Cross-Domain Sentiment Classification Based On Convolutional Neural Networks

The traditional sentiment classification method is based on a large amount of labeled data.The feasible method to overcome this deficiency is to implement sentiment classification of the target domain with the help of relevant source domain which has sufficient labelled samples.However,sentiment classification is closely related to the topics and domains,so the labeled data in one domain cannot simply be replicated in another domain.Deep learning can avoid the influence of feature engineering as much as possible.Convolution neural network also has a good adaptability,fault tolerance and the advantages of the self-learning,and has been successfully applied to many fields such as natural language processing.This paper focuses on bridging the gap and transferring sentiment knowledge between domains in cross-domain sentiment classification.The work mainly includes the following two aspects.(1)This paper proposes a cross-domain sentiment classification method based on domain-independent features.We quantify emotional polarity and the semantic consistency between domains for each words.Then the semantic consistency and emotional polarity is incorporated into the process to select the domain-independent features.The words with strong emotion and semantic consistency are selected as domain-independent features.Moreover,we extend the texts by domain-independent features to enhance the emotional semantic,and further train cross-domain sentiment classifier on these extended features.Experiments on Amazon product review datasets show that the method is effective.(2)This paper proposes a cross-domain sentiment classification method based on domain-specific features.First,we select high-quality domain-specific features based on word embedding and modified mutual information,and further accomplish cross-domain sentiment classification based on CNN and the domain-specific features.Finally,a series of experiments were conducted on the Amazon product review datasets,and the experimental results were analyzed.Experimental results show that the combination of domain-specific features and CNN can improve the performance.