Neural structural correspondence learning method based on features extending

Traditional neural structural correspondence learning only uses mutual information to select the pivot features, resulting in low quality of the selected pivot, which affects the cross-domain sentiment classification results. To solve this problem, we propose a features-extending neural structural correspondence learning model. The feature selection method are extended by the term frequency-inverse document frequency to obtain the pivot features that are more suitable for cross-domain sentiment classification. A mapping relationship between the pivot feature and the non-pivot feature is established by neural network, to reduce domain differences. The experiment results on Amazon product dataset show that the pivot features selected by the proposed method are more closely related to the sentiment polarity, and the average accuracy of the proposed method is 0.33% higher than that of traditional neural structural correspondence learning.