A mutual mean teacher framework for cross-domain aspect-based sentiment analysis

Aspect-based sentiment analysis is a fine-grained task that involves jointly extracting aspect terms and their corresponding sentiment polarities. However, due to the high cost of data labeling, obtaining annotated corpora for a specific target domain is often challenging. To address this issue, previous studies have attempted to transfer knowledge from a labeled source domain to the target domain by optimizing with pseudo-labels generated for the target domain data. These approaches allow both domains to be used to learn domain-invariant features. Nevertheless, such methods are susceptible to label noise, which hinders the extraction of domain-invariant features for the target domain. To mitigate the impact of error-prone pseudo-labels, we propose a mutual mean teacher framework for cross-domain aspect-based sentiment analysis. This framework generates pseudo-labels using a peer teacher network, thereby providing more reliable and robust pseudo-labels. Additionally, to develop a robust task classifier that performs well on both the source and target domains, we maximize the mutual information between the input token representations and the probability distributions of output labels, which helps prevent the model from making overly confident and incorrect predictions. Experiments conducted on ten different domain dataset pairs demonstrate that our proposed model exhibits competitive performance improvements compared to the current state-of-the-art model.