Soft labeling constraint for generalizing from sentiments in single domain

摘要

In this work, we deal with domain generalization in sentiment analysis. In traditional domain generalization systems, multiple source domains are used to generalize to a single target domain. However, we tackle the scenario where examples of sentiments from only one domain are available. Recent works have proposed to generate target domain examples from a single source domain by means of an adversarial training, ensuring that generated examples performs well on classifier trained on source domain. However, the inherent assumption is that domain shift is only due to covariate shift. In our work, we argue that, in realistic scenarios such as sentiment analysis, there is significant change in label distribution across domains as well. Subsequently, we propose a soft labeling formulation that provides better generalization and more robust classifiers across unseen sentiment domains. Experimental results on the Amazon-reviews benchmark dataset show the effectiveness of the proposed formulation.