Label-Review-Opinion Generation for Cross-Domain Aspect-Based Sentiment Analysis

Aspect-based sentiment analysis (ABSA) aims to extract all aspect terms from review texts and classify their sentiment polarity. Although supervised methods exhibit excellent performance on the ABSA task, review texts in new domains typically lack labels, and it is time-consuming and costly to annotate for them. Unsupervised domain adaptation can transfer the knowledge learned from the source domain to the target domain to alleviate the problem of insufficient fine-grained labeled data. In this paper, we propose a novel label-review-opinion generative model (LRO-Gen) for cross-domain ABSA task. Specifically, we generate pseudo-labels for target domain review texts. Then, the model uses syntactic similarity rules alongside the pseudo-labels to generate new target-specific labeled review texts, which narrows the gap between domains. Finally, the generative extraction model is used in place of the traditional extraction model to improve domain generalization. In particular, we strengthen the correlation between sentiment elements by transforming generating a pair of sentiment elements into a pseudo-opinion sentence. Experimental results on four datasets demonstrate that our model is more effective than previous cross-domain ABSA models.