Wasserstein based transfer network for cross-domain sentiment classification

Abstract

Automatic sentiment analysis of social media texts is of great significance for identifying people’s opinions that can help people make better decisions. Annotating data is time consuming and laborious, and effective sentiment analysis on domains lacking of labeled data has become a problem. Cross-domain sentiment classification is a promising task, which leverages the source domain data with rich sentiment labels to analyze the sentiment polarity of the target domain lacking supervised information. Most of the existing researches usually explore algorithms that select common features manually to bridge different domains. In this paper, we propose a Wasserstein based Transfer Network (WTN) to share the domain-invariant information of source and target domains. We benefit from BERT to achieve rich knowledge and obtain deep level semantic information of text. The recurrent neural network with attention is used to capture features automatically, and Wasserstein distance is applied to estimate feature representations of source and target domains, which could help to capture significant domain-invariant features by adversarial training. Extensive experiments on Amazon datasets demonstrate that WTN outperforms other state-of-the-art methods significantly. Especially, the model behaves more stable across different domains.