A large number of low-frequency words existed in Chinese short text classification, reasonable development and utilization could effectively improve the effect of text classification. Aiming at the problem that low-frequency words couldn't be effectively used in the research of text classification based on word vector, this paper proposed a method of data enhancement based on low-frequency words. Firstly, it used the data generated by the restricted text generation model to fine tune the word vector of low-frequency words, and then used a word vector construction algorithm to transfer the updated information of high-frequency words to low-frequency words, so that low-frequency words could obtain a more accurate word vector representation that conformed to the distribution of training set. Secondly, it introduced prior knowledge such as similar words and entity concepts to supplement context information. Finally, it used the improved Chi-square statistics to remove the obvious noise words, and designed the word attention layer to weight each word to reduce the impact of irrelevant noise on classification. Experiments on several basic classification models show that each basic model has been significantly improved after applying the improved method. It not only shows the effectiveness of the proposed method, but also shows that low-frequency words can improve the classification effect in short text classification task.

[2021-12-17] Accepted Paper /doc/2021.08.0385/v1

[2022-02-05] Printed Article /doc/2021.08.0385/v2