Performance optimization algorithm for CRF layer of Chinese sequence labeling model based on deep learning

The “deep learning model + CRF” model architecture such as Bi-LSTM-CRF has outstanding performance| in the tasks of part-of-speech tagging, named entity recognition, and semantic role tagging in the field of natural language processing, and has become the most popular algorithm for sequence tagging tasks. However, in practical applications, due to the strong fitting ability of the feature extractor in the deep learning model, the use efficiency of the CRF layer is relatively low; at the same time, the calculation cost of the CRF layer is relatively large. This paper proposes a performance optimization algorithm for the CRF layer in the sequence annotation task based on the deep learning model. The algorithm introduces five rules. First of all, the tag result of the feature extractor is checked by the five rules, if the adjacent tag is found to be logically incorrect, then use the CRF layer for logic correction. The experimental results show that the algorithm shows great advantages in the process of word segmentation, part-of-speech tagging and named entity recognition, a lot of computing time has been saved.