Cross-Domain Sentiment Classification with Mere Contrastive Learning and Improved Method

Cross-domain sentiment classification refers to training a classification model using labeled data in the source domain, and migrating to the target domain to perform sentiment classification on unlabeled data and meet the performance requirements, so as to deal with the scarcity of labeled data and the time-consuming manual labeling work. However, most previous studies use cross-entropy-based methods, and the cross-entropy loss will bring problems such as poor stability and weak generalization ability. In this paper, instead of using cross-entropy, we propose a model that uses only contrastive learning on BERT and applies mutual information maximization. Past research has devoted to exploring algorithms for manually selecting common features, and traditional supervised methods cannot meet the needs of deep learning. In this regard, we use the Wasserstein distance to estimate the feature representations of the source domain and the target domain to share domain-invariant information between the both domains, and learn domain-invariant features with an unsupervised adaptive method. Due to the scarcity of labels on the target domain, we apply mutual information maximization to balance the model's predictions. Experiments on datasets show that our model has an average accuracy of 92.40% on cross-domain sentiment classification and 93.03% on multi-domain sentiment classification, both higher than CLIM and COBE on cross-domain and multi-domain tasks. The visualization results more intuitively demonstrate the advanced performance of our model, and the ablation experiments clearly reflect the effectiveness of our method.