Enhancing Chinese Named Entity Recognition with Disentangled Expert Knowledge

Chinese Named Entity Recognition (NER) requires model identify entity boundaries in the sentence i.e., entity segmentation, and meanwhile assign entities to pre-defined categories, i.e., entity classification. Current NER tasks follows sequence tagging scheme and assign the characters to different labels by considering both segmentation position and entity categories. In such a scheme, the characters in the same entity will be regarded as different classes in the training process according to different positions. In fact, the knowledge of entity segmentation is shared across different entity categories, while entity category knowledge is relatively independent of entity segmentation. Such labeling scheme will lead to the entanglement of these two objectives, hindering the effective knowledge acquisition by the models. To address the entanglement issue and comprehensively extract useful knowledge of two objectives, we propose a novel framework that disentangle the original NER labels into two additional training labels for entity segmentation and entity classification respectively. Then we introduce two dedicated expert models to effectively extract specific knowledge from the disentangled labels. Afterwards, their predictions will be integrated into the original model as auxiliary knowledge, further enhancing the primary NER model learning process. We conduct experiments on three publicly available datasets to demonstrate the effectiveness of our proposed method.