Research on Chinese Intent Recognition Based on BERT pre-trained model

Abstract

As a sub-task in natural language understanding, intent recognition research plays an important role in it. The accuracy of intent recognition is directly related to the performance of semantic slot filling, the choice of data set, and the research that will affect subsequent dialogue systems. Considering the diversity in text representation, traditional machine learning has been unable to accurately understand the deep meaning of user texts. This paper uses a BERT pre-trained model in deep learning based on Chinese text knots, and then adds a linear classification to it. Using the downstream classification task to fine-tune the pre-trained model so that the entire model together maximizes the performance of the downstream task. This paper performs domain intent classification experiments on the Chinese text THUCNews dataset.Compared with recurrent neural network(RNN) and convolutional neural network(CNN) methods, this method can improve performance by 3 percentage points. Experimental results show that the BERT pre-trained model can provide better accuracy and recall of Chinese news text domain intent classification.