Cross-domain sentiment classification using deep learning approach

Miao Sun, Qi Tan, Runwei Ding, Hong Liu

Abstract: Deep learning, as a new unsupervised leaning algorithm, has strong capabilities to learn data representations. Previous work has shown that new features learned by deep learning algorithm help to improve the accuracy of cross-domain classification. In this paper, we firstly propose a modified version of marginalized stacked denoising autoencoders (mSDA). We call it mSDA++ algorithm, which can learn excellent and low-dimensional features for training classifier. In addition, we combine mSDA with EASYADAPT algorithm to further improve the accuracy of cross-domain classification. Then we use SVM, mSDA, mSDA++, and EA+mSDA algorithms to do the cross-domain sentiment classification experiments on Amazon benchmark dataset. The results show that EA+mSDA algorithm attains the best accuracy. Besides, the mSDA++ algorithm can accelerate the subsequent calculation and reduce the data storage space.