Abstract:Feature space has the high-dimensional problem in text mining. This paper presented a new description method of text feature based on word clustering. The purpose is to mine semantic association between words using machine learning, then to construct the concept dictionary in specific areas dynamically, finally to describe the text feature with the concept constructed. This method analyzes the co-occurrence of words in training corpus firstly, without using theme dictionary, then generates word cluster expressed in seed words which represents a concept of theme by word clustering, finally takes the seed words as text features. The experimental results indicate that this method not only reduces dimensionality of feature space but also overcomes the limitations of the concept in HowNet, and improve the performance of text categorization.