Explicit Modeling the Context for Chinese NER

Anonymous

Abstract: Named entity recognition (NER) is the foundation of many natural language processing tasks. Current NER models have achieved promising results. But as pointed by several studies, they fail with a high ratio on generalization tests such as invariance test because they heavily rely on name information. So, we propose a context module to explicitly model the contextual information, and a trainable balance factor is designed to incorporate the result of context module. To learn this factor, we propose several tailored data augmentation strategies to generate synthetic labels for it. These approaches help the model learn whether it should focus on the context. Our method achieves on average 1.2\% absolute improvement of F1 than BERT-CRF on three datasets. Moreover, our method performs on par with the best solutions who rely heavily on external features besides BERT. We also conduct invariance test to analyse the effect of the context information. The source code of our model and augmentation strategies will be available at anonymous.url.