Abstract: Sentiment analysis is to extract sentimental polarity from subjective texts. However, sentiment analysis often suffers from the problems that the labeled samples are insufficient and sentimental vocabularies vary widely from domain to domain. Thus, this paper proposes a cross-domain sentiment classification method based on Convolutional Neural Networks, which accomplishes unsupervised sentiment analysis in the target domain based on the sufficient labeled samples in the source domain. Firstly, we quantify the emotional polarity of words, and define the semantic consistency between domains based on the word vectors. Then, the words with strong emotion and semantic consistency are selected as shared words between domains. Secondly, the text feature is extracted by Convolution Neural Networks, and we extend the texts in source domain with shared words based on their polarity. Thirdly, the target domain texts are classified based on classifier trained on extended texts. Finally, the experimental results on the amazon data set show that our method can improve the performance of cross-domain sentiment classification.