Dropout-Based Ensemble Dual Discriminator for Cross-Domain Sentiment Classification.

The main task of Cross-Domain Sentiment Classification is to train a well-performing classification model by using labeled source domain data, and then transfer the model to unlabeled target domain data, thereby solving the expensive labor consumption and domain shift caused by a large number of labels resulting performance degradation. Most mainstream adversarial domain adaptation methods are based on a single discriminator, which ignores the uneven distribution of labels between domains and multiple modalities of data and tends to cause negative transfer and poor generalization performance. We propose a Dropout-based ensemble dual discriminator for Cross-Domain Sentiment Classification. We functionally decouple the single discriminator by using two forms of text data, and replace it with a positive sentiment discriminator and a negative sentiment discriminator. A dynamic set of discriminators will be obtained by random deactivation of the discriminator network neurons, then the feature extractor has to extract richer and more realistic domain-invariant features to fool the discriminator and mitigate the mode collapse phenomenon. To solve the problem of class imbalance in a large number of unlabeled data samples, we use mutual information maximization to train sentiment classifiers to ensure that label predictions are distributed in a reasonably balanced state. We conduct full experiments on the Amazon and Airlines datasets. Experiments showed that our proposed model achieves state-of-the-art cross-domain sentiment classification performance.