GAST: Graph Adaptive Semantic Transfer Model for Cross‑Domain Sentiment Analysis

Cross‑domain sentiment analysis aims to transfer sentiment knowledge from a well‑labeled source domain to a sparsely labeled target domain by exploiting shared linguistic features. Traditional approaches either ignore explicit linguistic cues such as POS tags and dependency structures or treat the problem as a black‑box transfer task.

The paper first reviews relevant work, including DANN (ICML 2015) and IATN (AAAI 2019), which use domain classifiers and attention mechanisms but lack explicit syntactic modeling. It then motivates the need for richer semantic representations, explicit POS information, and syntactic graph structures.

The proposed GAST model consists of two main modules: a POS‑based Transformer that incorporates POS‑tag embeddings alongside word embeddings, and a Hybrid Graph Attention Network (HGAT) that models dependency graphs generated by CoreNLP. Two graph‑modeling strategies are explored—relation aggregation and relation activation—where the latter adaptively weights syntactic relations via scaled dot‑product attention.

Training jointly optimizes three losses: sentiment classification, domain classification (as in DANN), and a self‑supervised feature alignment loss that encourages extraction of domain‑shared features. Experiments on four domains (books, DVD, electronics, kitchen) demonstrate that GAST achieves SOTA performance on 12 cross‑domain sentiment datasets, with ablation studies confirming the contribution of each module.

Case studies visualizing attention maps show that GAST effectively captures sentiment words, their contextual modifiers, and syntactic dependencies, leading to more interpretable predictions. Efficiency analysis reveals that GAST reaches comparable performance with only 40% of the training data required by IATN.

The conclusion emphasizes that GAST successfully addresses the “what to learn” and “how to learn” questions in cross‑domain sentiment analysis by explicitly modeling POS tags and dependency graphs, and outlines future work applying the framework to massive unlabeled review data in real‑world e‑commerce scenarios.

A short Q&A addresses the applicability of the framework to other cross‑domain NLP tasks, the trade‑off between large pre‑trained models and explicit linguistic priors, and the impact of dependency‑graph quality on results.