive experiments conducted over the labeled source domain MSRA and the unlabeled target domain Social Media and News show that our approach outperforms state-of-the-arts, and achieves the improvement of F1 score by 8.7% over the baseline.