Abstract: The LDA method does not take the input space into consideration effectively, when making topic label to each word in the original space, it holds the non-action words, which affects the probability distribution of the topic extremely. In order to overcome this insufficiency, a new mRMR_LDA algorithm is proposed in this paper. The mRMR maps the input space to the low dimensional space, and filters the non-action words, which makes LDA perform topic label in a simpler and clearer space, so that it can achieve a more precise topic distribution. The classification accuracy of the 20 Newsgroup corpus and the corpus of Fudan University is improved by 1.53% and 1.18% respectively using the proposed algorithm. Experimental results show that the mRMR_LDA model has a better performance in text classification.

Key words: Latent Dirichlet Allocation（LDA）, minimum Redundancy Maximum Relevance（mRMR）, text categorization

摘要： LDA没有考虑到输入，在原始的输入空间上对每一个词进行主题标签，因保留非作用词，而影响了主题概率分布。针对这种情况提出了一种mRMR_LDA算法，预先使用mRMR特征选择算法将输入空间映射到低维空间，过滤掉非作用词，使得LDA能在更简洁和更清晰的空间上进行主题标签，得到更精确的主题分布。对20 Newsgroups语料库和复旦大学语料库进行分类，分类精度分别提高了1.53%和1.18%，实验结果表明提出的mRMR_LDA模型在文本分类中有较好的分类性能。

关键词: 潜在狄利克雷分配, 最小冗余最大相关, 文本分类