Short Text Classification Based on LDA and SVM

Abstract

Short text classification is different from traditional document in its high sparse features, strong context dependency and noise, etc. To solve the problem, this paper proposes a kind of classification method based on LDA(Latent Dirichlet Allocation) and SVM(Vector Space Model). We use LDA to model the topic of text corpus, infer parameter by Gibbs sampling algorithm, calculate the model parameter indirectly, and determine the optimal number of topics through Bayes standard method. Then each document is represented as the probability distribution of fixed implied topic set, the latent topic-text matrix is obtained and SVM classifier is trained on this matrix. Compared with other classification methods, the experiments results verify the effectiveness and superiority, the effect of feature dimension reduction is obviously, and the value of Macro-P, Macro-R, Macro-F and Precision is improved efficiently.