Random Division: An Effective Method for Chinese Text Classification

As a fundamental part of natural language processing, text classification is the backbone of tasks and applications such as machine translation and classification. Among the text classification tasks of all languages, the one for Chinese appears to be one of the most challenging due to the complex structures and expressions within the nature of Chinese. Researchers generally require a significant amount of data for model training and tuning, while most of the time, that desired amount of data cannot be fulfilled and satisfied. Given the circumstances, we propose an effective data enhancement technique to lower the demand for data. The central principle is as follows: Randomize the acquired word vectors and tokens from tokenizing the text based on a certain density level (i.e., every group contains five words), then use the randomized results as data input. During the above process, a considerable number of data variations would be generated, easing the demand for data. From the experiments, we tested our theory on multiple Chinese natural language processing datasets and received signs of improvements in model performance across all the datasets used, thus proving the validity of the previously mentioned method.