DeComa: Dictionary Enhanced Chinese NER Based on Confidence Estimation

Many recent Chinese NER approaches have adopted Dictionary Enhanced and achieved state-of-the-art results. However, the Dictionary Enhanced approach introduces noise because the recall strategy of the augmented set is string matching rather than semantic matching, which causes the incorporation of non-normal distribution of noise in the character vector. To solve this problem, we proposed a semantically adaptive Dictionary Enhanced augmentation structure that not only losslessly incorporates dictionary information into the word vector, but also has faster inference speed. In addition, we proposed a label discrimination structure based on confidence estimation to address the problem that the NER model cannot be widely used in real industrial scenarios due to the lack of training data. Using this strategy, our model not only improves the F1 value by up to 2.9% on the publicly available corpus in academia, but also achieves the best results on the noisy dataset after artificial random perturbation.