Short text manifold representation based on AutoEncoder network

A short text manifold representation method based on AutoEncoder network was proposed for the sparsity and the problem of the curse of dimensionality of short text. The main idea is to extract manifold features of short text for non-linear dimensionality reduction from AutoEncoder Network by reconstructing text data and finding manifold mapping at first. Then extend short text and get the optimum manifold representation model by tuning the mapping based the global pair-wise pattern between label and its Multi-document in High-dimensional observation space. The Short text manifold representation can be obtained using the model. Combined with SVM, KNN, Naïve-Bayes, the method can get better classification results than VSM, LDA and LSI. The Macro_F₁ of the method can be over 97.8%. The experimental result indicates the manifold representation can describes features of short text more accurately and non-sparse, leading to a significant improvement of the classification.