Introduction
Sentiment analysis models often struggle with domain shifts. A model trained on movie reviews may fail when applied to electronics reviews because the sentiment-bearing vocabulary changes—for instance, "unpredictable" is positive for a thriller but negative for a battery. This challenge, known as Cross-Domain Sentiment Classification (CDSC), typically requires large amounts of labeled data for each new domain. While adversarial methods like Domain-Adversarial Neural Networks (DANN) attempt to learn domain-invariant features, they often neglect category-specific boundaries, potentially mixing "positive" and "negative" signals during alignment.
The Symmetric Adversarial Transfer Network (SATNet) addresses this by employing a symmetric architecture and a two-level alignment strategy. It aims to align both the overall domain distributions and the specific sentiment categories simultaneously to ensure sentiment labels remain meaningful across domains.
Methodology
SATNet consists of a feature extractor ($G$), two sentiment classifiers ($F^s$ for source, $F^t$ for target), and a domain discriminator ($D$). Its architecture is defined by several key innovations:
- Symmetric Design: The domain discriminator $D$ shares neurons with the sentiment classifiers. This architecture forces the network to learn features that are simultaneously useful for classification and domain distinction.
- Two-Level Alignment:
- Domain-Level: Adversarial training is used to minimize the domain discrepancy, making features from different domains indistinguishable to the discriminator.
- Category-Level: The model enforces consistency between the outputs of the source and target classifiers, ensuring that sentiment categories (e.g., "positive") are mapped correctly across domains.
- Entropy Minimization: To leverage unlabeled target data, the model minimizes prediction entropy, encouraging "confident" predictions and sharpening decision boundaries.
- Two-Step Training: The training process alternates between minimizing classification/domain errors and maximizing domain confusion, resulting in a robust, domain-invariant feature space.
Experimental Setup
The model was evaluated using the Amazon reviews dataset, a standard CDSC benchmark. The dataset includes reviews from four domains: Books (B), DVD (D), Electronics (E), and Kitchen (K). The researchers conducted experiments on 12 transfer tasks (e.g., $B \to K$, where the model is trained on Books and tested on Kitchen).
Results and Discussion
SATNet demonstrated superior performance compared to several strong baselines, including DANN, mSDA, and ACAN.
As shown in Table 1, SATNet achieved the highest average accuracy of 82.31% across all 12 tasks. It particularly excelled in specific transfers, such as $E \to K$, where it reached 88.00% accuracy.
Ablation studies highlighted the importance of each component. Removing the domain-level loss (SATNet-D) or the entropy minimization (SATNet-M) resulted in a noticeable drop in performance, confirming that both domain alignment and category-level consistency are critical for effective transfer. While SATNet did not achieve the highest score in every single task (e.g., ACAN performed better on $B \to K$), its overall consistency and highest average accuracy suggest superior generalization capabilities.
Conclusion
SATNet provides a robust framework for cross-domain sentiment classification by combining symmetric adversarial training with two-level distribution alignment. While the current implementation uses a simple CNN feature extractor, future work could explore scaling the architecture with Transformers like BERT or RoBERTa. Additionally, testing the model on more diverse datasets beyond consumer products would further validate its effectiveness in handling extreme domain shifts.
Summary updated