A Two-Stage Multi-task Learning-Based Method for Selective Unsupervised Domain Adaptation

Multi-task learning and domain adaption has been successfully applied in a number of real-world NLP applications, such as sentiment analysis. However, major challenges rising from real-world applications include that there may be little or even no labels in some domains, and there are large shifts between domains. In this paper, we consider the multi-source unsupervised domain adaptation (MS-UDA) for sentiment analysis (SA), which is a popular application scenario in MS-UDA. In general, the variance among different domains can result in huge semantic gaps among domains. To alleviate this issue for MS-UDA, we propose a two-stage domain adaptation framework: (1) a supervised multi-task learning is applied to extract domain-common features and domain-specific features from all the available source domains; (2) an adversarial training over the target domain will be conducted to generate target-specific features from the one of the most similar source domain. Experimental results on two sentiment analysis datasets have demonstrated that the promising performance of our framework, which not only outperforms unsupervised state-of-the-art competitors but also is very close to supervised methods and even better than supervised methods on some tasks.