| Popis: |
Named entity recognition (NER) has beed studied well recent years. However, the diversity problem that arises while multiple labels are decoded into entities has been ignored in most studies. Concretely, in the process of decoding labels into entities, the same label prediction accuracy may lead to different entity prediction accuracy. This paper propose label consistency loss (LCL) to address the multi-label decoding diversity problem while considering the global label prediction accuracy. Specifically, the proposed method consists of two loss functions: global loss function and entity loss function. The global loss function is dedicated to improve the global label prediction accuracy by calculating the global error between predicted labels and gold labels. Entity loss is dedicated to solve the diversity problem by achieving labels in same entity are in the same prediction score distribution. To verify the effectiveness of LCL, this paper propose L2E metric which denote the conversion rate of labels to entities. Apply LCL to baseline and conduct experiments on two datasets. The experiment results show the F1 performance of baseline is improved by 1.16% percent and 0.17% percent, the L2E is improved by 1.52% and 0.03% on two datasets. |
