An Adaptive Hybrid Framework for Cross-domain Aspect-based Sentiment Analysis

Yan Zhou, Fuqing Zhu, Pu Song, Jizhong Han, Tao Guo, Songlin HuAAAI 2021 · OpenReview

Abstract

Cross-domain aspect-based sentiment analysis aims to utilize the useful knowledge in a source domain to extract aspect terms and predict their sentiment polarities in a target domain. Recently, methods based on adversarial training have been applied to this task and achieved promising results. In such methods, both the source and target data are utilized to learn domain-invariant features through deceiving a domain discriminator. However, the task classifier is only trained on the source data, which causes the aspect and sentiment information lying in the target data can not be exploited by the task classifier. In this paper, we propose an Adaptive Hybrid Framework (AHF) for cross-domain aspect-based sentiment analysis. We integrate pseudo-label based semi-supervised learning and adversarial training in a unified network. Thus the target data can be used not only to align the features via the training of domain discriminator, but also to refine the task classifier. Furthermore, we design an adaptive mean teacher as the semi-supervised part of our network, which can mitigate the effects of noisy pseudo labels generated on the target data. We conduct experiments on four public datasets and the experimental results show that our framework significantly outperforms the state-of-the-art methods.

Introduction

Aspect-Based Sentiment Analysis (ABSA) is a fine-grained sentiment analysis task that involves identifying specific "aspects" (e.g., "battery life" in a laptop review) and determining the sentiment expressed toward them. While supervised models have reached high performance, they are heavily dependent on large-scale labeled datasets. In many real-world scenarios, we encounter a "domain shift" where labeled data exists for a source domain (e.g., Restaurants) but only unlabeled data is available for a target domain (e.g., Laptops).

Previous cross-domain approaches primarily relied on adversarial training to align feature distributions across domains. However, these methods often fail to utilize the target domain data to train the task classifier itself, missing out on domain-specific nuances. The Adaptive Hybrid Framework (AHF) addresses this by combining adversarial feature alignment with a semi-supervised "Mean Teacher" approach to refine the classifier using unlabeled target data.

Methodology

The AHF architecture integrates three core components to bridge the domain gap effectively:

  1. Adversarial Training: A domain discriminator coupled with a Gradient Reversal Layer (GRL) encourages the feature extractor to produce domain-invariant representations, ensuring that features from both source and target domains are indistinguishable to the discriminator.
  2. Mean Teacher Framework: To leverage unlabeled target data, the model employs a Student-Teacher setup. The Teacher network's weights are updated as an exponential moving average (EMA) of the Student's weights. The Teacher generates pseudo-labels for target data, which then serve as training targets for the Student.
  3. Adaptive Mechanisms: To handle the inherent noise in pseudo-labels and the class imbalance typical of sequence labeling (where non-aspect 'O' tags dominate), AHF introduces dynamic confidence thresholds and similarity-weighted losses.
The architecture of the Adaptive Hybrid Framework (AHF), showing the interaction between the Student network, Teacher network, and Domain Discriminator.

Experimental Setup

The framework was evaluated across four public benchmark datasets: Restaurant (R), Laptop (L), Device (D), and Service (S). By pairing these, the authors created 10 distinct transfer tasks (e.g., Restaurant → Laptop). Performance was measured using the F1 score, comparing AHF against several state-of-the-art baselines in cross-domain ABSA.

Results and Discussion

AHF achieved a significant performance boost, with an average F1 score of 37.67% across all transfer pairs, outperforming the previous state-of-the-art (SAL) by approximately 3%.

A key factor in this success is the adaptive thresholding mechanism. In sequence labeling for ABSA, the 'O' (non-aspect) tag is far more frequent than aspect-related tags. A static threshold would lead the model to over-predict 'O' tags. As shown in the training dynamics, AHF maintains separate, evolving thresholds for 'O' and aspect labels, allowing the model to capture rare aspect terms more accurately as training progresses.

Dynamic evolution of confidence thresholds for 'O' labels versus aspect labels during the training process.

Ablation studies further confirmed that both the domain discriminator and the similarity-weighted loss are essential. Removing these components led to noticeable performance degradation, proving that the hybrid nature of the framework is critical for successful domain adaptation.

Conclusion

The Adaptive Hybrid Framework demonstrates that target domain data is valuable not just for feature alignment, but also for direct classifier refinement via semi-supervised learning. By addressing class imbalance through adaptive thresholds and using a stable Teacher network for pseudo-labeling, AHF sets a new benchmark for cross-domain ABSA. Future work may focus on mitigating "negative transfer" where adversarial alignment incorrectly maps unrelated features across domains.

Summary updated

Abstract

Cross-domain aspect-based sentiment analysis aims to utilize the useful knowledge in a source domain to extract aspect terms and predict their sentiment polarities in a target domain. Recently, methods based on adversarial training have been applied to this task and achieved promising results. In such methods, both the source and target data are utilized to learn domain-invariant features through deceiving a domain discriminator. However, the task classifier is only trained on the source data, which causes the aspect and sentiment information lying in the target data can not be exploited by the task classifier. In this paper, we propose an Adaptive Hybrid Framework (AHF) for cross-domain aspect-based sentiment analysis. We integrate pseudo-label based semi-supervised learning and adversarial training in a unified network. Thus the target data can be used not only to align the features via the training of domain discriminator, but also to refine the task classifier. Furthermore, we design an adaptive mean teacher as the semi-supervised part of our network, which can mitigate the effects of noisy pseudo labels generated on the target data. We conduct experiments on four public datasets and the experimental results show that our framework significantly outperforms the state-of-the-art methods.
BibTeX
@inproceedings{zhou2021adaptive,
  author = {Yan Zhou and Fuqing Zhu and Pu Song and Jizhong Han and Tao Guo and Songlin Hu},
  title = {{An Adaptive Hybrid Framework for Cross-domain Aspect-based Sentiment Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=KVpoK_0Dvr},
}
Figure