Latent Layer Causal Structure Enhancement for Cross-Domain Named Entity Recognition

Named Entity Recognition (NER) is a key task in natural language processing, aiming to recognize and classify entity information in unstructured text. Existing cross-domain named entity recognition (CD-NER) models usually utilize transfer learning methods such as multi-task learning or comparison learning to alleviate the problem of data imbalance. However, these methods ignore spurious correlations due to differences in data distribution. On this basis, this paper proposes a CD-NER model based on latent causal structural alignment, which first utilizes the powerful generative capabilities of large language models to extend and enhance the dataset, and then, aligns the semantic features through the latent structural alignment mechanism to obtain cross-domain invariant features, and utilizes the causal module, which exploits the causal relationship between features and labels through the causal intervention mechanism to mitigate data distribution bias and data imbalance problems. Finally, experiments were conducted on different CD-NER datasets. The results show that the average performance of C-NER is improved compared to the current cross-domain models, and has strong adaptability and generalization ability in different domains.