Abstract: In text feature spaces，feature dimensions are usually 10，000 and more，which restrict choice of classification algorithms and reduce performance of classification algorithms，also make classifiers hardly design，so feature selection is necessary to avoid curse of dimensionality.A syntaxic feature selection method is presented.The method firstly uses an optimized document frequency to select feature to filter out some terms to reduce the sparsity of feature spaces，and then employs an attribute reduction algorithm based on rough set and gray correlation to eliminate redundancy，so can acquire the feature subset which are more representative.The experimental results show that the syntaxic method is effective.