Chinese Name Entity Recognition with Label Semantics

Named Entity Recognition (NER) requires effective information capture from annotated instances and the transfer of useful knowledge from external resources. In this paper, we utilize the semantic knowledge embedded in the label names to provide the model with additional signals and enriched priors. In our experiment, we employ two distinct BERT encoders to generate semantic representations for both the label names and the text. This approach allows us to leverage the semantic information embedded in the label names, providing the model with additional signals and enriched priors. To optimize the model, we employ a specialized loss function called Focal Loss, which can address the challenges of our particular task and align with our objectives. The label semantics approach has demonstrated improved performance across multiple Chinese NER datasets, providing evidence for the effectiveness of our work.