Aspect-Oriented Unsupervised Domain Adaptation for Polarity Classification

This paper proposes a novel method of unsupervised domain adaptation for Aspect Based Sentiment Analysis. Unlike previous studies, the domain is not defined as a genre of texts but as a target aspect for which the polarity should be identified. The labeled training data of the target domain is constructed by two methods. One is automatic labeling, where polarity labels of sentences in the target domain are determined by a classifier trained from labeled data in the source domain. The other is Cross-Aspect Review Generation. After extracting domain-specific sentiment words and keywords for each source and target domain, new labeled sentences in the target domain are generated by substituting sentiment words and keywords in sentences in the source domain. Then, the generated sentence is filtered out when its fluency measured by the pseudo-log-likelihood score is too low. Finally, a polarity classifier in the target domain is obtained by fine-tuning Bidirectional Encoder Representations from Transformers (BERT) using the constructed labeled data. Experimental results show that the average accuracy of our proposed method is 65.8% and 80.0% on the restaurant and laptop datasets respectively, which significantly outperforms the baselines.