LDA topic model is an effective tool for text feature extraction. Although the topic feature is obtained through the co-occurrence of the term in the document level, which transfers the term space into the topic space, the redundant topic is included in the process of generating the document. As to the redundant topic shortage during topic feature extraction by LDA, this paper proposes an LDA topic model reduction algorithm NRS-LDA based on neighborhood rough set. Use the neighborhood rough set to construct the topic decision system. By pre-setting the number of topics, calculate the importance of each topic; rank according to the importance degree and delete the topics of low importance. Apply the NRS-LDA algorithm to the K-means text clustering problem and compare it with the traditional extraction algorithm of text feature and with the improved algorithm. The experimental results show that the proposed NRS-LDA method can obtain higher clustering accuracy.