Improved Deep Belief Network to Feature Extraction in Chinese Text Classification

In the field of text classification, the classical text categorization method uses bag-of-words (BOW)model as a classification feature. The BOW model only focuses on the word frequency, but ignores the connection between words, so the BOW model has defects of insufficient expression of the text. The word embedding model maps each individual word to the new vector space, which makes up for the defects of isolated vocabulary and insufficient text expression. As a result, using the word embedding can improve the accuracy of the text classification. In order to avoid the dimension explosion, this paper uses keyword-based word embedding to represent Chinese text. In the feature extraction process, we use the combination of deep belief network and deep Boltzmann machine to extract word vector features of the text. This improved deep belief network feature extraction method further improves the accuracy of text classification.